Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science 6.685 Electric Machines Class Notes 1: Electromagnetic Forces c 2003 James L. Kirtley Jr. September 5, 2005

Introduction

Bearings

Stator Stator Conductors Rotor

Air Gap Rotor Conductors

Shaft

End Windings

Figure 1: Form of Electric Machine This section of notes discusses some of the fundamental processes involved in electric machinery. In the section on energy conversion processes we examine the two major ways of estimating electromagnetic forces: those involving thermodynamic arguments (conservation of energy) and eld methods (Maxwells Stress Tensor). But rst it is appropriate to introduce the topic by describing a notional rotating electric machine. Electric machinery comes in many dierent types and a strikingly broad range of sizes, from those little machines that cause cell phones and pagers to vibrate (yes, those are rotating electric machines) to turbine generators with ratings upwards of a Gigawatt. Most of the machines with which we are familiar are rotating, but linear electric motors are widely used, from shuttle drives in weaving machines to equipment handling and amusement park rides. Currently under development are large linear induction machines to be used to launch aircraft. It is our purpose in this subject to develop an analytical basis for understanding how all of these dierent machines work. We start, however, with a picture of perhaps the most common of electric machines.

Electric Machine Description:

Figure 1 is a cartoon drawing of a conventional induction motor. This is a very common type of electric machine and will serve as a reference point. Most other electric machines operate in 1

a fashion which is the same as the induction machine or which dier in ways which are easy to reference to the induction machine. Most (but not all!) machines we will be studying have essentially this morphology. The rotor of the machine is mounted on a shaft which is supported on some sort of bearing(s). Usually, but not always, the rotor is inside. I have drawn a rotor which is round, but this does not need to be the case. I have also indicated rotor conductors, but sometimes the rotor has permanent magnets either fastened to it or inside, and sometimes (as in Variable Reluctance Machines) it is just an oddly shaped piece of steel. The stator is, in this drawing, on the outside and has windings. With most of the machines we will be dealing with, the stator winding is the armature, or electrical power input element. (In DC and Universal motors this is reversed, with the armature contained on the rotor: we will deal with these later). In most electrical machines the rotor and the stator are made of highly magnetically permeable materials: steel or magnetic iron. In many common machines such as induction motors the rotor and stator are both made up of thin sheets of silicon steel. Punched into those sheets are slots which contain the rotor and stator conductors. Figure 2 is a picture of part of an induction machine distorted so that the air-gap is straightened out (as if the machine had innite radius). This is actually a convenient way of drawing the machine and, we will nd, leads to useful methods of analysis.

Stator Core Stator Conductors In Slots

Air Gap

Rotor Conductors In Slots

Figure 2: Windings in Slots What is important to note for now is that the machine has an air gap g which is relatively small (that is, the gap dimension is much less than the machine radius r). The air-gap also has a physical length l. The electric machine works by producing a shear stress in the air-gap (with of course side eects such as production of back voltage). It is possible to dene the average airgap shear stress, which we will refer to as . Total developed torque is force over the surface area times moment (which is rotor radius): T = 2r 2 < > Power transferred by this device is just torque times speed, which is the same as force times

surface velocity, since surface velocity is u = r:

Pm = T = 2r < > u If we note that active rotor volume is , the ratio of torque to volume is just: T =2< > Vr Now, determining what can be done in a volume of machine involves two things. First, it is clear that the volume we have calculated here is not the whole machine volume, since it does not include the stator. The actual estimate of total machine volume from the rotor volume is actually quite complex and detailed and we will leave that one for later. Second, we need to estimate the value of the useful average shear stress. Suppose both the radial ux density Br and the stator surface current density Kz are sinusoidal ux waves of the form: Br = 2B0 cos (p t) Kz = 2K0 cos (p t) Note that this assumes these two quantities are exactly in phase, or oriented to ideally produce torque, so we are going to get an optimistic bound here. Then the average value of surface traction is: 1 2 < >= Br Kz d = B0 K0 2 0 The magnetic ux density that can be developed is limited by the characteristics of the magnetic materials (iron) used. Current densities are a function of technology and are typically limited by how much eort can be put into cooling and the temperature limits of insulating materials. In practice, the range of shear stress encountered in electric machinery technology is not terribly broad: ranging from a few kPa in smaller machines to about 100 kPa in very large, well cooled machines. It is usually said that electric machines are torque producing devices, meaning tht they are dened by this shear stress mechanism and by physical dimensions. Since power is torque times rotational speed, high power density machines necessarily will have high shaft speeds. Of course there are limits on rotational speed as well, arising from centrifugal forces which limit tip velocity. Our rst step in understanding how electric machinery works is to understand the mechanisms which produce forces of electromagnetic origin.

Energy Conversion Process:

In a motor the energy conversion process can be thought of in simple terms. In steady state, electric power input to the machine is just the sum of electric power inputs to the dierent phase terminals: Pe = Mechanical power is torque times speed: Pm = T 3

vi ii

Electro-

Electric Power In

Mechanical Converter

Mechanical Power Out

Losses: Heat, Noise, Windage,...

Figure 3: Energy Conversion Process And the sum of the losses is the dierence: Pd = Pe Pm It will sometimes be convenient to employ the fact that, in most machines, dissipation is small enough to approximate mechanical power with electrical power. In fact, there are many situations in which the loss mechanism is known well enough that it can be idealized away. The thermodynamic arguments for force density take advantage of this and employ a conservative or lossless energy conversion system.

3.1

Energy Approach to Electromagnetic Forces:

f + v Magnetic Field System x

Figure 4: Conservative Magnetic Field System To start, consider some electromechanical system which has two sets of terminals, electrical and mechanical, as shown in Figure 4. If the system stores energy in magnetic elds, the energy stored depends on the state of the system, dened by (in this case) two of the identiable variables: ux (), current (i) and mechanical position (x). In fact, with only a little reection, you should be able to convince yourself that this state is a single-valued function of two variables and that the energy stored is independent of how the system was brought to this state. Now, all electromechanical converters have loss mechanisms and so are not themselves conservative. However, the magnetic eld system that produces force is, in principle, conservative in the 4

sense that its state and stored energy can be described by only two variables. The history of the system is not important. It is possible to chose the variables in such a way that electrical power into this conservative system is: d P e = vi = i dt Similarly, mechanical power out of the system is: Pm = fe dx dt

The dierence between these two is the rate of change of energy stored in the system: dWm = Pe Pm dt It is then possible to compute the change in energy required to take the system from one state to another by:
a

Wm (a) Wm (b) =

id f e dx

where the two states of the system are described by a = (a , xa ) and b = (b , xb ) If the energy stored in the system is described by two state variables, and x, the total dierential of stored energy is: Wm Wm d + dx dWm = x and it is also: dWm = id f e dx So that we can make a direct equivalence between the derivatives and: fe = Wm x

In the case of rotary, as opposed to linear, motion, torque T e takes the place of force f e and angular displacement takes the place of linear displacement x. Note that the product of torque and angle has the same units as the product of force and distance (both have units of work, which in the International System of units is Newton-meters or Joules. In many cases we might consider a system which is electricaly linear, in which case inductance is a function only of the mechanical position x. (x) = L(x)i In this case, assuming that the energy integral is carried out from = 0 (so that the part of the integral carried out over x is zero), Wm = This makes

1 1 2 d = L(x) 2 L(x)

1 1 f e = 2 2 x L(x) 5

Note that this is numerically equivalent to 1 f e = i2 L(x) 2 x This is true only in the case of a linear system. Note that substituting L(x)i = too early in the derivation produces erroneous results: in the case of a linear system it produces a sign error, but in the case of a nonlinear system it is just wrong. 3.1.1 Example: simple solenoid

Consider the magnetic actuator shown in cartoon form in Figure 5. The actuator consists of a circular rod of ferromagnetic material (very highly permeable) that can move axially (the xdirection) inside of a stationary piece, also made of highly permeable material. A coil of N turns carries a current I. The rod has a radius R and spacing from the at end of the stator is the variable dimension x. At the other end there is a radial clearance between the rod and the stator g. Assume g R. If the axial length of the radial gaps is = R/2, the area of the radial gaps is the same as the area of the gap between the rod and the stator at the variable gap.
R/2

g C L

N turns

Figure 5: Solenoid Actuator

The permeances of the variable width gap is:
P1 = 0 R2 x

and the permeance of the radial clearance gap is, if the gap dimension is small compared with the radius: 20 R 0 R2 P2 = = g g The inductance of the coil system is: L= Magnetic energy is: 6 P1 P2 0 R2 N 2 N2 = N2 = R1 + R2 P2 + P2 x+g

Wm =

id =

2 x + g 1 2 = 0 2 L(x) 2 0 R2 N 2

And then, of course, force of electric origin is: fe = Here that is easy to carry out: 1 d 1 = dx L 0 R2 N 2 So that the force is: f e (x) = 1 2 0 2 0 R2 N 2 Wm 2 d 1 = 0 x 2 dx L(x)

Given that the system is to be excited by a current, we may at this point substitute for ux: = L(x)i = and then total force may be seen to be: fe = 0 R2 N 2 i2 (x + g)2 2 0 R2 N i x+g

The force is negative in the sense that it tends to reduce x, or to close the gap. 3.1.2 Multiply Excited Systems

There may be (and in most electric machine applications there will be) more than one source of electrical excitation (more than one coil). In such systems we may write the conservation of energy expression as: dWm =

ik dk f e dx

which simply suggests that electrical input to the magnetic eld energy storage is the sum (in this case over the index k) of inputs from each of the coils. To nd the total energy stored in the system it is necessary to integrate over all of the coils (which may and in general will have mutual inductance). Wm =

i d

Of course, if the system is conservative, Wm (1 , 2 , . . . , x) is uniquely specied and so the actual path taken in carrying out this integral will not aect the value of the resulting energy. 7

3.1.3

Coenergy

We often will describe systems in terms of inductance rather than its reciprocal, so that current, rather than ux, appears to be the relevant variable. It is convenient to derive a new energy variable, which we will call co-energy, by:
Wm =

i ii Wm

so that force produced is:

and in this case it is quite easy to show that the energy dierential is (for a single mechanical variable) simply: dWm = k dik + f e dx
k

fe =

Wm x

3.2

Example: Synchronous Machine

Stator Gap Rotor B F A C B A C F F F C A B C A B

Figure 6: Cartoon of Synchronous Machine Consider a simple electric machine as pictured in Figure 6 in which there is a single winding on a rotor (call it the eld winding and a polyphase armature with three identical coils spaced at uniform locations about the periphery. We can describe the ux linkages as: a = La ia + Lab ib + Lab ic + M cos(p)if b = Lab ia + La ib + Lab ic + M cos(p c f 2 )if 3 2 = Lab ia + Lab ib + La ic + M cos(p + )if 3 2 2 = M cos(p)ia + M cos(p )ib + M cos(p + ) + Lf if 3 3

It is assumed that the ux linkages are sinusoidal functions of rotor position. As it turns out, many electrical machines work best (by many criteria such as smoothness of torque production) 8

if this is the case, so that techniques have been developed to make those ux linkages very nearly sinusoidal. We will see some of these techniques in later chapters of these notes. For the moment, we will simply assume these dependencies. In addition, we assume that the rotor is magnetically round, which means the stator self inductances and the stator phase to phase mutual inductances are not functions of rotor position. Note that if the phase windings are identical (except for their angular position), they will have identical self inductances. If there are three uniformly spaced windings the phase-phase mutual inductances will all be the same. Now, this system can be simply described in terms of coenergy. With multiple excitation it is important to exercise some care in taking the coenergy integral (to ensure that it is taken over a valid path in the multi-dimensional space). In our case there are actually ve dimensions, but only four are important since we can position the rotor with all currents at zero so there is no contribution to coenergy from setting rotor position. Suppose the rotor is at some angle and that the four currents have values ia0 , ib0 , ic0 and if 0 . One of many correct path integrals to take would be:
Wm =

ia0

0 ib0

La ia dia (Lab ia0 + La ib ) dib (Lab ia0 + Lab ib0 + La ic ) dic

if 0

+ + + The result is:

0 ic0

2 2 )ib0 + M cos(p + )ic0 + Lf if dif M cos(p)ia0 + M cos(p 3 3

Wm =

1 2 La ia0 + i2 + i2 + Lab (iao ib0 + iao ic0 + ico ib0 ) b0 co 2 2 1 2 ) + ic0 cos(p + ) + Lf i2 0 +M if 0 ia0 cos(p) + ib0 cos(p f 3 3 2

Since there are no variations of the stator inductances with rotor position , torque is easily given by:
Wm 2 2 Te = = pM if 0 ia0 sin(p) + ib0 sin(p ) + ico sin(p + ) 3 3

3.2.1

Current Driven Synchronous Machine

Now assume that we can drive this thing with currents: ia0 = Ia cos t ib0 ic0 if 0 2 = Ia cos t 3 2 = Ia cos t + 3 = If

and assume the rotor is turning at synchronous speed: p = t + i Noting that cos x sin y
=
1 2

sin(x y) +
1 sin(x + y), we nd the torque expression above to be: 2 1 1 sin i + sin (2t + i ) 2 2 1 1 sin i + sin 2t + i + 2 2 1 1 + sin i + sin 2t + i + 2 2

Te = pM Ia If

4 3 4 3

The sine functions on the left add and the ones on the right cancel, leaving: 3 Te = pM Ia If sin i 2 And this is indeed one way of looking at a synchronous machine, which produces steady torque if the rotor speed and currents all agree on frequency. Torque is related to the current torque angle i . As it turns out such machines are not generally run against current sources, but we will take up actual operation of such machines later.

Field Descriptions: Continuous Media

While a basic understanding of electromechanical devices is possible using the lumped parameter approach and the principle of virtual work as described in the previous section, many phenomena in electric machines require a more detailed understanding which is aorded by a continuum approach. In this section we consider a elds-based approach to energy ow using Poyntings Theorem and then a elds based description of forces using the Maxwell Stress Tensor. These techniques will both be useful in further analysis of what happens in electric machines.

4.1

Field Description of Energy Flow: Poytings Theorem

B E = t

Start with Faradays Law:

and Amperes Law:

Multiplying the rst of these by H and the second by E and taking the dierence:
B E J H E E H = E H = H dt

H =J

On the left of this expression is the divergence of electromagnetic energy ow: S =EH 10

 Here, S is the celebrated Poynting ow which describes power in an electromagnetic eld sysstem. (The units of this quantity is watts per square meter in the International System). On the right hand side are two terms: H B is rate of change of magnetic stored energy. The second dt J looks a lot like power dissipation. We will discuss each of these in more detail. For the term, E moment, however, note that the divergence theorem of vector calculus yields:

volume

Sdv =

nda S

that is, the volume integral of the divergence of the Poynting energy ow is the same as the Poynting energy ow over the surface of the volume in question. This integral becomes:

n S da =

B EJ +H t volume

dv

which is simply a realization that the total energy ow into a region of space is the same as the volume integral over that region of the rate of change of energy stored plus the term that looks like dissipation. Before we close this, note that, if there is motion of any material within the system, we can use the empirical expression for transformation of electric eld between observers moving with respect to each other. Here the primed frame is moving with respeect to the unprimed frame with the velocity v v E = E + B This transformation describes, for example, the motion of a charged particle such as an electron under the inuence of both electric and magnetic elds. Now, if we assume that there is material motion in the system we are observing and if we assign to be the velocity of that material, so that v is measured in a frame in which thre is no material motion (that is the frame of the material E itself), the product of electric eld and current density becomes: v v v E J = E B J = E J B J = E J + J B

In the last step we used the fact that in a scalar triple product the order of the scalar (dot) and vector (cross) products can be interchanged and that reversing the order of terms in a vector (cross) product simply changes the sign of that product. Now we have a ready interpretation for what we have calculated: If the primed coordinate system is actually the frame of material motion, 1 E J = |J |2 which is easily seen to be dissipation and is positive denite if material conductivity is positive. The last term is obviously conversion of energy from electromagnetic to mechanical form: J B =F v v where we have now identied force density to be:

F =J B 11

This is the Lorentz Force Law, which describes the interaction of current with magnetic eld to produce force. It is not, however, the complete story of force production in electromechanical systems. As we learned earlier, changes in geometry which aect magnetic stored energy can also produce force. Fortunately, a complete description of electromechanical force is possible using only magnetic elds and that is the topic of our next section.

4.2

Field Description of Forces: Maxwell Stress Tensor

Forces of electromagnetic origin, because they are transferred by electric and magnetic elds, are the result of those elds and may be calculated once the elds are known. In fact, if a surface can be established that fully encases a material body, the force on that body can be shown to be the integral of force density, or traction over that surface. The traction derived by taking the cross product of surface current density and ux density on the air-gap surface of a machine (above) actually makes sense in view of the empirically derived Lorentz Force Law: Given a (vector) current density and a (vector) ux density. This is actually enough to describe the forces we see in many machines, but since electric machines have permeable magnetic material and since magnetic elds produce forces on permeable material even in the absence of macroscopic currents it is necessary to observe how force appears on such material. A suitable empirical expression for force density is:
1 H H F = J B 2

where H is the magnetic eld intensity and is the permeability.

Now, note that current density is the curl of magnetic eld intensity, so that:
F
1 H H H H 2 1 H H = H H 2

And, since:

force density is: F

1 H H = H H H H 2

1 1 = H H H H H H 2 2 1 H H = H H 2
k

This expression can be written by components: the component of force in the ith dimension is: Fi = Hk xk

Hi xi

1 2 H 2 k k

The rst term can be written as:

Hk

xk

Hi =

xk 12

Hk Hi Hi

xk

Hk

The last term in this expression is easily shown to be divergence of magnetic ux density, which is zero: Hk = 0 B = xk k Using this, we can write force density in a more compact form as: Fk = xi

2 Hn Hi Hk ik 2 n

where we have used the Kroneker delta ik = 1 if i = k, 0 otherwise. Note that this force density is in the form of the divergence of a tensor: Fk = or Tik xi

F =T

In this case, force on some object that can be surrounded by a closed surface can be found by using the divergence theorem: f=

vol

F dv =

vol

T dv =

T da n

or, if we note surface traction to be i = total force in direction i is just: f=

Tik nk , where n is the surface normal vector, then the

k

i da =

Tik nk da

The interpretation of all of this is less dicult than the notation suggests. This eld description of forces gives us a simple picture of surface traction, the force per unit area on a surface. If we just integrate this traction over the area of some body we get the whole force on the body. Note one more thing about this notation. Sometimes when subscripts are repeated as they are here the summation symbol is omitted. Thus we would write i = k Tik nk = Tik nk .

4.3

Example: Linear Induction Machine

Figure 7 shows a highly simplied picture of a single sided linear induction motor. This is not how most linear induction machines are actually built, but it is possible to show through symmetry arguments that the analysis we can carry out here is actually valid for other machines of this class. This machine consists of a stator (the upper surface) which is represented as a surface current on the surface of a highly permeable region. The moving element consists of a thin layer of conducting material on the surface of a highly permeable region. The moving element (or shuttle) has a velocity u with respect to the stator and that motion is in the x direction. The stator surface current density is assumed to be: Kz = Re K z ej(tkx) 13

111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000

s
K s

y u x

Figure 7: Simple Model of single sided linear induction machine Note that we are ignoring some important eects, such as those arising from nite length of the stator and of the shuttle. Such eects can be quite important, but we will leave those until later, as they are what make linear motors interesting. Viewed from the shuttle for which the dimension in the direction of motion is x x ut , the relative frequency is: t kx = ( ku) t kx = s t kx

Now, since the shuttle surface can support a surface current and is excited by magnetic elds which are in turn excited by the stator currents, it is reasonable to assume that the form of rotor current is the same as that of the stator: Ks = Re K s ej(s tkx ) Amperes Law is, in this situation: g which is, in complex amplitudes: Hy = Kz + Ks jkg Hy = Kz + Ks x

The y- component of Faradays Law is, assuming the problem is uniform in the z- direction: js B y = jkE z s 0 H y k A bit of algebraic manipulation yields expressions for the complex amplitudes of rotor surface current and gap magnetic eld: E = z
j 0k2sgs

or

Ks = Hy =

0 1 + j muk2 gs s

Kz

Kz j 0 kg 1 + j muk2 gs s 14

To nd surface traction, the Maxwell Stress Tensor can be evaluated at a surface just below the stator (on this surface the x- directed magnetic eld is simply H x = K z . Thus the traction is x = Txy = 0 Hx Hy and the average of this is: < x >= This is: < x >=
0 Re H x H y 2

Now, if we consider electromagnetic power ow (Poyntings Theorem): in the y- direction: Sy = Ez Hx And since in the frame of the shuttle E = s 0 H y z k
< Sy >= 0 s s

2 0 s s 0 1 |K z | k2 g 2 kg 1 + 0 s s 2 k2 g

Similarly, evaluated in the frame of the stator: < Sy >= < x > k This shows what we already suspected: the electromagnetic power ow from the stator is the force density on the shuttle times the wave velocity. The electromagnetic ower ow into the shuttle is the same force density times the slip velocity. The dierence between these two is the power converted to mechanical form and it is the force density times the shuttle velocity.

1 s 0 s k2 g 2 < x > 2 |K z | = k 2 k kg 1 + 0 s s
k2 g

4.4

Rotating Machines

The use of this formulation in rotating machines is a bit tricky because, at lest formally, directional vectors must have constant identity if an integral of forces is to become a total force. In cylindrical coordinates, of course, the directional vectors are not of constant identity. However, with care and understanding of the direction of traction and how it is integrated we can make use of the MST approach in rotating electric machines. Now, if we go back to the case of a circular cylinder and are interested in torque, it is pretty clear that we can compute the circumferential force by noting that the normal vector to the cylinder is just the radial unit vector, and then the circumferential traction must simply be: = 0 H r H Assuming that there are no uxes inside the surface of the rotor, simply integrating this over the surface gives azimuthal force. In principal this is the same as surrounding the surface of the rotor by a continuum of innitely small boxes, one surface just outside the rotor and with a normal facing outward, the other surface just inside with normal facing inward. (Of course the MST is zero on this inner surface). Then multiplying by radius (moment arm) gives torque. The last step is to note that, if the rotor is made of highly permeable material, the azimuthal magnetic eld just outside the rotor is equal to surface current density. 15

Generalization to Continuous Media

Now, consider a system with not just a multiplicity of circuits but a continuum of current-carrying paths. In that case we could identify the co-energy as:
Wm

area

( )dJ d a a

where that area is chosen to cut all of the current carrying conductors. This area can be picked to be perpedicular to each of the current laments since the divergence of current is zero. The ux is calculated over a path that coincides with each current lament (such paths exist since current has zero divergence). Then the ux is: ( ) = a

n B d

Now, if we use the vector potential A for which the magnetic ux density is: B =A the ux linked by any one of the current laments is: ( ) = a

A d

where d is the path around the current lament. This implies directly that the coenergy is:
Wm

area

dJ a A d d

Now: it is possible to make d coincide with d and be parallel to the current laments, so that: a
Wm =

vol

A dJdv

5.1

Permanent Magnets

Permanent magnets are becoming an even more important element in electric machine systems. Often systems with permanent magnets are approached in a relatively ad-hoc way, made equivalent to a current that produces the same MMF as the magnet itself. The constitutive relationship for a permanent magnet relates the magnetic ux density B to magnetic eld H and the property of the magnet itself, the magnetization M . B = 0 H + M

Now, the eect of the magnetization is to act as if there were a current (called an amperian current) with density: J = M Note that this amperian current acts just like ordinary current in making magnetic ux density. Magnetic co-energy is: Wm = A dM dv vol 16

 Next, note the vector identity C D = D C C D Now,

Wm =

vol

A dM dv +

vol

A dM dv

Then, noting that B = A:

Wm =

s A dM d +

vol

B dM dv

The rst of these integrals (closed surface) vanishes if it is taken over a surface just outside the magnet, where M is zero. Thus the magnetic co-energy in a system with only a permanent magnet source is Wm = B dM dv vol Adding current carrying coils to such a system is done in the obvious way.

17

Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science 6.685 Electric Machines Class Notes 2 Magnetic Circuit Basics c 2003 James L. Kirtley Jr. September 5, 2005

Introduction

Magnetic Circuits oer, as do electric circuits, a way of simplifying the analysis of magnetic eld systems which can be represented as having a collection of discrete elements. In electric circuits the elements are sources, resistors and so forth which are represented as having discrete currents and voltages. These elements are connected together with wires and their behavior is described by network constraints (Kirkhos voltage and current laws) and by constitutive relationships such as Ohms Law. In magnetic circuits the lumped parameters are called Reluctances (the inverse of Reluctance is called Permeance). The analog to a wire is referred to as a high permeance magnetic circuit element. Of course high permeability is the analog of high conductivity. By organizing magnetic eld systems into lumped parameter elements and using network constraints and constitutive relationships we can simplify the analysis of such systems.

Electric Circuits

First, let us review how Electric Circuits are dened. We start with two conservation laws: conservation of charge and Faradays Law. From these we can, with appropriate simplifying assumptions, derive the two fundamental circiut constraints embodied in Kirkhos laws.

2.1

KCL

Conservation of charge could be written in integral form as: df dv = 0 (1) volume dt This simply states that the sum of current out of some volume of space and rate of change of free charge in that space must be zero. Now, if we dene a discrete current to be the integral of current density crossing through a part of the surface:

J nda +

ik =

surfacek

J nda

(2)

and if we assume that there is no accumulation of charge within the volume (in ordinary circuit theory the nodes are small and do not accumulate charge), we have:

J nda =

ik = 0

(3)

which holds if the sum over the index k includes all current paths into the node. This is, of course, KCL.

2.2

KVL

Faradays Law is, in integral form: d B nda (4) dt where the closed loop in the left hand side of the equation is the edge of the surface of the integral on the right hand side. Now if we dene voltage in the usual way, between points a and b for element k:

E d =

vk =

bk

ak

E d

(5)

Then, if we assume that the right-hand side of Faradays Law (that is, magnetic induction) is zero, the loop equation becomes:

vk = 0

(6)

This works for circuit analysis because most circuits do not involve magnetic induction in the loops. However, it does form the basis for much head scratching over voltages encountered by ground loops.

2.3

Constitutive Relationship: Ohms Law

Many of the materials used in electric circuits carry current through a linear conduction mechanism. That is, the relationship between electric eld and electric current density is J = E (7)

Suppose, to start, we can identify a piece of stu which has constant area and which is carrying current over some nite length, as shown in Figure 1. Assume this rod is carrying current density J (We wont say anything about how this current density managed to get into the rod, but assume that it is connected to something that can carry current (perhaps a wire....). Total current carried by the rod is simply I = |J|A and then voltage across the element is: v=

E d =

I A

from which we conclude the resistance is R= V = I A

Figure 1: Simple Rod Shaped Resistor Of course we can still employ the lumped parameter picture even with elements that are more complex. Consider the annular resistor shown in Figure 2. This is an end-on view of something which is uniform in cross-section and has depth D in the direction you cant see. Assume that the inner and outer elements are very good conductors, relative to the annular element in between. Assume further that this element has conductivity and inner and outer radii Ri and Ro , respectively.

+
v

Electrodes

Resistive Material

Figure 2: Annular Resistor Now, if the thing is carrying current from the inner to the outer electrode, current density would be: J = ir Jr (r) = Electric eld is Er = Then voltage is v=

Ro Ri

I 2Dr

Jr I = 2Dr Ro I log 2D Ri
Ro Ri

Er (r) =

so that we conclude the resistance of this element is R= log

2D

Magnetic Circuit Analogs

In the electric circuit, elements for which voltage and current are dened are connected together by elements thought of as wires, or elements with zero or negligible voltage drop. The interconnection points are nodes. In magnetic circuits the analogous thing occurs: elements for which magnetomotive force and ux can be dened are connected together by high permeability magnetic circuit elements (usually iron) which are the analog of wires in electric circuits.

3.1

Analogy to KCL

Gauss Law is:

B nda = 0

(8)

which means that the total amount of ux coming out of a region of space is always zero. Now, we will dene a quantity which is sometimes called simply ux or a ux tube. This might be thought to be a collection of ux lines that can somehow be bundled together. Generally it is the ux that is identied with a magnetic circuit element. Mathematically it is: k =

B nda

(9)

In most cases, ux as dened above is carried in magnetic circuit elements which are made of high permeability material, analogous to the wires of high conductivity material which carry current in electric circuits. It is possible to show that ux is largely contained in such high permeability materials. If all of the ux tubes out of some region of space (node) are considered in the sum, they must add to zero: k = 0 (10)
k

3.2

Analogy to KVL: MMF

Amperes Law is

H d =

Where, as for Faradays Law, the closed contour on the left is the periphery of the (open) surface on the right. Now we dene what we call Magnetomotive Force in direct analog to Electromotive , Force, (voltage). Fk =

bk

J nda

(11)

ak

H d

(12)

Further, dene the current enclosed by a loop to be: F0 = Then the analogy to KVL is:
k

J nda

(13)

Fk = F0 4

Note that the analog is not exact as there is a source term on the right hand side whereas KVL has no source term. Note also that sign counts here. The closed integral is taken in such direction so that the positive sense of the surface enclosed is positive (upwards) when the surface is to the left of the contour. (This is another way of stating the celebrated right hand rule: if you wrap your right hand around the contour with your ngers pointing in the direction of the closed contour integration, your thumb is pointing in the positive direction for the surface).

3.3

Analog to Ohms Law: Reluctance

Consider a gap between two high permeability pieces as shown in Figure 3. If we assume that their permeability is high enough, we can assume that there is no magnetic eld H in them and so the MMF or magnetic potential is essentially constant, just like in a wire. For the moment, assume that the gap dimension g is small and uniform over the gap area A. Now, assume that some ux is owing from one of these to the other. That ux is = BA where B is the ux density crossing the gap and A is the gap area. Note that we are ignoring fringing elds in this simplied analysis. This neglect often requires correction in practice. Since the permeability of free space is 0 , (assuming the gap is indeed lled with free space), magnetic eld intensity is B H= 0 and gap MMF is just magnetic eld intensity times gap dimension. This, of course, assumes that the gap is uniform and that so is the magnetic eld intensity: F = B g 0

Which means that the reluctance of the gap is the ratio of MMF to ux: R= F g = 0 A

y Area A g
Figure 3: Air Gap

3.4

Simple Case

Consider the magnetic circuit situation shown in Figure 4. Here there is a piece of highly permeable material shaped to carry ux across a single air-gap. A coil is wound through the window in the magnetic material (this shape is usually referred to as a core). The equivalent circuit is shown in Figure 5.

Region 1

Region 2 I

Figure 4: Single air-capped Core Note that in Figure 4, if we take as the positive sense of the closed loop a direction which goes vertically upwards through the leg of the core through the coil and then downwards through the gap, the current crosses the surface surrounded by the contour in the positive sense direction.

+
F = N I

Figure 5: Equivalent Circuit

3.5

Flux Connement

The gap in this case has the same reluctance as computed earlier, so that the ux in the gap is simply = N I . Now, by focusing on the two regions indicated we might make a few observations R about magnetic circuits. First, consider region 1 as shown in Figure 6.

Figure 6: Flux Connement Boundary: This is Region 1

In this picture, note that magnetic eld H parallel to the surface must be the same inside the material as it is outside. Consider Amperes Law carried out about a very thin loop consisting of the two arrows drawn at the top boundary of the material in Figure 6 with very short vertical paths joining them. If there is no current singularity inside that loop, the integral around it must be zero which means the magnetic eld just inside must be the same as the magnetic eld outside. Since the material is very highly permeable and B = H, and highly permeable means is very large, unless B is really large, H must be quite small. Thus the magnetic circuit has small magnetic eld H and therfore ux densities parallel to and just outside its boundaries aer also small.

B is perpendicular

Figure 7: Gap Boundary At the surface of the magnetic material, since the magnetic eld parallel to the surface must be very small, any ux lines that emerge from the core element must be perpendicular to the surface as shown for the gap region in Figure 7. This is true for region 1 as well as for region 2, but note that the total MMF available to drive elds across the gap is the same as would produce eld lines from the area of region 1. Since any lines emerging from the magnetic material in region 1 would have very long magnetic paths, they must be very weak. Thus the magnetic circuit material largely connes ux, with only the relatively high permeance (low reluctance) gaps carrying any substantive amount of ux.

3.6

Example: C-Core

Consider a gapped c-core as shown in Figure 8. This is two pieces of highly permeable material shaped generally like Cs. They have uniform depth in the direction you cannot see. We will call that dimension D. Of course the area A = wD, where w is the width at the gap. We assume the two gaps have the same area. Each of the gaps will have a reluctance R= g 0 A

Suppose we wind a coil with N turns on this core as shown in Figure 9. Then we put a current I in that coil. The magnetic circuit equivalent is shown in Figure 10. The two gaps are in series and, of course, in series with the MMF source. Since the two uxes are the same and the MMFs add: F0 = N I = F1 + F2 = 2R and then = NI 0 AN I = 2R 2g 7

g Area A

Figure 8: Gapped Core

N Turns I

Figure 9: Wound, Gapped Core and corresponding ux density in the gaps would be: By = 0 N I 2g

3.7

Example: Core with Dierent Gaps

As a second example, consider the perhaps oddly shaped core shown in Figure 11. Suppose the gap on the right has twice the area as the gap on the left. We would have two gap reluctances: g g R2 = R1 = 0 A 20 A Since the two gaps are in series the ux is the same and the total reluctance is R= Flux in the magnetic circuit loop is = F 2 0 AN I = R 3 g 2 0 N I = A 3 g 3 g 2 0 A

and the ux density across, say, the left hand gap would be: By =

Figure 10: Equivalent Magnetic Circuit

N Turns I

Figure 11: Wound, Gapped Core: Dierent Gaps

Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science 6.685 Electric Machines Class Notes 3: Eddy Currents, Surface Impedances and Loss Mechanisms September 5, 2005 c 2005 James L. Kirtley Jr.

Introduction

Losses in electric machines arise from conduction and magnetic hysteresis. Conduction losses are attributed to straightforward transport conduction and to eddy currents. Transport losses are relatively easy to calculate so we will not pay them much attention. Eddy currents are more interesting and result in frequency dependent conduction losses in machines. Eddy currents in linear materials can often be handled rigorously, but eddy currents in saturating material are more dicult and are often handled in a heuristic fashion. We present here both analytical and semi-emiprical ways of dealing with such losses. We start with surface impedance: the ratio of electric eld to surface current. This is important not just in calculating machine losses, but also in describing how some machines operate.

Surface Impedance of Uniform Conductors

The objective of this section is to describe the calculation of the surface impedance presented by a layer of conductive material. Two problems are considered here. The rst considers a layer of linear material backed up by an innitely permeable surface. This is approximately the situation presented by, for example, surface mounted permanent magnets and is probably a decent approximation to the conduction mechanism that would be responsible for loss due to asynchronous harmonics in these machines. It is also appropriate for use in estimating losses in solid rotor induction machines and in the poles of turbogenerators. The second problem, which we do not work here but simply present the previously worked solution, concerns saturating ferromagnetic material.

2.1

Linear Case

The situation and coordinate system are shown in Figure 1. The conductive layer is of thicknes T and has conductivity and permeability 0 . To keep the mathematical expressions within bounds, we assume rectilinear geometry. This assumption will present errors which are small to the extent that curvature of the problem is small compared with the wavenumbers encountered. We presume that the situation is excited, as it would be in an electric machine, by a current sheet of the form Kz = Re Kej(tkx) In the conducting material, we must satisfy the diusion equation: 2 H = 0 H t

Conductive Region

Hx

Kz
y x

Permeable Surface
Figure 1: Axial View of Magnetic Field Problem In view of the boundary condition at the back surface of the material, taking that point to be y = 0, a general solution for the magnetic eld in the material is: Hx = Re A sinh yej(tkx) Hy where the coecient satises: 2 = j0 + k2 and note that the coecients above are chosen so that H has no divergence. Note that if k is small (that is, if the wavelength of the excitation is large), this spatial coecient becomes 1+j = where the skin depth is: 2 = 0 Faradays law: E = gives: Now: the surface current is just K s = H x so that the equivalent surface impedance is: Z= Ez = j0 coth T H x B t

k = Re j A cosh yej(tkx)

E z = 0 H y k

A pair of limits are interesting here. Assuming that the wavelength is long so that k is negligible, then if T is small (i.e. thin material), Z j0 1 = 2T T 2

On the other hand as T ,

1+j Next it is necessary to transfer this surface impedance across the air-gap of a machine. So, assume a new coordinate system in which the surface of impedance Z s is located at y = 0, and we wish to determine the impedance Z = E z /H x at y = g. In the gap there is no current, so magnetic eld can be expressed as the gradient of a scalar potential which obeys Laplaces equation: Z H =

and Ignoring a common factor of ej(tkx) , we can express H in the gap as: H x = jk + eky + eky

2 = 0

H y = k + eky eky At the surface of the rotor, or E z = H x Z s

0 + = jkZ s + + and then, at the surface of the stator,

kg kg E + e e Z = z = j0 k + ekg + ekg Hx

A bit of manipulation is required to obtain: Z = j0 k

ekg (0 jkZ s ) ekg (0 + jkZ s ) ekg (0 jkZ s ) + ekg (0 + jkZ s ) 0 k2 g

It is useful to note that, in the limit of Z s , this expression approaches the gap impedance Zg = j and, if the gap is small enough that kg 0, Z Z g ||Z s

Iron

Electric machines employ ferromagnetic materials to carry magnetic ux from and to appropriate places within the machine. Such materials have properties which are interesting, useful and problematical, and the designers of electric machines must deal with this stu. The purpose of this note is to introduce the most salient properties of the kinds of magnetic materials used in electric machines. We will be concerned here with materials which exhibit magnetization: ux density is something other than B = 0 H. Generally, we will speak of hard and soft magnetic materials. Hard materials are those in which the magnetization tends to be permanent, while soft materials are used in magnetic circuits of electric machines and transformers. Since they are related we will nd ourselves talking about them either at the same time or in close proximity, even though their uses are widely disparite.

3.1

Magnetization:

It is possible to relate, in all materials, magnetic ux density to magnetic eld intensity with a consitutive relationship of the form: B = 0 H + M

where magnetic eld intensity H and magnetization M are the two important properties. Now, in linear magnetic material magnetization is a simple linear function of magnetic eld: M = m H so that the ux density is also a linear function: B = 0 (1 + m ) H Note that in the most general case the magnetic susceptibility m might be a tensor, leading to ux density being non-colinear with magnetic eld intensity. But such a relationship would still be linear. Generally this sort of complexity does not have a major eect on electric machines.

3.2

Saturation and Hysteresis

In useful magnetic materials this nice relationship is not correct and we need to take a more general view. We will not deal with the microscopic picture here, except to note that the magnetization is due to the alignment of groups of magnetic dipoles, the groups often called domaines. There are only so many magnetic dipoles available in any given material, so that once the ux density is high enough the material is said to saturate, and the relationship between magnetic ux density and magnetic eld intensity is nonlinear. Shown in Figure 2, for example, is a saturation curve for a magnetic sheet steel that is sometimes used in electric machinery. Note the magnetic eld intensity is on a logarithmic scale. If this were plotted on linear coordinates the saturation would appear to be quite abrupt. At this point it is appropriate to note that the units used in magnetic eld analysis are not always the same nor even consistent. In almost all systems the unit of ux is the weber (Wb), 4

Figure 2: Saturation Curve: Commercial M-19 Silicon Iron

Courtesy of United States Steel Corporation. (U.S. Steel). U.S. Steel accepts no liability for reliance on any information contained in the graphs shown above.

Remanent Flux Density B r

Flux Density

Saturation Flux Density Bs Magnetic Field Saturation Field H s

Coercive Field Hc

Figure 3: Hysteresis Curve Nomenclature which is the same as a volt-second. In SI the unit of ux density is the tesla (T), but many people refer to the gauss (G), which has its origin in CGS. 10,000 G = 1 T. Now it gets worse, because there is an English system measure of ux density generally called kilo-lines per square inch. This is because in the English system the unit of ux is the line. 108 lines is equal to a weber. Thus a Tesla is 64.5 kilolines per square inch. The SI and CGS units of ux density are easy to reconcile, but the units of magnetic eld are a bit harder. In SI we generally measure H in amperes/meter (or ampere-turns per meter). Often, however, you will see magnetic eld represented as Oersteds (Oe). One Oe is the same as the magnetic eld required to produce one gauss in free space. So 79.577 A/m is one Oe. In most useful magnetic materials the magnetic domaines tend to be somewhat sticky, and a more-than-incremental magnetic eld is required to get them to move. This leads to the property called hysteresis, both useful and problematical in many magnetic systems. Hysteresis loops take many forms; a generalized picture of one is shown in Figure 3. Salient features of the hysteresis curve are the remanent magnetization Br and the coercive eld Hc . Note that the actual loop that will be traced out is a function of eld amplitude and history. Thus there are many other minor loops that might be traced out by the B-H characteristic of a piece of material, depending on just what the elds and uxes have done and are doing. Hysteresis is important for two reasons. First, it represents the mechanism for trapping magnetic ux in a piece of material to form a permanent magnet. We will have more to say about that anon. Second, hysteresis is a loss mechanism. To show this, consider some arbitrary chunk of material for which we can characterize an MMF and a ux: F = NI =

H d

V dt = N 6

Area

B dA

Energy input to the chunk of material over some period of time is

w=

V Idt =

F d =

H d

dB dA dt

Now, imagine carrying out the second (double) integral over a continuous set of surfaces which are perpendicular to the magnetic eld H. (This IS possible!). The energy becomes: w=

t

H dBdvol dt

and, done over a complete cycle of some input waveform, that is: w = Wm =

vol

Wm dvol

H dB

That last expression simply expresses the area of the hysteresis loop for the particular cycle. Generally, for most electric machine applications we will use magnetic material characterized as soft, having as narrow a hysteresis loop (and therefore as low a hysteretic loss) as possible. At the other end of the spectrum are hard magnetic materials which are used to make permanent magnets. The terminology comes from steel, in which soft, annealed steel material tends to have narrow loops and hardened steel tends to have wider loops. However permanent magnet technology has advanced to the point where the coercive forces possible in even cheap ceramic magnets far exceed those of the hardest steels.

3.3

Conduction, Eddy Currents and Laminations:

Steel, being a metal, is an electrical conductor. Thus when time varying magnetic elds pass through it they cause eddy currents to ow, and of course those produce dissipation. In fact, for almost all applications involving soft iron, eddy currents are the dominant source of loss. To reduce the eddy current loss, magnetic circuits of transformers and electric machines are almost invariably laminated, or made up of relatively thin sheets of steel. To further reduce losses the steel is alloyed with elements (often silicon) which poison the electrical conductivity. There are several approaches to estimating the loss due to eddy currents in steel sheets and in the surface of solid iron, and it is worthwhile to look at a few of them. It should be noted that this is a hard problem, since the behavior of the material itself is dicult to characterize.

3.4

Complete Penetration Case

Consider the problem of a stack of laminations. In particular, consider one sheet in the stack represented in Figure 4. It has thickness t and conductivity . Assume that the skin depth is much greater than the sheet thickness so that magnetic eld penetrates the sheet completely. Further, assume that the applied magnetic ux density is parallel to the surface of the sheets: 2B0 ejt B = iz Re 7

y t z x

Figure 4: Lamination Section for Loss Calculation Now we can use Faradays law to determine the electric eld and therefore current density in the sheet. If the problem is uniform in the x- and z- directions, E x = j0 B0 y Note also that, unless there is some net transport current in the x- direction, E must be antisymmetric about the center of the sheet. Thus if we take the origin of y to be in the center, electric eld and current are: E x = jB0 y Local power dissipated is

J x = jB0 y

|J|2 To nd average power dissipated we integrate over the thickness of the lamination:
2 P (y) = 2 B0 y 2 =

2 < P >= t

t 2

2 2 P (y)dy = 2 B0 t

t 2

y 2 dy =

1 2 2 2 B0 t 12

Pay attention to the orders of the various terms here: power is proportional to the square of ux density and to the square of frequency. It is also proportional to the square of the lamination thickness (this is average volume power dissipation). As an aside, consider a simple magnetic circuit made of this material, with some length and area A, so that volume of material is A. Flux lined by a coil of N turns would be: = N = N AB0 and voltage is of course just V = jwL. Total power dissipated in this core would be: Pc = A V2 1 2 2 2 B0 t = 12 Rc

where the equivalent core resistance is now

Rc = A 12N2
t2 8

B0

Figure 5: Idealized Saturating Characteristic

3.5

Eddy Currents in Saturating Iron

The same geometry holds for this pattern, although we consider only the one-dimensional problem (k 0). The problem was worked by McLean and his graduate student Agarwal [2] [1]. They assumed that the magnetic eld at the surface of the at slab of material was sinusoidal in time and of high enough amplitude to saturate the material. This is true if the material has high permeability and the magnetic eld is strong. What happens is that the impressed magnetic eld saturates a region of material near the surface, leading to a magnetic ux density parallel to the surface. The depth of the region aected changes with time, and there is a separating surface (in the at problem this is a plane) that moves away from the top surface in response to the change in the magnetic eld. An electric eld is developed to move the surface, and that magnetic eld drives eddy currents in the material. Assume that the material has a perfectly rectangular magnetization curve as shown in Figure 5, so that ux density in the x- direction is: Bx = B0 sign(Hx ) The ux per unit width (in the z- direction) is: = and Faradays law becomes:

0

Bx dy

t while Amperes law in conjunction with Ohms law is: Ez = Hx = Ez y Now, McLean suggested a solution to this set in which there is a separating surface at depth below the surface, as shown in Figure 6 . At any given time: Hx = Hs (t) 1 + Jz = Ez = Hs

y x

Bs
Separating Surface

Bs Penetration Depth

Figure 6: Separating Surface and Penetration Depth That is, in the region between the separating surface and the top of the material, electric eld Ez is uniform and magnetic eld Hx is a linear function of depth, falling from its impressed value at the surface to zero at the separating surface. Now: electric eld is produced by the rate of change of ux which is: Ez = = 2Bx t t Eliminating E, we have: Hs 2 = t Bx and then, if the impressed magnetic eld is sinusoidal, this becomes: d 2 H0 = | sin t| dt B0 This is easy to solve, assuming that = 0 at t = 0, =

t 2H0 sin B0 2

Now: the surface always moves in the downward direction (as we have drawn it), so at each half cycle a new surface is created: the old one just stops moving at a maximum position, or penetration depth: 2H0 = B0 This penetration depth is analogous to the skin depth of the linear theory. However, it is an absolute penetration depth. The resulting electric eld is: Ez = t 2H0 cos 2 0 < t <

This may be Fourier analyzed: noting that if the impressed magnetic eld is sinusoidal, only the time fundamental component of electric eld is important, leading to: Ez = 8 H0 (cos t + 2 sin t + . . .) 3 10

Complex surface impedance is the ratio between the complex amplitude of electric and magnetic eld, which becomes: E 8 1 (2 + j) Zs = z = Hx 3 Thus, in practical applications, we can handle this surface much as we handle linear conductive surfaces, by establishing a skin depth and assuming that current ows within that skin depth of 16 the surface. The resistance is modied by the factor of 3 and the power factor of this surface is about 89 % (as opposed to a linear surface where the power factor is about 71 %. Agarwal suggests using a value for B0 of about 75 % of the saturation ux density of the steel.

Semi-Empirical Method of Handling Iron Loss

Neither of the models described so far are fully satisfactory in describing the behavior of laminated iron, because losses are a combination of eddy current and hysteresis losses. The rather simple model employed for eddy currents is precise because of its assumption of abrupt saturation. The hysteresis model, while precise, would require an empirical determination of the size of the hysteresis loops anyway. So we must often resort to empirical loss data. Manufacturers of lamination steel sheets will publish data, usually in the form of curves, for many of their products. Here are a few ways of looking at the data. A low frequency ux density vs. magnetic eld (saturation) curve was shown in Figure 2. Included with that was a measure of the incremental permeability = dB dH

In some machine applications either the total inductance (ratio of ux to MMF) or incremental inductance (slope of the ux to MMF curve) is required. In the limit of low frequency these numbers may be useful. For designing electric machines, however, a second way of looking at steel may be more useful. This is to measure the real and reactive power as a function of magnetic ux density and (sometimes) frequency. In principal, this data is immediately useful. In any well-designed electric machine the ux density in the core is distributed fairly uniformly and is not strongly aected by eddy currents, etc. in the core. Under such circumstances one can determine the ux density in each part of the core. With that information one can go to the published empirical data for real and reactive power and determine core loss and reactive power requirements. Figure 7 shows core loss and apparent power per unit mass as a function of (RMS) induction for 29 gage, fully processed M-19 steel. The two left-hand curves are the ones we will nd most useful. P denotes real power while Pa denotes apparent power. The use of this data is quite straightforward. If the ux density in a machine is estimated for each part of the machine and the mass of steel calculated, then with the help of this chart a total core loss and apparent power can be estimated. Then the eect of the core may be approximated with a pair of elements in parallel with the terminals, with: Rc = Xc = q|V |2 P q|V |2 Q 11

Figure 7: Real and Apparent Loss: M19, Fully Processed, 29 Ga

Courtesy of United States Steel Corporation. (U.S. Steel). U.S. Steel accepts no liability for reliance on any information contained in the graphs shown above.

12

M-19, 29 Ga, Fully Processed

100

10

Loss, W/Lb

Flux Density
1 0.1 T 0.3 T 0.5 T 0.1 0.7 T 1.0 T

0.01 10 100 Frequency, Hz 1000 10000

Figure 8: Steel Sheet Core Loss Fit vs. Flux Density and Frequency Q =

2 Pa P 2

Where q is the number of machine phases and V is phase voltage. Note that this picture is, strictly speaking, only valid for the voltage and frequency for which the ux density was calculated. But it will be approximately true for small excursions in either voltage or frequency and therefore useful for estimating voltage drop due to exciting current and such matters. In design program applications these parameters can be re-calculated repeatedly if necessary. Looking up this data is a bit awkward for design studies, so it is often convenient to do a curve t to the published data. There are a large number of possible ways of doing this. One method that has been found to work reasonably well for silicon iron is an exponential t: P P0

B B0

f f0

This t is appropriate if the data appears on a log-log plot to lie in approximately straight lines. Figure 8 shows such a t for the same steel sheet as the other gures. For apparent power the same sort of method can be used. It appears, however, that the simple exponential t which works well for real power is inadequate, at least if relatively high inductions are to be used. This is because, as the steel saturates, the reactive component of exciting current rises rapidly. I have had some success with a double exponential t: VA VA0

B B0

+ VA1

B B0

13

Table 1: Exponential Fit Parameters for Two Steel Sheets 29 Ga, Fully Processed M-19 M-36 1T 1T 60 Hz 60 Hz 0.59 0.67 1.88 1.86 1.53 1.48 1.08 1.33 .0144 .0119 1.70 2.01 16.1 17.2

Base Flux Density Base Frequency Base Power (w/lb) Flux Exponent Frequency Exponent Base Apparent Power 1 Base Apparent Power 2 Flux Exponent Flux Exponent

B0 f0 P0 B F V A0 V A1 0 1

To rst order the reactive component of exciting current will be linear in frequency.

References
[1] W. MacLean, Theory of Strong Electromagnetic Waves in Massive Iron, Journal of Applied Physics, V.25, No 10, October, 1954 [2] P.D. Agarwal, Eddy-Current Losses in Solid and Laminated Iron, Trans. AIEE, V. 78, pp 169-171, 1959

14

Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science 6.685 Electric Machines Class Notes 4: Elementary Synchronous Machine Models c 2005 James L. Kirtley Jr. September 14, 2005

Introduction

The objective here is to develop a simple but physically meaningful model of the synchronous machine, one of the major classes of electric machine. We can look at this model from several dierent directions. This will help develop an understanding of analysis of machines, particularly in cases where one or another analytical picture is more appropriate than others. Both operation and sizing will be of interest here. Along the way we will approach machine windings from two points of view. On the one hand, we will approximate windings as sinusoidal distributions of current and ux linkage. Then we will take a concentrated coil point of view and generalize that into a more realistic and useful winding model.

Physical Picture: Current Sheet Description

Consider this simple picture. The machine consists of a cylindrical rotor and a cylindrical stator which are coaxial and which have sinusoidal current distributions on their surfaces: the outer surface of the rotor and the inner surface of the stator.

Stator

K zs K zr
g
 z

  r

Rotor

Figure 1: Elementary Machine Model: Axial View The rotor and stator bodies are made of highly permeable material (we approximate this as being innite for the time being, but this is something that needs to be looked at carefully later).

We also assume that the rotor and stator have current distributions that are axially (z) directed and sinusoidal:
S Kz R Kz

= KS cos p = KR cos p ( )

Here, the angle is the physical angle of the rotor. The current distribution on the rotor is xed with respect to the rotor. Now: assume that the air-gap dimension g is much less than the radius: g << R. It is not dicult to show that with this assumption the radial ux density Br is nearly uniform across the gap (i.e. not a function of radius) and obeys:
R K S + Kz Br = 0 z R g

Then the radial magnetic ux density for this case is simply: Br = 0 R (KS sin p + KR sin p ( )) pg

Now it is possible to compute the traction on rotor and stator surfaces by recognizing that the surface current distributions are the azimuthal magnetic elds: at the surface of the stator, S R H = Kz , and at the surface of the rotor, H = Kz . So at the surface of the rotor, traction is: = Tr = 0 R (KS sin p + KR sin p ( )) KR cos p ( ) pg

The average of that is simply: < >= 0 R KS KR sin p 2pg

The same exercise done at the surface of the stator yields the same results (with opposite sign). To nd torque, use: 0 R3 T = 2R2 < >= KS KR sin p pg We pause here to make a few observations: 1. For a given value of surface currents Ks and Kr, torque goes as the third power of linear dimension. That implies that the the achieved shear stres is constant with machine size. And the ratio of machine torque density to machine volume is constant. 2. If, on the other hand, gap is held constant, torque goes as the fourth power of machine volume. Since the volume of the machine goes as the third power, this implies that torque capability goes as the 4/3 power of machine volume. 3. Actually, this understates the situation since the assumed surface current densities are the products of volume current densities and winding depth, which one would expect to increase with machine size. As machine radius grows one would expect both stator and rotor surface current densities to grow. Thus machine torque (and power) densities tend to increase somewhat faster than linearly with machine volume. 2

4. The current distributions want to align with each other. In actual practice what is done is to generate a stator current distribution which is not static as implied here but which rotates in space: S Kz = KS cos (p t) and this pulls the rotor along. 5. For a given pair of current distributions there is a maximum torque that can be sustained, but as long as the torque that is applied to the rotor is less than that value the rotor will adjust to the correct angle.

Continuous Approximation to Winding Patterns:

Now lets try to produce those surface current distributions with physical windings. In fact we cant do exactly that yet, but we can approximate a physical winding with a turns distribution that would look like: nS = nR = NS cos p 2R NR cos p ( ) 2R

Note that this implies that NS and NR are the total number of turns on the rotor and stator. i.e.: p

nS Rd = NS

Then the surface current densities are as we assumed above, with: KS = N S IS 2R KR = NR IR 2R

So far nothing is dierent, but with an assumed number of turns we can proceed to computing inductances. It is important to remember what these assumed winding distributions mean: they are the density of wires along the surface of the rotor and stator. A positive value implies a wire with sense in the +z direction, a negative value implies a wire with sense in the -z direction. That is, if terminal current for a winding is positive, current is in the +z direction if n is positive, in the -z direction if n is negative. In fact, such a winding would be made of elementary coils with one half (the negatively going half) separated from the other half (the positively going half) by a physical angle of /p. So the ux linked by that elemental coil would be: i () =

/p

0 Hr ( )Rd

So, if only the stator winding is excited, radial magnetic eld is: Hr = N S IS sin p 2gp

and thus the elementary coil ux is: i () = 0 NS IS R cos p p2 g

Now, this is ux linked by an elementary coil. To get ux linked by a whole winding we must add up the ux linkages of all of the elementary coils. In our continuous approximation to the real coil this is the same as integrating over the coil distribution: S = p This evaluates fairly easily to: S = 0

2p

2p

i ()nS ()Rd

2 RNS Is 4 gp2

which implies a self-inductance for the stator winding of: LS = 0

2 RNS 4 gp2

The same process can be used to nd self-inductance of the rotor winding (with appropriate changes of spatial variables), and the answer is: LR = 0
2 RNR 4 gp2

To nd the mutual inductance between the two windings, excite one and compute ux linked by the other. All of the expressions here can be used, and the answer is: M () = 0 RNS NR cos p 4 gp2

Now it is fairly easy to compute torque using conventional methods. Assuming both windings are excited, magnetic coenergy is:
Wm =

1 1 2 2 LS IS + LR IR + M ()IS IR 2 2

and then torque is:

T =
RNS NR Wm = 0 IS IR sin p 4 gp

and then substituting for NS IS and NR IR :

NS IS = 2RKS

NR IR = 2RKR we get the same answer for torque as with the eld approach: T = 2R2 < >= 4 0 R3 KS KR sin p pg

Classical, Lumped-Parameter Synchronous Machine:

Now we are in a position to examine the simplest model of a polyphase synchronous machine. Suppose we have a machine in which the rotor is the same as the one we were considering, but the stator has three separate windings, identical but with spatial orientation separated by an electrical angle of 120 = 2/3. The three stator windings will have the same self- inductance (La ). With a little bit of examination it can be seen that the three stator windings will have mutual inductance, and that inductance will be characterized by the cosine of 120 . Since the physical angle between any pair of stator windings is the same, 1 Lab = Lac = Lbc = La 2 There will also be a mutual inductance between the rotor and each phase of the stator. Using M to denote the magnitude of that inductance: M Maf Mbf Mcf = 0 RNa Nf 4 gp2 = M cos (p) 2 3 2 = M cos p + 3 = M cos p

We show in Chapter 1 of these notes that torque for this system is: 2 T = pM ia if sin (p) pM ib if sin p 3

2 pM ic if sin p + 3

Balanced Operation:

Now, suppose the machine is operated in this fashion: the rotor turns at a constant velocity, the eld current is held constant, and the three stator currents are sinusoids in time, with the same amplitude and with phases that dier by 120 degrees. p = t + i if = If 2 3 2 = I cos t + 3

ia = I cos (t) ib = I cos t ic

Straightforward (but tedious) manipulation yields an expression for torque: 3 T = pM IIf sin i 2 5

Operated in this way, with balanced currents and with the mechanical speed consistent with the electrical frequency (p = ), the machine exhibits a constant torque. The phase angle i is called the torque angle, but it is important to use some caution, as there is more than one torque angle. Now, look at the machine from the electrical terminals. Flux linked by Phase A will be: a = La ia + Lab ib + Lac ic + M If cos p Noting that the sum of phase currents is, under balanced conditions, zero and that the mutual phase-phase inductances are equal, this simplies to: a = (La Lab ) ia + M If cos p = Ld ia + M If cos p where we use the notation Ld to denote synchronous inductance. Now, if the machine is turning at a speed consistent with the electrical frequency we say it is operating synchronously, and it is possible to employ complex notation in the sinusoidal steady state. Then, note: ia = I cos (t + i ) = Re Iejt+i If , we can write an expression for the complex amplitude of ux as: a = Re a ejt where we have used this complex notation:

I = Ieji If = If ejm

Now, if we look for terminal voltage of this system, it is: va =

da = Re ja ejt dt

This system is described by the equivalent circuit shown in Figure 2.

jXd

I
+

E af
-

V
-

Figure 2: Round Rotor Synchronous Machine Equivalent Circuit where the internal voltage is: E af = jM If ejm 6

Now, if that is connected to a voltage source (i.e. if V is xed), terminal current is: I= V Eaf ej jXd

where Xd = Ld is the synchronous reactance. Then real and reactive power (in phase A) are: P + jQ = = = This makes real and reactive power: Pa = Qa = 1 V Eaf sin 2 Xd 1V2 1 V Eaf cos 2 Xd 2 Xd 1 VI 2 V Eaf ej 1 V 2 jXd 1 |V |2 1 V Eaf ej 2 jXd 2 jXd

If we consider all three phases, real power is P = 3 V Eaf sin 2 Xd

Now, at last we need to look at actual operation of these machines, which can serve either as motors or as generators. Vector diagrams that describe operation as a motor and as a generator are shown in Figures 3 and 4, respectively.

Ia V V Ia jXdIa jXdIa Eaf Eaf

Figure 3: Motor Operation, Over- and Under- Excited

Operation as a generator is not much dierent from operation as a motor, but it is common to make notations with the terminal current given the opposite (generator) sign. 7

Eaf

Eaf jXdIg Ig jXdIg

Ig

Figure 4: Generator Operation, Over- and Under- Excited

Reconciliation of Models

We have determined that we can predict its power and/or torque characteristics from two points of view : rst, by knowing currents in the rotor and stator we could derive an expression for torque vs. a power angle: 3 T = pM IIf sin i 2 From a circuit point of view, it is possible to derive an expression for power: P = 3 V Eaf sin 2 Xd

and of course since power is torque times speed, this implies that: T = 3 pV Eaf 3 V Eaf sin = sin 2 Xd 2 Xd

In this section of the notes we will, rst of all, reconcile these notions, look a bit more at what they mean, and then generalize our simple theory to salient pole machines as an introduction to two-axis theory of electric machines.

6.1

Torque Angles:

Then, noting that terminal voltage V = t , Ea = M If and Xd = Ld , straightforward substitution yields: 3 3 pV Eaf sin = pM IIf sin i 2 Xd 2

Figure 5 shows a vector diagram that shows operation of a synchronous motor. It represents the MMFs and uxes from the rotor and stator in their respective positions in space during normal operation. Terminal ux is chosen to be real, or occupy the horizontal position. In motor operation the rotor lags by angle , so the rotor ux M If is shown in that position. Stator current is also shown, and the torque angle between it and the rotor, i is also shown. Now, note that the dotted line OA, drawn perpendicular to a line drawn between the stator ux Ld I and terminal ux t , has length: |OA| = Ld I sin i = t sin

LdI

MIf

Figure 5: Synchronous Machine Phasor Addition So the current- and voltage- based pictures do give the same result for torque.

Per-Unit Systems:

Before going on, we should take a short detour to look into per-unit systems, a notational device that, in addition to being convenient, will sometimes be conceptually helpful. The basic notion is quite simple: for most variables we will note a base quantity and then, by dividing the variable by the base we have a per-unit version of that variable. Generally we will want to tie the base quantity to some aspect of normal operation. So, for example, we might make the base voltage and current correspond with machine rating. If that is the case, then power base becomes: PB = 3VB IB and we can dene, in similar fashion, an impedance base: ZB = VB IB

Now, a little caution is required here. We have dened voltage base as line-neutral and current base as line current (both RMS). That is not necessary. In a three phase system we could very well have dened base voltage to have been line-line and base current to be current in a delta connected element: IB IB = VB = 3VB 3 In that case the base power would be unchanged but base impedance would dier by a factor of three: PB = VB IB ZB = 3ZB However, if we were consistent with actual impedances (note that a delta connection of elements of
impedance 3Z is equivalent to a wye connection of Z), the per-unit impedances of a given system
are not dependent on the particular connection. In fact one of the major advantages of using a
9

per-unit system is that per-unit values are uniquely determined, while ordinary variables can be line-line, line-neutral, RMS, peak, etc., for a large number of variations. Perhaps unfortunate is the fact that base quantities are usually given as line-line voltage and base power. So that: VB 1 VB V2 PB ZB = = B IB = = IB 3 IB PB 3VB Now, we will usually write per-unit variables as lower-case versions of the ordinary variables: v= V VB p= P PB etc.

Thus, written in per-unit notation, real and reactive power for a synchronous machine operating in steady state are: veaf veaf v2 sin q= cos p= xd xd xd These are, of course, in motor reference coordinates, and represent real and reactive power into the terminals of the machine.

Normal Operation:

The synchronous machine is used, essentially interchangeably, as a motor and as a generator. Note that, as a motor, this type of machine produces torque only when it is running at synchronous speed. This is not, of course, a problem for a turbogenerator which is started by its prime mover (e.g. a steam turbine). Many synchronous motors are started as induction machines on their damper cages (sometimes called starting cages). And of course with power electronic drives the machine can often be considered to be in synchronism even down to zero speed. As either a motor or as a generator, the synchronous machine can either produce or consume reactive power. In normal operation real power is dictated by the load (if a motor) or the prime mover (if a generator), and reactive power is determined by the real power and by eld current. Figure 6 shows one way of representing the capability of a synchronous machine. This picture represents operation as a generator, so the signs of p and q are reversed, but all of the other elements of operation are as we ordinarily would expect. If we plot p and q (calculated in the normal way) against each other, we see the construction at the right. If we start at a location q = v 2 /xd , (and remember that normally v = 1 per-unit) , then the locus of p and q is what would be obtained by swinging a vector of length veaf /xd over an angle . This is called a capability chart because it is an easy way of visualizing what the synchronous machine (in this case generator) can do. There are three easily noted limits to capability. The upper limit is a circle (the one traced out by that vector) which is referred to as eld capability. The second limit is a circle that describes constant |p + jq |. This is, of course, related to the magnitude of armature current and so this limit is called armature capability. The nal limit is related to machine stability, since the torque angle cannot go beyond 90 degrees. In actuality there are often other limits that can be represented on this type of a chart. For example, large synchronous generators typically have a problem with heating of the stator iron when they attempt to operate in highly underexcited conditions (q strongly negative), so that one will often see another limit that prevents the operation of the machine near its stability

10

Field Limit

Stator Limit P
1 X d

Stability Limit

Figure 6: Synchronous Generator Capability Diagram limit. In very large machines with more than one cooling state (e.g. dierent values of cooling hydrogen pressure) there may be multiple curves for some or all of the limits. Another way of describing the limitations of a synchronous machine is embodied in the Vee Curve. An example is shown in Figure 7 . This is a cross-plot of magnitude of armature current with eld current. Note that the eld and armature current limits are straightforward (and are the right-hand and upper boundaries, respectively, of the chart). The machine stability limit is what terminates each of the curves at the upper left-hand edge. Note that each curve has a minimum at unity power factor. In fact, there is yet another cross-plot possible, called a compounding curve, in which eld current is plotted against real power for xed power factor.

Salient Pole Machines: Two-Reaction Theory

So far, we have been describing what are referred to as round rotor machines, in which stator reactance is not dependent on rotor position. This is a pretty good approximation for large turbine generators and many smaller two-pole machines, but it is not a good approximation for many synchronous motors nor for slower speed generators. For many such applications it is more cost eective to wind the eld conductors around steel bodies (called poles) which are then fastened onto the rotor body, with bolts or dovetail joints. These produce magnetic anisotropies into the machine which aect its operation. The theory which follows is an introduction to two-reaction theory and consequently for the rotating eld transformations that form the basis for most modern dynamic analyses. Figure 8 shows a very schematic picture of the salient pole machine, intended primarily to show how to frame this analysis. As with the round rotor machine the stator winding is located in slots in the surface of a highly permeable stator core annulus. The eld winding is wound around steel pole pieces. We separate the stator current sheet into two components: one aligned with and one in quadrature to the eld. Remember that these two current components are themselves (linear) combinations of the stator phase currents. The transformation between phase currents and the d11

Vee Curves
1.2
1

Per-Unit Ia

0.8 0.6 0.4 0.2

0
0 0.5 1 1.5 2 2.5 3

Per-Unit Field Current Per-Unit Real Power: 0.00 0.32 0.64 0.96

Figure 7: Synchronous Machine Vee Curve and q- axis components is straightforward and will appear in Chapter 8 of these notes. The key here is to separate MMF and ux into two orthogonal components and to pretend that each can be treated as sinusoidal. The two components are aligned with the direct axis and with the quadrature axis of the machine. The direct axis is aligned with the eld winding, while the quadrature axis leads the direct by 90 degrees. Then, if is the angle between the direct axis and the axis of phase a, we can write for ux linking phase a: s = d cos q sin Then, in steady state operation, if Va =
da dt

and = t + ,

Va = s sin q cos which allows us to dene: Vd = q Vq = d one might think of the voltage vector as leading the ux vector by 90 degrees. Now, if the machine is linear, those uxes are given by: d = Ld Id + M If q = L q Iq Note that, in general, Ld = Lq . In wound-eld synchronous machines, usually Ld > Lq . The reverse is true for most salient (buried magnet) permanent magnet machines. 12

d axis

Iq

q axis

Id

Figure 8: Cartoon of a Salient Pole Synchronous Machine

V q V

V d Figure 9: Resolution of Terminal Voltage

Referring to Figure 9, one can resolve terminal voltage into these components: Vd = V sin Vq = V cos or: Vd = q = Lq Iq = V sin which is easily inverted to produce: Id = Iq V cos Eaf Xd V sin = Xq 13

Vq = d = Ld Id + M If = V cos

where Xd = Ld Xq = Lq Eaf = M If Now, we are working in ordinary variables (this discussion should help motivate the use of perunit!), and each of these variables is peak amplitude. Then, if we take up a complex frame of reference: V = Vd + jVq

I = Id + jIq complex power is: 3 3 P + jQ = V I = {(Vd Id + Vq Iq ) + j (Vq Id Vd Iq )} 2 2 or: P 3 = 2 3 2

V2 V Eaf sin + Xd 2

1 1 Xd Xq

sin 2

Q =

V2 2

1 1 + Xd Xq

V2 2

1 1 Xd Xq

V Eaf cos cos 2 Xd

I V

I d axix

j X Id d jX I q q axis E 1 E af

jX I q q

Figure 10: Phasor Diagram: Salient Pole Machine A phasor diagram for a salient pole machine is shown in Figure 10. This is a little dierent from the equivalent picture for a round-rotor machine, in that stator current has been separated into its d- and q- axis components, and the voltage drops associated with those components have been drawn separately. It is interesting and helpful to recognize that the internal voltage Eaf can be expressed as: Eaf = E1 + (Xd Xq ) Id 14

where the voltage E1 is on the quadrature axis. In fact, E1 would be the internal voltage of a round rotor machine with reactance Xq and the same stator current and terminal voltage. Then the operating point is found fairly easily: = tan E1 =

Xq I cos V + Xq I sin

(V + Xq I sin )2 + (Xq I cos )2

Power-Angle Curves
1.5

xd=2.2
1

xq = 1.6

0.5

Per-Unit

Round Rotor 0 -4 -3 -2 -1 0 1 2 3 4 Salient Rotor

-0.5

-1

-1.5 Torque Angle

Figure 11: Torque-Angle Curves: Round Rotor and Salient Pole Machines A comparison of torque-angle curves for a pair of machines, one with a round, one with a salient rotor is shown in Figure 11 . It is not too dicult to see why power systems analysts often neglect saliency in doing things like transient stability calculations.

10

Relating Rating to Size

It is possible, even with the simple model we have developed so far, to establish a quantitative relationship between machine size and rating, depending (of course) on elements such as useful ux and surface current density. To start, note that the rating of a machine (motor or generator) is: |P + jQ| = qV I where q is the number of phases, V is the RMS voltage in each phase and I is the RMS current. To establish machine rating we must establish voltage and current, and we do these separately. 15

10.1

Voltage
Na cos p 2R

Assume that our sinusoidal approximation for turns density is valid: na () = And suppose that working ux density is: Br () = B0 sin p( ) Now, to compute ux linked by the winding (and consequently to compute voltage), we rst compute ux linked by an incremental coil: i () = Then ux linked by the whole coil is: a = p

2p

+ p

Br ( )Rd

2p

i ()na ()Rd =

2RNa B0 cos p 4 p

This is instantaneous ux linked when the rotor is at angle . If the machine is operating at some electrical frequency with a phase angle so that p = t + , the RMS magnitude of terminal voltage is: B0 Va = 2RNa p4 2 Finally, note that the useful peak current density that can be used is limited by the fraction of machine periphery used for slots: B0 = Bs (1 s ) where Bs is the ux density in the teeth, limited by saturation of the magnetic material.

10.2

Current

The (RMS) magnitude of the current sheet produced by a current of (RMS) magnitude I is: Kz = q Na I 2 2R 2 qNa

And then the current is, in terms of the current sheet magnitude: I = 2RKz

Note that the surface current density is, in terms of area current density Js , slot space factor s and slot depth hs : Kz = s Js hs This gives terminal current in terms of dimensions and useful current density:
I= 4R
s hs Js qNa 16

10.3

Rating
Bs 2R2 s (1 s ) hs Js p 2

Assembling these expressions, machine rating becomes: |P + jQ| = qV I =

This expression is actually fairly easily interpreted. The product of slot factor times one minus slot factor optimizes rather quickly to 1/4 (when s = 1). We could interpret this as: |P = jQ| = As us where the interaction area is: As = 2R The surface velocity of interaction is: us = R = R p

and the fragment of expression which looks like traction is: Bs = hs Js s (1 s ) 2 Note that this is not quite traction since the current and magnetic ux may not be ideally aligned, and this is why the expression incorporates reactive as well as real power. This is not quite yet the whole story. The limit on Bs is easily understood to be caused by saturation of magnetic material. The other important element on shear stress density, hs Js is a little more involved.

10.4

Role of Reactance
I 0 R s hs Js 2 = V pg 1 s Bs p(1 s )Bs 0 Rs 2

The per-unit, or normalized synchronous reactance is: xd = Xd

While this may be somewhat interesting by itself, it becomes useful if we solve it for hs Ja : hs Ja = xd g

That is, if xd is xed, hs Ja (and so power) are directly related to air-gap g. Now, to get a limit on g, we must answer the question of how far the eld winding can throw eective air-gap ux? To understand this question, we must calculate the eld current to produce rated voltage, no-load, and then the excess of eld current required to accommodate load current. Under rated operation, per- unit eld voltage is: e2 = v 2 + (xd i)2 + 2xd i sin af Or, if at rated conditions v and i are both unity (one per- unit), then eaf =

1 + x2 + 2xd sin d 17

10.5

Field Winding

Thus, given a value for xd and , per- unit internal voltage eaf is also xed. Then eld current required can be calculated by rst estimating eld winding current for no-load operation. Br = and rated eld current is: If = If nl eaf or, required rated eld current is: N f If = 2gp(1 s )Bs eaf 0 0 Nf If nl 2gp

Next, If can be related to a eld current density: Nf If = NRS ARS Jf 2

where NRS is the number of rotor slots and the rotor slot area ARS is ARS = wR hR where hR is rotor slot height and wR is rotor slot width: wR = Then: Nf If = RR hR Jf Now we have a value for air- gap g: g= 0 RR hR Jf 2 p(1 s )Bs eaf 2R R NRS

This then gives us useful armature surface current density: xd R hs Js = hR Jf 2 2 eaf s We will not have a lot more to say about this. Note that the ratio of xd /eaf can be quite small (if the per-unit reactance is small), will never be a very large number for any practical machine, and is generally less than one. As a practical matter it is unusual for the per-unit synchronous reatance of a machine to be larger than about 2 or 2.25 per-unit. What this tells us should be obvious: either the rotor or the stator of a machine can produce the dominant limitation on shear stress density (and so on rating). The best designs are balanced, with both limits being reached at the same time.

18

Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science 6.685 Electric Machinery Class Notes 5: Winding Inductances c 2005 James L. Kirtley Jr. September 5, 2005

Introduction

The purpose of this document is to show how the inductances of windings in round- rotor machines with narrow air gaps may be calculated. We deal only with the idealized air- gap magnetic elds, and do not consider slot, end winding, peripheral or skew reactances. We do, however, consider the space harmonics of winding magneto-motive force (MMF).

Description of Stators
Back Iron

Slots

Slot Depression

Teeth

Figure 1: Stator Cross-Section Figure 1 shows a cartoon view of an axial cross-section of a twelve-slot stator. Actually, what is shown is the shape of a thin sheet of steel, or lamination that is used to make up the magnetic circuit. The iron is made of thin sheets to control eddy current losses. Thickness varies according to freuqency of operation, but in machines for 60 Hz (the vast bulk of machines made for industrial 1

use), lamination thickness is typically .014 (.355 mm). These are stacked to make the magnetic circuit of the appropriate length. Windings are carried in the slots of this structure. Figure 1 shows trapezoidal slots with teeth of approximately uniform cross-section over most of their length but wider extent near the air-gap. The tooth ends, in combination with the relatively narrow slot depression region, help control certain parasitic losses in the rotor of many machines by improving uniformity of the air-gap elds, increase the air-gap permeance and help hold the windings in the slots. It should be noted that large machines, with what are called form wound coils, have straight-sided rectangular slots and consequently teeth of non-uniform cross-section. The description that follows will hold for both types of machine.

A 1

A 2

C 3

C 4

B 5

B 6

A 7

A 8

C 9

C 10

B 11

B 12

Figure 2: Full-Pitched Winding To simplify the discussion, imagine the slot/tooth region to be straightened out as shown in Figure 2. This shows a three-phase, two-pole winding in the twelve slots. Such a winding would have two slots per pole per phase. One of the two coils of phase A would be wound in slots 1 and 7 (six slots apart).

A A 1

A C 2

C C 3

C B 4

B B 5

B A 6

A A 7

A C 8

C C 9

C B 10

B B 11

B A 12

Figure 3: Five-Sixths-Pitched Winding Machines are seldom wound as shown in Figure 2 for a variety of reasons. It is usually advantageous in reducing the length of the end turns and to reducing space harmonic eects in the machine (usually bad eects!) to wind the machine with short-pitched windings as shown in Figure 3. Each phase in this case consists of four coils (two per slot). The four coils of Phase A would span between slots 1 and 6, slots 2 and 7, slots 7 and 12 and slots 8 and 1. Each of these coil spans is ve slots, so this choice of winding pattern is referred to as Five-Sixths pitch. So this cartoon-gure machine stator (which could represent either a synchronous or induction motor or generator) has both breadth because there are more than one slots per pole per phase, and it may have the need for accounting for winding pitch. What follows in this note is a simple protocol for estimating the important air-gap elds and inductances.

Winding MMF

To start, consider the MMF of a full- pitch, concentrated winding as shown in schematic form in Figure 4. Assuming that the winding has a total of N turns over p pole- pairs, and is carrying current I the MMF is: 4 NI sin np (1) F = n 2p n=1 nodd This distribution is shown, as a function of angle in Figure 5. This leads directly to magnetic ux density in the air- gap: Br =

n=1 nodd

0 4 N I sin np g n 2p

(2)

Note that a real winding, which will most likely not be full- pitched and concentrated, will have a winding factor which is the product of pitch and breadth factors, to be discussed later.

Magnetic Circuit: Stator Rotor

NI p z

Air-Gap

Figure 4: Primitive Geometry Problem Now, suppose that there is a polyphase winding, consisting of more than one phase (we will use three phases), driven with one of two types of current. The rst of these is balanced, current: Ia = I cos(t) Ib = I cos(t Ic 2 ) 3 2 = I cos(t + ) 3

(3)

Conversely, we might consider Zero Sequence currents: Ia = Ib = Ic = I cos t 3

(4)

F( ) NI p

2p

3 2p

Figure 5: Air-Gap MMF Then it is possible to express magnetic ux density for the two distinct cases. For the balanced case: Br = where
n=1

Brn sin(np t)

(5)

The upper sign holds for n = 1, 7, ... The lower sign holds for n = 5, 11, ... all other terms are zero and 3 0 4 N I 2 g n 2p The zero- sequence case is simpler: it is nonzero only for the triplen harmonics: Brn = Br = 0 4 N I 3 (sin(np t) + sin(np + t)) g n 2p 2 n=3,9,...

(6)

(7)

Next, consider the ux from a winding on the rotor: that will have the same form as the ux produced by a single armature winding, but will be referred to the rotor position: Brf =

n=1 nodd which is, substituting =

t p ,

0 4 N I sin np g n 2p

(8)

Brf =

n=1 nodd

0 4 N I sin n(p t) g n 2p

(9)

The next step here is to nd the ux linked if we have some air- gap ux density of the form: Br =

n=1

Brn sin(np t) 4

(10)

Now, it is possible to calculate ux linked by a single-turn, full-pitched winding by:

= and, using (10), this is: = 2Rl
n=1 Brn

Br Rld

(11)

np

cos(t)

(12)

This allows us to compute self- and mutual- inductances, since winding ux is: = N (13)

The end of this is a set of expressions for various inductances. It should be noted that, in the real world, most windings are not full-pitched nor concentrated. Fortunately, these shortcomings can be accommodated by the use of winding factors. The simplest and perhaps best denition of a winding factor is the ratio of ux linked by an actual winding to ux that would have been linked by a full- pitch, concentrated winding with the same number of turns. That is: actual (14) kw = f ullpitch It is relatively easy to show, using reciprocity arguments, that the winding factors are also the ratio of eective MMF produced by an actual winding to the MMF that would have been produced by the same winding were it to be full- pitched and concentrated. The argument goes as follows: mutual inductance between any pair of windings is reciprocal. That is, if the windings are designated one and two, the mutual inductance is ux induced in winding one by current in winding two, and it is also ux induced in winding two by current in winding one. Since each winding has a winding factor that inuences its linking ux, and since the mutual inductance must be reciprocal, the same winding factor must inuence the MMF produced by the winding. The winding factors are often expressed for each space harmonic, although sometimes when a winding factor is referred to without reference to a harmonic number, what is meant is the space factor for the space fundamental. Two winding factors are commonly specied for ordinary, regular windings. These are usually called pitch and breadth factors, reecting the fact that often windings are not full pitched, which means that individual turns do not span a full electrical radians and that the windings occupy a range or breadth of slots within a phase belt. The breadth factors are ratios of ux linked by a given winding to the ux that would be linked by that winding were it full- pitched and concentrated. These two winding factors are discussed in a little more detail below. What is interesting to note, although we do not prove it here, is that the winding factor of any given winding is the product of the pitch and breadth factors: kw = kp kb (15) With winding factors as dened by (14) and the sections below, it is possible to dene winding inductances. For example, the synchronous inductance of a winding will be the apparent inductance of one phase when the polyphase winding is driven by a balanced set of currents as in (3). This is, approximately: 2 3 4 0 N 2 Rlkwn (16) Ld = 2 p2 gn2 n=1,5,7,... 5

This expression is approximate because it ignores the asynchronous interactions between higher order harmonics and the rotor of the machine. These are beyond the scope of this note. Zero- sequence inductance is the ratio of ux to current if a winding is excited by zero sequence currents, as in (4): 2 4 0 N 2 Rlkwn (17) L0 = 3 p2 gn2 n=3,9,... And then mutual inductance, as between a eld winding (f ) and an armature winding (a), is: M () =

n=1 nodd

4 0 Nf Na kf n kan Rl cos(np) p2 gn2

(18)

Winding Factors

Now we turn our attention to computing the winding factors for simple, regular winding patterns. We do not prove but only state that the winding factor can, for regular winding patterns, be expressed as the product of a pitch factor and a breadth factor, each of which can be estimated separately.

4.1

Pitch Factor

z r

Figure 6: Short-Pitched Coils Pitch factor is found by considering the ux linked by a less- than- full pitched winding. Consider the situation in which radial magnetic ux density is: Br = Bn sin(np t) A winding with pitch will link ux (see Figure 6: = Nl

+ 2p 2p 2p 2p

(19)

Bn sin(np t)Rd 6

(20)

Pitch refers to the angular displacement between sides of the coil, expressed in electrical radians. For a full- pitch coil = . The ux linked is: n n 2N lRBn = sin( ) sin( ) (21) np 2 2 Using the denition (14), the pitch factor is seen to be: kpn = sin n 2 (22)

4.2

Breadth Factor

Now for breadth factor. This describes the fact that a winding may consist of a number of coils, each linking ux slightly out of phase with the others. A regular winding will have a number (say m) coil elements, separated by electrical angle . (See Figure 7

Figure 7: Distributed Coils A full- pitch coil with one side at angle will, in the presence of magnetic ux density as described by (19), link ux: = Nl This is readily evaluated to be: =
2N lRBn Re ej(tn) np
p p p

Bn sin(np t)Rd

(23)

(24)

where in (24), complex number notation has been used for convenience in carrying out the rest of this derivation. What happens here is that the coils link uxes that dier in phase, so the addition of ux is as shown in vector form in Figure 8. 7

Individual Flux Linkages

Total Flux Linkage

Figure 8: Vector Flux Addition Now: if the winding is distributed into m sets of slots and the slots are evenly spaced, the angular position of each slot will be: i = i and the number of turns in each slot will be = The breadth factor is then simply: kb = Note that (27) can be written as:
m1 m ejn 2 jni kb = e m i=0

m1 2

(25)

N mp ,

so that actual ux linked will be: (26)

2N lRBn 1 m1 Re ej(tni ) np m i=0 1 m1 jn(i m1 ) 2 e m i=0

(27)

(28)

Now, focus on that sum. We know that any coverging geometric sum has a simple sum:
i=0

xi =

1 1x

(29)

and that a truncated sum is:

Then the sum in (28) can be written as:

m1 i=0

m1 i=0

i=0

(30)

i=m

ejni = 1 ejnm

i=0

ejni =

1 ejnm 1 ejn

(31)

Now, inserting the results of (31) into (28), and using the denitions for sine, the breadth factor is found: sin nm 2 (32) kbn = m sin n 2 8

4.3

Alternate Derivation of Breadth Factor

Most textbooks, if they bother to prove the Breadth Factor, use a geometric proof as shown in Figure 9.

C A B

m 2

Figure 9: Alternate Proof of Breadth Factor The short vectors (e.g. AC) represent the voltages induced in individual coils. In fact, what is shown in this gure is the same as is shown in Figure 8, but spread out to show the actual addition. Now, note that if each of the vectors is bisected by a line segment at right angles, all of those line segments meet at point O. The line segment that includes OB is one of these. Line segments that run from O to the ends of the vectors will have an angle from the bisectors of the vectors.
2 Similarly, the line segment OA has an angle of m with respect to the bisector of the resultant 2 voltage vector. Now, if we note F1 as the length of each of the individual coil voltage vectors and F as the length of the resultant sum, the length of half of the bisector is: AB = but then Then the resultant vector is: m F = OA sin 2 2 (33)

1 1 AC = F1 = OA sin 2 2 2 sin m 2 F
= 2AB
= mF1 m sin 2

(34)

(35)

Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science 6.685 Electric Machines Class Notes 6: DC (Commutator) and Permanent Magnet Machines c 2005 James L. Kirtley Jr.

September 5, 2005

Introduction

Virtually all electric machines, and all practical electric machines employ some form of rotating or alternating eld/current system to produce torque. While it is possible to produce a true DC machine (e.g. the Faraday Disk), for practical reasons such machines have not reached application and are not likely to. In the machines we have examined so far the machine is operated from an alternating voltage source. Indeed, this is one of the principal reasons for employing AC in power systems. The rst electric machines employed a mechanical switch, in the form of a carbon brush/commutator system, to produce this rotating eld. While the widespread use of power electronics is making brushless motors (which are really just synchronous machines) more popular and common, commutator machines are still economically very important. They are relatively cheap, particularly in small sizes, and they tend to be rugged and simple. You will nd commutator machines in a very wide range of applications. The starting motor on all automobiles is a series-connected commutator machine. Many of the other electric motors in automobiles, from the little motors that drive the outside rear-view mirrors to the motors that drive the windshield wipers are permanent magnet commutator machines. The large traction motors that drive subway trains and diesel/electric locomotives are DC commutator machines (although induction machines are making some inroads here). And many common appliances use universal motors: series connected commutator motors adapted to AC.

1.1

Geometry:

Stator Yoke Field Poles

Rotor

Armature Winding Field Winding

Figure 1: Wound-Field DC Machine Geometry

A schematic picture (cartoon) of a commutator type machine is shown in 1. The armature of this machine is on the rotor (this is the part that handles the electric power), and current is fed to the armature through the brush/commutator system. The interaction magnetic eld is provided (in this picture) by a eld winding. A permanent magnet eld is applicable here, and we will have quite a lot more to say about such arrangements below. Now, if we assume that the interaction magnetic ux density averages Br , and if there are Ca conductors underneath the poles at any one time, and if there are m parallel paths, then we may estimate torque produced by the machine by: Te = Ca RBr Ia m

where R and are rotor radius and length, respectively and Ia is terminal current. Note that Ca is not necessarily the total number of conductors, but rather the total number of active conductors (that is, conductors underneath the pole and therefore subject to the interaction eld). Now, if we note Nf as the number of eld turns per pole, the interaction eld is just: Br = 0 N f If g

leading to a simple expression for torque in terms of the two currents: Te = GIa If where G is now the motor coecient (units of N-m/ampere squared): G = 0 Ca Nf R m g

Now, lets go back and look at this from the point of view of voltage. Start with Faradays Law: E = B t

Integrating both sides and noting that the area integral of a curl is the edge integral of the quantity, we nd: B E d = t Now, that is a bit awkward to use, particularly in the case we have here in which the edge of the contour is moving (note we will be using this expression to nd voltage). We can make this a bit more convenient to use if we note: d dt

B n da =

B n da + t

v B d

where v is the velocity of the contour. This gives us a convenient way of noting the apparent electric eld within a moving object (as in the conductors in a DC machine): E = E + v B Now, note that the armature conductors are moving through the magnetic eld produced by the stator (eld) poles, and we can ascribe to them an axially directed electric eld: Ez = RBr 2

dl

Figure 2: Motion of a contour through a magnetic eld produces ux change and electric eld in the moving contour If the armature conductors are arranged as described above, with Ca conductors in m parallel paths underneath the poles and with a mean active radial magnetic eld of Br , we can compute a voltage induced in the stator conductors: Ca RBr m Note that this is only the voltage induced by motion of the armature conductors through the eld and does not include brush or conductor resistance. If we include the expression for eective magnetic eld, we nd that the back voltage is: Eb = Eb = GIf which leads us to the conclusion that newton-meters per ampere squared equals volt seconds per ampere. This stands to reason if we examine electric power into the interaction and mechanical power out: Pem = Eb Ia = Te Now, a more complete model of this machine would include the eects of armature, brush and lead resistance, so that in steady state operation: Va = Ra Ia + GIf Now, consider this machine with its armatucre connected to a voltage source and its eld operating at steady current, so that: Ia = Va GIf Ra

Then torque, electric power in and mechanical power out are:

Te = GIf Pe Pm Va GIf
Ra Va GIf = Va Ra Va GIf = GIf Ra

Now, note that these expressions dene three regimes dened by rotational speed. The two break points are at zero speed and at the zero torque speed: 0 = 3 Va GIf

Ra

+ Va

G I f
-

Figure 3: DC Machine Equivalent Circuit

Electrical

Mechanical

Figure 4: DC Machine Operating Regimes For 0 < < 0 , the machine is a motor: electric power in and mechanical power out are both positive. For higher speeds: 0 < , the machine is a generator, with electrical power in and mechanical power out being both negative. For speeds less than zero, electrical power in is positive and mechanical power out is negative. There are few needs to operate machines in this regime, short of some types of plugging or emergency braking in tractions systems.

1.2

Hookups:

We have just described a mode of operation of a commutator machine usually called separately excited, in which eld and armature circuits are controlled separately. This mode of operation is used in some types of traction applications in which the exibility it aords is useful. For example, some traction applications apply voltage control in the form of choppers to separately excited machines. Note that the zero torque speed is dependent on armature voltage and on eld current. For high torque at low speed one would operate the machine with high eld current and enough armature voltage to produce the requisite current. As speed increases so does back voltage, and eld current may need to be reduced. At any steady operating speed there will be some optimum mix of eld and armature currents to produced the required torque. For braking one could (and this is often done) re-connect the armature of the machine to a braking resistor and turn the machine into a generator. Braking torque is controlled by eld current. A subset of the separately excited machine is the shunt connection in which armature and eld are supplied by the same source, in parallel. This connection is not widely used any more: it does 4

Ra

G I f
-

Figure 5: Two-Chopper, separately excited machine hookup

+ V

Figure 6: Series Connection not yield any meaningful ability to control speed and the simple applications to which it used to be used are handled by induction machines. Another connection which is still widely used is the series connection, in which the eld winding is sized so that its normal operating current level is the same as normal armature current and the two windings are connected in series. Then: Ia = If = And then torque is: Te = V Ra + Rf + G

GV 2 (Ra + Rf + G)2

It is important to note that this machine has no zero-torque speed, leading to the possibility that an unloaded machine might accelerate to dangerous speeds. This is particularly true because the commutator, made of pieces of relatively heavy material tied together with non-conductors, is not very strong. Speed control of series connected machines can be achieved with voltage control and many appliances using this type of machine use choppers or phase control. An older form of control used in traction applications was the series dropping resistor: obviously not a very ecient way of controlling the machine and not widely used (except in old equipment, of course).

A variation on this class of machine is the very widely used universal motor, in which the stator and rotor (eld and armature) of the machine are both constructed to operate with alternating current. This means that both the eld and armature are made of laminated steel. Note that such a machine will operate just as it would have with direct current, with the only addition being the reactive impedance of the two windings. Working with RMS quantities: I = Te = V Ra + Rf + G + j (La + Lf ) G|V |2 (Ra + Rf + G)2 + (La + Lf )2

where is the electrical supply frequency. Note that, unlike other AC machines, the universal motor is not limited in speed to the supply frequency. Appliance motors typically turn substantially faster than the 3,600 RPM limit of AC motors, and this is one reason why they are so widely used: with the high rotational speeds it is possible to produce more power per unit mass (and more power per dollar).

1.3

Commutator:

The commutator is what makes this machine work. The brush and commutator system of this class of motor involves quite a lot of black art, and there are still aspects of how they work which are poorly understood. However, we can make some attempt to show a bit of what the brush/commutator system does. To start, take a look at the picture shown in Figure 7. Represented are a pair of poles (shaded) and a pair of brushes. Conductors make a group of closed paths. Current from one of the brushes takes two parallel paths. You can follow one of those paths around a closed loop, under each of the two poles (remember that the poles are of opposite polarity) to the opposite brush. Open commutator segments (most of them) do not carry current into or out of the machine.

Figure 7: Commutator and Current Paths A commutation interval occurs when the current in one coil must be reversed. (See Figure 8 In the simplest form this involves a brush bridging between two commutator segments, shorting out that coil. The resistance of the brush causes the current to decay. When the brush leaves the leading segment the current in the leading coil must reverse. We will not attempt to fully understand the commutation process in this type of machine, but we can note a few things. Resistive commutation is the process relied upon in small machines. 6

Figure 8: Commutator at Commutation When the current in one coil must be reversed (because it has left one pole and is approaching the other), that coil is shorted by one of the brushes. The brush resistance causes the current in the coil to decay. Then the leading commutator segment leaves the brush the current MUST reverse (the trailing coil has current in it), and there is often sparking.

1.4

Commutation
Stator Yoke Field Poles

Rotor

Armature Winding Field Winding Commutation Interpoles

Figure 9: Commutation Interpoles In larger machines the commutation process would involve too much sparking, which causes brush wear, noxious gases (ozone) that promote corrosion, etc. In these cases it is common to use separate commutation interpoles. These are separate, usually narrow or seemingly vestigal pole pieces which carry armature current. They are arranged in such a way that the ux from the interpole drives current in the commutated coil in the proper direction. Remember that the coil being commutated is located physically between the active poles and the interpole is therefore in the right spot to inuence commutation. The interpole is wound with armature current (it is in series with the main brushes). It is easy to see that the interpole must have a ux density proportional to the current to be commutated. Since the speed with which the coil must be commutated is proportional to rotational velocity and so is the voltage induced by the interpole, if the right 7

number of turns are put around the interpole, commutation can be made to be quite accurate.

1.5

Compensation:

Field Poles

PoleFace Compensation Winding

Rotor

Armature Winding Field Winding Commutation Interpoles

Figure 10: Pole Face Compensation Winding The analysis of commutator machines often ignores armature reaction ux. Obviously these machines DO produce armature reaction ux, in quadrature with the main eld. Normally, commutator machines are highly salient and the quadrature inductance is lower than direct-axis inductance, but there is still ux produced. This adds to the ux density on one side of the main poles (possibly leading to saturation). To make the ux distribution more uniform and therefore to avoid this saturation eect of quadrature axis ux, it is common in very highly rated machines to wind compensation coils: essentially mirror-images of the armature coils, but this time wound in slots in the surface of the eld poles. Such coils will have the same number of ampere-turns as the armature. Normally they have the same number of turns and are connected directly in series with the armature brushes. What they do is to almost exactly cancel the ux produced by the armature coils, leaving only the main ux produced by the eld winding. One might think of these coils as providing a reaction torque, produced in exactly the same way as main torque is produced by the armature. A cartoon view of this is shown in Figure 10.

Permanent Magnets in Electric Machines

Of all changes in materials technology over the last several years, advances in permanent magnets have had the largest impact on electric machines. Permanent magnets are often suitable as replacements for the eld windings in machines: that is they can produce the fundamental interaction eld. This does three things. First, since the permanent magnet is lossless it eliminates the energy required for excitation, usually improving the eciency of the machine. Second, since eliminating the excitation loss reduces the heat load it is often possible to make PM machines more compact. Finally, and less appreciated, is the fact that modern permanent magnets have very large coercive

force densities which permit vastly larger air gaps than conventional eld windings, and this in turn permits design exibility which can result in even better electric machines. These advantages come not without cost. Permanent magnet materials have special characteristics which must be taken into account in machine design. The highest performance permanent magnets are brittle ceramics, some have chemical sensitivities, all are sensitive to high temperatures, most have sensitivity to demagnetizing elds, and proper machine design requires understanding the materials well. These notes will not make you into seasoned permanent magnet machine designers. They are, however, an attempt to get started, to develop some of the mathematical skills required and to point to some of the important issues involved.

2.1

Permanent Magnets:

Hysteresis Loop: Permanent Magnet

1 0.8 0.6 0.4 0.2 Tesla -400 -300 -200 0 -100 -0.2 0 -0.4 -0.6 -0.8 -1 Kilo Am peres/Meter 100 200 300 400

Figure 11: Hysteresis Loop Of Ceramic Permanent Magnet Permanent magnet materials are, at core, just materials with very wide hysteresis loops. Figure 11 is an example of something close to one of the more popular ceramic magnet materials. Note that this hysteresis loop is so wide that you can see the eect of the permeability of free space. It is usual to display only part of the magnetic characteristic of permanent magnet materials (see Figure 12), the third quadrant of this picture, because that is where the material is normally operated. Note a few important characteristics of what is called the demagnetization curve. The remanent ux density Br , is the value of ux density in the material with zero magnetic eld H. The coercive eld Hc is the magnetic eld at which the ux density falls to zero. Shown also on the curve are loci of constant energy product. This quantity is unfortunately named, for although it has the same units as energy it represents real energy in only a fairly general sense. It is the product of ux density and eld intensity. As you already know, there are three commonly used systems of units for magnetic eld quantities, and these systems are often mixed up to form very confusing units. We will try to stay away from the English system of units in which eld intensity H is measured in amperes per inch and ux density B in lines (actually, usually kilolines) per square inch. In CGS units ux density is measured in Gauss (or kilogauss) and magnetic eld intensity in Oersteds. And in SI the unit of ux density is the Tesla, which is one Weber per square meter, and the unit of eld intensity is the Ampere per meterOf these, only the last one, A/m is . obvious. A Weber is a volt-second. A Gauss is 104 Tesla. And, nally, an Oersted is that eld 9

Demagnetization Curve
0.5 0.45 0.4 0.35 B, Tesla 0.3 0.25 0.2 0.15 0.1

Energy Product Loci

Br

Hc
-250 -200 -150 -100 -50

0.05 0 0

H, kA/m

Figure 12: Demagnetization Curve intensity required to produce one Gauss in the permeability of free space. Since the permeability of free space 0 = 4 107 Hy/m, this means that one Oe is about 79.58 A/m. Commonly, the energy product is cited in MgOe (Mega-Gauss-Oersted)s. One MgOe is equal to 7.958kJ/m3 . A commonly used measure for the performance of a permanent magnet material is the maximum energy product, the largest value of this product along the demagnetization curve. To start to understand how these materials might be useful, consider the situation shown in Figure 13: A piece of permanent magnet material is wrapped in a magnetic circuit with eectively innite permeability. Assume the thing has some (nite) depth in the direction you cant see. Now, if we take Amperes law around the path described by the dotted line, H d = 0 since there is no current anywhere in the problem. If magnetization is upwards, as indicated by the arrow, this would indicate that the ux density in the permanent magnet material is equal to the remanent ux density (also upward). A second problem is illustrated in Figure 14, in which the same magnet is embedded in a magnetic circuit with an air gap. Assume that the gap has width g and area Ag. The magnet has height hm and area Am. For convenience, we will take the positive reference direction to be up (as we see it here) in the magnet and down in the air-gap. Thus we are following the same reference direction as we go around the Amperes Law loop. That becomes: H d = Hm hm + Hg g Now, Gauss law could be written for either the upper or lower piece of the magnetic circuit. Assuming that the only substantive ux leaving or entering the magnetic circuit is either in the magnet or the gap: B d A = Bm Am 0 Hg Ag Solving this pair we have: Bm = 0 Ag hm Hm = 0 Pu Hm Am g 10

Permanent Magnet

Magnetic Circuit,

Figure 13: Permanent Magnet in Magnetic Circuit

Permanent Magnet g hm

Magnetic Circuit,

Figure 14: Permanent Magnet Driving an Air-Gap This denes the unit permeance, essentially the ratio of the permeance facing the permanent magnet to the internal permeance of the magnet. The problem can be, if necessary, solved graphically, since the relationship between Bm and Hm is inherently nonlinear, as shown in Figure 15 load line analysis of a nonlinear electronic circuit. Now, one more cut at this problem. Note that, at least for fairly large unit permeances the slope of the magnet characteristic is fairly constant. In fact, for most of the permanent magnets used in machines (the one important exception is the now rarely used ALNICO alloy magnet), it is generally acceptable to approximate the demagnitization curve with: Bm = m Hm + M0 Here, the magnetization M0 is xed. Further, for almost all of the practical magnet materials the magnet permeability is nearly the same as that of free space (m 0 ). With that in mind, consider the problem shown in Figure 16, in which the magnet lls only part of a gap in a magnetic circuit. But here the magnet and gap areas are essentially the same. We could regard the magnet as simply a magnetization. 11

 0u

Figure 15: Load Line, Unit Permeance Analysis

Permanent Magnet

Magnetic Circuit,

Figure 16: Surface Magnet Primitive Problem In the region of the magnet and the air-gap, Amperes Law and Gauss law can be written: H = 0

0 Hm + M0

= 0

0 H g = 0 Now, if in the magnet the magnetization is constant, the divergence of H in the magnet is zero. Because there is no current here, H is curl free, so that everywhere:
2

= 0

H =

That is, magnetic eld can be expressed as the gradient of a scalar potential which satises Laplaces equation. It is also pretty clear that, if we can assign the scalar potential to have a value of zero anywhere on the surface of the magnetic circuit it will be zero over all of the magnetic circuit (i.e. at both the top of the gap and the bottom of the magnet). Finally, note that we cant actually assume that the scalar potential satises Laplaces equation everywhere in the problem. In fact the divergence of M is zero everywhere except at the top surface of the magnet where it is singular! In fact, we can note that there is a (some would say ctitious) magnetic charge density: m = M 12

At the top of the magnet there is a discontinuous change in M and so the equivalent of a magnetic surface charge. Using Hg to note the magnetic eld above the magnet and Hm to note the magnetic eld in the magnet, 0 Hg = 0 (Hm + M0 ) m = M0 = Hg Hm and then to satisfy the potential condition, if hm is the height of the magnet and g is the gap: gHg = hm Hm Solving, hm hm + g Now, one more observation could be made. We would produce the same air-gap ux density if we regard the permanent magnet as having a surface current around the periphery equal to the magnetization intensity. That is, if the surface current runs around the magnet: Hg = M0 K = M0 This would produce an MMF in the gap of: F = K hm and then since the magnetic eld is just the MMF divided by the total gap: Hg = hm F = M0 hm + g hm + g

The real utility of permanent magnets comes about from the relatively large magnetizations: numbers of a few to several thousand amperes per meter are common, and these would translate into enormous current densities in magnets of ordinary size.

3 Simple Permanent Magnet Machine Structures: Machines

Commutator

Figure 17 is a cartoon picture of a cross section of the geometry of a two-pole commutator machine using permanent magnets. This is actually the most common geometry that is used. The rotor (armature) of the machine is a conventional, windings-in-slots type, just as we have already seen for commutator machines. The eld magnets are fastened (often just bonded) to the inside of a steel tube that serves as the magnetic ux return path. Assume for the purpose of rst-order analysis of this thing that the magnet is describable by its remanent ux density Br and has permeability of 0 . First, we will estimate the useful magnetic ux density and then will deal with voltage generated in the armature.

3.1

Interaction Flux Density

Using the basics of the analysis presented above, we may estimate the radial magnetic ux density at the air-gap as being: Br Bd = 1 1 + Pc 13

Back Iron h
m

g R

Rotor

Permanent Magnets

Figure 17: PM Commutator Machine where the eective unit permeance is: Pc = fl hm Ag ff g Am

A book on this topic by James Ireland suggests values for the two fudge factors: 1. The leakage factor fl is cited as being about 1.1. 2. The reluctance factor ff is cited as being about 1.2. We may further estimate the ratio of areas of the gap and magnet by: R+ g Ag 2 = m Am R + g + h2 Now, there are a bunch of approximations and hand wavings in this expression, but it seems to work, at least for the kind of machines contemplated. A second correction is required to correct the eective length for electrical interaction. The reason for this is that the magnets produce fringing elds, as if they were longer than the actual stack length of the rotor (sometimes they actually are). This is purely empirical, and Ireland gives a value for eective length for voltage generation of: e = where = + 2N R , and the empirical coecient A hm N log 1 + B B R

where B = 7.4 9.0 A = 0.9 14 hm R

3.1.1

Voltage:

It is, in this case, simplest to consider voltage generated in a single wire rst. If the machine is running at angular velocity , speed voltage is, while the wire is under a magnet, vs = RBr Now, if the magnets have angular extent m the voltage induced in a wire will have a waveform as shown in Figure 18: It is pulse-like and has the same shape as the magnetic eld of the magnets.

vs

m t

Figure 18: Voltage Induced in One Conductor The voltage produced by a coil is actually made up of two waveforms of exactly this form, but separated in time by the coil throw angle. Then the total voltage waveform produced will be the sum of the two waveforms. If the coil thrown angle is larger than the magnet angle, the two voltage waveforms add to look like this: There are actually two coil-side waveforms that add with a slight phase shift.

vc

0m 0m
Figure 19: Voltage Induced in a Coil If, on the other hand, the coil thrown is smaller than the magnet angle, the picture is the same, only the width of the pulses is that of the coil rather than the magnet. In either case the average voltage generated by a coil is: v = RNs Bd where is the lesser of the coil throw or magnet angles and Ns is the number of series turns in the coil. This gives us the opportunity to develop the number of active turns: Ca C = Ns = tot m m 15

Here, Ca is the number of active conductors, Ctot is the total number of conductors and m is the number of parallel paths. The motor coecient is then: K= Re Ctot Bd m

3.2

Armature Resistance

The last element we need for rst-order prediction of performance of the motor is the value of armature resistance. The armature resistance is simply determined by the length and area of the wire and by the number of parallel paths (generally equal to 2 for small commutator motors). If we note Nc as the number of coils and Na as the number of turns per coil, Ns = Total armature resistance is given by: Ra = 2w t Ns m Nc Na m

where w is the resistivity (per unit length) of the wire: w = 1

2 w 4 d
w

(dw is wire diameter, w is wire conductivity and t is length of one half-turn). This length depends on how the machine is wound, but a good rst-order guess might be something like this: t + R

16

Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science 6.685 Electric Machines Class Notes 7: Permanent Magnet Brushless DC Motors c 2005 James L. Kirtley Jr. September 5, 2005

Introduction

This document is a brief introduction to the design evaluation of permanent magnet motors, with an eye toward servo and drive applications. It is organized in the following manner: First, we describe three dierent geometrical arrangements for permanent magnet motors: 1. Surface Mounted Magnets, Conventional Stator, 2. Surface Mounted Magnets, Air-Gap Stator Winding, and 3. Internal Magnets (Flux Concentrating). After a qualitative discussion of these geometries, we will discuss the elementary rating parameters of the machine and show how to arrive at a rating and how to estimate the torque and power vs. speed capability of the motor. Then we will discuss how the machine geometry can be used to estimate both the elementary rating parameters and the parameters used to make more detailed estimates of the machine performance. Some of the more involved mathematical derivations are contained in appendices to this note.

Motor Morphologies

There are, of course, many ways of building permanent magnet motors, but we will consider only a few in this note. Actually, once these are understood, rating evaluations of most other geometrical arrangements should be fairly straightforward. It should be understood that the rotor inside vs. rotor outside distinction is in fact trivial, with very few exceptions, which we will note.

2.1

Surface Magnet Machines

Figure 1 shows the basic magnetic morphology of the motor with magnets mounted on the surface of the rotor and an otherwise conventional stator winding. This sketch does not show some of the important mechanical aspects of the machine, such as the means for fastening the permanent magnets to the rotor, so one should look at it with a bit of caution. In addition, this sketch and the other sketches to follow are not necessarily to a scale that would result in workable machines. This gure shows an axial section of a four-pole (p = 2) machine. The four magnets are mounted on a cylindrical rotor core, or shaft, made of ferromagnetic material. Typically this would simply be a steel shaft. In some applications the magnets may be simply bonded to the steel. For applications in which a glue joint is not satisfactory (e.g. for high speed machines) some sort of rotor banding or retaining ring structure is required.

Stator Winding in Slots

Rotor Core (Shaft)

Stator Core

AirGap Rotor Magnets

Figure 1: Axial View of a Surface Mount Motor The stator winding of this machine is conventional, very much like that of an induction motor, consisting of wires located in slots in the surface of the stator core. The stator core itself is made of laminated ferromagnetic material (probably silicon iron sheets), the character and thickness of the sheets determined by operating frequency and eciency requirements. They are required to carry alternating magnetic elds, so must be laminated to reduce eddy current losses. This sort of machine is simple in construction. Note that the operating magnetic ux density in the air-gap is nearly the same as in the magnets, so that this sort of machine cannot have air-gap ux densities higher than that of the remanent ux density of the magnets. If low cost ferrite magnets are used, this means relatively low induction and consequently relatively low eciency and power density. (Note the qualier relatively here!). Note, however, that with modern, high performance permanent magnet materials in which remanent ux densities can be on the order of 1.2 T, air-gap working ux densities can be on the order of 1 T. With the requirement for slots to carry the armature current, this may be a practical limit for air-gap ux density anyway. It is also important to note that the magnets in this design are really in the air gap of the machine, and therefore are exposed to all of the time- and space- harmonics of the stator winding MMF. Because some permanent magnets have electrical conductivity (particularly the higher performance magnets), any asynchronous elds will tend to produce eddy currents and consequent losses in the magnets.

2.2

Interior Magnet or Flux Concentrating Machines

Interior magnet designs have been developed to counter several apparent or real shortcomings of surface mount motors: Flux concentrating designs allow the ux density in the air-gap to be higher than the ux density in the magnets themselves. 2

 In interior magnet designs there is some degree of shielding of the magnets from high order space harmonic elds by the pole pieces. There are control advantages to some types of interior magnet motors, as we will show anon. Essentially, they have relatively large negative saliency which enhances ux weakening for high speed operation, in rather direct analogy to what is done in DC machines. Some types of internal magnet designs have (or claim) structural advantages over surface mount magnet designs.
Rotor Pole Pieces

Armature in Slots

Rotor Magnets

Stator Core

Nonmagnetic Rotor Core (shaft)

Figure 2: Axial View of a Flux Concentrating Motor The geometry of one type of internal magnet motor is shown (crudely) in Figure 2. The permanent magnets are oriented so that their magnetization is azimuthal. They are located between wedges of magnetic material (the pole pieces) in the rotor. Flux passes through these wedges, going radially at the air- gap, then azimuthally through the magnets. The central core of the rotor must be non-magnetic, to prevent shorting out the magnets. No structure is shown at all in this drawing, but quite obviously this sort of rotor is a structural challenge. Shown is a six-pole machine. Typically, one does not expect ux concentrating machines to have small pole numbers, because it is dicult to get more area inside the rotor than around the periphery. On the other hand, a machine built in this way but without substantial ux concentration will still have saliency and magnet shielding properties. A second morphology for an internal magnet motor is shown in Figure 3. This geometry has been proposed for highly salient synchronous machines without permanent magnets: such machines would run on the saliency torque and are called synchronous reluctance motors. however, the saliency slots may be lled with permanent magnet material, giving them some internally generated ux as well. The rotor iron tends to short out the magnets, so that the bridges around the ends of the permanent magnets must be relatively thin. They are normally saturated. 3

Stator Core Stator Slots

Air Gap

Rotor

Saliency Slots

Figure 3: Axial View of Internal Magnet Motor At rst sight, these machines appear to be quite complicated to analyze, and that judgement seems to hold up.

2.3

Air Gap Armature Windings

Shown in Figure 4 is a surface-mounted magnet machine with an air-gap, or surface armature winding. Such machines take advantage of the fact that modern permanent magnet materials have very low permeabilities and that, therefore, the magnetic eld produced is relatively insensitive to the size of the air-gap of the machine. It is possible to eliminate the stator teeth and use all of the periphery of the air-gap for windings. Not shown in this gure is the structure of the armature winding. This is not an issue in conventional stators, since the armature is contained in slots in the iron stator core. The use of an air-gap winding gives opportunities for economy of construction, new armature winding forms such as helical windings, elimination of cogging torques, and (possibly) higher power densities.

Zeroth Order Rating

In determining the rating of a machine, we may consider two separate sets of parameters. The rst set, the elementary rating parameters, consist of the machine inductances, internal ux linkage and stator resistance. From these and a few assumptions about base and maximum speed it is possible to get a rst estimate of the rating and performance of the motor. More detailed performance estimates, including eciency in sustained operation, require estimation of other parameters. We will pay more attention to that rst set of parameters, but will attempt to show how at least some of the more complete operating parameters can be estimated.

Stator Winding

Rotor Core (Shaft)

Stator Core

111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000

Rotor Magnets

Figure 4: Axial View of a PM Motor With an Air-Gap Winding

3.1

Voltage and Current: Round Rotor

To get started, consider the equivalent circuit shown in Figure 5. This is actually the equivalent circuit which describes all round rotor synchronous machines. It is directly equivalent only to some of the machines we are dealing with here, but it will serve to illustrate one or two important points.

Ia + Vt -

Ea

Figure 5: Synchronous Machine Equivalent Circuit What is shown here is the equivalent circuit of a single phase of the machine. Most motors are three-phase, but it is not dicult to carry out most of the analysis for an arbitrary number of phases. The circuit shows an internal voltage Ea and a reactance X which together with the terminal current I determine the terminal voltage V . In this picture armature resistance is ignored. If the machine is running in the sinusoidal steady state, the major quantities are of the form: Ea = a cos (t + ) Vt = V cos t 5

 V jXI Ea

Figure 6: Phasor Diagram For A Synchronous Machine Ia = I cos (t ) The machine is in synchronous operation if the internal and external voltages are at the same frequency and have a constant (or slowly changing) phase relationship (). The relationship between the major variables may be visualized by the phasor diagram shown in Figure 3.1. The internal voltage is just the time derivative of the internal ux from the permanent magnets, and the voltage drop in the machine reactance is also the time derivative of ux produced by armature current in the air-gap and in the leakage inductances of the machine. By convention, the angle is positive when current I lags voltage V and the angle is positive then internal voltage Ea leads terminal voltage V . So both of these angles have negative sign in the situation shown in Figure 3.1. If there are q phases, the time average power produced by this machine is simply: q P = V I cos 2 For most polyphase machines operating in what is called balanced operation (all phases doing the same thing with uniform phase dierences between phases), torque (and consequently power) are approximately constant. Since we have ignored power dissipated in the machine armature, it must be true that power absorbed by the internal voltage source is the same as terminal power, or: q P = Ea I cos ( ) 2
Since in the steady state:
P = T p where T is torque and /p is mechanical rotational speed, torque can be derived from the terminal quantities by simply: q T = p a I cos ( ) 2 In principal, then, to determine the torque and hence power rating of a machine it is only necessary to determine the internal ux, the terminal current capability, and the speed capability of the rotor. In fact it is almost that simple. Unfortunately, the model shown in Figure 5 is not quite complete for some of the motors we will be dealing with, and we must go one more level into machine theory. 6

3.2

A Little Two-Reaction Theory

The material in this subsection is framed in terms of three-phase (q = 3) machine theory, but it is actually generalizable to an arbitrary number of phases. Suppose we have a machine whose three-phase armature can be characterized by internal uxes and inductance which may, in general, not be constant but is a function of rotor position. Note that the simple model we presented in the previous subsection does not conform to this picture, because it assumes a constant terminal inductance. In that case, we have: ph = Lph I ph + R (1)

where R is the set of internally produced uxes (from the permanent magnets) and the stator winding may have both self- and mutual- inductances. Now, we nd it useful to do a transformation on these stator uxes in the following way: each armature quantity, including ux, current and voltage, is projected into a coordinate system that is xed to the rotor. This is often called the Parks Transformation. For a three phase machine it is:

Where the transformation and its inverse are:

ua ud uq = udq = T uph = T ub uc u0

(2)

cos cos( 23 ) cos( + 23 ) 2 T = sin sin( 2 ) sin( + 23 ) 3 3 1 1 1 2 2 2

(3)

T 1

It is easy to show that balanced polyphase quantities in the stationary, or phase variable frame, translate into constant quantities in the so-called d-q frame. For example: Ia = I cos t 2 ) 3 2 Ic = I cos(t + ) 3 = t + 0 Ib = I cos(t maps to: Id = I cos 0 Iq = I sin 0 Now, if = t + 0 , the transformation coordinate system is chosen correctly and the d- axis will correspond with the axis on which the rotor magnets are making positive ux. That happens 7

cos sin 1 = cos( 2 ) sin( 2 ) 1 3 3 2 2 cos( + 3 ) sin( + 3 ) 1

(4)

if, when = 0, phase A is linking maximum positive ux from the permanent magnets. If this is the case, the internal uxes are: aa = f cos ab = f cos( ac 2 ) 3 2 = f cos( + ) 3

Now, if we compute the uxes in the d-q frame, we have: dq = Ldq I dq + R = T Lph T 1 I dq + R (5)

Now: two things should be noted here. The rst is that, if the coordinate system has been chosen as described above, the ux induced by the rotor is, in the d-q frame, simply: f R = 0 0

(6)

That is, the magnets produce ux only on the d- axis. The second thing to note is that, under certain assumptions, the inductances in the d-q frame are independent of rotor position and have no mutual terms. That is: Ldq = T Lph T 1 Ld 0 0 Lq 0 = 0 0 0 L0

(7)

The assertion that inductances in the d-q frame are constant is actually questionable, but it is close enough to being true and analyses that use it have proven to be close enough to being correct that it (the assertion) has held up to the test of time. In fact the deviations from independence on rotor position are small. Independence of axes (that is, absence of mutual inductances in the d-q frame) is correct because the two axes are physically orthogonal. We tend to ignore the third, or zero axis in this analysis. It doesnt couple to anything else and has neither ux nor current anyway. Note that the direct- and quadrature- axis inductances are in principle straightforward to compute. They are direct axis the inductance of one of the armature phases (corrected for the fact of multiple phases) with the rotor aligned with the axis of the phase, and quadrature axis the inductance of one of the phases with the rotor aligned 90 electrical degrees away from the axis of that phase. Next, armature voltage is, ignoring resistance, given by: d d ph = T 1 dq dt dt and that the transformed armature voltage must be: V ph = 8 (8)

V dq = T V ph d = T (T 1 dq ) dt d d (9) dq + (T T 1 )dq = dt dt The second term in this expresses speed voltage. A good deal of straightforward but tedious manipulation yields: T d 1 T = dt

0
d dt

d dt 0 0

The direct- and quadrature- axis voltage expressions are then:

dd q Vd = dt dq Vq = + d dt where d = dt Instantaneous power is given by: P = Va Ia + Vb Ib + Vc Ic

0 0 0

(10)

(11) (12)

(13)

Using the transformations given above, this can be shown to be: 3 3 (14) P = Vd Id + Vq Iq + 3V0 I0 2 2 which, in turn, is: 3 dd dq d0 3 Id + Iq ) + 3 I0 (15) P = (d Iq q Id ) + ( dt dt 2 dt 2 Then, noting that = p and that (15) describes electrical terminal power as the sum of shaft power and rate of change of stored energy, we may deduce that torque is given by: q T = p(d Iq q Id ) (16) 2 Note that we have stated a generalization to a q- phase machine even though the derivation given here was carried out for the q = 3 case. Of course three phase machines are by far the most common case. Machines with higher numbers of phases behave in the same way (and this generalization is valid for all purposes to which we put it), but there are more rotor variables analogous to zero axis. Now, noting that, in general, Ld and Lq are not necessarily equal, d = Ld Id + f q = Lq Iq then torque is given by: q T = p (f + (Ld Lq ) Id ) Iq 2 9 (17) (18) (19)

3.3

Finding Torque Capability

For high performance drives, we will generally assume that the power supply, generally an inverter, can supply currents in the correct spatial relationship to the rotor to produce torque in some reasonably eective fashion. We will show in this section how to determine, given a required torque (or if the torque is limited by either voltage or current which we will discuss anon), what the values of Id and Iq must be. Then the power supply, given some means of determining where the rotor is (the instantaneous value of ), will use the inverse Parks transformation to determine the instantaneous valued required for phase currents. This is the essence of what is known as eld oriented control, or putting stator currents in the correct location in space to produce the required torque. Our objective in this section is, given the elementary parameters of the motor, nd the capability of the motor to produce torque. There are three things to consider here: Armature current is limited, generally by heating, A second limit is the voltage capability of the supply, particularly at high speed, and If the machine is operating within these two limits, we should consider the optimal placement of currents (that is, how to get the most torque per unit of current to minimize losses). Often the discussion of current placement is carried out using, as a tool to visualize what is going on, the Id , Iq plane. Operation in the steady state implies a single point on this plane. A simple illustration is shown in Figure 7. The thermally limited armature current capability is represented as a circle around the origin, since the magnitude of armature current is just the length of a vector from the origin in this space. In general, for permanent magnet machines with buried magnets, Ld < Lq , so the optimal operation of the machine will be with negative Id . We will show how to determine this optimum operation anon, but it will in general follow a curve in the Id , Iq plane as shown. Finally, an ellipse describes the voltage limit. To start, consider what would happen if the terminals of the machine were to be short-circuited so that V = 0. If the machine is operating at suciently high speed so that armature resistance is negligible, armature current would be simply: Id = Iq = 0 Now, loci of constant ux turn out to be ellipses around this point on the plane. Since terminal ux is proportional to voltage and inversely proportional to frequency, if the machine is operating with a given terminal voltage, the ability of that voltage to command current in the Id , Iq plane is an ellipse whose size shrinks as speed increases. To simplify the mathematics involved in this estimation, we normalize reactances, uxes, currents and torques. First, let us dene the base ux to be simply b = f and the base current Ib to be the armature capability. Then we dene two per-unit reactances: xd = xq = 10
L d Ib b L q Ib b (20) (21) f Ld

Voltage Limit Loci

Speed = Base Speed > Base

Optimal Torque Locus

q Armature Current Limit

id Short Circuit Point

Figure 7: Limits to Operation Next, dene the base torque to be:

q Tb = p b Ib 2 and then, given per-unit currents id and iq , the per-unit torque is simply: te = (1 (xq xd ) id ) iq (22)

It is fairly straightforward (but a bit tedious) to show that the locus of current-optimal operation (that is, the largest torque for a given current magnitude or the smallest current magnitude for a given torque) is along the curve: 1 i2 a +2 2 4 (xq xd ) 1 i2 a 2 2 4 (xq xd )
2

id =

1 2 (xq xd ) + 1 2 (xq xd )

1 4 (xq xd ) 1 4 (xq xd )

+
2

i2 a 2 i2 a 2

(23)

iq =

(24)

The rating point will be the point along this curve when ia = 1, or where this curve crosses the armature capability circle in the id , iq plane. It should be noted that this set of expressions only works for salient machines. For non-salient machines, of course, torque-optimal current is on the q-axis. In general, for machines with saliency, the per-unit torque will not be unity at the rating, so that the rated, or Base Speed torque is not the Base torque, but: Tr = Tb te where te is calculated at the rating point (that is, ia = 1 and id and iq as per (23) and (24)). (25)

11

For suciently low speeds, the power electronic drive can command the optimal current to produce torque up to rated. However, for speeds higher than the Base Speed, this is no longer true. Dene a per-unit terminal ux: V = b Operation at a given ux magnitude implies: 2 = (1 + xd id )2 + (xq iq )2 which is an ellipse in the id , iq plane. The Base Speed is that speed at which this ellipse crosses the point where the optimal current curve crosses the armature capability. Operation at the highest attainable torque (for a given speed) generally implies d-axis currents that are higher than those on the optimal current locus. What is happening here is the (negative) d-axis current serves to reduce eective machine ux and hence voltage which is limiting q-axis current. Thus operation above the base speed is often referred to as ux weakening. The strategy for picking the correct trajectory for current in the id , iq plane depends on the value of the per-unit reactance xd . For values of xd > 1, it is possible to produce some torque at any speed. For values of xd < 1, there is a speed for which no point in the armature current capability is within the voltage limiting ellipse, so that useful torque has gone to zero. Generally, the maximum torque operating point is the intersection of the armature current limit and the voltage limiting ellipse: id = iq = xd 2 x2 xq d 1 i2 d xd 2 2
x
x
q d
2

x2 2 + 1 q 2 x2
x
q d

(26) (27)

It may be that there is no intersection between the armature capability and the voltage limiting ellipse. If this is the case and if xd < 1, torque capability at the given speed is zero. If, on the other hand, xd > 1, it may be that the intersection between the voltage limiting ellipse and the armature current limit is not the maximum torque point. To nd out, we calculate the maximum torque point on the voltage limiting ellipse. This is done in the usual way by dierentiating torque with respect to id while holding the relationship between id and iq to be on the ellipse. The algebra is a bit messy, and results in: id 3xd (xq xd ) x2 d = 4x2 (xq
xd ) d 1 xq 2 (1 + xd id )2 3xd (xq xd ) x2 d 4x2 (xq
xd ) d
2

(xq xd ) ( 2 1) + xd 2 (xq xd ) x2 d

(28) (29)

iq =

Ordinarily, it is probably easiest to compute (28) and (29) rst, then test to see if the currents are outside the armature capability, and if they are, use (26) and (27). These expressions give us the capability to estimate the torque-speed curve for a machine. As an example, the machine described by the parameters cited in Table 1 is a (nominal) 3 HP, 4-pole, 3000 RPM machine. The rated operating point turns out to have the following attributes: 12

Table 1: Example Machine

D- Axis Inductance Q- Axis Inductance Internal Flux Armature Current 2.53 mHy 6.38 mHy 58.1 mWb 30 A

Table 2: Operating Characteristics of Example Machine

Per-Unit D-Axis Current At Rating Point Per-Unit Q-Axis Current At Rating Point Per-Unit D-Axis Reactance Per-Unit Q-Axis Reactance Rated Torque (Nm) Terminal Voltage at Base Point (V) id iq xd xq Tr -.5924 .8056 1.306 3.294 9.17 97

The loci of operation in the Id , Iq plane is shown in Figure 8. The armature current limit is shown only in the second and third quadrants, so shows up as a semicircle. The two ellipses correspond with the rated point (the larger ellipse) and with a speed that is three times rated (9000 RPM). The torque-optimal current locus can be seen running from the origin to the rating point, and the higher speed operating locus follows the armature current limit. Figure 9 shows the torque/speed and power/speed curves. Note that this sort of machine only approximates constant power operation at speeds above the base or rating point speed.

Parameter Estimation

We are now at the point of estimating the major parameters of the motors. Because we have a number of dierent motor geometries to consider, and because they share parameters in not too orderly a fashion, this section will have a number of sub-parts. First, we calculate ux linkage, then reactance.

4.1

Flux Linkage

Given a machine which may be considered to be uniform in the axial direction, ux linked by a single, full-pitched coil which spans an angle from zero to /p, is: =
0
p

Br Rld

where Br is the radial ux through the coil. And, if Br is sinusoidally distributed this will have a peak value of 2RlBr p = p 13

PM Brushless Machine Current Loci 60

40

20 Q-Axis Current (A)

-20

-40

-60 -80 -60 -40 -20 0 D-Axis Current (A)

20

Figure 8: Operating Current Loci of Example Machine

PM Brushless Machine

10
8
Torque, N-m

6
4
2
0 0
1000 2000 3000 4000 5000 6000 7000 8000 9000

4000
Power, Watts

3000
2000
1000
0 0

1000

2000

3000

4000 5000 Speed, RPM

6000

7000

8000

9000

Figure 9: Torque- and Power-Speed Capability

14

Now, if the actual winding has Na turns, and using the pitch and breadth factors derived in Appendix 1, the total ux linked is simply: f = where kw = kp kb kp = sin 2 sin m 2 kb = m sin 2 The angle is the pitch angle, = 2p Np Ns 2RlB1 Na kw p (30)

where Np is the coil span (in slots) and Ns is the total number of slots in the stator. The angle is the slot electrical angle: 2p = Ns Now, what remains to be found is the space fundamental magnetic ux density B1 . In Appendix 2 it is shown that, for magnets in a surface-mount geometry, the magnetic eld at the surface of the magnetic gap is: B1 = 0 M1 kg where the space-fundamental magnetization is: M1 = pm Br 4 sin 0 2 (31)

where Br is remanent ux density of the permanent magnets and m is the magnet angle. and where the factor that describes the geometry of the magnetic gap depends on the case. For magnets inside and p = 1, kg =
p1 Rs 2p Rs 2p Ri

p p 2p 1p 1p p+1 p+1 Ri R1 R 2 R2 R1 + p+1 p1 1 2 2 Rs R i 1 R2 2 2 2 R2 R1 + Ri log 2 R1

For magnets inside and p = 1, kg =

For the case of magnets outside and p = 1: kg =

p1 Ri 2p Rs 2p Ri

p p p+1 p+1 1p 1p R2p R1 R2 R2 R1 + p+1 p1 s

and for magnets outside and p = 1, 15

kg =

1 2 2 Rs R i

1 R2 2 2 2 R2 R1 + Rs log 2 R1

Where Rs and Ri are the outer and inner magnetic boundaries, respectively, and R2 and R1 are the outer and inner boundaries of the magnets. Note that for the case of a small gap, in which both the physical gap g and the magnet thickness hm are both much less than rotor radius, it is straightforward to show that all of the above expressions approach what one would calculate using a simple, one-dimensional model for the permanent magnet: kg hm g + hm

This is the whole story for the winding-in-slot, narrow air-gap, surface magnet machine. For airgap armature windings, it is necessary to take into account the radial dependence of the magnetic eld.

4.2

Air-Gap Armature Windings

With no windings in slots, the conventional denition of winding factor becomes dicult to apply. If, however, each of the phase belts of the winding occupies an angular extent w , then the equivalent to (31) is: w sin p 2 kw = w p 2 Next, assume that the density of conductors within each of the phase belts of the armature winding is uniform, so that the density of turns as a function of radius is: N (r) =
2 Rwo

2Na r 2 Rwi

This just expresses the fact that there is more azimuthal room at larger radii, so with uniform density the number of turns as a function of radius is linearly dependent on radius. Here, Rwo and Rwi are the outer and inner radii, respectively, of the winding. Now it is possible to compute the ux linked due to a magnetic eld distribution:
Rwo

f =

Rwi

2lNa kw r 2r 0 Hr (r)dr 2 R2 p Rwo wi

(32)

Note the form of the magnetic eld as a function of radius expressed in 80 and 81 of the second appendix. For the winding outside case it is:
2p Hr = A r p1 + Rs r p1

Then a winding with all its turns concentrated at the outer radius r = Rwo would link ux: c = 2lRwo kw 2lRwo kw 2p p1 p1 0 Hr (Rwo ) = 0 A Rwo + Rs Rwo p p 16

Carrying out (32), it is possible, then, to express the ux linked by a thick winding to the ux that would have been linked by a radially concentrated winding at its outer surface by: kt = where, for the winding outside, p = 2 case: kt = 2 2 ) (1 + 2p ) (1 x 1 x2+p 2p 1 x2p + 2+p 2p (33) f c

where we have used the denitions = Rwo /Rs and x = Rwi /Rwo . In the case of winding outside, p = 2, 1 x4 4 2 log x (34) kt = 2 ) (1 + 2p ) (1 x 4 In a very similar way, we can dene a winding factor for a thick winding in which the reference radius is at the inner surface. (Note: this is done because the inner surface of the inside winding is likely to be coincident with the inner ferromagnetic surface, as the outer surface of the outer winding ls likely to be coincident with the outer ferromagnetic surface). For p = 2: kt = and for p = 2: kt = 2x2 (1 x2 ) (1 + 2p ) 1 x4 (x)4 log x 4 (36) 2xp (1 x2 ) (1 + 2p ) 1 x2+p 1 x2p + (x)2p 2+p 2p (35)

where = Ri /Rwi So, in summary, the ux linked by an air-gap armature is given by: f = 2RlB1 Na kw kt p (37)

where B1 is the ux density at the outer radius of the physical winding (for outside winding machines) or at the inner radius of the physical winding (for inside winding machines). Note that the additional factor kt is a bit more than one (it approaches unity for thin windings), so that, for small pole numbers and windings that are not too thick, it is almost correct and in any case conservative to take it to be one.

4.3

Interior Magnet Motors:

For the ux concentrating machine, it is possible to estimate air-gap ux density using a simple reluctance model. The air- gap permeance of one pole piece is: ag = 0 l where p is the angular width of the pole piece. 17 Rp g

And the incremental permeance of a magnet is: m = 0 hm l wm

The magnet sees a unit permeance consisting of its own permeance in series with one half of each of two pole pieces (in series) : u = Magnetic ux density in the magnet is: Bm = B 0 And then ux density in the air gap is: Bg = 2hm 2hm wm Bm = B 0 Rp 4ghm + Rp wm u 1 + u Rp wm ag = m 4g hm

The space fundamental of that can be written as: B1 = 4 pp wm sin B0 m 2 2g

1
1+
w m p R g 4 hm

where we have introduced the shorthand:

m =

The ux linkage is then computed as before: f = 2RlB1 Na kw p (38)

4.4

Winding Inductances

The next important set of parameters to compute are the d- and q- axis inductances of the machine. We will consider three separate cases, the winding-in-slot, surface magnet case, which is magnetically round, or non-salient, the air-gap winding case, and the ux concentrating case which is salient, or has dierent direct- and quadrature- axis inductances. 4.4.1 Surface Magnets, Windings in Slots

In this conguration there is no saliency, so that Ld = Lq . There are two principal parts to inductance, the air-gap inductance and slot leakage inductance. Other components, including end turn leakage, may be important in some congurations, and they would be computed in the same way as for an induction machine. If magnet thickness is not too great, we may make the narrow air-gap assumption, in which case the fundamental part of air-gap inductance is: Ld1 =
2 2 q 4 0 Na kw lRs 2 p2 (g + hm )

(39)

18

Here, g is the magnetic gap, including the physical rotational gap and any magnet retaining means that might be used. hm is the magnet thickness. Since the magnet thickness is included in the air-gap, the air-gap permeance may not be very large, so that slot leakage inductance may be important. To estimate this, assume that the slot shape is rectangular, characterized by the following dimensions: hs height of the main portion of the slot ws width of the top of the main portion of the slot hd height of the slot depression wd slot depression opening Of course not all slots are rectangular: in fact in most machines the slots are trapezoidal in shape to maintain teeth cross-sections that are radially uniform. However, only a very small error (a few percent) is incurred in calculating slot permeance if the slot is assumed to be rectangular and the top width is used (that is the width closest to the air-gap). Then the slot permeance is, per unit length: hd 1 hs + P = 0 3 ws wd Assume for the rest of this discussion a standard winding, with m slots in each phase belt (this assumes, then, that the total number of slots is Ns = 2pqm), and each slot holds two halfcoils. (A half-coil is one side of a coil which, of course, is wound in two slots). If each coil has Nc turns (meaning Na = 2pmNc ) , then the contribution to phase self-inductance of one 2 slot is, if both half-coils are from the same phase, 4lPNc . If the half-coils are from dierent 2 phases, then the contribution to self inductance is lPNc and the magnitude of the contribution to 2 . (Some caution is required here. For three phase windings the mutual mutual inductance is lPNc inductance is negative, so are the senses of the currents in the two other phases, so the impact of mutual leakage is to increase the reactance. This will be true for other numbers of phases as well, even if the algebraic sign of the mutual leakage inductance is positive, in which case so will be the sense of the other- phase current.) We will make two other assumptions here. The standard one is that the winding coil throw, or span between sides of a coil, is Ns Nsp . Nsp is the coil short pitch. The other is that each 2p phase belt will overlap with, at most two other phases: the ones on either side in sequence. This last assumption is immediately true for three- phase windings (because there are only two other phases. It is also likely to be true for any reasonable number of phases. Noting that each phase occupies 2p(m Nsp ) slots with both coil halves in the same slot and 2pNsp slots in which one coil half shares a slot with each of two dierent phases, we can write down the two components of slot leakage inductance, self- and mutual:
2 Las = 2pl (m Nsp ) (2Nc )2 + 2Nsp Nc

2 Lam = 2plNsp Nc

For a three- phase machine, then, the total slot leakage inductance is:
2 La = Las Lam = 2plPNc (4m Nsp )

For a uniform, symmetric winding with an odd number of phases, it is possible to show that the eective slot leakage inductance is: La = Las 2Lam cos 19 2 q

Total synchronous inductance is the sum of air-gap and leakage components: so far this is:
Ld = Ld1 + La 4.4.2 Air-Gap Armature Windings

It is shown in Appendix 1 that the inductance of a single-phase of an air-gap winding is: La =

n

Lnp

where the harmonic components are: Lk = 8 + +

2 2 0 lkwn Na k(1 x2 )2

1 x2k 2k
2

(4 k2 ) (1 2k ) +

1 x2+k
2

(2 + k)2 (1 2k ) 1 2k x2+k

2k 1 xk+2

(4 k2 ) ( 2k 1)

1 x2k

(2 k)2 ( 2k 1)

2k 1 x2k

where we have used the following shorthand coecients: x = = = Rwi Rwo Ri Rs Rwo Rs

k 1 x2 4 k2 2

This ts into the conventional inductance framework: Ln =

2 2 4 0 Na Rs Lkwn ka N 2 p2 g

if we assign the thick armature coecient to be: ka = 1 x2k 2k 1 x2+k 1 2gk Rwo (1 x2 )2 (4 k2 ) (1 2k ) + + (2 + k)2 (1 2k ) 1 2k x2+k 2k 1 xk+2
2

(4 k2 ) ( 2k 1)

1 x2k

(2 k)2 ( 2k 1) k 1 4 k2 2 x2

2k 1 x2k

20

and k = np and g = Rs Ri is the conventionally dened air gap. If the aspect ratio Ri /Rs is not too far from unity, neither is ka . In the case of p = 2, the fundamental component of ka is: 2gk 4 4 1 x4 1 1 x4 2 4 + x4 1 4 ka = log x + 4 (log x)2 + 2 Rwo (1 x2 ) 8 4 (1 4 ) (1 4 ) 16 (1 4 )
2

For a q-phase winding, a good approximation to the inductance is given by just the rst space harmonic term, or: 2 2 q 4 0 Na Rs Lkwn ka Ld = 2 n 2 p2 g 4.4.3 Internal Magnet Motor

The permanent magnets will have an eect on reactance because the magnets are in the main ux path of the armature. Further, they aect direct and quadrature reactances dierently, so that the machine will be salient. Actually, the eect on the direct axis will likely be greater, so that this type of machine will exhibit negative saliency: the quadrature axis reactance will be larger than the direct- axis reactance. A full- pitch coil aligned with the direct axis of the machine would produce ux density: Br = 0 Na I 2g 1 +
Rp wm 4g hm

Note that only the pole area is carrying useful ux, so that the space fundamental of radial ux density is: B1 = 0 Na I 4 sin pm 2 2g 1 + wm Rp
hm 4g

Then, since the ux linked by the winding is: a = 2RlNa kw B1 p

The d- axis inductance, including mutual phase coupling, is (for a q- phase machine): Ld =
2 2 pp q 4 0 Na Rlkw m sin 2g 2 p 2

The quadrature axis is quite dierent. On that axis, the armature does not tend to push ux through the magnets, so they have only a minor eect. What eect they do have is due to the fact that the magnets produce a space in the active air- gap. Thus, while a full- pitch coil aligned with the quadrature axis will produce an air- gap ux density: Br = the space fundamental of that will be: 21 0 N I g

B1 =

0 N I 4 pt 1 sin g 2

where t is the angular width taken out of the pole by the magnets. So that the expression for quadrature axis inductance is: Lq =
2 2 q 4 0 Na Rlkw 2 p2 g

1 sin

pt 2

Current Rating and Resistance

The last part of machine rating is its current capability. This is heavily inuenced by cooling methods, for the principal limit on current is the heating produced by resistive dissipation. Generally, it is possible to do rst-order design estimates by assuming a current density that can be handled by a particular cooling scheme. Then, in an air-gap winding: we Ja 2 and note that, usually, the armature lls the azimuthal space in the machine:
2 2 Na Ia = Rwo Rwi

2qwe = 2 For a winding in slots, nearly the same thing is true: if the rectangular slot model holds true: 2qNa Ia = Ns hs ws Js where we are using Js to note slot current density. Now, suppose we can characterize the total slot area by a space factor s which is the ratio between total slot area and the annulus occupied by the slots: for the rectangular slot model: s = Ns hs ws 2 2 Rwo Rwi

where Rwi = R+hd and Rwo = Rwi +hs in a normal, stator outside winding. In this case, Ja = Js s and the two types of machines can be evaluated in the same way. It would seem apparent that one would want to make s as large as possible, to permit high currents. The limit on this is that the magnetic teeth between the conductors must be able to carry the air-gap ux, and making them too narrow would cause them to saturate. The peak of the time fundamental magnetic eld in the teeth is, for example, Bt = B 1 where wt is the width of a stator tooth: wt = so that Bt 2(R + hd ) ws Ns B1 1 s 22 2R Ns wt

5.1

Resistance

Winding resistance may be estimated as the length of the stator conductor divided by its area and its conductivity. The length of the stator conductor is: lc = 2lNa fe where the end winding factor fe is used to take into account the extra length of the end turns (which is usually not negligible). The area of each turn of wire is, for an air-gap winding : Aw =
2 2 we Rwo Rwi w 2 Na

where w , the packing factor relates the area of conductor to the total area of the winding. The resistance is then just: 2 4lNa Ra = 2 2 we Rwo Rwi w and, of course, is the conductivity of the conductor. For windings in slots the expression is almost the same, simply substituting the total slot area: Ra =
2 2qlNa Ns hs ws w

The end turn allowance depends strongly on how the machine is made. One way of estimating what it might be is to assume that the end turns follow a roughly circular path from one side of the machine to the other. The radius of this circle would be, very roughly, Rw /p, where Rw is the average radius of the winding: Rw (Rwo + Rwi )/2 Then the end-turn allowance would be: fe = 1 + Rw pl

Appendix 1: Air-Gap Winding Inductance

In this appendix we use a simple two-dimensional model to estimate the magnetic elds and then inductances of an air-gap winding. The principal limiting assumption here is that the winding is uniform in the z direction, which means it is long in comparison with its radii. This is generally not true, nevertheless the answers we will get are not too far from being correct. The style of analysis used here can be carried into a three-dimensional, or quasi-three dimensional domain to get much more precise answers, at the expense of a very substantial increase in complexity. The coordinate system to be used is shown in Figure 10. To maintain generality we have four radii: Ri and Rs are ferromagnetic boundaries, and would of course correspond with the machine shaft and the stator core. The winding itself is carried between radii R1 and R2 , which correspond with radii Rwi and Rwo in the body of the text. It is assumed that the armature is carrying a current in the z- direction, and that this current is uniform in the radial dimension of the armature. If a single phase of the armature is carrying current, that current will be: Jz0 =
we 2

N a Ia 2 2 R2 R 1

23

Winding

Outer Magnetic Boundary 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 R 11111111111111111111111111111 00000000000000000000000000000 s 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 R 11111111111111111111111111111 00000000000000000000000000000 i 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 R 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 1 11111111111111111111111111111 00000000000000000000000000000 Inner Magnetic R 11111111111111111111111111111 00000000000000000000000000000 Boundary 2 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000

Figure 10: Coordinate System for Inductance Calculation over the annular wedge occupied by the phase. The resulting distribution can be fourier analyzed, and the n-th harmonic component of this will be (assuming the coordinate system has been chosen appropriately): 4 Na Ia 4 we = Jzn = Jz0 sin n 2 2 kwn n 2 R2 R 1 where the n-th harmonic winding factor is: kwn = sin n we 2 n we 2

and note that we is the electrical winding angle: we = pw Now, it is easiest to approach this problem using a vector potential. Since the divergence of ux density is zero, it is possible to let the magnetic ux density be represented by the curl of a vector potential: B =A Taking the curl of that: A = 0 J = A 2 A and using the Coulomb gage we have a reasonable tractable partial dierential equation in the vector potential: 2 A = 0 J Now, since in our assumption there is only a z- directed component of J, we can use that one component, and in circular cylindrical coordinates that is: 1 Az 1 2 r + 2 2 Az = 0 Jz r r r r 24 A =0

For this problem, all variables will be varying sinusoidally with angle, so we will assume that angular dependence ejk . Thus: k2 1 Az r 2 Az = 0 Jz r r r r This is a three-region problem. Note the regions as: i w o Ri < r < R1 R1 < r < R2 R2 < r < Rs (40)

For i and o, the current density is zero and an appropriate solution to (40) is: Az = A+ r k + A r k In the region of the winding, w, a particular solution must be used in addition to the homogeneous solution, and Az = A+ r k + A r k + Ap where, for k = 2, Ap = or, if k = 2, Ap = 0 Jz r 2 4 k2

1 0 Jz r 2 log r 4 4

And, of course, the two pertinent components of the magnetic ux density are: Br = B 1 Az r Az = r

Next, it is necessary to match boundary conditions. There are six free variables and correspondingly there must be six of these boundary conditions. They are the following: At the inner and outer magnetic boundaries, r = Ri and r = Rs , the azimuthal magnetic eld must vanish. At the inner and outer radii of the winding itself, r = R1 and r = R2 , both radial and azimuthal magnetic eld must be continuous. These conditions may be summarized by:
k1 k1 kAi Ri kAi Ri = 0 +

k1 k1 = 0 kAo Rs kAo Rs + 0 Jz R2 k1 k1 k1 o k1 Aw R2 + Aw R2 = A+ R2 + Ao R2 + 4 k2

25

20 Jz R2 4 k2 0 Jz R1 k1 k1 Aw R1 + Aw R1 + 4 k2 20 Jz R1 k1 k1 kAw R1 + kAw R1 + + 4 k2
k1 w k1 + kAw R2 + kA R2 +

k1 k1 = kAo R2 + kAo R2 + k1 k1 i = A+ R1 + Ai R1 k1 k1 = kAi R1 + kAi R1 +

Note that we are carrying this out here only for the case of k = 2. The k = 2 case may be obtained by substituting its particular solution in at the beginning or by using LHopitals rule on the nal solution. This set may be solved (it is a bit tedious but quite straightforward) to yield, for the winding region: Az = 0 Jz 2k +

2+k 2+k
2k 2k 2k 2k R2 R1 Rs R 2 R i R 1 + 2k 2k 2k 2k (2 + k) Rs Ri (2 k) Rs Ri 2k 2k R2 R 1

r k

2k 2k (2 k) Ri Rs

2k 2+k 2k 2+k Rs R2 Ri R1 2k 2k (2 + k) Ri Rs

2k 2 r k r 4 k2

Now, the inductance linked by any single, full-pitched loop of wire located with one side at azimuthal position and radius r is: i = 2lAz (r, ) To extend this to the whole winding, we integrate over the area of the winding the incremental ux linked by each element times the turns density. This is, for the n-th harmonic of ux linked: n = 4lkwn Na 2 2 R2 R 1
R2 R1

Az (r)rdr

Making the appropriate substitutions for current into the expression for vector potential, this becomes: n =
2 2 8 0 lkwn Na Ia 2 2 k R2 R 1 2 2+k 2+k 2k 2k 2k 2k R2 R 1 Rs R2 Ri R1 + 2k 2k 2k 2k (2 k) Rs Ri (2 + k) R2 Ri k+2 k+2 R2 R 1 k+2

2k 2k (2 k) Ri Rs

2k 2k R2 R 1

2+k 2k 2+k 2k 2k 4 4 R2k R2 Ri R1 R2 R1 2k R2 R1 + s 2k 2k 4 k2 4 (2 + k) R2k Rs i

Appendix 2: Permanent Magnet Field Analysis

This section is a eld analysis of the kind of radially magnetized, permanent magnet structures commonly used in electric machinery. It is a fairly general analysis, which will be suitable for use with either surface or in-slot windings, and for the magnet inside or the magnet outside case. This is a two-dimensional layout suitable for situations in which eld variation along the length of the structure is negligible.

26

Layout

The assumed geometry is shown in Figure 11. Assumed iron (highly permeable) boundaries are at radii Ri and Rs . The permanent magnets, assumed to be polarized radially and alternately (i.e. North-South ...), are located between radii R1 and R2 . We assume there are p pole pairs (2p magnets) and that each magnet subsumes an electrical angle of me . The electrical angle is just p times the physical angle, so that if the magnet angle were me = , the magnets would be touching.

Magnets

Outer Magnetic Boundary 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 R 11111111111111111111111111111 00000000000000000000000000000 s 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 R 11111111111111111111111111111 00000000000000000000000000000 i 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 11111111111111111111111111111 00000000000000000000000000000 111111111111 000000000000 R 11111111111111111111111111111 00000000000000000000000000000 000000000000 111111111111 1 11111111111111111111111111111 00000000000000000000000000000 Inner Magnetic R 11111111111111111111111111111 00000000000000000000000000000 Boundary 2 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000

Figure 11: Axial View of Magnetic Field Problem If the magnets are arranged so that the radially polarized magnets are located around the azimuthal origin ( = 0), the space fundamental of magnetization is: M = ir M0 cos p where the fundamental magnitude is: M0 = 4 me Brem sin 2 0 (42) (41)

and Brem is the remanent magnetization of the permanent magnet. Since there is no current anywhere in this problem, it is convenient to treat magnetic eld as the gradient of a scalar potential: H = The divergence of this is: 2 = H 27 (44) (43)

Since magnetic ux density is divergence-free, B =0 we have: H = M or: (46) (45)

1 (47) 2 = M = M0 cos p r Now, if we let the magnetic scalar potential be the sum of particular and homogeneous parts: = p + h (48)

1 2 p = M0 cos p (49) r We can nd a suitable solution to the particular part of this in the region of magnetization by trying: (50) p = Cr cos p Carrying out the Laplacian on this: 1 2 p = Cr 2 2 p2 cos p = M0 cos p r which works if = 1, in which case: p = M0 r cos p 1 p2 (52) (51)

where 2 h = 0, then:

Of course this solution holds only for the region of the magnets: R1 < r < R2 , and is zero for the regions outside of the magnets. A suitable homogeneous solution satises Laplaces equation, 2 h = 0, and is in general of the form: (53) h = Ar p cos p + Br p cos p Then we may write a trial total solution for the ux density as: Ri < r < R1 = R1 < r < R2 = R2 < r < Rs = A1 r p + B1 r p cos p M0 r cos p A2 r p + B2 r p + 1 p2 A3 r p + B3 r p cos p (54) (55) (56)

The boundary conditions at the inner and outer (assumed innitely permeable) boundaries at r = Ri and r = Rs require that the azimuthal eld vanish, or = 0, leading to:
2p B1 = Ri A1 2p Rs A3

(57) (58)

B3 =

28

At the magnet inner and outer radii, H and Br must be continuous. These are: H = Br These become, at r = R1 :

p1 2p p1 pA1 R1 Ri R1 p1 2p p1 pA1 R1 + Ri R1 p1 p1 = p A2 R1 + B2 R1 p p1 p1 = p A2 R1 B2 R1

1 r + Mr = 0 r

(59) (60)

M0 1 p2 M0 + M0 1 p2 M0 1 p2 M0 + M0 1 p2

(61) (62)

and at r = R2 :
p1 2p p1 pA3 R2 Rs R2 p1 2p p1 pA3 R2 + Rs R2 p1 p1 = p A2 R2 + B2 R2 p p1 p1 = p A2 R2 B2 R2

(63) (64)

Some small-time manipulation of these yields:

p 2p p A1 R1 Ri R1 p 2p p A1 R1 + Ri R1 p 2p p A3 R2 Rs R2 p 2p p A3 R2 + Rs R2

M0 1 p2 M0 p p = A2 R1 B2 R1 + pR1 1 p2 M0 p p = A2 R2 + B2 R2 + R2 1 p2 M0 p p = A2 R2 B2 R2 + pR2 1 p2
p p = A2 R1 + B2 R1 + R1

(65) (66) (67) (68)

Taking sums and dierences of the rst and second and then third and fourth of these we obtain:
p p 2A1 R1 = 2A2 R1 + R1 M0 2p p 2A1 Ri R1 p 2A3 R2 2p p 2A3 Rs R2

p p and then multiplying through by appropriate factors (R2 and R1 ) and then taking sums and dierences of these, p p p p (A1 A3 ) R1 R2 = (R1 R2 R2 R1 ) 2p p p 2p A1 Ri A3 Rs R1 R2 =

1+p 1 p2 p1 p = 2B2 R1 + R1 M0 1 p2 1+p p = 2A2 R2 + R2 M0 1 p2 p1 p = 2B2 R2 + R2 M0 1 p2

(69) (70) (71) (72)

M0 p + 1 2 1 p2 p p M0 p 1 R1 R2 R2 R1 2 1 p2

(73) (74)

29

Dividing through by the appropriate groups: A1 A3 =

2p 2p A1 Ri A3 Rs = p p R1 R2 R2 R1 M0 1 + p p p R1 R2 2 1 p2

(75) (76)

p p R1 R2 R2 R1 M0 p 1 p p 2 1 p2 R1 R2

2p and then, by multiplying the top equation by Rs and subtracting: 2p 2p A1 Rs Ri = p p R1 R2 R2 R1 M0 1 + p p p R1 R2 2 1 p2 2p Rs p p R1 R2 R2 R1 M0 p 1 p p 2 1 p2 R1 R2

(77)

This is readily solved for the eld coecients A1 and A3 : A1 = A3 = M0 2 2

2p Rs 2p Ri 2p Ri

p+1 p1 1p 1p 1+p 1+p 2p Rs + 2 R2 R 1 R1 R 2 21 p p 1 1 1 1p 1p 2p 1+p 1+p R1 R 2 Ri R2 R1 1p 1+p

(78) (79)

M0
2p Rs

Now, noting that the scalar potential is, in region 1 (radii less than the magnet),
2p = A1 (r p Ri r p ) cos p 2p = A3 (r p Rs r p ) cos p

r < R1 r > R2

and noting that p(p + 1)/(p2 1) = p/(p 1) and p(p 1)/(p2 1) = p/(p + 1), magnetic eld is: r < R1 M0 2
2p Rs

(80)
2p Ri

Hr =

Hr =

r > R2 M0 2
2p Rs

p p 1p 1p 1+p 1+p 2p R1 R 2 R2 R1 Rs + p1 p+1 p p 1p 1p 2p 1+p 1+p R2 R1 R1 R 2 Ri + p1 p+1

2p r p1 + Ri r p1 cos p

(81)
2p Ri 2p r p1 + Rs r p1 cos p

The case of p = 1 appears to be a bit troublesome here, but is easily handled by noting that: lim p R2 1p 1p = log R1 R2 p1 R1

p1

Now: there are a number of special cases to consider.

For the iron-free case, Ri 0 and R2 , this becomes, simply, for r < R1 :
Hr = M0 p 1p 1p r p1 cos p R1 R 2 2 p1 30 (82)

Note that for the case of p = 1, the limit of this is

Hr = and for r > R2 :
Hr = M0 p p+1 p+1 R2 R1 r (p+1) cos p 2 p+1 R2
M0 log cos 2 R1

For the case of a machine with iron boundaries and windings in slots, we are interested in the elds at the boundaries. In such a case, usually, either Ri = R1 or Rs = R2 . The elds are: at the outer boundary: r = Rs : Hr = M0
p1 Rs 2p Rs 2p Ri

p p 1p 1p p+1 p+1 R2p R1

R2 R2 R1 + p+1 p1 i

cos p

or at the inner boundary: r = Ri :

Hr = M0
p1 Ri 2p Rs 2p Ri

p p p+1 p+1 1p 1p R2p R1 R2 + R2 R1 p+1 p1 s

cos p

31

Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science 6.685 Electric Machinery Class Notes 8: Analytic Design Evaluation of Induction Machines c 2005 James L. Kirtley Jr.

January 12, 2006

Introduction

Induction machines are perhaps the most widely used of all electric motors. They are generally simple to build and rugged, oer reasonable asynchronous performance: a manageable torque-speed curve, stable operation under load, and generally satisfactory eciency. Because they are so widely used, they are worth understanding. In addition to their current economic importance, induction motors and generators may nd application in some new applications with designs that are not similar to motors currently in commerce. An example is very high speed motors for gas compressors, perhaps with squirrel cage rotors, perhaps with solid iron (or perhaps with both). Because it is possible that future, high performance induction machines will be required to have characteristics dierent from those of existing machines, it is necessary to understand them from rst principles, and that is the objective of this document. It starts with a circuit theoretical view of the induction machine. This analysis is strictly appropriate only for wound-rotor machines, but leads to an understanding of more complex machines. This model will be used to explain the basic operation of induction machines. Then we will derive a model for squirrel-cage machines. Finally, we will show how models for solid rotor and mixed solid rotor/squirrel cage machines can be constructed. The view that we will take in this document is relentlessly classical. All of the elements that we will use are calculated from rst principles, and we do not resort to numerical analysis or empirical methods unless we have no choice. While this may seem to be seriously limiting, it serves our basic objective here, which is to achieve an understanding of how these machines work. It is our feeling that once that understanding exists, it will be possible to employ more sophisticated methods of analysis to get more accurate results for those elements of the machines which do not lend themselves to simple analysis. An elementary picture of the induction machine is shown in Figure 1. The rotor and stator are coaxial. The stator has a polyphase winding in slots. The rotor has either a winding or a cage, also in slots. This picture will be modied slightly when we get to talking of solid rotor machines, anon. Generally, this analysis is carried out assuming three phases. As with many systems, this generalizes to dierent numbers of phases with little diculty.

Induction Motor Transformer Model

The induction machine has two electrically active elements: a rotor and a stator. In normal operation, the stator is excited by alternating voltage. (We consider here only polyphase machines). The stator excitation creates a magnetic eld in the form of a rotating, or traveling wave, which induces currents in the circuits of the rotor. Those currents, in turn, interact with the traveling 1

Stator Core

Stator Winding in Slots Rotor Winding or Cage in Slots

Rotor

AirGap

Figure 1: Axial View of an Induction Machine wave to produce torque. To start the analysis of this machine, assume that both the rotor and the stator can be described by balanced, three phase windings. The two sets are, of course, coupled by mutual inductances which are dependent on rotor position. Stator uxes are (a , b , c ) and rotor uxes are (A , B , C ). The ux vs. current relationship is given by:

a b c A B C

L S MT SR

M SR LR

ia ib ic iA iB iC

(1)

where the component matrices are: La Lab Lab LS = Lab La Lab Lab Lab La LA LAB LAB LAB LR = LAB LA LAB LAB LA M cos(p + 2 ) M cos(p) 3 2 3 ) M cos(p 2

2 3 )

(2)

(3)

The mutual inductance part of (1) is a circulant matrix: M cos(p) = M cos(p M cos(p +

M SR

M cos(p M cos(p + 2 3 ) M cos(p)

2 3 ) 2 3 )

(4)

To carry the analysis further, it is necessary to make some assumptions regarding operation. To start, assume balanced currents in both the stator and rotor: ia = IS cos(t)
ib = IS cos(t ic = IS cos(t +
2 3 ) 2 3 )

(5)

iA = IR cos(R t + R )
iB = IR cos(R t + R iC = IR cos(R t + R + The rotor position can be described by = m t + 0

2 3 ) 2 3 )

(6)

(7)

Under these assumptions, we may calculate the form of stator uxes. As it turns out, we need only write out the expressions for a and A to see what is going on: a = (La Lab )Is cos(t) + M IR (cos(R t + R ) cos p(m + 0 ) (8) 2 2 2 2 + cos(R t + R + ) cos(p(m t + 0 ) ) + cos(R t + R ) cos(p(m t + 0 ) + ) 3 3 3
3 which, after reducing some of the trig expressions, becomes:
3 a = (La Lab )Is cos(t) + M IR cos((pm + R )t + R + p0 ) 2
Doing the same thing for the rotor phase A yields:
A = M Is (cos p(m t + 0 ) cos(t)) + cos(p(m t + 0 ) + cos(p(m t + 0 ) + 2 2 ) cos(t ) 3 3 (9)

(10)

2 2 ) cos(t + ) + (LA LAB )IR cos(R t + R ) 3 3

This last expression is, after manipulating:

3
A = M Is cos(( pm )t p0 ) + (LA LAB )IR cos(R t + R ) 2 (11)

These two expressions, 9 and 11 give expressions for uxes in the armature and rotor windings in terms of currents in the same two windings, assuming that both current distributions are sinusoidal in time and space and represent balanced distributions. The next step is to make another assumption, that the stator and rotor frequencies match through rotor rotation. That is: pm = R It is important to keep straight the dierent frequencies here: R m is stator electrical frequency is rotor electrical frequency is mechanical rotation speed (12)

so that pm is electrical rotation speed. To refer rotor quantities to the stator frame (i.e. non- rotating), and to work in complex amplitudes, the following denitions are made: a = Re(a ejt ) A = Re(A ejR t ) ia = Re(I a ejt ) iA = Re(I A ejR t ) With these denitions, the complex amplitudes embodied in 56 and 64 become: 3 a = LS I a + M I A ej(R +p0 ) 2 (17) (13) (14) (15) (16)

3 A = M I a ejp0 + LR I A ejR (18) 2 There are two phase angles embedded in these expressions: 0 which describes the rotor physical phase angle with respect to stator current and R which describes phase angle of rotor currents with respect to stator currents. We hereby invent two new rotor variables: AR = A ejp) I AR = I A ej(p0 +R ) (19) (20)

These are rotor ux and current referred to armature phase angle. Note that AR and I AR have the same phase relationship to each other as do A and I A . Using 19 and 20 in 17 and 18, the basic ux/current relationship for the induction machine becomes: a AR = LS 3 2M
3 2M LR

Ia I AR

(21)

This is an equivalent single- phase statement, describing the ux/current relationship in phase a, assuming balanced operation. The same expression will describe phases b and c. Voltage at the terminals of the stator and rotor (possibly equivalent) windings is, then: V a = ja + Ra I a V AR = jR AR + RA I AR or: 3 V a = jLS I a + j M I AR + Ra I a 2 (22) (23) (24)

3 (25) V AR = jR M I a + jR LR I AR + RA I AR 2 To carry this further, it is necessary to go a little deeper into the machines parameters. Note that LS and LR are synchronous inductances for the stator and rotor. These may be separated into space fundamental and leakage components as follows: 4

LS = La Lab = LR = LA LAB =

2 2 3 4 0 RlNS kS + LSl 2g 2 p 2 2 3 4 0 RlNR kR + LRl 2 p2 g

(26) (27)

Where the normal set of machine parameters holds: R l g p N k S R Ll is rotor radius is active length
is the eective air- gap
is the number of pole- pairs
represents number of turns
represents the winding factor
as a subscript refers to the stator
as a subscript refers to the rotor
is leakage inductance

The two leakage terms LSl and LRl contain higher order harmonic stator and rotor inductances, slot inducances, end- winding inductances and, if necessary, a provision for rotor skew. Essentially, they are used to represent all ux in the rotor and stator that is not mutually coupled. In the same terms, the stator- to- rotor mutual inductance, which is taken to comprise only a space fundamental term, is: M= 4 0 RlNS NR kS kR p2 g (28)

Note that there are, of course, space harmonic mutual ux linkages. If they were to be included, they would hair up the analysis substantially. We ignore them here and note that they do have an eect on machine behavior, but that eect is second- order. Air- gap permeance is dened as: ag = so that the inductances are: 3 2 2 LS = ag kS NS + LSl 2 3 2 2 LR = ag kR NR + LRl 2 M = ag NS NR kS kR Here we dene slip s by: R = s 5 (33) (30) (31) (32) 4 0 Rl p2 g (29)

so that
s=1 Then the voltage balance equations become: V a = j 3 3 2 2 ag kS NS + LSl I a + j ag NS NR kS kR I AR + Ra I a 2 2 (35) (36) pm (34)

3 3 2 2 ag kR NR + LRl I AR + RA I AR V AR = js ag NS NR kS kR I a + js 2 2

At this point, we are ready to dene rotor current referred to the stator. This is done by assuming an eective turns ratio which, in turn, denes an equivalent stator current to produce the same fundamental MMF as a given rotor current: I2 = NR kR I NS kS AR (37)

Now, if we assume that the rotor of the machine is shorted so that V AR = 0 and do some manipulation we obtain: V a = j(XM + X1 )I a + jXM I 2 + Ra I a 0 = jXM I a + j(XM + X2 )I 2 + where the following denitions have been made: XM = 3 2 2 ag NS kS 2 (40) (41)
2

(38) (39)

R2 I s 2

X1 = LSl X2 = LRl R2 = RA NS kS NR kR NS kS NR kR
2

(42) (43)

These expressions describe a simple equivalent circuit for the induction motor shown in Figure 2. We will amplify on this equivalent circuit anon.

Ia R a

X1

X2 I2 < > Xm < R2 > < s

Figure 2: Equivalent Circuit

Operation: Energy Balance

Now we are ready to see how the induction machine actually works. Assume for the moment that Figure 2 represents one phase of a polyphase system and that the machine is operated under balanced conditions and that speed is constant or varying only slowly. Balanced conditions means that each phase has the same terminal voltage magnitude and that the phase dierence between phases is a uniform. Under those conditions, we may analyze each phase separately (as if it were a single phase system). Assume an RMS voltage magnitude of Vt across each phase. The gap impedance, or the impedance looking to the right from the right-most terminal of X1 is: R2 Zg = jXm ||(jX2 + ) (44) s A total, or terminal impedance is then Zt = jX1 + Ra + Zg and terminal current is It = Rotor current is found by using a current divider: I2 = It jXm jX2 + R2 + jXm s R2 s (47) Vt Zt (45) (46)

Air-gap power is then calculated (assuming a three-phase machine): Pag = 3|I2 |2 (48)

This is real (time-average) power crossing the air-gap of the machine. Positive slip implies rotor speed less than synchronous and positive air-gap power (motor operation). Negative slip means rotor speed is higher than synchronous, negative air-gap power (from the rotor to the stator) and generator operation. Now, note that this equivalent circuit represents a real physical structure, so it should be possible to calculate power dissipated in the physical rotor resistance, and that is: Ps = Pag s 7 (49)

(Note that, since both Pag and s will always have the same sign, dissipated power is positive.) The rest of this discussion is framed in terms of motor operation, but the conversion to generator operation is simple. The dierence between power crossing the air-gap and power dissipated in the rotor resistance must be converted from mechanical form: Pm = Pag Ps and electrical inputpower is: Pin = Pag + Pa where armature dissipation is: Pa = 3|It |2 Ra Output (mechanical) power is Pout = Pag Pw Where Pw describes friction, windage and certain stray losses which we will discuss later. And, nally, eciency and power factor are: = Pout Pin Pin 3Vt It (53) (52) (51) (50)

(54) (55)

cos =

3.1

Example of Operation

The following MATLAB script generates a torque-speed and power-speed curve for the simple induction motor model described above. Note that, while the analysis does not require that any of the parameters, such as rotor resistance, be independent of rotor speed, this simple script does assume that all parameters are constant.

3.2

Example

That MATLAB script has been run for a standard motor with parameters given in Table 1. Torque vs. speed and power vs. speed are plotted for this motor in Figure 3. These curves were generated by the MATLAB script shown above.

Squirrel Cage Machine Model

Now we derive a circuit model for the squirrel-cage motor using eld analytical techniques. The model consists of two major parts. The rst of these is a description of stator ux in terms of stator and rotor currents. The second is a description of rotor current in terms of air- gap ux. The result of all of this is a set of expressions for the elements of the circuit model for the induction machine. To start, assume that the rotor is symmetrical enough to carry a surface current, the fundamental of which is: K r = z Re K r ej(stp ) = z Re K r ej(tp) 8 (56)

% ------------------------------------------------------
% Torque-Speed Curve for an Induction Motor
% Assumes the classical model
% This is a single-circuit model
% Required parameters are R1, X1, X2, R2, Xm, Vt, Ns
% Assumed is a three-phase motor
% This thing does a motoring, full speed range curve
% Copyright 1994 James L. Kirtley Jr.
% -------------------------------------------------------
s = .002:.002:1; % vector of slip N = Ns .* (1 - s); % Speed, in RPM oms = 2*pi*Ns/60; % Synchronous speed Rr = R2 ./ s; % Rotor resistance Zr = j*X2 + Rr; % Total rotor impedance Za = par(j*Xm, Zr); % Air-gap impedance Zt = R1 + j*X1 +Za; % Terminal impedance Ia = Vt ./ Zt; % Terminal Current I2 = Ia .* cdiv (Zr, j*Xm); % Rotor Current Pag = 3 .* abs(I2) .^2 .* Rr; Air-Gap Power % Pm = Pag .* (1 - s); % Converted Power Trq = Pag ./ oms; % Developed Torque subplot(2,1,1) plot(N, Trq) title(Induction Motor); ylabel(N-m); subplot(2,1,2) plot(N, Pm); ylabel(Watts); xlabel(RPM);

Table 1: Example, Standard Motor

Rating Voltage Stator Resistance R1 Rotor Resistance R2 Stator Reactance X1 Rotor Reactance X2 Magnetizing Reactance Xm Synchronous Speed Ns 300 440 254 .73 .64 .06 .06 2.5 1200 kw VRMS, l-l VRMS, l-n RPM

Induction Motor
300
250
200
Nm 150
100
50
0
0 200 400 600 800 1000 1200

3 2.5
2

x 10

Watts

1.5
1
0.5
0
0 200 400 600 RPM 800 1000 1200

Figure 3: Torque and Power vs. Speed for Example Motor Note that in 56 we have made use of the simple transformation between rotor and stator coordinates: = m t (57) and that pm = r = (1 s) Here, we have used the following symbols: Kr s r m is is is is is complex amplitude of rotor surface current per- unit slip stator electrical frequency rotor electrical frequency rotational speed (58)

The rotor current will produce an air- gap ux density of the form: Br = Re B r ej(tp) where B r = j0 10 R K pg r (59)

(60)

Note that this describes only radial magnetic ux density produced by the space fundamental of rotor current. Flux linked by the armature winding due to this ux density is:
0

AR = lNS kS This yields a complex amplitude for AR :

Br ()Rd

(61)

AR = Re AR ejt where AR = 2l0 R2 NS kS Kr p2 g

(62)

(63)

Adding this to ux produced by the stator currents, we have an expression for total stator ux: a =
2 2 3 4 0 NS RlkS 2l0 R2 NS kS + LSl I a + Kr 2 p2 g p2 g

(64)

Expression 64 motivates a deniton of an equivalent rotor current I2 in terms of the space fundamental of rotor surface current density: I2 = R K 3 NS kS z (65)

Then we have the simple expression for stator ux: a = (Lad + LSl )I a + Lad I 2 where Lad is the fundamental space harmonic component of stator inductance: Lad =
2 2 3 4 0 NS kS Rl 2 p2 g

(66)

(67)

4.1

Eective Air-Gap: Carters Coecient

In induction motors, where the air-gap is usually quite small, it is necessary to correct the air-gap permeance for the eect of slot openings. These make the permeance of the air-gap slightly smaller than calculated from the physical gap, eectively making the gap a bit bigger. The ratio of eective to physical gap is: t+s (68) ge = g t + s gf () where f () = f s 2g = tan() log sec (69)

11

4.2

Squirrel Cage Currents

The second part of this derivation is the equivalent of nding a relationship between rotor ux and I2 . However, since this machine has no discrete windings, we must focus on the individual rotor bars. Assume that there are NR slots in the rotor. Each of these slots is carrying some current. If the machine is symmetrical and operating with balanced currents, we may write an expression for current in the kth slot as: ik = Re I k ejst where I k = Ie
j 2p N
R

(70)

(71)

and I is the complex amplitude of current in slot number zero. Expression 71 shows a uniform progression of rotor current phase about the rotor. All rotor slots carry the same current, but that current is phase retarded (delayed) from slot to slot because of relative rotation of the current wave at slip frequency. The rotor current density can then be expressed as a sum of impulses: Kz = Re

NR 1 k=0

The unit impulse function () is our way of approximating the rotor current as a series of impulsive currents around the rotor. This rotor surface current may be expressed as a fourier series of traveling waves:

1 j(r tk 2p ) 2k NR Ie ) ( R NR

(72)

Kz = Re
n=

K n ej(r tnp )

(73)

Note that in 73, we are allowing for negative values of the space harmonic index n to allow for reverse- rotating waves. This is really part of an expansion in both time and space, although we are considering only the time fundamental part. We may recover the nth space harmonic component of 73 by employing the following formula: K n =< 1
2 0

Kr (, t)ej(r tnp) d >

(74)

Here the brackets <> denote time average and are here beause of the two- dimensional nature of the expansion. To carry out 74 on 72, rst expand 72 into its complex conjugate parts: Kr = 1 2
NR 1 k=0

I j(r tk 2p ) I j(r tk 2p ) 2k NR NR ) e + e ( R R NR

(75)

If 75 is used in 74, the second half of 75 results in a sum of terms which time average to zero. The rst half of the expression results in:

12

Kn =

I 2R

2 NR 1 0 k=0

j 2pk jnp NR

2k )d NR

(76)

The impulse function turns the integral into an evaluation of the rest of the integrand at the impulse. What remains is the sum: Kn = I 2R
NR 1 k=0

j(n1) 2kp N
R

(77)

The sum in 77 is easily evaluated. It is:

NR 1 k=0

2kp(n1) NR

P NR if (n 1) NR = integer 0 otherwise

(78)

The integer in 78 may be positive, negative or zero. As it turns out, only the rst three of these (zero, plus and minus one) are important, because these produce the largest magnetic elds and therefore uxes. These are: (n 1) p NR = 1 or n = NR p p =0 =1 or n = 1 or n =
NR +p p

(79)

Note that 79 appears to produce space harmonic orders that may be of non- integer order. This is not really true: is is necessary that np be an integer, and 79 will always satisfy that condition. So, the harmonic orders of interest to us are one and n+ = n NR +1 p NR = 1 p (80) (81)

Each of the space harmonics of the squirrel- cage current will produce radial ux density. A surface current of the form: Kn = Re produces radial magnetic ux density:
Brn = Re B rn ej(r tnp ) where 0 NR I (84) 2npg In turn, each of the components of radial ux density will produce a component of induced voltage. To calculate that, we must invoke Faradays law: B rn = j 13

NR I j(r tnp ) e 2R

(82)

(83)

B t
The radial component of 85, assuming that the elds do not vary with z, is:
E = Br 1 Ez = R t Or, assuming an electric eld component of the form: Ezn = Re E n ej(r tnp)

(85)

(86)

(87)

Using 84 and 87 in 86, we obtain an expression for electric eld induced by components of airgap ux: r R En = (88) B np n E n = j 0 NR r R I 2g(np)2 (89)

Now, the total voltage induced in a slot pushes current through the conductors in that slot. We may express this by: E 1 + E n + E n+ = Z slot I (90)

Now: in 90, there are three components of air- gap eld. E1 is the space fundamental eld, produced by the space fundamental of rotor current as well as by the space fundamental of stator current. The other two components on the left of 90 are produced only by rotor currents and actually represent additional reactive impedance to the rotor. This is often called zigzag leakage inductance. The parameter Zslot represents impedance of the slot itself: resistance and reactance associated with cross- slot magnetic elds. Then 90 can be re-written as: E 1 = Z slot I + j 0 NR r R 2g 1 1 + I 2 (n+ p) (n p)2 (91)

To nish this model, it is necessary to translate 91 back to the stator. See that 65 and 77 make the link between I and I 2 : I2 = NR I 6NS kS 1 1 + 2 (n+ p) (n p)2 (92)

Then the electric eld at the surface of the rotor is: E1 = 6NS kS 3 0 NS kS R Z slot + jr NR g I2 (93)

This must be translated into an equivalent stator voltage. To do so, we use 88 to translate 93 into a statement of radial magnetic eld, then nd the ux liked and hence stator voltage from that. Magnetic ux density is:

14

Br = =

pE 1 r R 6NS kS p NR R

Rslot 3 0 NS kS p + jLslot + j r g

1 1 + 2 (n+ p) (n p)2

I2

(94)

where the slot impedance has been expressed by its real and imaginary parts: Z slot = Rslot + jr Lslot Flux linking the armature winding is:
0

(95)

ag = NS kS lR Which becomes:

2p

Re B r ej(tp) d

(96)

ag = Re ag ejt where: ag = j Then air- gap voltage is: V ag = jag = = I 2 2NS kS lR Br p 2 2 12lNS kS R2 jLslot + NR s 2NS kS lR B r p

(97)

(98)

+ j

2 2 6 0 RlNS kS g

1 1 + 2 (n p)2 (n+ p)

(99)

Expression 99 describes the relationship between the space fundamental air- gap voltage V ag and rotor current I 2 . This expression ts the equivalent circuit of Figure 4 if the denitions made below hold: X2 I2

< > < R2 > < s

Figure 4: Rotor Equivalent Circuit

X2 R2

2 2 2 2 12lNS kS 6 0 RlNS kS = Lslot + NR g 2 2 12lNS kS = Rslot NR

1 1 + 2 (NR + p) (NR p)2

(100) (101)

15

The rst term in 100 expresses slot leakage inductance for the rotor. Similarly, 101 expresses rotor resistance in terms of slot resistance. Note that Lslot and Rslot are both expressed per unit length. The second term in 100 expresses the zigzag leakage inductance resulting from harmonics on the order of rotor slot pitch. Next, see that armature ux is just equal to air- gap ux plus armature leakage inductance. That is, 66 could be written as: a = ag + Lal I a (102)

4.3

Stator Leakage

There are a number of components of stator leakage Lal , each representing ux paths that do not directly involve the rotor. Each of the components adds to the leakage inductance. The most prominent components of stator leakage are referred to as slot, belt, zigzag, end winding, and .skew Each of these will be discussed in the following paragraphs. 4.3.1 Belt Leakage

Belt and zigzag leakage components are due to air- gap space harmonics. As it turns out, these are relatively complicated to estimate, but we may get some notion from our rst- order view of the machine. The trouble with estimating these leakage components is that they are not really independent of the rotor, even though we call them leakage. Belt harmonics are of order n = 5 and n = 7. If there were no rotor coupling, the belt harmonic leakage terms would be: Xag5 = Xag7 =
2 2 3 4 0 NS k5 Rl 2 52 p2 g

(103)

2 2 3 4 0 NS k7 Rl (104) 2 72 p2 g The belt harmonics link to the rotor, however, and actually appear to be in parallel with components of rotor impedance appropriate to 5p and 7p pole- pair machines. At these harmonic orders we can usually ignore rotor resistance so that rotor impedance is purely inductive. Those components are: 2 2 2 2 6 0 RlNS k5 12lNS k5 Lslot + NR g 2 2 2 2 12lNS k7 6 0 RlNS k7 Lslot + NR g

X2,5 = X2,7 =

1 1 + 2 (NR + 5p) (NR 5p)2 1 1 + 2 (NR + 7p) (NR 7p)2

(105) (106)

In the simple model of the squirrel cage machine, because the rotor resistances are relatively small and slip high, the eect of rotor resistance is usually ignored. Then the fth and seventh harmonic components of belt leakage are: X5 = Xag5 X2,5 X7 = Xag7 X2,7 16 (107) (108)

4.3.2

Zigzag Leakage

Stator zigzag leakage is from those harmonics of the orders pns = Nslots p where Nslots . Xz =
2 3 4 0 NS Rl 2 g

Note that these harmonic orders do not tend to be shorted out by the rotor cage and so no direct interaction with the cage is ordinarily accounted for. 4.3.3 Skew Leakage

kns + kns + 2 (Nslots + p) (Nslots p)2

(109)

In order to reduce saliency eects that occur because the rotor teeth will tend to try to align with the stator teeth, induction motor designers always use a dierent number of slots in the rotor and stator. There still may be some tendency to align, and this produces cogging torques which in turn produce vibration and noise and, in severe cases, can retard or even prevent starting. To reduce this tendency to cog, rotors are often built with a little skew, or twist of the slots from one end to the other. Thus, when one tooth is aligned at one end of the machine, it is un-aligned at the other end. A side eect of this is to reduce the stator and rotor coupling by just a little, and this produces leakage reactance. This is fairly easy to estimate. Consider, for example, a space-fundamental ux density Br = B1 cos p, linking a (possibly) skewed full-pitch current path: =
l 2 l 2 +p x 2p l 2p + p x l

B1 cos pRddx

Here, the skew in the rotor is electrical radians from one end of the machine to the other. Evaluation of this yields: 2B1 Rl sin 2 = p 2 Now, the dierence between what would have been linked by a non-skewed rotor and what is linked by the skewed rotor is the skew leakage ux, now expressible as: Xk = Xag 1 4.3.4 Stator Slot Leakage

sin 2 2 2

Currents in the stator slots produce uxes that link the stator conductors but not the rotor. To estimate these uxes, refer to the slot geometry shown in Figure 4.3.4. This shows a possibly unrealistic straight-sided stator slot. Typical in induction machines is for such slots to be trapezoidal in shape. A more careful eld analysis than we will do here shows that this analysis will be no more than a few percent in error if the slot width used in the calculation is the slot top (the end of the slot closest to the air-gap). There are ve important dimensions here: the slot height h, width w and the slot depression height d and width u, and (not shown) length . To estimate slot leakage inductance we assume some current in the slot, calculate the magnetic energy that results and then use the expression:

17

1 wm = L I 2 2 If there are N conductors in the slot, each carrying current I, the current density in the slot is: J= NI hw

Using Amperes Law around a loop (shown dotted in the gure), magnetic eld in the x direction at height y from the bottom of the slot is: Hx = In the slot depression that eld is: NI y w h

NI u Magnetic energy stored in the slot and slot depression are then conveniently calculated as: Hxd = 1 wm = L I 2 = w 2 Noting the slot permeance as:
h 0

1 1h d 1 2 2 + 0 Hx dy + ud Hxd = 0 N 2I 2 2 2 3w u P = 0

1h d + 3w u We have the total inductance of the slot to be: L = PN 2 For the purpose of this estimate we will assume an ordinary winding consisting of coils of Nc turns each. For such a winding if there are m slots per pole per phase and p pole pairs and if the winding is short-pitched by Nsp slots, there will be 2p(m Nsp ) slots per phase with two coils from the same phase and 2pNsp slots per phase sharing another phase. (We assume here a three phase machine). Then the self slot leakage inductance must be:
2 2 Ls = P 4Nc 2p(m Nsp ) + Nc 2pNsp

Since there are a total of pNsp mutual slots between each pair of phases, and the sense of the windings is opposite, the mutual component of slot leakage is:
2 Lm = PpNsp Nc

Total slot leakage is then:

2 L = Ls Lm = PpNc (8m 5Nsp )

Expressed in terms of the total number of stator turns, Na = 2pmNc , L = P

2 Na p

2 5 Nsp m 4 m2

18

1111 1 0000 0 111111 000000 1 11 1 0 00 0 1111 11111 0000 00000 11111111 00000000 1 1111 0 0000 1 1 1 0J0 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 111111 000000 1 0 1 0 111111 000000

y x h

Figure 5: Stator Slot Geometry for Leakage Calculation 4.3.5 End Winding Leakage

The nal component of leakage reactance is due to the end windings. This is perhaps the most dicult of the machine parameters to estimate, being essentially three-dimensional in nature. There are a number of ways of estimating this parameter, but for our purposes we will use a simplied parameter from Alger[1]: Xe =
2 14 q 0 RNa (p 0.3) 4 2 2 p2

As with all such formulae, extreme care is required here, since we can give little guidance as to when this expression is correct or even close. And we will admit that a more complete treatment of this element of machine parameter construction would be an improvement.

4.4

Stator Winding Resistance

Estimating stator winding resistance is fairly straightforward once end winding geometry is known. Total length of the armature winding is, per phase: w = Na 2 ( + e ) Estimating e , the length of one end winding, requires knowing how the winding is laid out and is beyond our scope here. (But once you see it you will know that length.) The area of the winding may be estimated by knowing wire diameter and how many strands are in parallel: Aw = 2 d Ninh 4 w 2Nc Aw Aslot 19

The area of the winding is related to slot area by a winding factor: a =

Winding resistance, per phase, is simply Ra = w Aw

where is wire conductivity. Note that conductivity of the materials used in induction machines is a function of temperature and so will be winding resistance (and rotor resistance for that matter). The Fitzgerald, Kingsley and Umans textbook[2] gives the following correction for resistance of copper: RT = R t T0 + T T0 + t

where RT and Rt are resistances at temperatures T and t. T0 = 234.5 for copper with basic conductivity of IACS (5.8 107 S/m)[3]. For aluminum with conductivity of 63% of IACS, T0 212.9 Temperatures are given in Celcius.

4.5

Harmonic Order Rotor Resistance and Stray Load Losses

It is important to recognize that the machine rotor sees each of the stator harmonics in essentially the same way, and it is quite straightforward to estimate rotor parameters for the harmonic orders, as we have done just above. Now, particularly for the belt harmonic orders, there are rotor currents owing in response to stator mmfs at fth and seventh space harmonic order. The resistances attributable to these harmonic orders are: R2,5 = R2,7 =
2 2 12lNs k5 Rslot,5 NR 2 2 12lNs k7 Rslot,7 NR

(110) (111)

The higher-order slot harmonics will have relative frequencies (slips) that are: sn = 1 (1 s)n n = 6k + 1 n = 6k 1 k an integer (112)

The induction motor electromagnetic interaction can now be described by an augmented magnetic circuit as shown in Figure 20. Note that the terminal ux of the machine is the sum of all of the harmonic uxes, and each space harmonic is excited by the same current so the individual harmonic components are in series. Each of the space harmonics will have an electromagnetic interaction similar to the fundamental: power transferred across the air-gap is:
2 Pem,n = 3I2,n

R2,n sn

Of course dissipation in each circuit is:

2 Pd,n = 3I2,n R2,n

20

leaving

R2,n (1 sn ) sn Note that this equivalent circuit has provision for two sets of circuits which look like cages. In fact one of these sets is for the solid rotor body if that exists. We will discuss that anon. There is also a provision (rc ) for loss in the stator core iron. Power deposited in the rotor harmonic resistance elements is characterized as stray load loss because it is not easily computed from the simple machine equivalent circuit.
2 Pm,n = 3I2,n

4.6

Slot Models

Some of the more interesting things that can be done with induction motors have to do with the shaping of rotor slots to achieve particular frequency-dependent eects. We will consider here three cases, but there are many other possibilities. First, suppose the rotor slots are representable as being rectangular, as shown in Figure 6, and assume that the slot dimensions are such that diusion eects are not important so that current in the slot conductor is approximately uniform. In that case, the slot resistance and inductance per unit length are: Rslot = 1 ws hs hs Lslot = 0 3ws (113) (114)

The slot resistance is obvious, the slot inductance may be estimated by recognizing that if the current in the slot is uniform, magnetic eld crossing the slot must be: Hy = I x ws hs
2

Then energy stored in the eld in the slot is simply: 1 L I 2 = ws 2 slot

hs 0

0 2

Ix ws hs

dx =

1 0 hs 2 I 6 ws

4.7

Deep Slots

Now, suppose the slot is not small enough that diusion eects can be ignored. The slot becomes deep to the extent that its depth is less than (or even comparable to) the skin depthfor conduction at slip frequency. Conduction in this case may be represented by using the Diusion Equation: 2 H = 0 H t

In the steady state, and assuming that only cross-slot ux (in the y direction) is important, and the only variation that is important is in the radial (x) direction: 2 Hy = js 0 Hy x2 21

wd

hd hs y

ws
Figure 6: Single Slot This is solved by solutions of the form: Hy = H e(1+j) where the skin depth is = 2 s 0
x

Since Hy must vanish at the bottom of the slot, it must take the form: Hy = Htop Since current is the curl of magnetic eld, Jz = Ez = Hy 1 + j cosh(1 + j) hs = Htop x sinh(1 + j) hs sinh(1 + j) x sinh(1 + j) hs

Then slot impedance, per unit length, is: 1 1+j hs Zslot = coth(1 + j) ws Of course the impedance (purely reactive) due to the slot depression must be added to this. It is possible to extract the real and imaginary parts of this impedance (the process is algebraically a bit messy) to yield: Rslot = 1 sinh 2 hs + sin 2 hs ws cosh 2 hs cos 2 hs

Lslot = 0

1 1 sinh 2 hs sin 2 hs hd + wd s ws cosh 2 hs cos 2 hs 22

4.8

Arbitrary Slot Shape Model

It is possible to obtain a better model of the behavior of rotor conductor slots by using simple numerical methods. In many cases rotor slots are shaped with the following objectives in mind: 1. A substantial part of the periphery of the rotor should be devoted to active conductor, for good running performance. 2. The magnetic iron of the rotor must occupy a certain fraction of the periphery, to avoid saturation. 3. For good starting performance, some means of forcing current to ow only in the top part of the rotor bar should be devised. Generally the rotor teeth, which make up part of the machines magnetic circuit, are of roughly constant width to avoid ux concentration. The rotor conductor bars are therefore tapered, with their narrow ends towards the center of the rotor. To provide for current concentration on starting they often have a starting bar at the outer periphery of the rotor with a much narrower region which has high inductance just below. The bulk of the rotor bar occupies the tapered region allowed between the teeth. This geometry is quite a bit more complicated than that described in the previous section. Note that, if we can describe the slot impedance per unit length as a function of frequency: Zs () = Rs () + jXs (), we can carry out the analysis of the machine as described previously. Thus our analysis is directed toward frequency response modeling of the rotor slot. Focusing then on a single slot, use the notation as described in Figure 7.

Ez [n]

w[n]

x = n x

Ez [n1]

x y z

Figure 7: Slot Geometry Notation The impedance per unit length is the ratio between slot current and axial electric eld: Zs = Ez I

For the purpose of this analysis we will use the symbol x as the radial distance from the bottom of the slot. Assume the slot can be divided radially into a number of regions or slices, each with 23

radial height x. We further assume that currents are axially (z) directed and that magnetic eld crosses the slot in the y direction. Under these assumptions the electric eld at the top of one of the slices is related to the electric eld at the bottom of the slice by magnetic eld crossing through the slice. Using the trapezoidal rule for integration: E z (x) E z (x x) = j0 The magnetic eld is simply: H y (x) = 1 w(x)
x 0

x H y (x) + H y (x x) 2

w(x)Ez (x)dx =

1 wn

nI n
i=1

where In is the total current owing in one slice. Note that this can be reformulated into a ladder network by again using the trapezoidal rule for integration: current owing in slice number n would be: x (wn E n + wn1 E n1 ) In = 2 Now the slot may be described as is shown in the ladder network of Figure 8. The incremental reactance of one slice is: Xn = 0 x and the resistance of a slice is: Rn =
L[n] 2

2 (wn + wn1 )

1 2 x wn + wn1
L[n1] 2 L[n1] 2

L[n] 2

R[n]

R[n1]

Figure 8: Slot Impedance Ladder Network The procedure is to start at the bottom of the slot, corresponding to the right-hand end of the ladder (the inductance at the bottom of the slot is innite so the rst slice has only the resistance), and building toward the top of the slot.

4.9

Multiple Cages

In some larger induction motors the rotor cage is built in such a way as to separate the functions of starting and running. The purpose of a deep slot is to improve starting performance of a motor. When the rotor is stationary, the frequency seen by rotor conductors is relatively high, and current crowding due to the skin eect makes rotor resistance appear to be high. As the rotor accelerates the frequency seen from the rotor drops, lessening the skin eect and making more use of the rotor conductor. This, then, gives the machine higher starting torque (requiring high resistance) without compromising running eciency. 24

This eect can be carried even further by making use of multiple cages such as is shown in , Figure 9. Here there are two conductors in a fairly complex slot. Estimating the impedance of this slot is done in stages to build up an equivalent circuit.
wd hd h2 hs w2 ws h1

w1

Figure 9: Double Slot Assume for the purposes of this derivation that each section of the multiple cage is small enough that currents can be considered to be uniform in each conductor. Then the bottom section may be represented as a resistance in series with an inductance: Ra = La = 1 w1 h1 0 h1 3 w1

The narrow slot section with no conductor between the top and bottom conductors will contribute an inductive impedance: hs Ls = 0 ws The top conductor will have a resistance: Rb = 1 w2 h2

Now, in the equivalent circuit, current owing in the lower conductor will produce a magnetic eld across this section, yielding a series inductance of Lb = 0 h2 w2

By analogy with the bottom conductor, current in the top conductor ows through only one third of the inductance of the top section, leading to the equivalent circuit of Figure 10, once the inductance of the slot depression is added on: hd Lt = 0 wd Now, this rotor bar circuit ts right into the framework of the induction motor equivalent circuit, shown for the double cage case in Figure 11, with R2a =
2 2 12lNS kS Ra NR

25

Lt

1 3 Lb

< > < Rb > <

2 3 Lb

Ls

La

< > < Ra > <

Figure 10: Equivalent Circuit: Double Bar

2 2 12lNS kS Rb NR 2 2 12lNS kS 2 ( Lb + Ls + La ) = NR 3 2 2 12lNS kS 1 = (Lt + Lb ) NR 3

R2b = X2a X2a

Ia R a

X1

X2b I2 X2a < < > > Xm < R2b < R2a > s > s < <

Figure 11: Equivalent Circuit: Double Cage Rotor

4.10

Rotor End Ring Eects

It is necessary to correct for end ring resistance in the rotor. To do this, we note that the magnitude of surface current density in the rotor is related to the magnitude of individual bar current by: 2R Iz = Kz (115) NR Current in the end ring is: IR = Kz R p (116)

Then it is straightforward to calculate the ratio between power dissipated in the end rings to power dissipated in the conductor bars themselves, considering the ratio of current densities and volumes. Assuming that the bars and end rings have the same radial extent, the ratio of current densities is: JR NR wr = Jz 2p lr 26 (117)

where wr is the average width of a conductor bar and lr is the axial end ring length. Now, the ratio of losses (and hence the ratio of resistances) is found by multiplying the square of current density ratio by the ratio of volumes. This is approximately: Rend = Rslot NR wr 2p lr
2

2R lr NR Rwr = NR l wr llr p2

(118)

4.11

Windage

Bearing friction, windage loss and fan input power are often regarded as elements of a black art. We approach them with some level of trepidation, for motor manufacturers seem to take a highly empirical view of these elements. What follows is an attempt to build reasonable but simple models for two eects: loss in the air gap due to windage and input power to the fan for cooling. Some caution is required here, for these elements of calculation have not been properly tested, although they seem to give reasonable numbers The rst element is gap windage loss. This is produced by shearing of the air in the relative rotation gap. It is likely to be a signigant element only in machines with very narrow air gaps or very high surface speeds. But these include, of course, the high performance machines with which we are most interested. We approach this with a simple couette ow model. Air-gap shear loss is approximately: Pw = 2R4 3 la f (119) where a is the density of the air-gap medium (possibly air) and f is the friction factor estimated , by: .0076 (120)
f= 1 4 Rn and the Reynolds NumberRn is Rg (121) Rn = air and air is the kinematic viscosity of the air-gap medium. The second element is fan input power. We base an estimate of this on two hypotheses. The rst of these is that the mass ow of air circulated by the fan can be calculated by the loss in the motor and an average temperature rise in the cooling air. The second hypothesis is the the pressure rise of the fan is established by the centrifugal pressure rise associated with the surface speed at the outside of the rotor. Taking these one at a time: If there is to be a temperature rise T in the cooling air, then the mass ow volume is: m= and then volume ow is just v= Pressure rise is estimated by centrifugal force: P = air 27 r p fan
2

Pd Cp T m air

then power is given by: Pfan = P v For reference, the properties of air are: Density Kinematic Viscosity Heat Capacity air air Cp 1.18 1.56 105 1005.7 kg/m2 m2 /sec J/kg

4.12

Magnetic Circuit Loss and Excitation

There will be some loss in the stator magnetic circuit due to eddy current and hysteresis eects in the core iron. In addition, particularly if the rotor and stator teeth are saturated there will be MMF expended to push ux through those regions. These eects are very dicult to estimate from rst principles, so we resort to a simple model. Assume that the loss in saturated steel follows a law such as: Pd = PB e B
f

B BB

(122)

This is not too bad an estimate for the behavior of core iron. Typically, f is a bit less than two (between about 1.3 and 1.6) and b is a bit more than two (between about 2.1 and 2.4). Of course this model is good only for a fairly restricted range of ux density. Base dissipation is usually expressed in watts per kilogram, so we rst compute ux density and then mass of the two principal components of the stator iron, the teeth and the back iron. In a similar way we can model the exciting volt-amperes consumed by core iron by something like: Qc = V a1 B BB
v1

+ V a2

B BB

v2

(123)

This, too, is a form that appears to be valid for some steels. Quite obviously it may be necessary to develop dierent forms of curve ts for dierent materials. Flux density (RMS) in the air-gap is: Br = Then ux density in the stator teeth is: Bt = B r wt + w1 wt (125) pVa 2RlNa k1 s (124)

where wt is tooth width and w1 is slot top width. Flux in the back-iron of the core is Bc = B r where dc is the radial depth of the core. R pdc (126)

28

One way of handling this loss is to assume that the core handles ux corresponding to terminal voltage, add up the losses and then compute an equivalent resistance and reactance: rc = xc = 3|Va |2 Pcore 3|Va |2 Qcore

then put this equivalent resistance in parallel with the air-gap reactance element in the equivalent circuit.

Solid Iron Rotor Bodies

Solid steel rotor electric machines (SSRM) can be made to operate with very high surface speeds and are thus suitable for use in high RPM situations. They resemble, in form and function, hysteresis machines. However, asynchronous operation will produce higher power output because it takes advantage of higher ux density. We consider here the interactions to be expected from solid iron rotor bodies. The equivalent circuits can be placed in parallel (harmonic-by-harmonic) with the equivalent circuits for the squirrel cage, if there is also a cage in the machine. To estimate the rotor parameters R2s and X2s , we assume that important eld quantities in the machine are sinusoidally distributed in time and space, so that radial ux density is: Br = Re B r ej(tp) and, similarly, axially directed rotor surface current is: Kz = Re K z ej(tp) Now, since by Faradays law: E = we have, in this machine geometry: B t (129) (128) (127)

1 Br Ez = R t

(130)

The transformation between rotor and stator coordinates is: = m t where m is rotor speed. Then: pm = r = (1 s) and Now, axial electric eld is, in the frame of the rotor, just: Ez = Re E z ej(tp) = Re E z ej(r 29
tp )

(131) (132)

(133) (134)

and Ez =

r R Br p

(135)

Of course electric eld in the rotor frame is related to rotor surface current by: Ez = Z sK z (136)

Now these quantities can be related to the stator by noting that air-gap voltage is related to radial ux density by: p (137) Br = V 2lNa k1 R ag The stator-equivalent rotor current is: I2 = R K 3 Na ka z (138)

Then we can nd stator referred, rotor equivalent impedance to be: Z2 = V ag 3 4 l 2 2 Ez N k = I2 2 R a a r K z (139)

Now, if rotor surface impedance can be expressed as: Z s = Rs + jr Ls then Z2 = where R2 = X2 = 34 l 2 2 N k Rs 2R a 1 34 l 2 2 N k Xs 2R a 1 (142) (143) R2 + jX2 s (140)

(141)

Now, to nd the rotor surface impedance, we make use of a nonlinear eddy-current model proposed by Agarwal. First we dene an equivalent penetration depth (similar to a skin depth): = 2Hm r B0 (144)

where is rotor surface material volume conductivity, B0 , saturation ux density is taken to be 75 % of actual saturation ux density and Hm = |K z | = 3 Na ka |I 2 | R (145)

Then rotor surface resistivity and surface reactance are: Rs = Xs 16 1 3 = .5Rs 30 (146) (147)

Note that the rotor elements X2 and R2 depend on rotor current I2 , so the problem is nonlinear. We nd, however, that a simple iterative solution can be used. First we make a guess for R2 and nd currents. Then we use those currents to calculate R2 and solve again for current. This procedure is repeated until convergence, and the problem seems to converge within just a few steps. Aside from the necessity to iterate to nd rotor elements, standard network techniques can be used to nd currents, power input to the motor and power output from the motor, torque, etc.

5.1

Solution

Not all of the equivalent circuit elements are known as we start the solution. To start, we assume a value for R2 , possibly some fraction of Xm , but the value chosen doesn not seem to matter much. The rotor reactance X2 is just a fraction of R2 . Then, we proceed to compute an air-gap impedance, just the impedance looking into the parallel combination of magnetizing and rotor branches: R2 (148) Zg = jXm ||(jX2 + ) s (Note that, for a generator, slip s is negative). A total impedance is then Zt = jX1 + R1 + Zg (149) and terminal current is It = Rotor current is just: I2 = It Vt Zt (150)

jXm jX2 + R2 s

(151)

Now it is necessary to iteratively correct rotor impedance. This is done by estimating ux density at the surface of the rotor using (145), then getting a rotor surface impedance using (146) and using that and (143 to estimate a new value for R2 . Then we start again with (148). The process drops through this point when the new and old estimates for R2 agree to some criterion.

5.2

Harmonic Losses in Solid Steel

If the rotor of the machine is constructed of solid steel, there will be eddy currents induced on the rotor surface by the higher-order space harmonics of stator current. These will produce magnetic elds and losses. This calculation assumes the rotor surface is linear and smooth and can be characterized by a conductivity and relative permeability. In this discussion we include two space harmonics (positive and negative going). In practice it may be necessary to carry four (or even more) harmonics, including both belt and zigzag order harmonics. Terminal current produces magnetic eld in the air-gap for each of the space harmonic orders, and each of these magnetic elds induces rotor currents of the same harmonic order. The magnetizing reactances for the two harmonic orders, really the two components of the zigzag leakage, are: Xzp = Xm
2 kp 2 2 Np k1

(152)

31

Xzn = Xm

2 kn 2 k2 Nn 1

(153)

where Np and Nn are the positive and negative going harmonic orders: For belt harmonics these orders are 7 and 5. For zigzag they are: Np = Nn = Ns + p p Ns p p (154) (155)

Now, there will be a current on the surface of the rotor at each harmonic order, and following 65, the equivalent rotor element current is: I 2p = I 2n = R K 3 Na kp p R K 3 Na kn n (156) (157)

These currents ow in response to the magnetic eld in the air-gap which in turn produces an axial electric eld. Viewed from the rotor this electric eld is: E p = sp RB p E n = sn RB n where the slip for each of the harmonic orders is: sp = 1 Np (1 s) sn = 1 + Np (1 s) and then the surface currents that ow in the surface of the rotor are: Kp = Kn = Ep Zsp En Zsn (162) (163) (160) (161) (158) (159)

where Zsp and Zsn are the surface impedances at positive and negative harmonic slip frequencies, respectively. Assuming a linear surface, these are, approximately: Zs = 1+j (164)

where is material restivity and the skin depth is = 2 s (165)

32

and s is the frequency of the given harmonic from the rotor surface. We can postulate that the appropriate value of to use is the same as that estimated in the nonlinear calculation of the space fundamental, but this requires empirical conrmation. The voltage induced in the stator by each of these space harmonic magnetic uxes is: Vp = Vn = 2Na kp lR Bp Np p 2Na kn lR Bn Nn p (166) (167)

Then the equivalent circuit impedance of the rotor is just: Z2p = Z2n =
2 2 3 4 Na kp l Zsp Vp = Ip 2 Np R s p 2 2 Vn 3 4 Na kn l Zsn = In 2 Nn R s n

(168)

(169)

The equivalent rotor circuit elements are now: R2p = R2n =

2 2 3 4 Na kp l 1 2 Np R p

(170) (171) (172) (173)

2 2 3 4 Na kn l 1 2 Nn R n 1 X2p = R2p 2 1 X2n = R2n 2

5.3

Stray Losses

So far in this document, we have outlined the major elements of torque production and consequently of machine performance. We have also discussed, in some cases, briey, the major sources of loss in induction machines. Using what has been outlined in this document will give a reasonable impression of how an induction machine works. We have also discussed some of the stray load losses: those which can be (relatively) easily accounted for in an equivalent circuit description of the machine. But there are other losses which will occur and which are harder to estimate. We do not claim to do a particularly accurate job of estimating these losses, and fortunately they do not normally turn out to be very large. To be accounted for here are: 1. No-load losses in rotor teeth because of stator slot opening modulation of fundamental ux density, 2. Load losses in the rotor teeth because of stator zigzag mmf, and 3. No-load losses in the solid rotor body (if it exists) due to stator slot opening modulation of fundamental ux density. 33

Note that these losses have a somewhat dierent character from the other miscellaneous losses we compute. They show up as drag on the rotor, so we subtract their power from the mechanical output of the machine. The rst and third of these are, of course, very closely related so we take them rst. The stator slot openings modulate the space fundamental magnetic ux density. We may estimate a slot opening angle (relative to the slot pitch): D = wd Ns 2wd Ns = 2r r 2 D sin 2

Then the amplitude of the magnetic eld disturbance is: BH = Br1

In fact, this ux disturbance is really in the form of two traveling waves, one going forward and one backward with respect to the stator at a velocity of /Ns . Since operating slip is relatively small, the two variations will have just about the same frequency as viewed from the rotor, so it seems reasonable to lump them together. The frequency is: H = Ns p

Now, for laminated rotors this magnetic eld modulation will aect the tips of rotor teeth. We assume (perhaps arbitrarily) that the loss due to this magnetic eld modulation can be estimated from ordinary steel data (as we estimated core loss above) and that only the rotor teeth, not any of the rotor body, are aected. The method to be used is straightforward and follows almost exactly what was done for core loss, with modication only of the frequency and eld amplitude. For solid steel rotors the story is only a little dierent. The magnetic eld will produce an axial electric eld: E z = R BH p and that, in turn, will drive a surface current Kz = Ez Zs

Now, what is important is the magnitude of the surface current, and since |Z s | = 1 + .52 Rs 1.118Rs , we can simply use rotor resistance. The nonlinear surface penetration depth is: = 2B0 H |K z |

A brief iterative substitution, re-calculating and then |K z | quickly yields consistent values for and Rs . Then the full-voltage dissipation is: Prs = 2Rl and an equivalent resistance is: Rrs = |K z |2

3|Va |2 Prs

34

Finally, the zigzag order current harmonics in the stator will produce magnetic elds in the air gap which will drive magnetic losses in the teeth of the rotor. Note that this is a bit dierent from the modulation of the space fundamental produced by the stator slot openings (although the harmonic order will be the same, the spatial orientation will be dierent and will vary with load current). The magnetic ux in the air-gap is most easily related to the equivalent circuit voltage on the nth harmonic: Bn = npvn 2lRNa kn

This magnetic eld variation will be substantial only for the zigzag order harmonics: the belt harmonics will be essentially shorted out by the rotor cage and those losses calculated within the equivalent circuit. The frequency seen by the rotor is that of the space harmonics, already calculated, and the loss can be estimated in the same way as core loss, although as we have pointed out it appears as a drag on the rotor.

6
6.1

Induction Motor Speed Control

Introduction

The inherent attributes of induction machines make them very attractive for drive applications. They are rugged, economical to build and have no sliding contacts to wear. The diculty with using induction machines in servomechanisms and variable speed drives is that they are hard to control, since their torque-speed relationship is complex and nonlinear. With, however, modern power electronics to serve as frequency changers and digital electronics to do the required arithmetic, induction machines are seeing increasing use in drive applications. In this chapter we develop models for control of induction motors. The derivation is quite brief for it relies on what we have already done for synchronous machines. In this chapter, however, we will stay in ordinary variables, skipping the per-unit normalization.

6.2

Volts/Hz Control

Remembering that induction machines generally tend to operate at relatively low per unitslip, we might conclude that one way of building an adjustable speed drive would be to supply an induction motor with adjustable stator frequency. And this is, indeed, possible. One thing to remember is that ux is inversely proportional to frequency, so that to maintain constant ux one must make stator voltage proportional to frequency (hence the name constant volts/Hz). However, voltage supplies are always limited, so that at some frequency it is necessary to switch to constant voltage control. The analogy to DC machines is fairly direct here: below some base speed, the machine is controlled in constant ux (volts/Hz) mode, while above the base speed, ux is inversely proportional to speed. It is easy to see that the maximum torque is then inversely to the square of ux, or therefore to the square of frequency. To get a rst-order picture of how an induction machine works at adjustable speed, start with the simplied equivalent network that describes the machine, as shown in Figure 12 Earlier in this chapter, it was shown that torque can be calculated by nding the power dissipated in the virtual resistance R2 /s and dividing by electrical speed. For a three phase machine, and assuming we are dealing with RMS magnitudes: 35

Ia R a

X1

X2 I2 < > Xm < R2 > < s

Figure 12: Equivalent Circuit

p R2 Te = 3 |I2 |2 s where is the electrical frequency and p is the number of pole pairs. It is straightforward to nd I2 using network techniques. As an example, Figure 13 shows a series of torque/speed curves for an induction machine operated with a wide range of input frequencies, both below and above its base frequency. The parameters of this machine are: Number of Phases Number of Pole Pairs RMS Terminal Voltage (line-line) Frequency (Hz) Stator Resistance R1 Rotor Resistance R2 Stator Leakage X1 Rotor Leakage X2 Magnetizing Reactance Xm 3 3 230 60 .06 .055 .34 .33 10.6

Strategy for operating the machine is to make terminal voltage magnitude proportional to frequency for input frequencies less than the Base Frequency, in this case 60 Hz, and to hold voltage constant for frequencies above the Base Frequency. For high frequencies the torque production falls fairly rapidly with frequency (as it turns out, it is roughly proportional to the inverse of the square of frequency). It also falls with very low frequency because of the eects of terminal resistance. We will look at this next.

6.3

Idealized Model: No Stator Resistance

Ignore, for the moment, R1 . An equivalent circuit is shown in Figure 14. It is fairly easy to show that, from the rotor, the combination of source, armature leakage and magnetizing branch can be replaced by its equivalent circuit, as shown in in Figure 15. In the circuit of Figure 15, the parameters are: V = V X Xm Xm + X1 = Xm ||X1 36

Induction Motor Torque

250

200

Nm

150

100

50

0 0

50

100

150 Speed, RPM

200

250

Figure 13: Induction Machine Torque-Speed Curves

X2 I2 < + > Xm < R2 > V < s

Ia X1

Figure 14: Idealized Circuit: Ignore Armature Resistance If the machine is operated at variable frequency , but the reactance is established at frequency B , current is: I= and then torque is Te = 3|I2 |2 |V |2 R2 3p R2 s = s (X1 + X2 )2 + ( R2 )2 s V j(X1 + X2 ) + B
R2 s

Now, if we note that what counts is the absolute slip of the rotor, we might dene a slip with respect to base frequency: r r B B s= = = sB B

37

Ia X1

X2 I2

< > < R2 > < s

Figure 15: Idealized Equivalent Then, if we assume that voltage is applied proportional to frequency: V = V0 and with a little manipulation, we get: |V0 |2 R2 3p sB Te = B (X1 + X2 )2 + ( R2 )2 s
B

This would imply that torque is, if voltage is proportional to frequency, meaning constant applied ux, dependent only on absolute slip. The torque-speed curve is a constant, dependent only on the dierence between synchronous and actual rotor speed. This is ne, but eventually, the notion of volts per Hz runs out because at some number of Hz, there are no more volts to be had. This is generally taken to be the base speed for the drive. Above that speed, voltage is held constant, and torque is given by: Te = |V |2 R2 3p sB B (X1 + X2 )2 + ( R2 )2 sB

The peak of this torque has a square-inverse dependence on frequency, as can be seen from Figure 16.

6.4

Peak Torque Capability

Assuming we have a smart controller, we are interested in the actual capability of the machine. At some voltage and frequency, torque is given by: Te = 3|I2 |2
p 3 |V |2 R2 R2 s = s ((X1 + X2 )( B ))2 + (R1 +

R2 2 s )

Now, we are interested in nding the peak value of that, which is given by the value of R2 s which maximizes power transfer to the virtual resistance. This is given by the matching condition: R2 = s
2 R1 + ((X1 + X2 )(

2 )) B

38

Induction Motor Torque

250

200

Nm

150

100

50

0 0

500

1500 1000 Speed, RPM

2000

Figure 16: Idealized Torque-Speed Curves: Zero Stator Resistance Then maximum (breakdown) torque is given by: Tmax =
3p 2 |V | 2 R1 + ((X1 + X2 )( B ))2 2 R1 + ((X1 + X2 )( ))2 )2 B

((X1 + X2 )( ))2 + (R1 + B

This is plotted in Figure 17. Just as a check, this was calculated assuming R1 = 0, and the results are plotted in gure 18. This plot shows, as one would expect, a constant torque limit region to zero speed.

6.5

Field Oriented Control

One of the more useful impacts of modern power electronics and control technology has enabled us to turn induction machines into high performance servomotors. In this note we will develop a picture of how this is done. Quite obviously there are many details which we will not touch here. The objective is to emulate the performance of a DC machine, in which (as you will recall), torque is a simple function of applied current. For a machine with one eld winding, this is simply: T = GIf Ia This makes control of such a machine quite easy, for once the desired torque is known it is easy to translate that torque command into a current and the motor does the rest. Of course DC (commutator) machines are, at least in large sizes, expensive, not particularly ecient, have relatively high maintenance requirements because of the sliding brush/commutator interface, provide environmental problems because of sparking and carbon dust and are environmentally sensitive. The induction motor is simpler and more rugged. Until fairly recently the induction motor has not been widely used in servo applications because it was thought to be hard to control. As we will show, it does take a little eort and even some computation to do the controls right, but this is becoming increasingly aordable. 39

Breakdown Torque
300

250

200 NewtonMeters

150

100

50

0 0

20

40

60 80 Drive Frequency, Hz

100

120

Figure 17: Torque-Capability Curve For An Induction Motor

6.6

Elementary Model:

We return to the elementary model of the induction motor. In ordinary variables, referred to the stator, the machine is described by ux-current relationships (in the d-q reference frame): dS dR qS qR = = LS M LS M M LR M LR idS idR iqS iqR

Note the machine is symmetric (there is no saliency), and since we are referred to the stator, the stator and rotor self-inductances include leakage terms: LS = M + LS

LR = M + LR The voltage equations are: ddS qS + rS idS dt dqS + dS + rS iqS dt ddR s qR + rR idR dt dqR + s dR + rR iqR dt 40

vdS vqS

= =

0 = 0 =

Breakdown Torque
300

250

NewtonMeters

200

150

100

50 0

20

40

60 80 Drive Frequency, Hz

100

120

Figure 18: Idealized Torque Capability Curve: Zero Stator Resistance Note that both rotor and stator have speed voltage terms since they are both rotating with respect to the rotating coordinate system. The speed of the rotating coordinate system is w with respect to the stator. With respect to the rotor that speed is , where wm is the rotor mechanical speed. Note that this analysis does not require that the reference frame coordinate system speed w be constant. Torque is given by:
3
T e = p (dS iqS qS idS ) 2

6.7

Simulation Model

As a rst step in developing a simulation model, see that the inversion of the ux-current relationship is (we use the d- axis since the q- axis is identical): idS = LR M dS dR LS LR M 2 LS LR M 2 LS M dS dR LS LR M 2 LS LR M 2

idR =

Now, if we make the following denitions (the motivation for this should by now be obvious): Xd = 0 LS Xkd = 0 LR Xad = 0 M
Xd = 0 LS

M2 LR

41

the currents become: idS = 0 Xad 0 dS X X dR Xd kd d Xd 0 Xad 0 dS X X dR Xkd Xd kd d

idR =

The q- axis is the same. Torque may be, with these calculations for current, written as: 3 0 Xad 3 Te = p (dS iqS qS idS ) = p (dS qR qS dR ) 2 2 Xkd Xd Note that the usual problems with ordinary variables hold here: the foregoing expression was written assuming the variables are expressed as peak quantities. If RMS is used we must replace 3/2 by 3! With these, the simulation model is quite straightforward. The state equations are: ddS dt dqS dt ddR dt dqR dt dm dt = VdS + qS RS idS = VqS dS RS iqS = s qR RR idR = s dR RS iqR = 1 (Te + Tm ) J

where the rotor frequency (slip frequency) is: s = pm For simple simulations and constant excitaion frequency, the choice of coordinate systems is arbitrary, so we can choose something convenient. For example, we might choose to x the coordinate system to a synchronously rotating frame, so that stator frequency = 0 . In this case, we could pick the stator voltage to lie on one axis or another. A common choice is Vd = 0 and Vq = V .

6.8

Control Model

If we are going to turn the machine into a servomotor, we will want to be a bit more sophisticated about our coordinate system. In general, the principle of eld-oriented control is much like emulating the function of a DC (commutator) machine. We gure out where the ux is, then inject current to interact most directly with the ux. As a rst step, note that because the two stator ux linkages are the sum of air-gap and leakage ux, dS qS = agd + LS idS = agq + LS iqS

42

This means that we can re-write torque as: 3 T e = p (agd iqS agq idS ) 2 Next, note that the rotor ux is, similarly, related to air-gap ux: agd = dR LR idR agq = qR LR iqR

Torque now becomes: 3 3 T e = p (dR iqS qR idS ) pLR (idR iqS iqR idS ) 2 2 Now, since the rotor currents could be written as: idR = iqR = That second term can be written as: idR iqS iqR idS = So that torque is now: Te = 3 LR p 1 2 LR 3 M (dR iqS qR idS ) = p (dR iqS qR idS ) 2 LR 1 (dR iqS qR idS ) LR M dR idS LR LR M qR iqS LR LR

6.9

Field-Oriented Strategy:

What is done in eld-oriented control is to establish a rotor ux in a known position (usually this position is the d- axis of the transformation) and then put a current on the orthogonal axis (where it will be most eective in producing torque). That is, we will attempt to set dR = 0 qR = 0 Then torque is produced by applying quadrature-axis current: Te = 3 M p 0 iqS 2 LR

The process is almost that simple. There are a few details involved in guring out where the quadrature axis is and how hard to drive the direct axis (magnetizing) current.

43

Now, suppose we can succeed in putting ux on the right axis, so that qR = 0, then the two rotor voltage equations are: 0 = 0 = Now, since the rotor currents are: idR = iqR = dR M idS LR LR M qR iqS LR LR ddR s qR + rR IdR dt dqR + s dR + rR IqR dt

The voltage expressions become, accounting for the fact that there is no rotor quadrature axis ux: 0 = M ddR dR + rR idS LR dt LR M iqS 0 = s dR rR LR

Noting that the rotor time constant is TR = we nd: TR ddR + dR = M idS dt M iqS s = TR dR LR rR

The rst of these two expressions describes the behavior of the direct-axis ux: as one would think, it has a simple rst-order relationship with direct-axis stator current. The second expression, which describes slip as a function of quadrature axis current and direct axis ux, actually describes how fast to turn the rotating coordinate system to hold ux on the direct axis. Now, a real machine application involves phase currents ia , ib and ic , and these must be derived from the model currents idS and iqs . This is done with, of course, a mathematical operation which uses a transformation angle . And that angle is derived from the rotor mechanical speed and computed slip: = (pm + s ) dt

A generally good strategy to make this sort of system work is to measure the three phase currents and derive the direct- and quadrature-axis currents from them. A good estimate of direct-axis ux is made by running direct-axis ux through a rst-order lter. The tricky operation involves dividing quadrature axis current by direct axis ux to get slip, but this is now easily done numerically (as 44

dR

ND
N

M 1 + STa M TR

T
* ia

ia
o

* d

T
* iq

-1

* ib

Amp

o o

ib ic

Motor Load

* c

Figure 19: Field Oriented Controller are the trigonometric operations required for the rotating coordinate system transformation). An elmentary block diagram of a (possbly) plausible scheme for this is shown in Figure 19. In this picture we start with commanded values of direct- and quadrature- axis currents, corresponding to ux and torque, respectively. These are translated by a rotating coordinate transformation into commanded phase currents. That transformation (simply the inverse Parks transform) uses the angle q derived as part of the scheme. In some (cheap) implementations of this scheme the commanded currents are used rather than the measured currents to establish the ux and slip. We have shown the commanded currents i , etc. as inputs to an Amplier. This might be a implemented as a PWM current-source, for example, and a tight loop here results in a rather high performance servo system.

References
[1] P.L. Alger, Induction Machines, Gordon and Breach, 1969 [2] A.E. Fitzgerals, C. Kingsley Jr., S.D. Umans, Electric Machinery, Sixth Edition, McGraw Hill, 2003

45

[3] D. Fink, H. W. Beaty, Standard Handbook for Electrical Engineers, Thirteenth Edition,
McGraw-Hill, 1993

46

r1

x1 jxag
< > < rc > <

xc jx2s5
< r > 2s5 < s > 5 <

jx2s
< r > 2s < s > <

jx2c
< r > 2c < s > <

jxa5

jx2c5
< r > 2c5 < s > 5 <

jxa7

jx2s7
< r > 2s7 < s > 7 <

jx2c7
< r > 2c7 < s > 7 <

jxam

jx2sm
< r > 2sm < s > m <

jx2cm
< r > 2cm < s > m <

jxap

jx2sp
< r2sp > < s > p <

jx2cp
< r2cp > < s > p <

Figure 20: Extended Equivalent Circuit

47

Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science 6.685 Electric Machinery Class Notes 9: Synchronous Machine Simulation Models c 2005 James L. Kirtley Jr. September 5, 2005

Introduction

In this document we develop models useful for calculating the dynamic behavior of synchronous machines. We start with a commonly accepted picture of the synchronous machine, assuming that the rotor can be fairly represented by three equivalent windings: one being the eld and the other two, the d- and q- axis damper windings, representing the eects of rotor body, wedge chain, amortisseur and other current carrying paths. While a synchronous machine is assumed here, the results are fairly directly applicable to induction machines. Also, extension to situations in which the rotor representation must have more than one extra equivalent winding per axis should be straightforward.

Phase Variable Model

To begin, assume that the synchronous machine can be properly represented by six equivalent windings. Four of these, the three armature phase windings and the eld winding, really are windings. The other two, representing the eects of distributed currents on the rotor, are referred to as the damper windings. Fluxes are, in terms of currents: ph R = Lph MT M LR Iph IR (1)

where phase and rotor uxes (and, similarly, currents) are: a = b c

ph

(2)

There are three inductance sub- matrices. The tances: La Lph = Lab Lac

f R = kd kq

(3)

rst of these describes armature winding inducLab Lac Lb Lbc Lbc Lc

(4)

where, for a machine that may have some saliency: La = La0 + L2 cos 2 Lb = La0 + L2 cos 2( (5) 2 ) (6) 3 2 Lc = La0 + L2 cos 2( + ) (7) 3 (8) Lab = Lab0 + L2 cos 2( ) 3 Lbc = Lab0 + L2 cos 2 (9) Lac = Lab0 + L2 cos 2( + ) (10) 3 Note that, in this last set of expressions, we have assumed a particular form for the mutual inductances. This is seemingly restrictive, because it constrains the form of phase- to- phase mutual inductance variations with rotor position. The coecient L2 is actually the same in all six of these last expressions. As it turns out, this assumption does not really restrict the accuracy of the model very much. We will have more to say about this a bit later. The rotor inductances are relatively simply stated: Lf Lf kd 0 LR = Lf kd Lkd 0 0 0 Lkq Lakd cos 2 3 ) Lakd cos( 2 3 ) Lakd cos( + Lakq sin 2 3 ) Lakq sin( 2 3 ) Lakq sin( +
2 3 ) 2 3 )

(11)

And the stator- to- rotor mutual inductances are: M cos M = M cos( M cos( +

(12)

Parks Equations

The rst step in the development of a suitable model is to transform the armature winding variables to a coordinate system in which the rotor is stationary. We identify equivalent armature windings in the direct and quadrature axes. The direct axis armature winding is the equivalent of one of the phase windings, but aligned directly with the eld. The quadrature winding is situated so that its axis leads the eld winding by 90 electrical degrees. The transformation used to map the armature currents, uxes and so forth onto the direct and quadrature axes is the celebrated Parks Transformation named after Robert H. Park, an early investigator into transient behavior , in synchronous machines. The mapping takes the form:

Where the transformation and its inverse are:

ua ud uq = udq = T uph = T ub uc u0

(13)

cos cos( 23 ) cos( + 23 ) 2 2 T = sin sin( 3 ) sin( + 23 ) 3 1 1 1 2 2 2

(14)

T 1

This transformation maps balanced sets of phase currents into constant currents in the d-q frame. That is, if rotor angle is = t + 0 , and phase currents are: Ia = I cos t Ib = I cos(t Ic Then the transformed set of currents is: Id = I cos 0 Iq = I sin 0 Now, we apply this transformation to (1) to express uxes and currents in the armature in the d-q reference frame. To do this, extract the top line in (1): ph = Lph I ph + M I R (16) 2 ) 3 2 = I cos(t + ) 3

cos sin 1 = cos( 2 ) sin( 2 ) 1 3 3 2 2 cos( + 3 ) sin( + 3 ) 1

(15)

The transformed ux is obtained by premultiplying this whole expression by the transformation matrix. Phase current may be obtained from d-q current by multiplying by the inverse of the transformation matrix. Thus: (17) dq = T Lph T 1 I dq + T M I R The same process carried out for the lower line of (1) yields: R = M T T 1 I dq + LR I R Thus the fully transformed version of (1) is: dq R =
dq 3 T LC 2

(18)

LC LR

Idq IR

(19)

If the conditions of (5) through (10) are satised, the inductance submatrices of (19) wind up being of particularly simple form. (Please note that a substantial amount of algebra has been left out here!) Ld 0 0 Lq 0 (20) Ldq = 0 0 0 L0 M LC = 0 0

Lakd 0 0 Lakq 0 0

(21)

Note that (19) through (21) express three separate sets of apparently independent ux/current relationships. These may be re-cast into the following form:

d kd = f q kq

Ld Lakd M Id 3 Lakd Lkd Lf kd Ikd 2 3 If Lf kd Lf 2M Lq 3 Lakq 2 Lakq Lkq Iq Ikq

(22)

(23) (24)

0 = L0 I0 Where the component inductances are: 3 Ld = La0 Lab0 + L2 2 3 Lq = La0 Lab0 L2 2 L0 = La0 + 2Lab0

(25) (26) (27)

Note that the apparently restrictive assumptions embedded in (5) through (10) have resulted in the very simple form of (21) through (24). In particular, we have three mutually independent sets of uxes and currents. While we may be concerned about the restrictiveness of these expressions, note that the orthogonality between the d- and q- axes is not unreasonable. In fact, because these axes are orthogonal in space, it seems reasonable that they should not have mutual ux linkages. The principal consequence of these assumptions is the de-coupling of the zero-sequence component of ux from the d- and q- axis components. We are not in a position at this time to determine the reasonableness of this. However, it should be noted that departures from this form (that is, coupling between the direct and zero axes) must be through higher harmonic elds that will not couple well to the armature, so that any such coupling will be weak. Next, armature voltage is, ignoring resistance, given by: d d = T 1 dq dt ph dt and that the transformed armature voltage must be: V ph = V dq = T V ph d = T (T 1 dq ) dt d d = dq + (T T 1 )dq dt dt A good deal of manupulation goes into reducing the second term of this, resulting in: T d 1 T = dt

(28)

(29)

0
d dt

d dt 0 0

0 0 0

(30)

This expresses the speed voltagethat arises from a coordinate transformation. The two voltage/ux relationships that are aected are: Vd = Vq = where we have used = dd q dt dq + d dt d dt (31) (32)

(33)

Power and Torque

P = V a Ia + V b Ib + V c Ic (34)

Instantaneous power is given by: Using the transformations given above, this can be shown to be: P = which, in turn, is: 3 3 dd dq d0 P = (d Iq q Id ) + ( Id + Iq ) + 3 I0 (36) 2 dt dt 2 dt Then, noting that electrical speed and shaft speed are related by = p and that (36) describes electrical terminal power as the sum of shaft power and rate of change of stored energy, we may deduce that torque is given by: 3 T = p(d Iq q Id ) 2 (37) 3 3 Vd Id + Vq Iq + 3V0 I0 2 2 (35)

Per-Unit Normalization

The next thing for us to do is to investigate the way in which electric machine system are normalized, or put into what is called a per-unit system. The reason for this step is that, when the voltage, current, power and impedance are referred to normal operating parameters, the behavior characteristics of all types of machines become quite similar, giving us a better way of relating how a particular machine works to some reasonable standard. There are also numerical reasons for normalizing performance parameters to some standard. The rst step in normalization is to establish a set of base quantities. We will be normalizing voltage, current, ux, power, impedance and torque, so we will need base quantities for each of these. Note, however, that the base quantities are not independent. In fact, for the armature, we need only specify three quantities: voltage (VB ), current (IB ) and frequency (0 ). Note that we do not normalize time nor frequency. Having done this for the armature circuits, we can derive each of the other base quantities:

 Base Power Base Impedance

3 PB = VB IB 2 ZB = VB IB VB 0

Base Flux Base Torque

B =

TB =

p PB 0

Note that, for our purposes, base voltage and current are expressed as peak quantities. Base voltage is taken on a phase basis (line to neutral for a wye connected machine), and base current is similarly taken on a phase basis, (line current for a wye connected machine). Normalized, or per-unit quantities are derived by dividing the ordinary variable (with units) by the corresponding base. For example, per-unit ux is: = 0 = VB B (38)

In this derivation, per- unit quantities will usually be designated by lower case letters. Two notable exceptions are ux, where we use the letter , and torque, where we will still use the upper case T and risk confusion. Now, we note that there will be base quantities for voltage, current and frequency for each of the dierent coils represented in our model. While it is reasonable to expect that the frequency base will be the same for all coils in a problem, the voltage and current bases may be dierent. We might write (22) as:

d kd = f

0 IdB Vdb Ld 0 IdB 3 Vkb 2 Lakd 0 IdB 3 Vf b 2 M

0 IkB Vdb Lakd 0 IkB Vkb Lkd 0 IkB Vf b Lf kd

0 I f B Vdb M 0 I f B Vkdb Lf kd 0 I f B Vf b Lf

id ikd

(39)

if

where i = I/IB denotes per-unit, or normalized current. Note that (39) may be written in simple form:

d xd xakd xad id xf kd ikd kd = xakd xkd xad xf kd xf f if

(40)

It is important to note that (40) assumes reciprocity in the normalized system. To wit, the following expressions are implied: xd = 0 IdB Ld VdB (41)

xkd = 0 xf =

xakd = = xad = = xf kd = = These in turn imply:

IkB Lkd VkB If B 0 Lf Vf B IkB 0 Lakd VdB 3 IdB 0 Lakd 2 VkB If B M 0 VdB 3 IdB 0 M 2 Vf B IkB 0 Lf kd Vf b If B Lf kd 0 Vkb

(42) (43)

(44)

(45)

(46)

3 VdB IdB = Vf B If B 2 3 VdB IdB = VkB IkB 2 Vf B If B = VkB IkB

(47) (48) (49)

These expressions imply the same power base on all of the windings of the machine. This is so because the armature base quantities Vdb and Idb are stated as peak values, while the rotor base 3 quantities are stated as DC values. Thus power base for the three- phase armature is 2 times the product of peak quantities, while the power base for the rotor is simply the product of those quantities. The quadrature axis, which may have fewer equivalent elements than the direct axis and which may have dierent numerical values, still yields a similar structure. Without going through the details, we can see that the per-unit ux/current relationship for the q- axis is: q kq = xq xakq xakq xkq iq ikq (50)

The voltage equations, including speed voltage terms, (31) and (32), may be augmented to reect armature resistance: dd (51) q + Ra Id Vd = dt dq Vq = d + + Ra Iq (52) dt The per-unit equivalents of these are: vd = 1 dd q + ra id 0 dt 0 7
(53)

vq =

1 dq + ra iq d + 0 0 dt

(54)

Ra Where the per-unit armature resistance is just ra = ZB Note that none of the other circuits in this model have speed voltage terms, so their voltage expressions are exactly what we might expect:

vf

vkd = vkq = v0 =

1 df + rf if 0 dt 1 dkd + rkd ikd 0 dt 1 dkq + rkq ikq 0 dt 1 d0 + ra i0 0 dt

(55) (56) (57) (58)

It should be noted that the damper winding circuits represent closed conducting paths on the rotor, so the two voltages vkd and vkq are always zero. Per-unit torque is simply: Te = d iq q id (59) Often, we need to represent the dynamic behavior of the machine, including electromechanical dynamics involving rotor inertia. If we note J as the rotational inertia constant of the machine system, the rotor dynamics are described by the two ordinary dierential equations: 1 d J p dt d dt = T e + T m = 0 (60) (61)

where T e and T m represent electrical and mechanical torques in ordinary variables. The angle represents rotor phase angle with respect to some synchronous reference. It is customary to dene an inertia constant which is not dimensionless but which nevertheless ts into the per-unit system of analysis. This is: H Or: H=
1 2J 0 p

Rotational kinetic energy at rated speed Base Power

2

(62)

PB

J 0 2pTB

(63)

Then the per-unit equivalent to (60) is: 2H d = Te + Tm 0 dt where now we use Te and Tm to represent per-unit torques. (64)

Equal Mutuals Base

In normalizing the dierential equations that make up our model, we have used a number of base quantities. For example, in deriving (40), the per-unit ux- current relationship for the direct axis, we used six base quantities: VB , IB , Vf B , If B , VkB and IkB . Imposing reciprocity on (40) results in two constraints on these six variables, expressed in (47) through (49). Presumably the two armature base quantities will be xed by machine rating. That leaves two more degrees of freedom in selection of base quantities. Note that the selection of base quantities will aect the reactance matrix in (40). While there are dierent schools of thought on just how to handle these degrees of freedom, a commonly used convention is to employ what is called the equal mutuals base system. The two degrees of freedom are used to set the eld and damper base impedances so that all three mutual inductances of (40) are equal: xakd = xf kd = xad (65) The direct- axis ux- current relationship becomes:

id xd xad xad d kd = xad xkd xad ikd if xad xad xf f

(66)

Equivalent Circuit

id ra

+ (0 vd + q ) -

+ d -

xal xad

xf l xkdl

< > < rkd > <

if rf

+ vf -

Figure 1: D- Axis Equivalent Circuit The ux- current relationship of (66) is represented by the equivalent circuit of Figure 1, if the leakage inductances are dened to be: xal = xd xad (67) (68) (69)

xkdl = xkd xad xf l = xf xad

Many of the interesting features of the electrical dynamics of the synchronous machine may be discerned from this circuit. While a complete explication of this thing is beyond the scope of this note, it is possible to make a few observations. The apparent inductance measured from the terminals of this equivalent circuit (ignoring resistance ra ) will, in the frequency domain, be of the form: x(s) = Pn (s) d (s) = xd id (s) Pd (s) (70)

Both the numerator and denominator polynomials in s will be second order. (You may convince yourself of this by writing an expression for terminal impedance). Since this is a diusion type circuit, having only resistances and inductances, all poles and zeros must be on the negative real axis of the s-plane. The per-unit inductance is, then: x(s) = xd
(1 + Td s)(1 + Td s) (1 + Tdo s)(1 + Tdo s)

(71)

The two time constants Td and Td are the reciprocals of the zeros of the impedance, which are the poles of the admittance. These are called the short circuittime constants. The other two time constants Tdo and Tdo are the reciprocals of the poles of the impedance, and so are called the open circuittime constants. We have cast this thing as if there are two sets of well- dened time constants. These are the transient time constants Td and Tdo , and the subtransient time constants Td and Tdo . In many cases, these are indeed well separated, meaning that: Td Td

(72) (73)

Tdo

Tdo

If this is true, then the reactance is described by the pole-zero diagram shown in Figure 2. Under this circumstance, the apparent terminal inductance has three distinct values, depending on frequency. These are the synchronous inductance, the transient inductance, and the subtransient inductance, given by:
Td Tdo T = xd d Tdo T T = xd d d Tdo Tdo

x = xd d x d

(74)

(75)

A Bode Plotof the terminal reactance is shown in Figure 3.

If the time constants are spread widely apart, they are given, approximately, by:

Tdo = Tdo =

xf 0 rf xkdl + xf l ||xad 0 rkd 10

(76) (77)

1 Td

1 Tdo

1
Td

1
Tdo

Figure 2: Pole-Zero Diagram For Terminal Inductance

log |x(j)|

1 Tdo

1 T
d

1 1 Tdo Td

log

Figure 3: Frequency Response of Terminal Inductance Finally, note that the three reactances are found simply from the model: xd = xal + xad x d xd = xal + xad ||xf l = xal + xad ||xf l ||xkdl (78) (79) (80)

Statement of Simulation Model

Now we can write down the simulation model. Actually, we will derive more than one of these, since the machine can be driven by either voltages or currents. Further, the expressions for permanent magnet machines are a bit dierent. So the rst model is one in which the terminals are all constrained by voltage. The state variables are the two stator uxes d , q , two damper uxes kd , kq , eld ux f , and rotor speed and torque angle . The most straightforward way of stating the model employs currents as auxiliary variables, and these are:

id xd xad xad ikd = xad xkd xad if xad xad xf iq ikq = xq xaq xaq xkq
1

d kd f

(81)

q kq

(82)

11

Then the state equations are: dd dt dq dt dkd dt dkq dt df dt d dt d dt and, of course, Te = d iq q id = 0 vd + q 0 ra id = 0 vq d 0 ra iq = 0 rkd ikd = 0 rkq ikq = 0 vf 0 rf if = 0 (Te + Tm ) 2H (83) (84) (85) (86) (87) (88) (89)

= 0

8.1

Statement of Parameters:

Note that often data for a machine may be given in terms of the reactances xd , x , xd , Tdo and d Tdo , rather than the elements of the equivalent circuit model. Note that there are four inductances in the equivalent circuit so we have to assume one. There is no loss in generality in doing so. Usually one assumes a value for the stator leakage inductance, and if this is done the translation is straightforward:

xad = xd xal xad (x xal ) d xf l = xad x + xal d 1 xkdl = 1 1 x

xal xad
d

1 xf

rf

rkd =

xf l + xad 0 Tdo xkdl + xad ||xf l 0 Tdo

8.2

Linearized Model

Often it becomes desirable to carry out a linearized analysis of machine operation to, for example, examine the damping of the swing mode at a particular operating point. What is done, then, is to assume a steady state operating point and examine the dynamics for deviations from that operating point that are small. The denition of small is really small enough that everything important appears in the rst-order term of a Taylor series about the steady operating point. Note that the expressions in the machine model are, for the most part, linear. There are, however, a few cases in which products of state variables cause us to do the expansion of the 12

Taylor series. Assuming a steady state operating point [d0 kd0 f 0 q0 kq0 0 0 ], the rstorder (small-signal) variations are described by the following set of equations. First, since the ux-current relationship is linear:

id1 xd xad xad ikd1 = xad xkd xad if 1 xad xad xf iq1 ikq1 =
xq xaq xaq xkq
1

q1 kq1

d1 kd1 f 1

(90)

(91)

Terminal voltage will be, for operation against a voltage source: Vd = V sin Vq = V cos

Then the dierential equations governing the rst-order variations are: dd1 dt
dq1
dt
dkd1
dt
dkq1
dt
df 1
dt d1 dt d1
= 1 dt Te = d0 iq1 + d1 iq0 q0 id1 q1 id0 = 0 V cos 0 1 + 0 q1 + 1 q0 0 ra id1 = 0 V sin 0 1 0 d1 1 d0 0 ra iq1 = 0 rkd ikd1 = 0 rkq ikq1 = 0 rf if 1 = 0 (Te1 + Tm1 ) 2H
(92) (93) (94) (95) (96) (97) (98)

8.3

Reduced Order Model for Electromechanical Transients

In many situations the two armature variables contribute little to the dynamic response of the machine. Typically the armature resistance is small enough that there is very little voltage drop across it and transients in the dierence between armature ux and the ux that would exist in the steady state decay rapidly (or are not even excited). Further, the relatively short armature time constant makes for very short time steps. For this reason it is often convenient, particularly when studying the relatively slow electromechanical transients, to omit the rst two dierential equations and set: d = vq = V cos q = vd = V sin (99) (100)

The set of dierential equations changes only a little when this approximation is made. Note, however, that it can be simulated with far fewer cycles if the armature time constant is short. 13

Current Driven Model: Connection to a System

The simulation expressions developed so far are useful in a variety of circumstances. They are, however, dicult to tie to network simulation programs because they use terminal voltage as an input. Generally, it is more convenient to use current as the input to the machine simulation and accept voltage as the output. Further, it is dicult to handle unbalanced situations with this set of equations. An alternative to this set would be to employ the phase currents as state variables. Eectively, this replaces d , q and 0 with ia , ib , and ic . The resulting model will, as we will show, interface nicely with network simulations. To start, note that we could write an expression for terminal ux, on the d- axis: d = x id + f d and here, of course, x = xal + xad ||xkdl ||xf l d This leads us to dene a ux behind subtransient reactance:
d =

xad ||xf l xad ||xkdl + kd xad ||xf l + xkdl xad ||xkdl + xf l

(101)

xad xkdl f + xad xf l kd xad xkdl + xad xf l + xkdl xf l

d = d + xd id

(102)

So that On the quadrature axis the situation is essentially the same, but one step easier if there is only one quadrature axis rotor winding: q = x iq + kq q where
Very often these uxes are referred to as voltage behind subtransient reactance, with d = eq = e . Then: and q d

xaq xaq + xkql

(103)

x = xal + xaq ||xkql q

d = x id + e q d q = x iq q
ed

(104) (105)

Now, if id and iq are determined, it is a bit easier to nd the other currents required in the simulation. Note we can write: kd f and this inverts easily: ikd if = xkd xad xad xf
1

xkd xad xad xf

ikd if

xad xad

id

(106)

kd f

xad xad

id

(107)

14

The quadrature axis rotor current is simply: ikq = xaq 1 kq iq xkq xkq (108)

The torque equation is the same, but since it is usually convenient to assemble the uxes behind subtransient reactance, it is possible to use:
Te = e iq + ed id + (xd xq )id iq q

(109)

Now it is necessary to consider terminal voltage. This is most conveniently cast in matrix notation. The vector of phase voltages is: v ph va = vb vc

(110)

Then, with similar notation for phase ux, terminal voltage is, ignoring armature resistance: 1 d ph 0 dt 1 d 1 T dq 0 dt

v ph = =

(111)

Note that we may dene the transformed vector of uxes to be: dq = x idq + e where the matrix of reactances shows orthogonality: x 0 0 d x = 0 x 0 q 0 0 x0 e q e = e d 0

(112)

(113)

and the vector of internal uxes is:

(114)

Now, of course, idq = T iph , so that we may re-cast (111) as: v ph = 1 d 1 T x T iph + T 1 e 0 dt (115)

Now it is necessary to make one assumption and one denition. The assumption, which is only moderately restrictive, is that subtransient saliency may be ignored. That is, we assume that x = x . The denition separates the zero sequence impedance into phase and neutral q d components:

15

x0 = x + 3xg d

(116)

Note that according to this denition the reactance xg accounts for any impedance in the neutral of the synchronous machine as well as mutual coupling between phases. Then, the impedance matrix becomes: x 0 0 0 0 0 d x = 0 x 0 + 0 0 0 d 0 0 3xg 0 0 x d x = x I + xg d where I is the identity matrix. Now the vector of phase voltages is: v ph = 1 d x i + T 1 xg T iph + T 1 e 0 dt d ph (119)

(117)

In compact notation, this is:

(118)

Note that in (119), we have already factored out the multiplication by the identity matrix. The next step is to carry out the matrix multiplication in the third term of (119). This operation turns out to produce a remarkably simple result: 1 1 1 1 T xg T = xg 1 1 1 1 1 1

(120)

The impact of this is that each of the three phase voltages has the same term, and that is related to the time derivative of the sum of the three currents, multiplied by xg . The third and nal term in (119) describes voltages induced by rotor uxes. It can be written as: 1 1 d de 1 d T 1 e = T 1 e + T 1 0 dt 0 dt 0 dt Now, the time derivative of the inverse transform is: 1 d 1 T 0 dt sin() cos() 0 sin( 23 ) cos( 2 ) 0 = 3 0 sin( + 23 ) cos( + 2 ) 0 3 xg d x dia d + (ia + ib + ic ) + e a 0 dt 0 dt di xd b xg d + (ia + ib + ic ) + eb 0 dt 0 dt x dic xg d d + (ia + ib + ic ) + ec 0 dt 0 dt 16

(121)

(122)

Now the three phase voltages can be extracted from all of this matrix algebra: va = vb = vc = (123) (124) (125)

Where the internal voltages are:

e = a (e sin() ed cos()) 0 q de 1 de 1 q + sin() d + cos() 0 dt 0 dt 2 2 ) e cos( )) = (eq sin( d 0 3 3 2 de 1 2 de 1 q ) + sin( ) d + cos( 0 3 dt 0 3 dt 2 2 ) e cos( + )) = (e sin( + q d 0 3 3 1 2 de 1 2 de q + cos( + ) + sin( + ) d 0 3 dt 0 3 dt

(126)

e b

(127)

e c

(128)

This set of expressions describes the equivalent circuit shown in Figure 4.

ia x d va vb ic xd vc ib xd

ea
+

e b g x e c
+ +

Figure 4: Equivalent Network Model

10

Restatement Of The Model

The synchronous machine model which uses the three phase currents as state variables may now be stated in the form of a set of dierential and algebraic equations: dkd dt dkq dt df dt d dt = 0 rkd ikd = 0 rkq ikq = 0 rf if = 0 17 (129) (130) (131) (132)

d dt where: ikd if and =

0 Tm + e iq + ed id q 2H
1

(133)

xkd xad xad xf ikq =

kd f

xad xad

id

xaq 1 kq iq xkq xkq

(It is assumed here that the dierence between subtransient reactances is small enough to be neglected.) The network interface equations are, from the network to the machine: 2 ) + ic cos( + 3 2 iq = ia sin() ib sin( ) ic sin( + 3 id = ia cos() + ib cos( and, in the reverse direction, from the machine to the network: e = a (e sin() ed cos()) 0 q de 1 de 1 q + sin() d + cos() 0 dt 0 dt 2 2 )) ) e cos( = (eq sin( d 0 3 3 1 2 de 2 de 1 q ) + ) d sin( + cos( 0 3 dt 0 3 dt 2 2 ) e cos( + )) = (e sin( + d 0 q 3 3 1 1 2 de 2 de q + cos( + ) + ) d sin( + 0 3 dt 0 3 dt 2 ) 3 2 ) 3

(134) (135)

(136)

e b

(137)

e c

(138)

And, of course, = 0 t + e q e d = =
d q

(139) (140) (141) (142) (143)

d = q =

xad xkdl f + xad xf l kd xad xkdl + xad xf l + xkdl xf l xaq kq xaq + xkql

18

11

Network Constraints

This model may be embedded in a number of networks. Dierent congurations will result in dierent constraints on currents. Consider, for example, the situation in which all of the terminal voltages are constrained, but perhaps by unbalanced (not entirely positive sequence) sources. In that case, the dierential equations for the three phase currents would be: x dia d 0 dt x dib d 0 dt x dic d 0 dt x + 2xg xg d + 3x (vb eb ) + (vc ec ) x + 3x xd g g d + 2x x xg g = (vb e ) d b + 3x (va ea ) + (vc ec ) x + 3x xd g g d + 2x xg xd g = (vc e ) (vb eb ) + (va ea ) c xd + 3xg xd + 3xg = (va e ) a (144) (145) (146)

12

Example: Line-Line Fault

We are not, however, constrained to situations dened in this way. This model is suitable for embedding into network analysis routines. It is also possible to handle many dierent situations directly. Consider, for example, the unbalanced fault represented by the network shown in Figure 5. This shows a line-line fault situation, with one phase still connected to the network.

va

ia ra

x d x d x d

ea
+

ib ra

e b g x e c
+ +

ra

Figure 5: Line-Line Fault Network Model In this situation, we have only two currents to worry about, and their dierential equations would be: dib dt dia dt = = 0 (e e 2ra ib ) b 2x c d 0 (va e ra ia ) a x + xg d (147) (148)

and, of course, ic = ib . Note that here we have included the eects of armature resistance, ignored in the previous section but obviously important if the results are to be believed. 19

13

Permanent Magnet Machines

Permanent Magnet machines are one state variable simpler than their wound-eld counterparts. They may be accurately viewed as having constant eld current. Assuming that we can dene the internal (eld) ux as: 0 = xad if 0 (149)

13.1

Model: Voltage Driven Machine

We have a reasonably simple expression for the rotor currents, in the case of a voltage driven machine: id ikd iq ikq = = xd xad xad xkd xq xaq xaq xkq
1

d 0 kd 0 q kq

(150) (151)

The simulation model then has six states: dd dt dq dt dkd dt dkq dt d dt d dt = 0 vd + q 0 ra id = 0 vq d 0 ra iq = 0 rkd ikd = 0 rkq ikq = 0 (d iq q id + Tm ) 2H (152) (153) (154) (155) (156) (157)

= 0

13.2

Curent-Driven Machine Model

In the case of a current-driven machine, rotor currents required in the simulation are: ikd = ikq = 1 (kd xad id 0 ) xkd 1 (kq xaq iq ) xkq (158) (159)

Here, the ux behind subtransient reactance is, on the direct axis:

d =

xkdl 0 + xad kd xad + xkdl

(160)

and the subtransient reactance is: x = xal + xad ||xkdl d 20 (161)

On the quadrature axis,

q =

xad kq xad + xkql

(162)

and x = xal + xaq ||xkql q In this case there are only four state equations: dkd dt dkq dt d dt d dt = 0 rkd ikd = 0 rkq ikq = 0 e iq + ed id + Tm 2H q (164) (165) (166) (167) (163)

= 0

The interconnections to and from the network are the same as in the case of a wound-eld machine: in the forward direction, from network to machine: 2 ) + ic cos( + 3 2 iq = ia sin() ib sin( ) ic sin( + 3 id = ia cos() + ib cos( and, in the reverse direction, from the machine to the network: e = a (e sin() ed cos()) 0 q de 1 de 1 q + sin() d + cos() 0 dt 0 dt 2 2 = (e sin( ) e cos( )) d 0 q 3 3 2 de 1 2 de 1 q ) + sin( ) d + cos( 0 3 dt 0 3 dt 2 2 )) ) e cos( + = (e sin( + d q 0 3 3 2 de 1 2 de 1 q ) + sin( + ) d + cos( + 0 3 dt 0 3 dt 2 ) 3 2 ) 3 (168) (169)

(170)

e b

(171)

e c

(172)

13.3

PM Machines with no damper

PM machines without much rotor conductivity may often behave as if they have no damper winding at all. In this case the model simplies even further. Armature currents are: id = iq = 1 (d 0 ) xd 1 q xq 21 (173) (174)

The state equations are: dd dt dq dt d dt d dt = 0 vd + q 0 ra id = 0 vq d 0 ra iq = 0 (d iq q id + Tm ) 2H (175) (176) (177) (178)

= 0

13.4

Current Driven PM Machines with no damper

In the case of no damper the machine becomes quite simple. There is no internal ux on the quadrature axis. Further, there are no time derivatives of the internal ux on the d- axis. The only machine state equations are mechanical: d dt d dt = 0 (0 iq + Tm ) 2H (179) (180)

= 0

The forward network interface is as before: 2 ) + ic cos( + 3 2 ) ic sin( + iq = ia sin() ib sin( 3 id = ia cos() + ib cos( 0 sin() 0 = 0 sin( 0 = 0 sin( + 0 2 ) 3 2 ) 3 (181) (182)

and, in the reverse direction, from the machine to the network, things are a bit simpler than before: e = a e b e c (183) 2 ) 3 2 ) 3 (184) (185) (186)

22

Characterization of Left-Handed Materials

Massachusetts Institute of Technology 6.635 lecture notes

Introduction
1. How are they realized? 2. Why the denomination Left-Handed? 3. What are their properties? 4. Does it really work?

It has already been shown (see previous classes) that rings, or split-rings, can realize a negative permeability ( < 0) over a certain frequency band. In addition to this, we need to realize a negative permittivity ( < 0). It has also been shown (see previous classes) that: lossless:
metal 2 p , 2

=1

where p =

ne2 0 me

(n: electron density, e: electron charge, me : eective mass of electrons). lossy:

metal 2 p =1 . ( + i)

A typical transmission curve looks like shown in Fig. 1.

T 1 no transmission ( < 0)

PSfrag replacements

transmission ( > 0)

Figure 1: Transmission curve for a plasma-like medium.

Section 2. Why left-handed?

With these characteristics, < 0 has been realized already at infrared frequencies (where metals behave like plasmas). Problem: how to realize it at GHz frequencies? Solution: by reducing n, the electron density. One way of doing this is to conne the electrons in space. This can be achieved by an array of rods for example, as shown in Fig. 2.

a
PSfrag replacements

Figure 2: Array of rods conning the electrons in space.

Note: it is important that the wires are thin, so as to reduce the radiation interaction and allow penetration into the structure. Eect of the wires: to reduce n to ne = n r2 . a2 (1)

Finally, note also that the rods have to be parallel to the electric eld. This, plus the (known already) fact that rings have to be perpendicular to the magnetic eld, gives an idea on how to realize physically LH metamaterials (see Fig. 3).

Why left-handed?

At this point, we have a metamaterial which can realize < 0, < 0 . We shall now see what does it imply on the electromagnetic elds. Let us write Maxwells curl equations for plane wave solutions and time harmonic notations: (2)

r r k E() = H() , r r k H() = E() .

(3a) (3b)

3
0.5 mm 5 mm

z
1.3 cm

s
FRONT BACK

(a)
1 mm

y x s

(b)

Figure 3: A realization for LH material.

In standard materials, Eq. (3) implies that the tryad (E, H, k) forms a right-handed system. However, under Eq. (2), we will have: (E, H, k) form a left-handed (LH) tryad. However, the time average Poynting power is still r < S() >= 1 r r {E() H ()} 2 (4)

and remains in the same direction so that we have the set up shown in Fig. 4. Characteristics: k is in the phase velocity direction. phase velocity and energy ux are in opposite directions.

Properties of LH media

Some know characteristics are: Reversed Doppler eect (track the phase), Reversed Cerenkov radiation (cf. 6.632),

3.1

Reversed Snells law

k
PSfrag replacements

H
Figure 4: Electric eld (E), magnetic eld (H), wave-vector (k) and Poynting power in an LH medium. (S)

Negative index of refraction. This last item is very signicant, and we shall spend some time discussing it. The index of refraction of a medium is dened as n=
r

r ,

(5)

or, writing explicitly the frequency dependence (cf. later), n() =

r () r () .

(6) < 0 and < 0),

For those frequencies inside the left-handed band (i.e. in the band where we can write: () < 0 () < 0 Eventually, we write n from Eq. (6): n= | ()()| ei = | ()()| . () = | ()| ei ,

(7a) (7b)

() = |()| ei ,

(8)

3.1

Reversed Snells law

An important consequence of this fact is the reversal of Snells law. Ray diagram:

n>0 PSfrag replacements i r r n<0

k diagram for an LH medium:

PSfrag replacements kx

k1

k2

kx S

k1z

k2z

3.2

Energy

Traditionally, the energy is given by W = E 2 + H 2 . (9)

What happens if < 0 and < 0? Is W < 0? Actually no, but this direct conclusion from Eq. (9) shows that this equation is not valid as is. In fact, these materials have to be modeled by frequency dispersive permittivity and permeability. In that case, the relation of Eq. (9) becomes (from Poyntings theorem): W = and we must have: ( ) > 0, () > 0. (11a) (11b) ( ) 2 () 2 E + H (10)

3.3

Properties of an LHM slab

When LH materials are studies as bulk materials, two models are commonly used for the permittivity/permeability: 1. Drude model:
2 ep , ( + ie ) 2 mp , ( + im )

=1

(12a)

r = 1

(12b)

which is schematically represented in Fig. 5.

Real(eps/eps0) 10
0.02 Imag(eps/eps0)

0.015

0.01

10
0.005

20

0.005

30
0.01

40 wp = 100 GHz wp = 266.5 GHz wp = 500 GHz wp = 1000 GHz 50 0 0.2 0.4 0.6 0.8 1 w/w0 1.2 1.4 1.6 1.8 2
0.015 wp = 100 GHz wp = 266.5 GHz wp = 500 GHz wp = 1000 GHz 0 0.2 0.4 0.6 0.8 1 w/w0 1.2 1.4 1.6 1.8 2

0.02

(a)

( ).

(b)

( ).

Figure 5: Permittivity for various values of ep in the Drude model (f0 = 30 GHz, ep = p ).

2. Resonant model:
2 2 ep eo , 2 2 eo + ie 2 2 mp mo , 2 2 mo + im

=1

(13a)

r = 1

(13b)

where (em)o are the electric/magnetic resonant frequencies and (em)p are the electric/magnetic plasma frequencies. An illustration of this model is given in Fig. 6.

3.3

Properties of an LHM slab

Let us consider the situation depicted in Fig. 7.

Real(eps/eps0) 100 wp = 266.5 GHz wp = 500 GHz

Imag(eps/eps0) 1 wp = 266.5 GHz wp = 500 GHz

80

0.8

60

0.6

40

0.4

20

0.2

20

0.2

40

0.4

60

0.6

80

0.8

100

1
0 0.2 0.4 0.6 0.8 1 w/w0 1.2 1.4 1.6 1.8 2

0.2

0.4

0.6

0.8

1 w/w0

1.2

1.4

1.6

1.8

(a)

( ).

(b)

( ).

Figure 6: Permittivity for various values of ep in the resonant model of Eq. (13a) (f0 = 30 GHz, ep = p , 0 = 100GHz).

PSfrag replacements

Figure 7: LH slab in free-space.

3.3

Properties of an LHM slab

Let us consider the case for which

= r = 1 (working at the right frequency).

1. From simple ray diagrams, using the reversed Snells law: PSfrag replacements x #0 #1 #2

S I1

I2 z

d We see that is the source is close enough to the slab (distance< d), the slab will produce two images, one inside the slab and one outside. The distance from source to the second image is S I2 = 2d . 2. Rigorous calculation: Let us consider a TE wave impinging on this slab. We write, for a single interface: y E0 =E0 eikx xit [eikz z + reikz z ] , y E1 =E0 eikx xit teikz z . (14a) (14b)

We need to match the boundary conditions and, for simplicity, we set the boundary to be at z = 0. We get: Tangential E eld: Tangential H eld (H =
1 i

eikz z + reikz z = teikz z . E): eikz z reikz z = 0 kz ikz z te . 1 k z

Upon solving, we get the reection/transmission coecient from free-space to the medium: 21 kz , 1 k z + 0 k z 1 k z 0 k z r= . 1 k z + 0 k z t= (15a) (15b)

In a similar way, the reection/transmission coecients from the medium to free-space are: 20 kz , 1 k z + 0 k z 0 k z 1 k z r = . 1 k z + 0 k z t = (16a) (16b)

In order to obtain the eld inside the slab, we shall compute the transmission coecient as:

T = teikz d t + teikz d r eikz d r eikz d t + tt (r )2 e5ikz d +

= tt eikz d
n=0

(r )n e2inkz d =

tt eikz d . 1 r 2 e2ikz d

(17)

Using the previous expressions, we obtain T = For the specic case of

r

= r = 1: T =

41 kz kz eikz d . (1 kz + 0 kz )2 (0 kz 1 kz )e2ikz d 4kz kz eikz d . (kz + kz )2 (kz + kz )e2ikz d

(18)

(19)

At this point, we need to dene kz and kz : Propagating waves (k > k ) kz kz Relation kz = kz =

2 k 2 k 2 k 2 k

Evanescent waves (k < k ) kz = i kz = i

2 k k 2 2 k k 2

kz = kz

kz = k z

Performing the calculation, we get: Propagating waves (kz = kz ): T = Evanescent waves (kz = kz ): T =
2 4kz eikz d = eikz d . 2 e2ikz d 4kz 2 4kz eikz d = eikz d . 2 4kz

10

Section 4. Does it really work?

Therefore, for all waves, we get: T = eikz d . (20)

Conclusions: Evanescent waves are amplied by the medium. Propagating waves are backward waves. Taking an innite amount of those creates a source. Two images can be formed, like shown in the ray diagram of the previous subsection.

Does it really work?

The theoretical predictions have been veried in experiments. To date, essentially two experiments have been done: 1. Prism, 2. Gaussian beam.

4.1

Prism experiment

First experiment (2001). Later reproduced.

LHM

RHM PSfrag replacements

4.2

Gaussian beam experiment

First publication was theoretical. Later experiments followed, but still need to be improved.

11

RHM

LHM

PSfrag replacements

Dispertion relations in Left-Handed Materials

Massachusetts Institute of Technology 6.635 lecture notes

Introduction

We know already the following properties of LH media: 1. 2.

r r

and r are frequency dispersive. and r are negative over a similar frequency band.

3. The tryad (E, H, k) is left-handed. 4. The index of refraction is negative. From the past lectures, we know that these materials can be realized by a succession of wires and rods: Periodic arrangement of rods: realizes a plasma medium with negative frequency band. The model for the permittivity is:
r r

over a certain

=1

2 ep . 2 + ie

(1)

Periodic arrangement of rings (split-rings) realizes a resonant r modeled as r = 1

2 F mp , 2 2 mo + im

(2)

where F is the fractional area of the unit cell occupied by the interior of the split-ring (F < 1). In the lossless case (e = m = 0), we can rewrite these two relations as:
2 2 ep , 2 2 2 2 b 2 0 F 2 r = = (1 F ) 2 2 2 , 2 0 0 r

(3a)
(3b)

where b = 0 / 1 F > 0 . Therefore:

Upon identifying the regions where relation for k:

2 2 ( 2 ep )( 2 b ) 1 k 2 = 2 = (1 F ) . 2 c 2 0
r

(4)

and r change signs, we can immediately get the

Section 2. Argument on n < 0

0 + +

b +

p + + +

r k2

The region [0 , b ], which also corresponds to r < 0 and r < 0, corresponds to positive, which means k real. Therefore, there is propagation in this band, but not in the adjacent ones. k2 It may still not be clear that k is negative, even if we write k = 2 = 2
2 2 r r 0 0

= k0 n .

(5)

A demonstration of the fact that n is negative follows.

2
2.1

Argument on n < 0
Complex Poynting theorem

We shall rst recall the derivation of the complex Poynting theorem and the signication of the various terms. We start from Maxwells curl equation E = i B , H = i D + J . (6a) (6b)

Upon multiplying Eq. (6a) by H and substracting the complex conjugate of Eq. (6b) mul tuplied by E we get: H EE H = (E H ) = i B H i D E J E = i[B H E D ] E J .

(7)

Upon rewriting, we get: E J = (E H ) + i[E D B H ] . (8)

On the right-hand side of the equation, the rst terms corresponds to the divergence of Poyting power, which is therefore positive. The second term relates to the complex EM energy, and is therefore also positive. Consequently, the left-hand side term must also be positive, and actually corresponds to the power supplied by J to the volume. We shall use this result hereafter.

2.2

1D wave equation

For the sake of simplication, let us work with a 1D problem. The wave equation
2

r r r E() + k 2 E() = iJ() ,

(9)

is rewritten with r E() = z E(x) , r J() = z j0 (x x0 ) , to yield 2 E(x) + k 2 E(x) = ij0 (x x0 ) . x2 The solution to this equation is E(x) = eik|xx0 | , where needs to be determined. From Eq. (12), we write: 1. First derivative: E(x) = ik |x x0 | eik|xx0 | . x x2 (10a) (10b)

(11)

(12)

(13)

2. Second derivative: 2 E(x) 2 =ik |x x0 | eik|xx0 | + (k 2 )( 2 |x x0 |)2 eik|xx0 | 2 2 x x x 2 ik|xx0 | = k e + 2ik(x x0 ) . Therefore: 2 E(x) + k 2 E(x) =2ik(x x0 ) = 2i k0 n (x x0 ) . x2 Comparing Eq. (11) to Eq. (15), we get = so that nally the solution is: E(x) = j0 0 r ik|xx0 | e . 2 n (17) j 0 0 r j0 = , 2k0 n 2 n (16) (15)

(14)

If we now compute the power supplied by the current J to the volume: P = 1 2

2 0 j 0 r E J dV = > 0. 4 n

(18)

The source must, on average, do positive work on the eld. Yet, in LH regime, we have r < 0 so that we must have n < 0 as well.

Section 3. Dispersion relations

Finally, we can also write the E eld as: E(x, t) = j0 ei(k0 n|xx0 |t) . 2 (19)

Thus, plane waves appear to propagate from and + to the source, seemingly running backward in time. Yet, the work done on the eld is positive so clearly the energy propagates outward from the source.

Dispersion relations

At this point, we know that n < 0 and k < 0. The dierence between phase and group velocity can be directly seen on the dispersion relation diagram. , kz kz vg =

v =

(20a)
1

(20b)

Free-space: k = Metamaterial:

where

= cte and = cte.

Let us take the following models

2 p , r 2 + ie 2 2 mp mo , r =1 2 2 mo + im

=1

(21a) (21b)

3.1

Lossless case (e = m = 0), mp = p

We rewrite
2 2 p , r 2 2 2 mp , r = 2 2 mo

(22a) (22b) (22c)

and plot the relations with

p mp mo e = m
k surface x 10 4 [rad/s]
10

= = = =

20e9 rad/s 20e9 rad/s 5e9 rad/s 0

3 2.5 2 x 10
6

Dispersion relation

k
2 0 k 2
z

1.5 1

0 x 10
6

2 0 2 k
x

0.5
6

x 10

Relative permittivity
50 300 200 100

10 [rad/s] x 10 Relative permeability

0 1 2 3 x 10 4
10

r
0 100 200 0

50

[rad/s]

[rad/s]

3 x 10

4
10

x 10

10

k surface (Gamma= 0)

3.5

2.5 [rad/s]

1.5

0.5

0 3

0 kz

2 x 10

3
6

3.1

Lossless case (e = m = 0), mp = p

Other cases follow. p mp mo e = m

k surface x 10 4 3 [rad/s] 2
10

= = = =

30e9 rad/s 20e9 rad/s 5e9 rad/s 0 rad/s

4 x 10
6

Dispersion relation

0 2 x 10
6

k
0 k 2
z

1 0 k
x

2 x 10
6

Relative permittivity
50 300 200 100 0

10 [rad/s] x 10 Relative permeability

r
0 100 0 1 2 3 x 10 4
10

50

200

[rad/s]

[rad/s]

3 x 10

4
10

x 10

10

k surface (Gamma= 0)

3.5

2.5 [rad/s]

1.5

0.5

0 4

0 k

3 x 10

4
6

p mp mo e = m
k surface x 10 4
10

= = = =

30e9 rad/s 20e9 rad/s 5e9 rad/s 10e7 rad/s

4 x 10
6

Dispersion relation

3 [rad/s] 2

0 2 x 10
6

k
0 kz 2 2 k
x

1 0 2 x 10
6

Relative permittivity
50 300 200 100 0

10 [rad/s] x 10 Relative permeability

r
0 100 0 1 2 3 x 10 4
10

50

200

[rad/s]

[rad/s]

3 x 10

4
10

x 10

10

k surface (Gamma= 100000000)

3.5

2.5 [rad/s]

1.5

0.5

0 4

0 k

3 x 10

4
6

3.1

Lossless case (e = m = 0), mp = p

p mp mo e = m
k surface x 10 4 [rad/s]
10

= = = =

30e9 rad/s 20e9 rad/s 5e9 rad/s 10e8 rad/s

2.5 2 1.5 x 10
6

Dispersion relation

k
1 0 kz 2 2 k
x

0 2 2 0 x 10
6

0.5 0

x 10

Relative permittivity
50 60 40 20 0

10 [rad/s] x 10 Relative permeability

r
0 20 0 1 2 3 x 10 4
10

50

40

[rad/s]

[rad/s]

3 x 10

4
10

x 10

10

k surface (Gamma= 1000000000)

3.5

2.5 [rad/s]

1.5

0.5

0 2.5

1.5

0.5

0 k

0.5

1.5

2 x 10

2.5
6

p mp mo e = m
k surface x 10 4 [rad/s]
10

= = = =

30e9 rad/s 20e9 rad/s 5e9 rad/s 10e9 rad/s

3 2.5 2 x 10
7

Dispersion relation

k
2 0 kz 2 2 kx 2 0 x 10
7

1.5 1

0 0.5 0 0 1 2 3 4 x 10
7

Relative permittivity
2 0 2 20 15 10

10 [rad/s] x 10 Relative permeability

4 6 8

0 1 2 3 x 10 4
10

r
5 0 5 0

[rad/s]

[rad/s]

3 x 10

4
10

x 10

10

k surface (Gamma= 1.000000e+10)

3.5

2.5 [rad/s]

1.5

0.5

0 4

0 k

3 x 10

4
7

10

3.1

Lossless case (e = m = 0), mp = p

p mp mo e = m
k surface x 10 4 [rad/s]
10

= = = =

30e9 rad/s 20e9 rad/s 5e9 rad/s 10e10 rad/s

5 4 3 x 10
7

Dispersion relation

k
2 2 1 0 kz 2 2 0 kx 2 4 x 10
7

0 4 x 10
7

Relative permittivity
0.966 0.966

10 [rad/s] x 10 Relative permeability

0.962

r
0 1 2 3 x 10 4
10

0.964

0.964

0.962

0.96

0.96

[rad/s]

[rad/s]

3 x 10

4
10

x 10

10

k surface (Gamma= 1.000000e+11)

3.5

2.5 [rad/s]

1.5

0.5

0 5

0 k

4 x 10

5
7

 ~ } { GG|g

n l

u u y x v az"

l p q

w av l n v l  ~  ~ } XdGv gg s x XdGv Ggg s x XdGv s w v o v w n v o n m l k m i8l m l ik m l k i 8uf t l ral i8jhe g q p m m l k i g f s

d u v gg8s 5BHaU5w (ga gXdGw gXGw Ugv Gw X 5yGw P"v ds 88digRdb aY u w u s u s s w s u s r v v u x r u v x r s u t r q p h f e c ` D A @ 9 7 3 1 4 3 1 ) ' % W "C 8"X5"2H#
kx
6 6

D C A @ 9 7 ""BP8V 5"20UBF# 4 3 1 ) ' % T

kx
0.5 1.5 2.5 3.5 0 5 x 10 0.5 1.5 2.5 3.5 0.5 1.5 2.5 3.5 0 5 0 5 1 [rad/s] 2 3 4 1 [rad/s] 2 3 4 1 [rad/s] 2 3 4 x 10 x 10
6

D C A @ 9 7 ""B58S 5"20(RF# 4 3 1 ) ' % Q

kx kz

D C A @ 9 7 I 4 3 1 ) ' % E ""BP855"2HGF#
kx kz

D C A @ 9 7 ""B586 5"20(&# 4 3 1 ) ' % $

kx k
z

!     " 
k
x

Plotting all the 3D curves on the same scale:

x 10

x 10

x 10

10

10

10

5 5 5

k surface (Gamma= 1.000000e+10)

k surface (Gamma= 100000000)

5 5 5 k surface (Gamma= 0) kz

kz
0 0 0

x 10

x 10

x 10

0.5

1.5

2.5

3.5

0 5

x 10 0.5 1.5 2.5 3.5 0.5 1.5 2.5 3.5 0 5 0 5 1 [rad/s] 2 3 4 1 [rad/s] 2 3 4 1 [rad/s] 2 3 4

x 10

x 10

x 10

x 10

x 10

10

10

10

5 5 5

k surface (Gamma= 1.000000e+11)

k surface (Gamma= 1000000000)

5 5 5

k surface (Gamma= 10000000)

kz
0 0 0

x 10

x 10

x 10

6 6 6

5 5 5

11

12

3.1

Lossless case (e = m = 0), mp = p

For simplicity, we can study the lossy case for e = m amd mo = 0 (although we dont really simulate the same medium, the fundamental behavior is similar, and simpler to carry out mathematically). The model therefore reads:
2 2 p + i . = r = 2 + i

(23)

We compute:
2 2 p + i 2 + i 2 [ 2 p + i] [ 2 i] = 4 + 2 2 2 ( 2 2 ) + 2 2 + i 2 p p = . 4 + 2 2

r r

(24a) The real part is given by: {

r r }

2 2 [ 2 (p 2 )] . 4 + 2 2

(25)

Losses have the eect to lower the plasma frequency to p =

2 p 2 .

(26)

In addition, we also see that if is very large, the plasma eect will completey dissapear (cf. dispersion relation for = 10e10 rad/s).

Greens functions for planarly layered media

Massachusetts Institute of Technology 6.635 lecture notes

Introduction: Greens functions

The Greens functions is the solution of the wave equation for a point source (dipole). For scalar problems, the wave equation is written as (k0 = ):
(
2 2 + k0 ) g(, r ) = ( r ) , r r

(1)

and the solution for an unbounded medium is: g(, r ) = r

r r eik0 | | . 4| r | r

(2)

From Maxwells equations in frequency domain with an eit dependency, the wave equation r for electric eld E() is:

2 r r E() k0 E() = i J() , r

=0 Therefore, the free-space dyadic Greens function satises

which solution is
G(, r ) = r
(check using (Ig) =
g

for source free case .

(3)

2 G(, r ) k0 G(, r ) = I ( r ) , r r r

(4)

I+

1 2 k0

g(, r ) , r

(5)

( g)I and Eq. (1)).

For the use of Greens functions in scattering problems, it is useful to express the Greens function in the same coordinates as the problem, which can be rectangular, cylindrical, spherical, etc. Here we shall concentrate on the rectangular representation (Cartesian).

Section 2. Cartesian coordinates

2
2.1

Cartesian coordinates
Scalar Greens function

The formulae are derived from Eq. (1) and the Fourier transform of the quantities: 1 (2)3 1 ( r ) = r (2)3 g(, r ) = r
+ +

r r dk eik( ) g(k) ,
r r dk eik( ) ,

(6a) (6b)

where k = kx x + ky y + kz z and dk = dkx dky dkz . Upon using Eq. (1), we write: (
2 2 + k0 ) + r r dk eik( ) g(k) = 2 + r r dk eik( ) ,

(7)

Introducing the dierential operator (

+

2 x2

2 y 2

2 ) z 2

we write:

2 + k0 )

r r dk eik( ) g(k) = =

+ +

dk (

r r 2 + k0 )eik( ) g(k)

r r 2 2 2 2 dk (kx ky kz + k0 )eik( ) g(k) r r dk eik( ) ,

= from which we conclude that g(k) =

(8)

2 kx

2 ky

1 2 . 2 + kz k0

(9)

Using Eq. (6a), we therefore need to evaluate the following integral: g(, r ) = r 1 (2)3
+

dkx dky dkz

1 r r ik( ) . 2 e 2 2 2 kx + k y + k z k 0

(10)

Note that Eq. (10) can be integrated along one of the three axis. In remote sensing application, the vertical axis is usually taken to be the z axis, (xy) being the transverse plane (planar components). We therefore choose to evaluate Eq. (10) along kz , and we split:

k = k x x + k y y + k z z = k + k z z , r = r + z z , r = r + z z .

(11a) (11b) (11c)

We will perform the integral of Eq. (10) in the complex plane, using Cauchys theorem and the Residue theorem. Before doing this, we have to be careful not to have divergent integrals. Since we integrate in kz , the condition is: lim eikz z < + . (12)

kz

If we write kz as kz = kz + ikz (kz R, kz R), we see that if z > 0, we have to choose kz > 0, which means that for complex plane integration, we need to deform the contour into the upper plane. if z < 0, we have to choose kz < 0, which corresponds to a deformation into the lower plane. In addition, we see from Eq. (10) that the integrand has a pole at
2 2 2 2 2 2 k0z = k 0 k x k y = k 0 k .

(13)

We therefore need to evaluate Eq. (10) via the Residue theorem. Calculus: let us just write the integral in dkz for z > 0:
+ 2 kz

1 1 r r r r ik( ) ik( ) = 2i Res 2 2 e 2 e k 0z kz k 0z kz k 0z r r = 2i lim eik ( ) eikz (zz ) kz k0z (kz k0z )(kz + k0z ) 1 ik ( ) ik0z (zz ) r r = 2i , e e 2k0z

(14)

so that g(, r ) = r i (2)2

dk

1 ik ( ) ik0z (zz ) r r , e e 2k0z

for z z > 0.

(15)

The treatment for z < 0 follows the same reasoning so that we write for all (z z ): (z z ) R, g(, r ) = r i (2)2

dk

1 ik ( ) ik0z |zz | r r . e e 2k0z

(16)

2.2

Dyadic Greens function

From Eq. (16) and Eq. (26), we can get the dyadic Greens functions. Note that is a dyadic operator (give a dyad when applied to a scalar) and can be exchanged with the integral sign. In addition, it only applies to the exponential terms so that we actually need to evaluate:
r r eik ( ) eik0z |zz |

(17)

2.2

Dyadic Greens function

or, by a simple change of variables:

r eik eik0z |z|

(18)

Calculus: Let us rst consider z > 0 and

r (eik eik0z |z| ) =

f (x, y, z) .

(19)

Various derivatives will be:

x x f (x, y, z) x y f (x, y, z) 2 = kx f (x, y, z) ,

= kx ky f (x, y, z) ,

... and identically for z < 0. At z = 0 however, 2 ik0z |z| = e ik0z eik0z |z| |z| z 2 z z = ik0z ik0z eik0z |z| |z| z
2

+ ik0z eik0z |z|

2 |z| z 2 (20)

2 = 2ik0z (z) k0z eik0z |z| .

Using these results, we write: 2 i g() = r z 2 (2)2 (z) = (2)2 1 ik 2 e r 2ik0z (z) k0z eik0z |z| 2k0z i r r dk eik 2 dk k0z eik eik0z |z| 8 i r ik ik0z |z| = () 2 r . e d k k0z e 8 dk

(21)

Again, all the other terms of function becomes:

applied to the integrand give k k so that the Greens 1 kk r I 2 eik for z > 0 , k0z k0 r 1 I K K eiK for z < 0 , d k 2 k0z k0

G(, r ) = z r z where

i () r 2 + 8 2 k0

d k

(22)

k = k x x + k y y + k 0z z , K = k x x + k y y k 0z z . Some notes:

(23a) (23b)

1. The Dirac delta function is known as the singularity of the Greens function and is important in calculating the elds in the source region. 2. The dierent signs ensure that the integral converges for evanescent waves, i.e. when 2 2 2 kx + k y > k 0 . 3. The square bracket in the expression of the Greens functions can be expressed in terms of superposition of TE and TM waves, as we shall see.

2.3

Superposition of TE and TM waves

Based on k, we can form an orthonormal system for TE/TM polarized waves: kz TE: e(k0z ) = = |k z | TM:

1
2 2 kx + ky

[ky y kx ] = x

1 (ky y kx ) , x k

(24a) (24b)

k 1 k0 h(k0z ) = e(k0z ) k = z (ky + y kx ) + z . x k0 ko k ko

The three vectors k, h and e form an orthonormal system, in which: e I = k k + e(k0z )(k0z ) + h(k0z )h(k0z ) . After translating to the origin, we get for the dyadic Greens functions:

(25)

() r i G(, r ) = z 2 + 2 r z k0 8 where

1 r r e e(k0z )(k0z ) + h(k0z )h(k0z ) eik( ) k0z r r 1 e(k0 )(k0 ) + h(k0 )h(k0 ) eiK( ) dk z z z z e k0z dk

for z > z , for z < z , (26)

e(k0z ) = e(k0z ) , e(k0z ) K h(k0z ) = . k0

(27a) (27b)

(Note that K, e(k0z ) and h(k0z ) form another orthonormal set of vectors about K).

2.4

Treatment of layered media

Depending upon the medium under study and the location of the source, the kernel of Eq. (26) will have to be modied. To make it more clear, we can gather the terms in the Greens function relative to the source (primed coordinates) and those relative to the source.

2.4

Treatment of layered media

G(, )=z rr z

for z < z , (28) If we now consider a layered medium problem, with an arbitrary number of layers and a source in the top region (incident wave), we write:

() r i 2 + 8 2 k0

d k d k

1 k 0z 1 k 0z

r r r r {[(k0z ) eik] e(k0z ) eik +[h(k0z ) eik] h(k0z ) eik } e r r {[(k0z ) eiK] e(k0z ) eiK r +[h(k0z ) eiK] h(k0z ) eiK r } e

for z > z ,

G(, r )i0 = r with

i 8 2

dkx dky

1 r r Ke e(k0z ) eiK + Kh h(k0z ) eiK k0z

(29)

1. For z < z , i = 0:

r r Ke =(k0z ) eiK + RT E e(k0z ) ek0z eik e r r Kh =h(k0z ) eiK + RT M h(k0z ) ek0z eik

(30a) (30b)

2. For region , i = :

Ke =A e(k z ) eik

+ B e(k z ) eiK + D h(k z ) e

(31a) (31b) (31c)

Kh =C h(k z ) e

r ik

r iK

3. For region t, i = t:

r Ke =T T E e(ktz ) eiKt

(32a) (32b) (32c)

Kh =T

TM

h(ktz ) e

r iKt

where k
z

x k =kx + y ky + z k z , K =kx + y ky z k z , x

2 2 = k 2 kx ky ,

(33a) (33b) (33c)

and the coecients A , B , C and D are determined from the boundary conditions. The boundary conditions apply to the tangential electric and magnetic elds. Thus, in r terms of Greens functions, we need to satisfy the continuity of z G(, r ) and z

G(, r ). Let us write this at the interface between media ( ) and ( + 1), by separating r the TE and TM components:

A eik z z + B eik z z = A k
z

+1 e +1

ik z z

+B

+1 e

ik z z +1 e ik z z

(34a) (34b) (34c) (34d)

A eik z z B eik z z = kz

+1 e

ik z z

B B +D

kz kz +1 A eik z z B eik z z = A k k +1 k
z

+1 e +1 e

ik z z

+1 e +1 e

ik z z

C eik z z + D eik z z = kz

+1

ik z z

ik z z

With the conditions in the rst and last layer as: A0 = R T E , At = 0 , B0 = 1 ,

TE

C0 = R T M , Ct = 0 ,

D0 = 1 ,
TM

(35a) (35b)

Bt = T

Dt = T

Before evaluating these coecients, we can build a recursive scheme to calculate the amplitudes from region to region + 1. For example, it is straightforward to build a propagation matrix for TE modes from Eq. (34a) and (34b): TE A eikz d A +1 eikz +1 d +1 (36) =V B eikz d B +1 eikz +1 d +1 A similar procedure of course applied to the TM modes: C +1 eikz +1 d +1 D +1 eikz +1 d +1 =V
TM

C eikz d D eikz d

(37)

In order to end up the recursive method, we have to express the reection and transmission coecient in the rst and last regions, respectively. We shall only illustrated this point here, as it has been developed in previous classes. Let us consider a plane wave incident from region 0, with its plane of incidence parallel to the (xy) plane. All elds vectors are independent on y, so that y = 0 in Maxwells equations. Thus, we can decompose the elds into their TE and TM components. We get in region : TE modes: H 1 E y, i z 1 Hz= E y, i x 2 + + 2 E x2 y
x

(38a) (38b)
y

= 0.

(38c)

2.4

Treatment of layered media

TM modes: H y, z 1 H y, E z = i x 2 + + 2 H x2 y
x

1 i

(39a) (39b)
y

= 0.

(39c)

For a TE wave inside the stratied medium: E H H

y x

=(A eik z z + B eik z z ) eikx x , k = z (A eik z z B eik z z ) eikx x , kx = (A eik z z + B eik z z ) eikx x .

(40a) (40b) (40c)

By matching the boundary conditions, and upon using the already known notation p R we get the recursive relation: [1 1/R2( +1) ] e2i[k( +1) z +k z ]d A e2ik z d , = + A B R ( +1) 1/R ( +1) e2ik( +1) z d + B +1 +1 with the limiting condition: At = 0, , Bt Example: for a two-layer medium (t = 2): R= R01 + R12 e2ik1z (d1 d0 ) 2ikz d0 e . 1 + R01 R12 e2ik1z (d1 d0 ) (44) A0 = R. B0 (43) (42)
( +1)

( +1)

kz ( +1) , +1 k z 1 p ( +1) = . 1 + p ( +1)

(41a) (41b)

Fundamentals of Model Theory

William Weiss and Cherie D'Mello
Department of Mathematics University of Toronto

c 1997 W.Weiss and C. D'Mello

Introduction
Model Theory is the part of mathematics which shows how to apply logic to the study of structures in pure mathematics. On the one hand it is the ultimate abstraction on the other, it has immediate applications to every-day mathematics. The fundamental tenet of Model Theory is that mathematical truth, like all truth, is relative. A statement may be true or false, depending on how and where it is interpreted. This isn't necessarily due to mathematics itself, but is a consequence of the language that we use to express mathematical ideas. What at rst seems like a de ciency in our language, can actually be shaped into a powerful tool for understanding mathematics. This book provides an introduction to Model Theory which can be used as a text for a reading course or a summer project at the senior undergraduate or graduate level. It is also a primer which will give someone a self contained overview of the subject, before diving into one of the more encyclopedic standard graduate texts. Any reader who is familiar with the cardinality of a set and the algebraic closure of a eld can proceed without worry. Many readers will have some acquaintance with elementary logic, but this is not absolutely required, since all necessary concepts from logic are reviewed in Chapter 0. Chapter 1 gives the motivating examples and we recommend that you read it rst, before diving into the more technical aspects of Chapter 0. Chapters 2 and 3 are selections of some of the most important techniques in Model Theory. The remaining chapters investigate the relationship between Model Theory and the algebra of the real and complex numbers. Thirty exercises develop familiarity with the de nitions and consolidate understanding of the main proof techniques. Throughout the book we present applications which cannot easily be found elsewhere in such detail. Some are chosen for their value in other areas of mathematics: Ramsey's Theorem, the Tarski-Seidenberg Theorem. Some are chosen for their immediate appeal to every mathematician: existence of in nitesimals for calculus, graph colouring on the plane. And some, like Hilbert's Seventeenth Problem, are chosen because of how amazing it is that logic can play an important role in the solution of a problem from high school algebra. In each case, the derivation is shorter than any which tries to avoid logic. More importantly, the methods of Model Theory display clearly the structure of the main ideas of the proofs, showing how theorems of logic combine with theorems from other areas of mathematics to produce stunning results. The theorems here are all are more than thirty years old and due in great part to the cofounders of the subject, Abraham Robinson and Alfred Tarski. However, we have not attempted to give a history. When we attach a name to a theorem, it is simply because that is what mathematical logicians popularly call it. The bibliography contains a number of texts that were helpful in the preparation of this manuscript. They could serve as avenues of further study and in addition, they contain many other references and historical notes. The more recent titles were added to show the reader where the subject is moving today. All are worth a look. This book began life as notes for William Weiss's graduate course at the University of Toronto. The notes were revised and expanded by Cherie D'Mello and

William Weiss, based upon suggestions from several graduate students. The electronic version of this book may be downloaded and further modi ed by anyone for the purpose of learning, provided this paragraph is included in its entirety and so long as no part of this book is sold for pro t.

Contents
Chapter 0. Models, Truth and Satisfaction Formulas, Sentences, Theories and Axioms Prenex Normal Form Chapter 1. Notation and Examples Chapter 2. Compactness and Elementary Submodels Compactness Theorem Isomorphisms, elementary equivalence and complete theories Elementary Chain Theorem Lowenheim-Skolem Theorems The Los-Vaught Test Every complex one-to-one polynomial map is onto Chapter 3. Diagrams and Embeddings Diagram Lemmas Every planar graph can be four coloured Ramsey's Theorem The Leibniz Principle and in nitesimals Robinson Consistency Theorem Craig Interpolation Theorem Chapter 4. Model Completeness Robinson's Theorem on existentially complete theories Lindstrom's Test Hilbert's Nullstellensatz Chapter 5. The Seventeenth Problem Positive de nite rational functions are the sums of squares Chapter 6. Submodel Completeness Elimination of quanti ers The Tarski-Seidenberg Theorem Chapter 7. Model Completions Almost universal theories Saturated models Blum's Test Bibliography Index
3

4 4 9 11 14 14 15 16 19 21 23 24 25 25 26 26 27 31 32 32 35 38 39 39 45 45 49 50 52 54 55 61 62

CHAPTER 0

Models, Truth and Satisfaction

We will use the following symbols: logical symbols: { the connectives ^ ,_ , : , ! , $ called \and", \or", \not", \implies" and \i " respectively { the quanti ers 8 , 9 called \for all" and \there exists" { an in nite collection of variables indexed by the natural numbers N v0 ,v1 , v2 , : : : { the two parentheses ), ( { the symbol = which is the usual \equal sign" constant symbols : often denoted by the letter c with subscripts function symbols : often denoted by the letter F with subscripts each function symbol is an m-placed function symbol for some natural number m 1 relation symbols : often denoted by the letter R with subscripts each relational symbol is an n-placed relation symbol for some natural number n 1. We now de ne terms and formulas. Definition 1. A term is de ned as follows: (1) a variable is a term (2) a constant symbol is a term (3) if F is an m-placed function symbol and t1 : : : tm are terms, then F (t1 : : : tm ) is a term. (4) a string of symbols is a term if and only if it can be shown to be a term by a nite number of applications of (1), (2) and (3). Remark. This is a recursive de nition. Definition 2. A formula is de ned as follows : (1) if t1 and t2 are terms, then t1 = t2 is a formula. (2) if R is an n-placed relation symbol and t1 : : : tn are terms, then R(t1 : : : tn ) is a formula. (3) if ' is a formula, then (:') is a formula (4) if ' and are formulas then so are (' ^ ), (' _ ), (' ! ) and (' $ ) (5) if vi is a variable and ' is a formula, then (9vi )' and (8vi )' are formulas (6) a string of symbols is a formula if and only if it can be shown to be a formula by a nite number of applications of (1), (2), (3), (4) and (5). Remark. This is another recursive de nition. :' is called the negation of '. ' ^ is called the conjunction of ' and . ' _ is called the disjunction of ' and . Definition 3. A subformula of a formula ' is de ned as follows:
4

0. MODELS, TRUTH AND SATISFACTION

(1) ' is a subformula of ' (2) if (: ) is a subformula of ' then so is (3) if any one of ( ^ ), ( _ ), ( ! ) or ( $ ) is a subformula of ' , then so are both and (4) if either (9vi ) or (8vi ) is a subformula of ' for some natural number i , then is also a subformula of ' (5) A string of symbols is a subformula of ', if and only if it can be shown to be such by a nite number of applications of (1), (2), (3) and (4). Definition 4. A variable vi is said to occur bound in a formula ' i for some subformula of ' either (9vi ) or (8vi ) is a subformula of '. In this case each occurance of vi in is said to be a bound occurance of vi . Other occurances of vi which do not occur bound in ' are said to be free. Exercise 1. Using the previous de nitions as a guide, de ne the substitution of a term t for a variable vi in a formula '. Example 1. F2 (v3 v4 ) is a term. ((8v3 )(v3 = v3 ^ v0 = v1 ) _ (9v0 )v0 = v0 ) is a formula. In this formula the variable v3 occurs bound, the variable v1 occurs free, but the variable v0 occurs both bound and free. Exercise 2. Reconsider Exercise 1 in light of substituting the above term for v0 in the formula. Definition 5. A language L is a set consisting of all the logical symbols with perhaps some constant, function and/or relational symbols included. It is understood that the formulas of L are made up from this set in the manner prescribed above. Note that all the formulas of L are uniquely described by listing only the constant, function and relation symbols of L. We use t(v0 : : : vk ) to denote a term t all of whose variables occur among v0 : : : vk . We use '(v0 : : : vk ) to denote a formula ' all of whose free variables occur among v0 : : : vk . Example 2. These would be formulas of any language : For any variable vi : vi = vi for any term t(v0 : : : vk ) and other terms t1 and t2 : t1 = t2 ! t(v0 : : : vi;1 t1 vi+1 : : : vk ) = t(v0 : : : vi;1 t2 vi+1 : : : vk ) for any formula '(v0 : : : vk ) and terms t1 and t2 : t1 = t2 ! '(v0 : : : vi;1 t1 vi+1 : : : vk ) $ '(v0 : : : vi;1 t2 vi+1 : : : vk ) Note the simple way we denote the substitution of t1 for vi . Definition 6. A model (or structure) A for a language L is an ordered pair hA Ii where A is a set and I is an interpretation function with domain the set of all constant, function and relation symbols of L such that: 1. if c is a constant symbol, then I (c) 2 A I (c) is called a constant

0. MODELS, TRUTH AND SATISFACTION

2. if F is an m-placed function symbol, then I (F ) is an m-placed function on 3. if R is an n-placed relation symbol, then I (R) is an n-placed relation on A. A is called the universe of the model A. We generally denote models with Gothic letters and their universes with the corresponding Latin letters in boldface. One set may be involved as a universe with many di erent interpretation functions of the language L. The model is both the universe and the interpretation function. Remark. The importance of Model Theory lies in the observation that mathematical objects can be cast as models for a language. For instance, the real numbers with the usual ordering < and the usual arithmetic operations, addition + and multiplication along with the special numbers 0 and 1 can be described as a model. Let L contain one two-placed (i.e. binary) relation symbol R0 , two two-placed function symbols F1 and F2 and two constant symbols c0 and c1 . We build a model by letting the universe A be the set of real numbers. The interpretation function I will map R0 to < , i.e. R0 will be interpreted as < . Similarly, I (F1 ) will be + , I (F2 ) will be , I (c0 ) will be 0 and I (c1 ) will be 1. So hA Ii is an example of a model. We now wish to show how to use formulas to express mathematical statements about elements of a model. We rst need to see how to interpret a term in a model. Definition 7. The value t x0 : : : xq ] of a term t(v0 : : : vq ) at x0 : : : xq in the model A is de ned as follows: 1. if t is vi then t x0 xq ] is xi , 2. if t is the constant symbol c, then t x0 : : : xq ] is I (c), the interpretation of c in A, 3. if t is F (t1 : : : tm ) where F is an m-placed function symbol and t1 : : : tm are terms, then t x0 : : : xq ] is G(t1 x0 : : : xq ] : : : tm x0 : : : xq ]) where G is the m-placed function I (F ), the interpretation of F in A. Definition 8. Suppose A is a model for a language L. The sequence x0 : : : xq satis es the formula '(v0 : : : vq ) all of whose free and bound variables are among v0 : : : vq , in the model A, written A j= ' x0 : : : xq ] provided we have: 1. if '(v0 : : : vq ) is the formula t1 = t2 , then A j= t1 = t2 x0 : : : xq ] means that t1 x0 : : : xq ] equals t2 x0 : : : xq ], 2. if '(v0 : : : vq ) is the formula R(t1 : : : tn ) where R is an n-placed relation symbol, then A j= R(t1 : : : tn ) x0 : : : xq ] means S (t1 x0 : : : xq ] : : : tn x0 : : : xq ]) where S is the n-placed relation I (R), the interpretation of R in A, 3. if ' is (: ), then A j= ' x0 : : : xq ] means not A j= x0 : : : xq ], 4. if ' is ( ^ ), then A j= ' x0 : : : xq ] means both A j= x0 : : : xq ] and A j= x0 : : : xq ], 5. if ' is ( _ ) then A j= ' x0 : : : xq ] means either A j= x0 : : : xq ] or A j= x0 : : : xq ],

0. MODELS, TRUTH AND SATISFACTION

6. if ' is ( ! ) then A j= ' x0 : : : xq ] means that A j= x0 : : : xq ] implies A j= x0 : : : xq ], 7. if ' is ( $ ) then A j= ' x0 : : : xq ] means that A j= x0 : : : xq ] i A j= x0 : : : xq ], 8. if ' is 8vi , then A j= ' x0 : : : xq ] means for every x 2 A A j= x0 : : : xi;1 x xi+1 : : : xq ], 9. if ' is 9vi , then A j= ' x0 : : : xq ] means for some x 2 A A j= x0 : : : xi;1 x xi+1 : : : xq ]: Exercise 3. Each of the formulas of Example 2 is satis ed in any model A for any language L by any (long enough) sequence x0 x1 : : : xq of A. This is where you test your solution to Exercise 2. We now prove two lemmas which show that the preceeding concepts are wellde ned. In the rst one, we see that the value of a term only depends upon the values of the variables which actually occur in the term. In this lemma the equal sign = is used, not as a logical symbol in the formal sense, but in its usual sense to denote equality of mathematical objects | in this case, the values of terms, which are elements of the universe of a model. Lemma 1. Let A be a model for L and let t(v0 : : : vp ) be a term of L. Let x0 : : : xq and y0 : : : yr be sequences from A such that p q and p r, and let xi = yi whenever vi actually occurs in t(v0 : : : vp ). Then t x0 : : : xq ] = t y0 : : : yr ] . Proof. We use induction on the complexity of the term t. 1. If t is vi then xi = yi and so we have t x0 : : : xq ] = xi = yi = t y0 : : : yr ] since p q and p r: 2. If t is the constant symbol c, then t x0 : : : xq ] = I (c) = t y0 : : : yr ] where I (c) is the interpretation of c in A. 3. If t is F (t1 : : : tm) where F is an m-placed function symbol, t1 : : : tm are terms and I (F ) = G, then t x0 : : : xq ] = G(t1 x0 : : : xq ] : : : tm x0 : : : xq ]) and t y0 : : : yr ] = G(t1 y0 : : : yr ] : : : tm y0 : : : yr ]). By the induction hypothesis we have that ti x0 : : : xq ] = ti y0 : : : yr ] for 1 i m since t1 : : : tm have all their variables among fv0 : : : vp g. So we have t x0 : : : xq ] = t y0 : : : yr ]. In the next lemma the equal sign = is used in both senses | as a formal logical symbol in the formal language L and also to denote the usual equality of mathematical objects. This is common practice where the context allows the reader to distinguish the two usages of the same symbol. The lemma con rms that satisfaction of a formula depends only upon the values of its free variables.

0. MODELS, TRUTH AND SATISFACTION

Lemma 2. Let A be a model for L and ' a formula of L, all of whose free and bound variables occur among v0 : : : vp . Let x0 : : : xq and y0 : : : yr (q r p) be two sequences such that xi and yi are equal for all i such that vi occurs free in '. Then A j= ' x0 : : : xq ] i A j= ' y0 : : : yr ] Proof. Let A and L be as above. We prove the lemma by induction on the complexity of '. 1. If '(v0 : : : vp ) is the formula t1 = t2 , then we use Lemma 1 to get: A j= (t1 = t2 ) x0 : : : xq ] i t1 x0 : : : xq ] = t2 x0 : : : xq ] i t1 y0 : : : yr ] = t2 y0 : : : yr ] i A j= (t1 = t2 ) y0 : : : yr ]: 2. If '(v0 : : : vp ) is the formula R(t1 : : : tn ) where R is an n-placed relation symbol with interpretation S , then again by Lemma 1, we get: A j= R(t1 : : : tn ) x0 : : : xq ] i S (t1 x0 : : : xq ] : : : tn x0 : : : xq ]) i S (t1 y0 : : : yr ] : : : tn y0 : : : yr ]) i A j= R(t1 : : : tn ) y0 : : : yr ]: 3. If ' is (: ), the inductive hypothesis gives that the lemma is true for . So, A j= ' x0 : : : xq ] i not A j= x0 : : : xq ] i not A j= y0 : : : yr ] i A j= ' y0 : : : yr ]: 4. If ' is ( ^ ), then using the inductive hypothesis on and we get A j= ' x0 : : : xq ] i both A j= x0 : : : xq ] and A j= x0 : : : xq ] i both A j= y0 : : : yr ] and A j= y0 : : : yr ] i A j= ' y0 : : : yr ]: 5. If ' is ( _ ) then A j= ' x0 : : : xq ] i either A j= x0 : : : xq ] or A j= x0 : : : xq ] i either A j= y0 : : : yr ] or A j= y0 : : : yr ] i A j= ' y0 : : : yr ]: 6. If ' is ( ! ) then A j= ' x0 : : : xq ] i A j= x0 : : : xq ] implies A j= x0 : : : xq ] i A j= y0 : : : yr ] implies A j= y0 : : : yr ] i A j= ' y0 : : : yr ]: 7. If ' is ( $ ) then A j= ' x0 : : : xq ] i we have A j= x0 : : : xq ] i A j= x0 : : : xq ] i we have A j= y0 : : : yr ] i A j= y0 : : : yr ] i A j= ' y0 : : : yr ]: 8. If ' is 8vi , then A j= ' x0 : : : xq ] i for every z 2 A A j= x0 : : : xi;1 z xi+1 : : : xq ] i for every z 2 A A j= y0 : : : yi;1 z yi+1 : : : yr ] i A j= ' y0 : : : yr ]: The inductive hypothesis uses the sequences x0 : : : xi;1 z xi+1 : : : xq and y0 : : : yi;1 z yi+1 : : : yr with the formula .

0. MODELS, TRUTH AND SATISFACTION

9. If ' is 9vi , then A j= ' x0 : : : xq ] i for some z 2 A A j= x0 : : : xi;1 z xi+1 : : : xq ] i for some z 2 A A j= y0 : : : yi;1 z yi+1 : : : yr ] i A j= ' y0 : : : yr ]: The inductive hypothesis uses the sequences x0 : : : xi;1 z xi+1 : : : xq and y0 : : : yi;1 z yi+1 : : : yr with the formula .
Definition 9. A sentence is a formula with no free variables.

In this case we say: A satis es ' or A is a model of ' or ' holds in A or ' is true in A If ' is a sentence of L, we write j= ' to mean that A j= ' for every model A for L. Intuitively then, j= ' means that ' is true under any relevant interpretation (model for L). Alternatively, no relevant example (model for L) is a counterexample to ' | so ' is true. Lemma 3. Let '(v0 : : : vq ) be a formula of the language L. There is another formula '0 (v0 : : : vq ) of L such that 1. '0 has exactly the same free and bound occurances of variables as '. 2. '0 can possibly contain :, ^ and 9 but no other connective or quanti er. 3. j= (8v0 ) : : : (8vq )(' $ '0 ) Exercise 4. Prove the above lemma by induction on the complexity of '. (1) no variable occuring in ' occurs both free and bound, (2) no bound variable occuring in ' is bound by more than one quanti er, and (3) in the written order, all of the quanti ers preceed all of the connectives. Remark. (3) is equivalent to saying that in the constructed order, all of the connective steps preceed all of the quanti er steps. Lemma 4. Let '(v0 : : : vp ) be any formula of a language L. There is a formula ' of L which has the following properties: 1. ' is in prenex normal form 2. ' and ' have the same free occurances of variables, and 3. j= (8v0 ) : : : (8vp )(' $ ' ) Exercise 5. Prove this lemma by induction on the complexity of '. There is a notion of rank on prenex formulas | the number of alternations of quanti ers. The usual formulas of elementary mathematics have prenex rank 0, i.e. no alternations of quanti ers. For example: (8x)(8y)(2xy x2 + y2 ):
Definition 10. A formula ' is said to be in prenex normal form whenever

A since by the previous lemma, it doesn't matter which sequence from A we use.

If ' is a sentence, we can write A j= ' without any mention of a sequence from

0. MODELS, TRUTH AND SATISFACTION

10

However, the ; de nition of a limit of a function has prenex rank 2 and is much more di cult for students to comprehend at rst sight: 8 9 8x((0 < ^ 0 < jx ; aj < ) ! jF (x) ; Lj < ): A formula of prenex rank 4 would make any mathematician look twice.

CHAPTER 1

Notation and Examples

Although the formal notation for formulas is precise, it can become cumbersome and di cult to read. Con dent that the reader would be able, if necessary, to put formulas into their formal form, we will relax our formal behaviour. In particular, we will write formulas any way we want using appropriate symbols for variables, constant symbols, function and relation symbols. We will omit parentheses or add them for clarity. We will use binary function and relation symbols between the arguments rather than in front as is the usual case for \plus", \times" and \less than". Whenever a language L has only nitely many relation, function and constant symbols we often write, for example: L = f< R0 + F1 c0 c1 g omitting explicit mention of the logical symbols (including the in nitely many variables) which are always in L. Correspondingly we may denote a model A for L as: A = hA < S0 + G1 a0 a1 i where the interpretations of the symbols in the language L are given by I (<) = <, I (R0 ) = S0 , I (+) = + , I (F1 ) = G1 , I (c0 ) = a0 and I (c1 ) = a1 . Example 3. R = hR < + 0 1i and Q = hQ < + 0 1i, where R is the reals, Q the rationals , are models for the language L = f< + 0 1g. Here < is a binary relation symbol, + and are binary function symbols, 0 and 1 are constant symbols whereas <, + , , 0, 1 are the well known relations, arithmetic functions and constants. Similarly, C = hC + 0 1i, where C is the complex numbers, is a model for the language L = f+ 0 1g. Note the exceptions to the boldface convention for these popular sets. Example 4. Here L = f< + 0 1g, where < is a binary relation symbol, + and are binary function symbols and 0 and 1 are constant symbols. The following formulas are sentences. 1. (8x):(x < x) 2. (8x)(8y):(x < y ^ y < x) 3. (8x)(8y)(8z )(x < y ^ y < z ! x < z ) 4. (8x)(8y)(x < y _ y < x _ x = y) 5. (8x)(8y)(x < y ! (9z )(x < z ^ z < y)) 6. (8x)(9y)(x < y) 7. (8x)(9y)(y < x) 8. (8x)(8y)(8z )(x + (y + z ) = (x + y) + z ) 9. (8x)(x + 0 = x)
11

1. NOTATION AND EXAMPLES

12

(8x)(9y)(x + y = 0) (8x)(8y)(x + y = y + x) (8x)(8y)(8z )(x (y z ) = (x y) z ) (8x)(x 1 = x) (8x)(x = 0 _ (9y)(y x = 1) (8x)(8y)(x y = y x) (8x)(8y)(8z )(x (y + z ) = (x y) + (y z )) 0 6= 1 (8x)(8y)(8z )(x < y ! x + z < y + z ) (8x)(8y)(8z )(x < y ^ 0 < z ! x z < y z ) for each n 1 we have the formula (8x0 )(8x1 ) (8xn )(9y)(xn yn + xn;1 yn;1 + 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20.
z }|

where, as usual, yk abbreviates y y y The latter formulas express that each polynomial of degree n has a root. The following formulas express the intermediate value property for polynomials of degree n: if the polynomial changes sign from w to z , then it is zero at some y between w and z . 21. for each n 1 we have (8x0 ) : : : (8xn )(8w)(8z ) (xn wn + xn;1 wn;1 + + x1 w + x0 ) (xn z n + xn;1 z n;1 + + x1 z + x0 ) < 0 ! (9y )(((w < y ^ y < z ) _ (z < y ^ y < w)) ^ (xn y n + xn;1 y n;1 + + x1 y + x0 = 0))] The most fundamental concept is that of a sentence being true when interpreted in a model A. We write this as A j= , and we extend this concept in the following de nitions. Definition 11. If is a set of sentences, A is said to be a model of , written A j= , whenever A j= for each 2 . is said to be satis able i there is some A such that A j= . Definition 12. A theory T is a set of sentences. If T is a theory and is a sentence, we write T j= whenever we have that for all A if A j= T then A j= . We say that is a consequence of T . A theory is said to be closed whenever it contains all of its consequences. Definition 13. If A is a model for the language L, the theory of A, denoted by ThA, is de ned to be the set of all sentences of L which are true in A, f of L : A j= g This is one way that a theory can arise. Another way is through axioms. Definition 14. T is said to be a set of axioms for T whenever j= for every in T in this case we write j= T . Remark. We will generally assume our theories are closed and we will often describe theories by specifying a set of axioms . The theory will then be all consequences of . Example 5. We will consider the following theories and their axioms:

+ x1 y + x0 = 0 _ xn = 0)
{

1. NOTATION AND EXAMPLES

13

1. The theory of Linear Orderings (LOR) which has as axioms sentences 1-4 from Example 4. 2. The theory of Dense Linear Orders (DLO) which has as axioms all the axioms of LOR, and sentence 5, 6 and 7 of Example 4. 3. The theory of Fields (FEI) which has as axioms sentences 8-17 from Example 4. 4. The theory of Ordered Fields (ORF) which has as axioms all the axioms of FEI, LOR and sentences 18 and 19 from Example 4. 5. The theory of Algebraically Closed Fields (ACF) which has as axioms all the axioms of FEI and all sentences from 20 of Example 4, i.e. in nitely many sentences, one for each n 1. 6. The theory of Real Closed Ordered Fields (RCF) which has as axioms all the axioms of ORF, and all sentences from 21 of Example 4, i.e. in nitely many sentences, one for each n 1. Exercise 6. Show that : 1. Q j= DLO 2. R j= RCF using the Intermediate Value theorem 3. C j= ACF using the Fundamental Theorem of Algebra where Q, R and C are as in Example 3. Remark. The theory of Real Closed Ordered Fields is sometimes axiomatized di erently. All the axioms of ORF are retained, but the sentences from 21 of Example 4, which amount to an Intermediate Value Property, are replaced by the sentences from 20 for odd n and the sentence (8x)(0 < x ! (9y)y2 = x) which states that every positive element has a square root. A signi cant amount of algebra would then be used to verify the Intermediate Value Property from these axioms.

CHAPTER 2

Compactness and Elementary Submodels

Theorem 1. The Compactness Theorem (Malcev) A set of sentences is satis able i every nite subset is satis able. Proof. There are several proofs. We only point out here that it is an easy consequence of the following, a theorem which appears in all elementary logic texts: Proposition. The Completeness Theorem (Godel, Malcev) A set of sentences is consistent i it is satis able. Although we do not here formally de ne \consistent", it does mean what you think it does. In particular, a set of sentences is consistent if and only if each nite subset is consistent. Remark. The Compactness Theorem is the only one for which we do not give a complete proof. If the reader has not previously seen the Completeness Theorem, there are other proofs of the Compactness Theorem which may be more easily absorbed: set theoretic (using ultraproducts), topological (using compact spaces, hence the name) or Boolean algebraic. However these topics are too far a eld to enter into the proofs here. We will use the Compactness Theorem as a starting point | in fact, all that follows can be seen as its corollaries. Exercise 7. Suppose T is a theory for the language L and is a sentence of L such that T j= . Prove that there is some nite T 0 T such that T 0 j= . Recall that T j= i T f: g is not satis able. Definition 15. If L, and L0 are two languages such that L L0 we say that 0 is an expansion of L and L is a reduction of L0 L Definition 16. Given a model A for the language L, we can expand it to a model A0 of L0 by giving appropriate interpretations to the symbols in L0 n L. We say that A0 is an expansion of A to L0 and that A is a reduct of A0 to L. We also use the notation A0 jL for the reduct of A0 to L. Theorem 2. If a theory T has arbitrarily large nite models, then it has an in nite model. Proof. Consider new constant symbols ci for i 2 N , the usual natural numbers, and expand from L, the language of T , to L0 = L fci : i 2 N g. Let = T f:ci = cj : i 6= j i j 2 N g: We rst show that every nite subset of has a model by interpreting the nitely many relevant constant symbols as di erent elements in an expansion of some nite model of T . Then we use compactness to get a model A0 of .
14

2. COMPACTNESS AND ELEMENTARY SUBMODELS

15

The model that we require is for the language L, so we take A to be the reduct of A0 to L.
L = f+

< 0 1g which are true in hN + < 0 1i, the standard model which we all learned in school. Theorem 3. (T. Skolem) There exist non-standard models of number theory. Proof. Add a new constant symbol c to L. Consider
z }|

Example 6. Number theory is ThhN + < 0 1i, the set of all sentences of

Th(hN + < 0 1i ) f1 + 1 + + 1 < c : n 2 N g and use the Compactness Theorem. The interpretation of the constant symbol c will not be a natural number.
Definition 17. Two models A and A0 for L are said to be isomorphic whenever there is a bijection f : A ! A0 such that

1. for each n-placed relation symbol R of L and corresponding interpretations S of A and S 0 of A0 we have S (x1 : : : xn ) i S 0 (f (x1 ) : : : f (xn )) for all x1 : : : xn in A 2. for each n-placed function symbol F of L and corresponding interpretations G of A and G0 of A0 we have f (G(x1 : : : xn )) = G0 (f (x1 ) : : : f (xn )) for all x1 : : : xn in A 3. for each constant symbol c of L and corresponding constant elements a of A and a0 of A0 we have f (a) = a0 . We write A = A0 . This is an equivalence relation. Definition 18. Two models A and A0 for L are said to be elementarily equivalent whenever we have that for each sentence of L A j= i A0 j= We write A A0 . This is another equivalence relation. Exercise 8. Suppose f : A ! A0 is an isomorphism and ' is a formula such that A j= ' a0 : : : ak ] for some a0 : : : ak from A then A0 j= ' f (a0 ) : : : f (ak )]. Use this to show that A = A0 implies A A0 . Definition 19. A model A0 is called a submodel of A i 6= A0 A and 1. each n-placed relation S 0 of A0 is the restriction to A0 of the corresponding relation S of A, i. e. S 0 = S \ (A0 )n 2. each m-placed function G0 of A0 is the restriction to A0 of the corresponding function G of A, i. e. G0 = G (A0 )m 3. each constant of A0 is the corresponding constant of A. We write A0 A. Definition 20. Let A and B be two models for L. We say A is an elementary submodel of B and B is an elementary extension of A and we write A B whenever 1. A B and

2. COMPACTNESS AND ELEMENTARY SUBMODELS

16

2. for all formulas '(v0 : : : vk ) of L and all a0 : : : ak 2 A A j= ' a0 : : : ak ] i B j= ' a0 : : : ak ]: Exercise 9. Prove that if A B then A B and A B. Example 7. Let N be the usual natural numbers with < as the usual ordering. Let B = hN < i and A = hN n f0g <i be models for the language with one binary relation symbol <. Then A B and A B in fact A = B. But we do not have A B A satis es a formula that says 1 is the least element of the ordering, but B does not. So we see that being an elementary submodel is a very strong condition indeed. Nevertheless, later in the chapter we will obtain many examples of elementary submodels. Exercise 10. Show that if A B and B C then A C and if A B and B C then A C. Definition 21. A chain of models for a language L is an increasing sequence of models A0 A1 An n 2 N: The union of the chain is de ned to be the model A = fAn : n 2 N g where the universe of A is A = fAn : n 2 N g and: 1. each relation S on A is the union of the corresponding relations Sn of An S = fSn : n 2 N g, i.e. the relation extending each Sn 2. each function G on A is the union of the corresponding functions Gn of An G = fGn : n 2 N g, i.e. the function extending each Gn 3. all the models An and A have the same constant elements. Note that each An A. Example 8. For each n 2 N , let An = f;n ;n + 1 ;n + 2 : : : 0 1 2 3 : : : g Z . Let An = hAn i. Each An A0 , but we don't have A0 fAn : n 2 N g. Remark. To be sure, what is de ned here is a chain of models indexed by the natural numbers N . More generally, a chain of models could be indexed by any ordinal. However we will not need the concept of an ordinal at this point.
Definition 22. An elementary chain is a chain of models fAn : n 2 N g such that for each m < n we have Am An . Theorem 4. (Tarski's Elementary Chain Theorem) Let fAn : n 2 N g be an elementary chain. For all n 2 N we have An fAn : n 2 N g Proof. Denote the union of the chain by A. We have Ak A for each k 2 N . Claim. If t is a term of the language L and a0 : : : ap are in Ak , then the value of the term t a0 : : : ap ] in A is equal to the value in Ak . Proof of Claim. We prove this by induction on the complexity of the term.

2. COMPACTNESS AND ELEMENTARY SUBMODELS

17

1. If t is the variable vi then both values are just ai . 2. If t is the constant symbol c then the values are equal because c has the same interpretation in A and in Ak . 3. If t is F (t1 : : : tm ) where F is a function symbol and t1 : : : tm are terms such that each value ti a0 : : : ap ] is the same in both A and Ak , then the value F (t1 : : : tm ) a0 : : : ap ] in A is G(t1 a0 : : : ap ] : : : tm a0 : : : ap ]) where G is the interpretation of F in A and the value of F (t1 : : : tm ) a0 : : : ap ] in Ak is Gk (t1 a0 : : : ap ] : : : tm a0 : : : ap ]) where Gk is the interpretation of F in Ak . But Gk is the restriction of G to Ak so these values are equal. In order to show that each Ak A it will su ce to prove the following statement for each formula '(v0 : : : vp ) of L. \ For all k 2 N and all a0 : : : ap in Ak : A j= ' a0 : : : ap ] i Ak j= ' a0 : : : ap ]:" Claim. The statement is true whenever ' is t1 = t2 where t1 and t2 are terms. Proof of Claim. Fix k 2 N and a0 : : : ap in Ak . A j= (t1 = t2 ) a0 : : : ap ] i t1 a0 : : : ap ] = t2 a0 : : : ap ] in A i t1 a0 : : : ap ] = t2 a0 : : : ap ] in Ak i Ak j= (t1 = t2 ) a0 : : : ap ]:
Claim. The statement is true whenever ' is R(t1 : : : tn ) where R is a relation symbol and t1 : : : tn are terms. Proof of Claim. Fix k 2 N and a0 : : : ap in Ak . Let S be the interpretation of R in A and Sk be the interpretation in Ak Sk is the restriction of S to Ak . A j= R(t1 : : : tn ) a0 : : : ap ] i S (t1 a0 : : : ap ] : : : tn a0 : : : ap ]) i Sk (t1 a0 : : : ap ] : : : tn a0 : : : ap ]) i Ak j= R(t1 : : : tn ) a0 : : : ap ]

' is : .

Claim. If the statement is true when ' is , then the statement is true when Proof of Claim. Fix k 2 N and a0 : : : ap in Ak . A j= (: ) a0 : : : ap ] i not A j= a0 : : : ap ] i not Ak j= a0 : : : ap ] i Ak j= (: ) a0 : : : ap ]:

2. COMPACTNESS AND ELEMENTARY SUBMODELS

18

Claim. If the statement is true when ' is 1 and when ' is 2 then the statement is true when ' is 1 ^ 2 . Proof of Claim. Fix k 2 N and a0 : : : ap in Ak . A j= ( 1 ^ 2 ) a0 : : : ap ] i A j= 1 a0 : : : ap ] and A j= 2 a0 : : : ap ] i Ak j= 1 a0 : : : ap ] and Ak j= 2 a0 : : : ap ] i Ak j= ( 1 ^ 2 ) a0 : : : ap ]:

is 9vi .

Claim. If the statement is true when ' is then the statement is true when '

A = fAj : j 2 Ng.

Proof of Claim. Fix k 2 N and a0 : : : ap in Ak . Note that

A j= 9vi a0 : : : ap ] i A j= 9vi a0 : : : aq ] where q is the maximum of i and p (by Lemma 2),

i A j= a0 : : : ai;1 a ai+1 : : : aq ] for some a 2 A i A j= a0 : : : ai;1 a ai+1 : : : aq ] for some a 2 Al for some l k i Al j= a0 : : : ai;1 a ai+1 : : : aq ] since the statement is true for , i Al j= 9vi a0 : : : aq ] i Ak j= 9vi a0 : : : aq ] since Ak Al i Ak j= 9vi a0 : : : ap ] (by Lemma 2): By induction on the complexity of ', we have proven the statement for all formulas ' which do not contain the connectives _, ! and $ or the quanti er 8. To verify the statement for all ' we use Lemma 3. Let ' be any formula of L. By Lemma 3 there is a formula which does not use _, !, $ nor 8 such that j= (8v0 ) : : : (8vp )(' $ ): Now x k 2 N and a0 : : : ap in Ak . We have A j= (' $ ) a0 : : : ap ] and Ak j= (' $ ) a0 : : : ap ]: A j= ' a0 : : : ap ] i A j= a0 : : : ap ] i Ak j= a0 : : : ap ] i Ak j= ' a0 : : : ap ] which completes the proof of the theorem.
Lemma 5. (The Tarski-Vaught Condition) Let A and B be models for L with A B. The following are equivalent: (1) A B

2. COMPACTNESS AND ELEMENTARY SUBMODELS

19

(2) for any formula (v0 : : : vq ) and any i q and any a0 : : : aq from A: if there is some b 2 B such that B j= a0 : : : ai;1 b ai+1 : : : aq ] then we have some a 2 A such that B j= a0 : : : ai;1 a ai+1 : : : aq ]: Proof. Only the implication (2) ) (1) requires a lot of proof. We will prove that for each formula '(v0 : : : vp ) and all a0 : : : ap from A we will have: A j= ' a0 : : : ap ] i B j= ' a0 : : : ap ] by induction on the complexity of ' using only the negation symbol :, the connective ^ and the quanti er 9 (recall Lemma 3). 1. The cases of formulas of the form t1 = t2 and R(t1 : : : tn ) come immediately from the fact that A B. 2. For negation: suppose ' is : and we have it for , then A j= ' a0 : : : ap ] i not A j= a0 : : : ap ] i not B j= a0 : : : ap ] i B j= ' a0 : : : ap ]: 3. The ^ case proceeds similarly. 4. For the 9 case we consider ' as 9vi . If A j= 9vi a0 : : : ap ], then the inductive hypothesis for and the fact that A B ensure that B j= 9vi a0 : : : ap ]. It remains to show that if B j= ' a0 : : : ap ] then A j= ' a0 : : : ap ]. Assume B j= 9vi a0 : : : ap ]. By Lemma 2, B j= 9vi a0 : : : aq ] where q is the maximum of i and p. By the de nition of satisfaction, there is some b 2 B such that B j= a0 : : : ai;1 b ai+1 : : : aq ]: By (2), there is some a 2 A such that B j= a0 : : : ai;1 a ai+1 : : : aq ]: By the inductive hypothesis on , for that same a 2 A, A j= a0 : : : ai;1 a ai+1 : : : aq ]: By the de nition of satisfaction, A j= 9vi a0 : : : aq ]: Finally, by Lemma 2, A j= a0 : : : ap ]. Recall that jBj is used to represent the cardinality, or size, of the set B. Note that since any language L contains in nitely many variables, jLj is always in nite, but may be countable or uncountable depending on the number of other symbols. We often denote an arbitrary in nite cardinal by the lower case Greek letter . Theorem 5. (Downward Lowenheim-Skolem Theorem) Let B be a model for L and let be any cardinal such that jLj < jBj. Then B has an elementary submodel A of cardinality . Furthermore if X B and jX j , then we can also have X A.

2. COMPACTNESS AND ELEMENTARY SUBMODELS

20

Since jLj and each formula of L is a nite string of symbols from L, there are at most many formulas of L. So there are at most elements of B that need to be added to each Xn and so, without loss of generosity each jXn j = . Let A = fXn : n 2 Ng then jAj = . Since A is closed under functions from B and contains all constants from B, A gives rise to a submodel A B. The Tarski-Vaught Condition is used to show that A B.
Theorem 6. (The Upward Lowenheim-Skolem Theorem) Let A be an in nite model for L and be any cardinal such that jLj and jAj < . Then A has an elementary extension of cardinality . Proof. For each a 2 A, let ca be a new constant symbol let L0 = L fca : a 2 Ag: Note that sentences of L0 are just formulas of L with all free variables replaced by constant symbols. In addition, add many new constant symbols to L0 to make L00 . De ne to be the following set of sentences of L00 : f:d = d0 : d and d0 are distinct new constant symbols of L00 n L0 g f : is a sentence of L0 obtained from the formula '(v0 : : : vp ) of L by replacing each free variable vi by the constant symbol cai and A j= ' a0 : : : ap ]g By interpreting ca as a and the d's as distinct elements of A we can transform A into a model of any nite subset of . Using the Compactness Theorem, we obtain a model D00 for L00 such that D00 j= . Note that D00 has size at least and that for any sentence of L0 , D00 j= i 2 . Obtain a model C00 for L00 from D00 by simply switching elements of the universe of D00 with A to ensure that for each a 2 A the interpretation of ca in C00 is a. Hence the universe of C00 contains A and C00 j= . Let C be the reduct of C00 to L. The following argument will show that A C. Let ' be any formula of L and a0 : : : ap any elements from A. Let be the sentence of L0 formed by replacing free occurances of vi with cai . We have A j= ' a0 : : : ap ] i 2 i D00 j= i C00 j= i C j= ' a0 : : : ap ]: However, C may have size strictly larger than . In this case we obtain our nal B by using the previous theorem to get B C with A B. It is now straightforward to conclude that A B.

Proof. Without loss of generosity assume jX j = . We recursively de ne sets Xn for n 2 N such that X = X0 X1 Xn and such that for each formula '(v0 : : : vp ) of L and each i p and each a0 : : : ap from Xn such that B j= 9vi ' a0 : : : ap ] we have x 2 Xn+1 such that B j= ' a0 : : : ai;1 x ai+1 : : : ap ]:

2. COMPACTNESS AND ELEMENTARY SUBMODELS

21

Definition 23. A theory T for a language L is said to be complete whenever for each sentence of L either T j= or T j= : . Lemma 6. A theory T for L is complete i any two models of T are elementarily equivalent. Proof. ()) easy. (() easy. Definition 24. A theory T is said to be categorical in cardinality whenever any two models of T of cardinality are isomorphic. We also say that T is categorical. The most interesting cardinalities in the context of categorical theories are @0 , the cardinality of countably in nite sets, and @1, the rst uncountable cardinal. Exercise 11. Show that DLO is @0 -categorical. There are two well-known proofs. One uses a back-and-forth construction of an isomorphism. The other constructs, by recursion, an isomorphism from the set of dyadic rational numbers between 0 and 1: n f m : m is a positive integer and n is an integer 0 < n < 2m g 2 onto a countable dense linear order without endpoints. Now use the following theorem to show that DLO is complete. Theorem 7. (The Los-Vaught Test) Suppose that a theory T has only in nite models for a language L and that T is -categorical for some cardinal jLj. Then T is complete. Proof. We will show that any two models of T are elementarily equivalent. Let A of cardinality 1 , and B of cardinality 2 , be two models of T . If 1 < use the Upward Lowenheim-Skolem Theorem to get A0 such that 0 j = and A A0 . jA If 1 > use the Downward Lowenheim-Skolem Theorem to get A0 such that 0 j = and A0 A. jA Either way, we can get A0 such that jA0 j = and A0 A. Similarly, we can get B0 such that jB0 j = and B0 B. Since T is -categorical, A0 = B0 . Hence A B. Recall that the characteristic of a eld is the prime number p such that

1+1+ +1=0 provided that such a p exists, and, if no such p exists the eld has characteristic 0. All of our best-loved elds: Q , R and C have characteristic 0. On the other hand, elds of characteristic p include the nite eld of size p (the prime Galois eld). Theorem 8. The theory of algebraically closed elds of characteristic 0 is complete. Proof. We use the Los-Vaught Test and the following Lemma. Lemma 7. Any two algebraically closed elds of characteristic 0 and cardinality @1 are isomorphic.

}|

2. COMPACTNESS AND ELEMENTARY SUBMODELS

22

Proof. Let A be such a eld containing the rationals Q = hQ + 0 1i as a prime sub eld. In a manner completely analogous to nding a basis for a vector space, we can nd a transcendence basis for A, that is, an indexed subset fa : 2 I g A such that A is the algebraic closure of the sub eld A0 generated by fa : 2 I g but no a is in the algebraic closure of the sub eld generated by the rest: fa : 2 I and 6= g. Since the sub eld generated by a countable subset would be countable and the algebraic closure of a countable sub eld would also be countable, we must have that the transcendence base is uncountable. Since jAj = @1 , the least uncountable cardinal, we must have in fact that jI j = @1 . Now let B be any other algebraically closed eld of characteristic 0 and size @1 . As above, obtain a transcendence basis fb : 2 J g with jJ j = @1 and its generated sub eld B0 . Since jI j = jJ j, there is a bijection g : I ! J which we can use to build an isomorphism from A to B. Since B has characteristic 0, a standard theorem of algebra gives that the rationals are isomorphically embedded into B. Let this embedding be: f : Q ,! B: We extend f as follows: for each 2 I , let f (a ) = bg( ) , which maps the transcendence basis of A into the transcendence basis of B. We now extend f to map A0 onto B0 as follows: Each element of A0 is given by p(a 1 : : : a m ) q(a 1 : : : a m ) where p and q are polynomials with rational coe cients and the a's come, of course, from the transcendence basis. Let f map such an element to p(bg( 1 ) : : : bg( m ) ) q(bg( 1 ) : : : bg( m ) ) where p and q are polynomials whose coe cients are the images under f of the rational coe cients of p and q. The nal extension of f to all of A and B comes from the uniqueness of algebraic closures. Remark. Lemma 7 is also true when 0 is replaced by any xed characteristic and @1 by any uncountable cardinal. Theorem 9. Let H be a set of sentences in the language of eld theory which are true in algebraically closed elds of arbitrarily high characteristic. Then H holds in some algebraically closed eld of characteristic 0. Proof. A eld is a model in the language f+ 0 1g of the axioms of eld theory. Let ACF be the set of axioms for the theory of algebraically closed elds see Example 5. For each n 2, let n denote the sentence

Let = ACF H f n : n 2g Let 0 be any nite subset of and let m be the largest natural number such that m 2 0 or let m = 1 by default.

:(1 + 1 +

}|

+ 1) = 0

2. COMPACTNESS AND ELEMENTARY SUBMODELS

23

Let A be an algebraically closed eld of characteristic p > m such that A j= H then in fact A j= 0 . So by compactness there is B such that B j= . B is the required eld.
Corollary 1. Let C denote, as usual, the complex numbers. Every one-toone polynomial map f : C m ! C m is onto. Proof. A polynomial map is a function of the form f (x1 : : : xm ) = hp1 (x1 : : : xm ) : : : pm(x1 : : : xm )i where each pi is a polynomial in the variables x1 : : : xm . We call max f degree of pi : i mg the degree of f . Let L be the language of eld theory and let m n be the sentence of L which expresses that \each polynomial map of m variables of degree < n which is one-toone is also onto". We wish to show that there are algebraically closed elds of arbitrarily high characteristic which satisfy H = f m n : m n 2 N g. We will then apply Theorem 9, Theorem 8, Lemma 6 and Exercise 6 and be nished. Let p be any prime and let Fp be the prime Galois eld of size p. The algebraic ~ closure Fp is the countable union of a chain of nite elds Fp = A0 A1 A2 Ak Ak+1 obtained by recursively adding roots of polynomials. ~ + We nish the proof by showing that each hFpm 0 1i satis es H. m ~ ~ Given any polynomial map f : (Fp ) ! (Fp ) which is one-to-one, we show ~ that f is also onto. Given any elements b1 : : : bm 2 Fp , there is some Ak containing b1 : : : bm as well as all the coe cients of f. Since f is one-to-one, f Am : Am ! Am is a one-to-one polynomial map. k k k Hence, since Am is nite, f Am is onto and so there are a1 : : : am 2 Ak such k k that f (a1 : : : am ) = hb1 : : : bmi. Therefore f is onto. Thus, for each prime number p and each m n 2 N , m n holds in a eld of ~ characteristic p, i.e. hFp + 0 1i satis es H.

It is a signi cant problem to replace \one-to-one" with \locally one-to-one".

CHAPTER 3

Diagrams and Embeddings

Let A = hA Ii be a model for a language L and X A. Expand L to = L fca : a 2 X g by adding new constant symbols to L. We can expand A to a model AX = hA I 0 i for LX by choosing I 0 extending I such that I 0 (ca ) = a for each a 2 X . We sometimes write this as hA cx ix2X . We often deal with the case X = A, to obtain AA . Exercise 12. Let A and B be models for L with X A B. Prove: (i) if A B then AX BX . (ii) if A = B then AX = BX . (iii) if A B then AX BX . Hint: A j= ' a1 : : : ap ] i AA j= ' where ' is the sentence of LA formed by replacing each free occurance of vi with cai . Definition 25. Let A be a model for L. 1. The elementary diagram of A is ThAA , the set of all sentences of LA which hold in AA . 2. The diagram of A, denoted by 4A, is the set of all those sentences in ThAA without quanti ers. Remark. There is a notion of atomic formula, which is a formula of the form t1 = t2 or R(t1 : : : tn ) where t1 : : : tn are terms. Sometimes 4A is de ned to be the set of all atomic formulas and negations of atomic formulas which occur in ThAA . However this is not substantially di erent from De nition 25, since the reader can quickly show that for any model B, B j= 4A in one sense i B j= 4A in the other sense. Definition 26. A is said to be isomorphically embedded in B whenever 1. there is a model C and an isomorphism f such that f : A ! C and C B or 2. there is a model D and an isomorphism g such that A D and g : D ! B Exercise 13. Prove that, in fact, 1 and 2 are equivalent conditions. Definition 27. A is said to be elementarily embedded in B whenever 1. there is a model C and an isomorphism f such that f : A ! C and C B
LX

24

3. DIAGRAMS AND EMBEDDINGS

25

2. there is a model D and an isomorphism g such that A D and g : D ! B Exercise 14. Again, prove that, in fact, 1 and 2 are equivalent.
Theorem 10. (The diagram lemmas) Let A and B be models for L.

1. A is isomorphically embedded in B i B can be expanded to a model of 4A . 2. A is elementarily embedded in B i B can be expanded to a model of Th(AA ). Proof. We sketch the proof of 1. ()) If f is as in 1 of De nition 26 above, then hB f (a)ia2A j= 4A. (() If hB ba ia2A j= 4A , then let f (a) = ba .
Exercise 15. Give a careful proof of part 2 of the theorem.

We now apply these notions to graph theory and to calculus. The natural language for graph theory has one binary relation symbol which we call E (to suggest the word \edge"), and the following two axioms: (8x)(8y)E (x y) $ E (y x) (8x):E (x x). A graph is, of course, a model of graph theory. Corollary 2. Every planar graph can be four coloured. Proof. We will have to use the famous result of Appel and Haken that every nite planar graph can be four coloured. Model Theory will take us from the nite to the in nite. We recall that a planar graph is one that can be embedded, or drawn, in the usual Euclidean plane and to be four coloured means that each vertex of the graph can be assigned one of four colours in such a way that no edge has the same colour for both endpoints. Let A be an in nite planar graph. Introduce four new unary relation symbols: R G B Y (for red, green, blue and yellow). We wish to prove that there is some expansion A0 of A such that A0 j= where is the sentence in the expanded language: (8x) R(x) _ G(x) _ B (x) _ Y (x)] ^ (8x) R(x) ! :(G(x) _ B (x) _ Y (x))] ^ : : : ^ (8x)(8y ):(R(x) ^ R(y ) ^ E (x y )) ^ which will ensure that the interpretations of R G B and Y will four colour the graph. Let = 4A f g. Any nite subset of has a model, based upon the appropriate nite subset of A. By the compactness theorem, we get B j= . Since B j= , the interpretations of R G B and Y four colour it. By the diagram lemma A is isomorphically embedded in the reduct of B, and this isomorphism delivers the four-colouring of A. A graph with the property that every pair of vertices is connected with an edge is called complete. At the other extreme, a graph with no edges is called discrete. A very important theorem in nite combinatorics says that most graphs contain an

3. DIAGRAMS AND EMBEDDINGS

26

example of one or the other as a subgraph. A subgraph of a graph is, of course, a submodel of a model of graph theory. Corollary 3. (Ramsey's Theorem) For each n 2 N there is an r 2 N such that if G is any graph with r vertices, then either G contains a complete subgraph with n vertices or a discrete subgraph with n vertices.
Proof. We follow F. Ramsey who began by proving an in nite version of the theorem (also called Ramsey's Theorem). Claim. Each in nite graph G contains either an in nite complete subgraph or an in nite discrete subgraph. Proof of Claim. By force of logical necessity, there are two possiblities: (1) there is an in nite X G such that for all x 2 X there is a nite Fx X such that E (x y) for all y 2 X n Fx , (2) for all in nite X G there is a x 2 X and an in nite Y X such that :E (x y ) for all y 2 Y . If (1) occurs, we recursively pick x1 2 X , x2 2 X n Fx1 , x3 2 X n (Fx1 Fx2 ), etc, to obtain an in nite complete subgraph. If (2) occurs we pick x0 2 G and Y0 G with the property and then recursively choose x1 2 Y0 and Y1 Y0 , x2 2 Y1 and Y2 Y1 and so on, to obtain an in nite discrete subgraph.

We now use Model Theory to go from the in nite to the nite. Let be the sentence, of the language of graph theory, asserting that there is no complete subgraph of size n. (8x1 : : : 8xn ) :E (x1 x2 ) _ :E (x1 x3 ) _ _ :E (xn;1 xn )]: Let be the sentence asserting that there is no discrete subgraph of size n. (8x1 : : : 8xn ) E (x1 x2 ) _ E (x1 x3 ) _ _ E (xn;1 xn )]: Let T be the set consisting of , and the axioms of graph theory. If there is no r as Ramsey's Theorem states, then T has arbitrarily large nite models. By Theorem 2, T has an in nite model, contradicting the claim. The following theorem of A. Robinson nally solved the centuries old problem of in nitesimals in the foundations of calculus. Theorem 11. (The Leibniz Principle) There is an ordered eld R called the hyperreals, containing the reals R and an in nitesimal number such that any statement about the reals which holds in R also holds in R. Proof. Let R be hR + < 0 1 i. We will make the statement of the theorem precise by proving that there is some model B, in the same language L as R and with the universe called R , such that R B and there is b 2 R such that 0 < b < a for each positive a 2 R.

3. DIAGRAMS AND EMBEDDINGS

27

For each real number a, we introduce a new constant symbol ca , and in addition another new constant symbol d. Let be the set of sentences in the expanded language given by: ThRR f0 < d < ca : a is a positive real g We can obtain a model C j= by the compactness theorem. Let C0 be the reduct of C to L. By the elementary diagram lemma R is elementarily embedded in C0 , and so there is a model B for L such that C0 = B and R B.
R is said to be in nitesimal whenever ;r < x < r for each positive r 2 R. 0 is in nitesimal. Two elements x y 2 R are said to be in nitely close, written x y whenever x ; y is in nitesimal. Note: x is in nitesimal i x 0. An element x 2 R is said to be nite whenever ;r < x < r for some positive r 2 R. Else it is in nite. Each nite x 2 R is in nitely close to some real number, called the standard part of x, written st(x). To di ; erentiate f , for each M x 2 R generate M y = f (x+ M x) ; f (x). Then My f 0(x) = st Mx whenever this exists and is the same for each in nitesimal M x 6= 0. The increment lemma states that if y = f (x) is di erentiable at x and M x 0, then M y = f 0 (x) M x + " M x for some in nitesimal ". Proofs of the usual theorems of calculus are now easier. The following theorem is considered one of the most fundamental results of mathematical logic. We give a detailed proof. Theorem 12. (Robinson Consistency Theorem) Let L1 and L2 be two languages with L = L1 \L2 . Suppose T1 and T2 are satis able theories in L1 and L2 respectively. Then T1 T2 is satis able i there is no sentence of L such that T1 j= and T2 j= : . Proof. The direction ) is easy and motivates the whole theorem. We begin the proof in the ( direction. Our goal is to show that T1 T2 is satis able. The following claim is a rst step. Claim. T1 f sentences of L : T2 j= g is satis able. Proof of Claim. Using the compactness theorem and considering conjunctions, it su ces to show that if T1 j= 1 and T2 j= 2 with 2 a sentence of L, then f 1 2 g is satis able. But this is true, since otherwise we would have 1 j= : 2 and hence T1 j= : 2 and so : 2 would be a sentence of L contradicting our hypothesis.

x2

Remark. This idea is extremely useful in understanding calculus. An element

The basic idea of the proof from now on is as follows. In order to construct a model of T1 T2 we construct models A j= T1 and B j= T2 and an isomorphism f : AjL ! BjL between the reducts of A and B to the language L, witnessing that AjL = BjL. We then use f to carry over interpretations of symbols in L1 n L from A to B , giving an expansion B of B to the language L1 L2 . Then, since B jL1 = A and B jL2 = B we get B j= T1 T2 . The remainder of the proof will be devoted to constructing such an A, B and f . A and B will be constructed as unions of elementary chains of An's and Bn 's while f will be the union of fn : An ,! Bn .

3. DIAGRAMS AND EMBEDDINGS

28

We begin with n = 0, the rst link in the elementary chain. Claim. There are models A0 j= T1 and B0 j= T2 with an elementary embedding f0 : A0 jL ,! B0 jL. Proof of Claim. Using the previous claim, let A0 j= T1 f sentences of L : T2 j= g We rst wish to show that Th(A0 jL)A0 T2 is satis able. Using the compactness theorem, it su ces to prove that if 2 Th(A0 jL)A0 then T2 f g is satis able. For such a let ca0 : : : can be all the constant symbols from LA0 n L which appear in . Let ' be the formula of L obtained by replacing each constant symbol cai by a new variable ui . We have A0 jL j= ' a0 : : : an ] and so A0 jL j= 9u0 : : : 9un ' By the de nition of A0 , it cannot happen that T2 j= :9u0 : : : 9un ' and so there is some model D for L2 such that D j= T2 and D j= 9u0 : : : 9un '. So there are elements d0 : : : dn of D such that D j= ' do : : : dn ]. Expand D to a model D for L2 LA0 , making sure to interpret each cai as di . Then D j= , and so D j= T2 f g. Let B0 j= Th(A0 jL)A0 T2 . Let B0 be the reduct of B0 to L2 clearly B0 j= T2 . Since B0 jL can be expanded to a model of Th(A0 jL)A0 , the Elementary Diagram Lemma gives an elementary embedding f0 : A0 jL ,! B0 jL and nishes the proof of the claim. The other links in the elementary chain are provided by the following result.
Claim. For each n elementary embedding

0 there are models An+1 j= T1 and Bn+1 j= T2 with an

such that

fn+1 : An+1 jL ,! Bn+1 jL

An An+1 Bn Bn+1 fn+1 extends fn and Bn range of fn+1 : A0 A1 An An+1

#f0

B0

#f1

B1

#fn

Bn

Bn+1

#fn+1

The proof of this claim will be discussed shortly. Assuming the claim, let S S S A = n2N An , B = n2N Bn and f = n2N fn . The Elementary Chain Theorem gives that A j= T1 and B j= T2 . The proof of the theorem is concluded by simply verifying that f : AjL ! BjL is an isomorphism. The proof of the claim is long and quite technical it would not be inappropriate to omit it on a rst reading. The proof, of course, must proceed by induction on

3. DIAGRAMS AND EMBEDDINGS

29

n. The case of a general n is no di erent from the case n = 0 which we state and prove in some detail. Claim. There are models A1 j= T1 and B1 j= T2 with an elementary embedding f1 : A1 jL ,! B1 jL such that A0 A1 , B0 B1 , f1 extends f0 and B0 range of f1. A0 A1
B0 B1 Proof of Claim. Let A+ be the expansion of A0 to the language L+ = L1 0 1 fca : a 2 A0 g formed by interpreting each ca as a;2 A0 A+ is just another notation 0 for (A0 )A0 . The elementary diagram of A+ is Th A+ A+ . Let B0 be the expansion 0 0 0 of B0 jL to the language L = L fca : a 2 A0 g fcb : b 2 B0 g formed by interpreting each ca as f0 (a) 2 B0 and each cb as b 2 B0 .
We wish to prove that Th A+ A+ ThB0 is satis able. By the compactness 0 ;0 theorem it su ces to prove that Th A+ A+ f g is satis able for each in ThB0 . 0 0 For such a sentence , let ca0 : : : cam cb0 : : : cbn be all those constant symbols occuring in but not in L. Let '(u0 : : : um w0 : : : wm ) be the formula of L obtained from by replacing each constant symbol cai by a new variable ui and each constant symbol cbi by a new variable wi . We have B0 j= so B0 jL j= ' f0 (a0 ) : : : f0 (am ) b0 : : : bn ] So B0 jL j= 9w0 : : : 9wn ' f0 (a0 ) : : : f0 (am )] Since f0 is an elementary embedding we have : A0 jL j= 9w0 : : : 9wn ' a0 : : : am ] Let '(w0 : : : wn ) be the formula of L+ obtained by replacing occurances of ui ^ 1 in '(u0 : : : um w0 : : : wn ) by cai then A+ j= 9w0 : : : 9wn '. So, of course, ^ 0 ; + A0 A+ j= 9w0 : : : 9wn ' ^ 0 and this means that there are d0 : : : dn in A+ = A0 such that 0 + ) + j= ' d0 : : : dn ]: (A0 A0 ^ ; + We can now expand A0 A+ to a model D by interpreting each cbi as di to obtain 0 ; D j= and so Th A+ A+ f g is satis able. 0 0 ; Let E j= Th A+ A+ ThB0 . By the elementary diagram lemma A+ is ele0 0 0 mentarily embedded into EjL+ . So there is a model A+ for L+ with A+ A+ and 1 1 1 0 1 an isomorphism g : A+ ! EjL+ . Using g we expand A+ to a model A01 isomorphic 10 1 1 to E. Let A1 denote A1 jL we have A1 j= ThB0 . We now wish to prove that Th (A1 )A1 Th B+ B+ is satis able, where B+ is 0 0 0 the common expansion of B0 and B0 to the language + L2 = L2 fca : a 2 A0 g fcb : b 2 B0 g:
; ;

#f0

#f1

3. DIAGRAMS AND EMBEDDINGS

30

By the compactness theorem, it su ces to show that ; Th B+ B+ f g 0 0

is satis able for each in Th (A1 )A1 . Let cx0 : : : cxn be all those constant symbols which occur in but are not in L . Let (u0 : : : un ) be the formula of L obtained from by replacing each cxi with a new variable ui . Since (A1 )A1 j= we have A1 j= x0 : : : xn ], and so A1 j= 9u0 : : : 9un : Also A1 j= ThB0 and ThB0 is a complete theory in the language L hence 9u0 : : : 9un is in ThB0 . Thus B0 j= 9u0 : : : 9un and so ; + B0 B+ j= 9u0 : : : 9un 0 and therefore there are b0 : : : bn in B+ = B0 such that 0 ; + B0 B+ j= b0 : : : bn]: 0
;

We can now expand B+ B+ to a model F by interpreting each cxi as bi then 0 0 ; F j= and Th B+ B+ f g is satis able. 0 0 ; Let G j= Th (A1 )A1 Th B+ B+ . By the elementary diagram lemma B+ is 0 0 0 + . So there is a model B+ for L+ with B+ B+ elementarily embedded into GjL2 1 2 0 1 and an isomorphism h : B+ ! GjL+ . Using h we expand B+ to a model B01 1 2 1 isomorphic to G. Let B1 denote B01 jL . Again by the elementary diagram lemma A1 is elementarily embedded into B1 . Let this be denoted by f1 : A1 ,! B1 : Let a 2 A0 we will show that f0 (a) = f1 (a). By de nition we have B0 j= (v0 = ca ) f0 (a)] and so B+ j= (v0 = ca ) f0 (a)]: Since B+ B+ , 0 0 1 B+ j= (v0 = ca ) f0 (a)] and so B1 j= (v0 = ca ) f0 (a)]: Now A+ j= (ca = v1 ) a] 1 0 and A+ A+ so A+ j= (ca = v1 ) a] so A1 j= (ca = v1 ) a]: Since f1 is elementary, 0 1 1 B1 j= (ca = v1 ) f1 (a)] so B1 j= (v0 = v1 ) f0 (a) f1 (a)] and so f0 (a) = f1(a). Thus f1 extends f0. Let b 2 B0 we will prove that b = f1 (a) for some a 2 A1 . By de nition we have: B0 j= (v0 = cb ) b] so B+ j= (v0 = cb ) b]. Since B+ B+ B+ j= (v0 = cb ) b] 0 0 1 1 so B1 j= (v0 = cb ) b]. On the other hand, since (9v1 )(v1 = cb ) is always satis ed, we have: A1 j= (9v1 )(v1 = cb ) so there is a 2 A1 such that A1 j= (v1 = cb ) a]: Since f1 is elementary, B1 j= (v1 = cb ) f1 (a)] so B1 j= (v0 = v1 ) b f1 (a)] so b = f1(a). Thus B0 range of f1 . We now let A1 be A+jL1 and let B1 be B+ jL2 . We get A0 A1 and B0 B1 1 1 and f1 : A1 jL ! B1 jL remains an elementary embedding. This completes the proof of the claim.
Exercise 16. The Robinson Consistency Theorem was originally stated as:

3. DIAGRAMS AND EMBEDDINGS

31

Let T1 and T2 be satis able theories in languages L1 and L2 respectively and let T T1 \T2 be a complete theory in the language L1 \ L2 . Then T1 T2 is satis able in the language L1 L2 . Show that this is essentially equivalent to our version in Theorem 12 by rst proving that this statement follows from Theorem 12 and then also proving that this statement implies Theorem 12. Of course, for this latter argument you are looking for a proof much shorter than our proof of Theorem 12 however it will help to use the rst claim of our proof in your own proof.
Theorem 13. (Craig Interpolation Theorem) Let ' and be sentences such that ' j= . Then there exists a sentence , called the interpolant, such that ' j= and j= and every relation, function or constant symbol occuring in also occurs in both ' and . Exercise 17. Show that the Craig Interpolation Theorem follows quickly from the Robinson Consistency Theorem. Also, use the Compactness Theorem to show that Theorem 12 follows quickly from Theorem 13.

CHAPTER 4

Model Completeness
The quanti er 8 is sometimes said to be the universal quanti er and the quanti er 9 to be the existential quanti er. A formula ' is said to be quanti er free whenever no quanti ers occur in '. A formula ' is said to be universal whenever it is of the form 8x0 : : : 8xk where is quanti er free. A formula ' is said to be existential whenever it is of the form 9x0 : : : 9xk where is quanti er free. A formula ' is said to be universal-existensial whenever it is of the form 8x0 : : : 8xk 9y0 : : : 9yk where is quanti er free. We extend these notions to theories T whenever each axiom of T has the property. Remark. Note that each quanti er free formula ' is trivially equialent to the existential formula 9vi ' where vi does not occur in '. Exercise 18. Let A and B be models for L with A B. Verify the following four statements: (i) A B i BA j= Th(AA ) i AA j= Th(BA ): (ii) A B i BA j= 4A i BA j= for each quanti er free of Th(AA ): (iii) A B i BA j= for each existential of Th(AA ): (iv) A B i AA j= for each universal of Th(BA ): Definition 28. A model A of a theory T is said to be existentially closed if whenever A B and B j= T , we have AA j= for each existential of Th(BA ): Remark. If A is existentially closed and A0 = A then A0 is also existentially closed. Definition 29. A theory T is said to be model complete whenever T 4A is complete in the language LA for each model A of T .
Theorem 14. ( A. Robinson ) Let T be a theory in the language L. The following are equivalent: (1) T is model complete, (2) T is existentially complete, i.e. each model of T is existentially closed. (3) for each formula '(v0 : : : vp ) of L there is a universal formula (v0 : : : vp ) such that T j= (8v0 : : : 8vp )(' $ ) (4) for all models A and B of T , A B implies A B. Proof. (1) ) (2): Let A j= T and B j= T with A B. Clearly AA j= 4A it is also easy to see that BA j= 4A. Now by (1), T 4A is complete and both AA and BA are models of this theory so they are elementarily equivalent.
32

4. MODEL COMPLETENESS

33

So let be any sentence of LA (existential or otherwise). If BA j= then AA j= and (2) follows. (2) ) (3): Lemma 4 shows that it su ces to prove it for formulas ' in prenex normal form. We do this by induction on the prenex rank of ' which is the number of alternations of quanti ers in '. The rst step is prenex rank 0. Where only universal quanti ers are present the result is trivial. The existential formula case is non-trivial it is the following claim: Claim. For each existential formula '(v0 : : : vp ) of L there is a universal formula (v0 : : : vp ) such that T j= (8v0 ) : : : (8vp )(' $ ) Proof of Claim. Add new constant symbols c0 : : : cp to L to form L = L fc0 : : : cp g and to form a sentence ' of L obtained by replacing each instance of vi in ' with the corresponding ci ' is an existential sentence. It su ces to prove that there is a universal sentence of L such that T j= ' $ . Let ; = funiversal sentences of L such that T j= ' ! g We hope to prove that there is some 2 ; such that T j= ! ' . Note, however, that any nite conjunction 1 ^ 2 ^ ^ n of sentences from ; is equivalent to a sentence in ; which is simply obtained from 1 ^ 2 ^ ^ n by moving all the quanti ers to the front. Thus it su ces to prove that there are nitely many sentences 1 2 : : : n from ; such that T j= 1 ^ 2 ^ ^ n!' : If no such nite set of sentences existed, then each would be satis able. By the compactness theorem, T ; able. Therefore it just su ces to prove that T ; j= ' .
T f 1

: : : ng

f:' g

f:' g

would be satis-

In order to prove that T ; j= ' , let A be any model of T ;. Form LA as usual but ensure that fc0 : : : cp g fca : a 2 Ag so that L LA . Let = T f' g 4A: We wish to show that is satis able. By the compactness theorem it su ces to consider T f' g where is a conjunction of nitely many sentences of 4A. Let be the formula obtained from by exchanging each constant symbol from LA n L occuring in for a new variable ua . So A j= 9ua0 : : : 9uam (ua0 : : : uam ): But then A is not a model of the universal sentence 8ua0 : : : 8uam : (ua0 : : : uam ). Recalling that A j= ;, we are forced to conclude that this universal sentence is not in ; and so not a consequence of T f' g. So T f' g f9ua0 : : : 9uam (ua0 : : : uam )g must be satis able, and any model of this can be expanded to a model of T f' g. Thus is satis able for the language LA .

4. MODEL COMPLETENESS

34

Let C j= . By the diagram lemma, there is a model B for L such that AjL B and B = CjL. Thus both AjL and B are models of T . Furthermore BA = C so that BA j= ' . We now use (2) to get that (AjL)A j= ' . But (AjL)A is just AA and so A j= ' . This means T ; j= ' and nishes the proof of the claim. We will now do the general cases for the proof of the induction on prenex rank. There are two cases, corresponding to the two methods available for increasing the number of alternations of quanti ers: (a) the addition of universal quanti ers (b) the addition of existential quanti ers. For the case (a), suppose '(v0 : : : vp ) is 8w0 : : : 8wm (v0 : : : vp w0 : : : wm ) and has prenex rank lower than ' so that we have by the inductive hypothesis that there is a quanti er free formula (v0 : : : vp w0 : : : wm x0 : : : xn ) with new variables x0 : : : xn such that T j= (8v0 : : : 8vp 8w0 : : : 8wm )( $ 8x0 : : : 8xn ) Therefore, case (a) is concluded by noticing that this gives us T j= (8v0 : : : 8vp )(8w0 : : : 8wm $ 8w0 : : : 8wm 8x0 : : : 8xn ): Exercise 19. Check this step using the de nition of satisfaction. For case (b), suppose '(v0 : : : vp ) is 9w0 : : : 9wn (v0 : : : vp w0 : : : wm ) and has prenex rank less than '. Here we will use the inductive hypothesis on : which of course also has prenex rank less than '. We obtain a quanti er free formula (v0 : : : vp w0 : : : wm x0 : : : xn ) with new variables x0 : : : xn such that T j= (8v0 : : : 8vp 8w0 : : : 8wm )(: $ 8x0 : : : 8xn ) So T j= (8v0 : : : 8vp )(8w0 : : : 8wm : $ 8w0 : : : 8wm 8x0 : : : 8xn ) And T j= (8v0 : : : 8vp)(9w0 : : : 9wm $ 9w0 : : : 9wm 9x0 : : : 9xn : ) Now 9w0 : : : 9wm 9x0 : : : 9xn : is an existential formula, so by the claim there is a universal formula such that T j= (8v0 : : : 8vp )(9w0 : : : 9wm 9x0 : : : 9xn : $ ): Hence T j= (8v0 : : : 8vp )(9w0 : : : 9wn $ ) which is the nal result which we needed. (3) ) (4) Let A j= T and B j= T with A B. Let ' be a formula of L and let a0 : : : ap be in A such that B j= ' a0 : : : ap ] Obtain a universal formula such that T j= (8v0 : : : 8vp )(' $ ) so B j= a0 : : : ap ] Since A B A j= a0 : : : ap ]

4. MODEL COMPLETENESS

35

and so A j= ' a0 : : : ap ] and A B. (4) ) (1): Let A j= T . We will show that T 4A is complete by showing that for any model B for LA , B j= T 4A implies B j= Th(AA ). Letting B be such a model, we note that BjL, the restriction of B to L, can be expanded to B, a model of 4A. So by the diagram lemma A is isomorphically embedded in BjL. Furthermore, by checking the proof of the diagram lemma we can ensure that there is an f : A ,! BjL such that for each a 2 A, f (a) is the interpretation of ca in B. (Recall that LA = L fca : a 2 Ag). Moreover, as in Exercise 13, there is a model D for L such that A D and an isomorphism g : D ! BjL with the property that for each a 2 A, g(a) is the interpretation of ca in B. Now let 2 Th(AA ), so that AA j= . Let '(u0 : : : uk ) be the formula of L obtained by replacing each occurance of cai in by the new variable ui . We have A j= ' a0 : : : ak ] Since A D we can use (4) to get A D and so we have D j= ' a0 : : : ak ]. With the isomorphism g we get that BjL j= ' g(a0 ) : : : g(ak )] and since g(ai ) is the interpretation of cai in B we have B j= . Thus B j= Th(AA ) and this proves (1). Example 9. We will see later that the theory ACF is model complete. But ACF is not complete because the characteristic of the algebraically closed feild can vary among models of ACF and the assertion that \I have characteristic p" can easily be expressed as a sentence of the language of ACF. Exercise 20. Suppose that T is a model complete theory in L and that either 1. any two models of T are isomorphically embedded into a third or 2. there is a model of T which is isomorphically embedded in any other. Then prove that T is complete. Example 10. Let N be the usual natural numbers and < the usual ordering. Let B = hN <i and A = hN n f0g <i be models for the language with one binary relation symbol <. ThA is, of course, complete, but it is not model complete because it is not existentially complete. In fact the model A is not existentially closed because B j= ThA and A B and BA j= (9v0 )(v0 < c1 ) where c1 is the constant symbol with interpretation 1. But AA does not satisfy this existential sentence.
Theorem 15. (Lindstrom's Test) Let T be a theory in a countable language L such that (1) all models of T are in nite, (2) the union of any chain of models of T is a model of T , and (3) T is -categorical for some in nite cardinal . Then T is model complete.

4. MODEL COMPLETENESS

36

Proof. W.L.O.G. we assume that T is satis able. We use conditions (1) and (2) to prove the following: Claim. T has existentially closed models of each in nite size . Proof of Claim. By the Lowenheim-Skolem Theorems we get A0 j= T with jA0 j = . We recursively construct a chain of models of T of size A0 A1 : : : An An+1 with the property that if B j= T and An+1 B and is an existential sentence of Th(BAn ), then (An+1 )An j= . Suppose An is already constructed we will construct An+1. Let n be a maximally large set of existential sentences of LAn such that for each nite 0 n there is a model C for LAn such that C j= 0 T 4An By compactness T n 4An has a model D and without loss of generosity An D. By the Downward Lowenheim-Skolem Theorem we get E such that An E, jEj = and E D. Let An+1 = EjL we will show that An+1 has the required properties. Since E D, E j= T 4An and so A An+1 (See Exercise 18). Let B j= T with An+1 B and be an existential sentence of Th(BAn ) we will show that (An+1 )An j= . Since n consists of existential sentences and D E (An+1 )An BAn we have (see Exercise 18) that BAn j= n . The maximal = property of n then forces to be in n because if 2 n then there must be some nite 0 for which there is no C such that C j= 0 f g T 4An but BAn n is such a C! Now since 2 n and E D j= n we must have E = (An+1 )An j= : Now let A be the union of the chain. By hypothesis A j= T . It is easy to check that jAj = . To check that A is existentially closed, let B j= T with A B and let be an existential sentence of ThBA . Since can involve only nitely many constant symbols, is a sentence of LAn for some n 2 N . Thus An+1 A B gives that (An+1 )An j= . Since is existential (see Exercise 18 again) we get that A j= . This completes the proof of the claim. We now claim that T is model complete using Theorem 14 by showing that every model A of T is existentially closed. There are three cases to consider: 1. jAj = 2. jAj > 3. jAj < where T is -categorical. Case (1). Let A be an existentially closed model of T of size . Then there is an isomorphism f : A ! A . Hence A is existentially closed. Case (2). Let be an existential sentence of LA and B j= T such that A B and BA j= . Let X = fa 2 A : ca occurs in g. By the Downward LowenheimSkolem Theorem we can nd A0 such that A0 A, X A0 and jA0 j = . Now by Case (1) A0 is existentially closed and we have A0 B and in LA0 so A0A0 j= . But since 2 Th(A0A0 ) and A0 A we have AA j= .

4. MODEL COMPLETENESS

37

Case (3). Let and B be as in case (2). By the Upward Lowenheim-Skolem Theorem we can nd A0 such that A A0 and jA0j = . By case (1) A0 is existentially closed. Claim. There is a model B0 such that A0 B0 and BA B0A . Assuming this claim, we have B0 j= T and B0A j= and by the fact that A0 is existentially closed we have A0A0 j= . Since A A0 we have AA j= . The following lemma implies the claim and completes the proof of the theorem. Lemma 8. Let A, B and A0 be models for L such that A B and A A0 . Then there is a model B0 for L such that A0 B0 and BA B0A . Proof. Let A, B, A0 and L be as above. Note that since A B we have

BA j= 4A and so AA BA . Let be a sentence from 4A0 . Let fdj : 0 j mg be the constant symbols from LA0 n LA appearing in . Obtain a quanti er free formula '(u0 : : : um) of LA by exchanging each dj in with a new variable ui . Since A0 0 j= we have A A0A j= 9u0 : : : 9um': Since A A0 we have AA A0A and so AA j= 9u0 : : : 9um ': Since AA BA , BA j= 9u0 : : : 9um ': Hence for some b0 : : : bm in B, BA j= ' b0 : : : bm ]. Expand BA to be a model BA for the w language LA fdj : 0 j mg by interpreting each dj as bj . Then BA j= and so Th(BA ) f g is satis able. This shows that ThBA is satis able for each nite subset 4A0 . By the Compactness Theorem there is a model C j= 4A0 ThBA . Using the Diagram Lemma for the language LA we obtain a model B0 for L such that A0A B0A and B0A = CjLA . Hence B0A j= ThBA and so B0A BA .
Exercise 21. Suppose A

A0 are models for L. Prove that for each sentence of LA , if 4A0 j= then 4A j= . Exercise 22. Prove that if T has a universal-existential set of axioms, then the union of a chain of models of T is also a model of T . Remark. The converse of this last exercise is also true it is called the ChangLos- Suszko Theorem. Theorem 16. The following theories are model complete: 1. dense linear orders without endpoints. (DLO) 2. algebraically closed elds. (ACF) Proof. (DLO): This theory has a universal existential set of axioms so that it is closed under unions of chains. It is @0 -categorical (by Exercise 11) so Lindstrom's test applies. (ACF): We rst prove that for any xed characteristic p, the theory of algebraically closed elds of characteristic p is model complete. The proof is similar to that for DLO, with @1 -categoricity (Lemma 7 ). Let A B be algebraically closed elds. They must have the same characteristic p. Therefore A B.
Corollary 4. Any true statement about the rationals involving only the usual ordering is also true about the reals.

4. MODEL COMPLETENESS

38

Proof. Let A = hQ <1 i and B = hR <2 i where <1 and <2 are the usual 1 2 orderings. The precise version of this corollary is: A B. This follows from Theorem 14 and Theorem 16 and the easy facts that A j= DLO, B j= DLO and A B. The reader will appreciate the power of these theorems by trying to prove A B directly, without using them.

Let be a nite system of polynomial equations and inequations in several variables with coe cients in the eld A. If has a solution in some eld extending A then has a solution in the algebraic closure of A. Proof. Let be the existential sentence of the language LA which asserts the fact that there is a solution of . Suppose has a solution in a eld B with A B. Then BA j= . So B0A j= where B0 is the algebraic closure of B. Let A0 be the algebraic closure of A. Since A B, we have A0 B0 . By Theorem 16, ACF is model complete, so A0 B0 . Hence A0A B0A and 0 j= . AA

Corollary 5. (Hilbert's Nullstellensatz)

CHAPTER 5

The Seventeenth Problem

We will give a complete proof later that RCF, the theory of real closed ordered elds, is model complete. However, by assuming this result now, we can give a solution to the seventeenth problem of the list of twenty-three problems of David Hilbert's famous address to the 1900 International Congress of Mathematicians in Paris. Corollary 6. (E. Artin) Let q(x1 : : : xn ) be a rational function with real coe cients, which is positive definite. i.e. q(a1 : : : an ) 0 for all a1 : : : an 2 R Then there are nitely many rational functions with real coe cients f1(x1 : : : xn ), : : : , fm (x1 : : : xn ) such that

q(x1 : : : xn ) =

m X j =1

(fj (x1 : : : xn ))2

We give a proof of this theorem after a sequence of lemmas. The rst lemma just uses calculus to prove the special case of the theorem in which q is a polynomial in only one variable. This result probably motivated the original question. Lemma 9. A positive de nite real polynomial is the sum of squares of real polynomials. Proof. We prove this by induction on the degree of the polynomial. Let p(x) 2 R x] with degree deg(p) 2 and p(x) 0 for all real x. Let p(a) = minfp(x) : x 2 Rg, so p(x) = (x ; a)q(x) + p(a) and p0 (a) = 0 for some polynomial q. But p0 (a) = (x ; a)q0 (x) + q(x)] x=a = q(a) so q(a) = 0 and q(x) = r(x)(x ; a) for some polynomial r(x). So p(x) = p(a) + (x ; a)2 r(x): For all real x we have (x ; a)2 r(x) = p(x) ; p(a) 0: Since r is continuous, r(x) 0 for all real x, and deg(r) = deg(p) ; 2. So, by P induction r(x) = n=1 (ri (x))2 where each ri (x) 2 R x]. i So p(x) = p(a) +
n X i=1 39

(x ; a)2 (ri (x))2 :

5. THE SEVENTEENTH PROBLEM

40

i.e. p(x) =

hp

p(a) +

i2

n X i=1

(x ; a)ri (x)]2 :

The following lemma shows why we deal with sums of rational functions rather than sums of polynomials. Lemma 10. x4 y2 + x2 y4 ; x2 y2 + 1 is positive de nite, but not the sum of squares of polynomials. Proof. Let the polynomial be p(x y). A little calculus shows that the mini26 mum value of p is 27 and con rms that p is positive de nite. Suppose

p(x y) =

l X i=1

(qi (x y))2

where qi (x y) are polynomials, each of which is the sum of terms of the form axm yn . First consider powers of x and the largest exponent m which can occur in any of the qi . Since no term of p contains x6 or higher powers of x, we see that we must have m 2. Considering powers of y similarly gives that each n 2. So each qi (x y) is of the form: ai x2 y2 + bi x2 y + ci xy2 + di x2 + ei y2 + fi xy + gi x + hi y + ki for some coe cients ai bi ci di ei fi gi hi and ki . Comparing coe cients of x4 y4 in p and the sum of the qi2 gives 0=
l X i=1

a2 i

so each ai = 0. Comparing the coe cients of x4 and y4 gives that each di = 0 = ei . Now comparing the coe cients of x2 and y2 gives that each gi = 0 = hi . Now comparing the coe cients of x2 y2 gives
;1 =

l X i=1

fi2

which is impossible.

The next lemma is easy but useful. Lemma 11. The reciprocal of a sum of squares is a sum of squares. Proof. For example A 2+ B 2 1 = A2 + B 2 = A2 + B 2 (A2 + B 2 )2 A2 + B 2 A2 + B 2 The following lemma is an algebraic result of E. Artin and O. Schreier, who invented the theory of real closed elds.

5. THE SEVENTEENTH PROBLEM

41

Lemma 12. Let A = hA + < A 0 1 i be an ordered eld such that each positive element of A is the sum of squares of elements of A. Let B be a eld containing the reduct of A to f+ 0 1g as a sub eld and such that zero is not the sum of nonzero squares in B. Let b 2 B n A be such that b is not the sum of squares of elements of B. Then there is an ordering <B extending <A on B such that b <B 0. Proof. It su ces to nd a set P B of \positive elements" of B such that (1) ;b 2 P (2) 0 2 P = (3) c2 2 P for each c 2 B (4) P is closed under + and (5) for any c 2 B n f0g either c 2 P or ;c 2 P . Once P has been obtained, we de ne <B as follows: c1 <B c2 i c2 ; c1 2 P: For each a 2 A, if 0 <A a then a is a sum of squares and so by (3) and (4) a 2 P . Thus <B extends < A . So that all that remains to do is to construct such a P . The rst approximation to P is P0 .

Let P0 = :

8 l <X

We claim that (1), (2), (3) P (4) holdP P0 . (1) and (3) are obvious. In and for order to verify (2), note that if m d2 b = li=1 c2 , then by the previous lemma i j =1 j about reciprocals of sums of squares, b would be a sum of squares. Now (4) holds by de nition of P0 , noting that c2 (;d2 b) = ;(ci dj )2 b and i j (;d2 b)(;d2 b) = (dj dk b)2 . j k We now construct larger and larger versions of P0 to take care of requirement (5). We do this in the following way. Suppose P0 P1 , P1 satis es (1), (2), (3) and (4), and c 2 P1 . We de ne P2 to be: = fp(;c) : p is a polynomial with coe cients in P1 g: It is easy to see that ;c 2 P2 , P1 P2 and that (1), (3) and (4) hold for P2 . To show that (2) holds for P2 we suppose that p(;c) = 0 and bring forth a contradiction. Considering even and odd exponents we obtain: p(x) = q(x2 ) + xr(x2 ) for some polynomials q and r with coe cients in P1 . If r(c2 ) = 0 then q(c2 ) = 0. But q(c2 ) is in P1 , which is a contradiction. On the other hand, if r(c2 ) 6= 0 then 0 = p(;c) = q(c2 ) ; cr(c2 ) which means that 2 c = q(c2 ) r(c2 ) r(12 ) c and since each of the factors on the right hand side is in P1 we get a contradiction. Now we need:

i=1

c2 ; i

m X j =1

d2 b : l m 2 N ci 2 B dj 2 B not all zero j

9 =

5. THE SEVENTEENTH PROBLEM

42

Lemma 13. Every ordered eld can be embedded as a submodel of a real closed ordered eld. Proof. It su ces to prove that for every ordered eld A there is an ordered eld B such that A B and for each natural number n 1, B j= n where n is the sentence in the language of eld theory which formally states: If p is a polynomial of degree at most n and w < y such that p(w) < 0 < p(y) then there is an x such that w < x < y and p(x) = 0. Consider the statement called IH(n): For any ordered eld E there is an ordered eld F such that E F and E j= n . IH(1) is true since any ordered eld E j= 1 . We will prove below that for each n, IH(n) implies IH(n + 1). Given our model A j= ORF, we will then be able to construct a chain of models:

A B1 B2 : : : Bn Bn+1 such that each Bn j= ORF f n g. Let B be the union of the chain. Since the theory ORF is preserved under unions of chains (see Exercise 22), B j= ORF. Furthermore, the nature of the sentences n allows us to conclude that for each n, B j= n and so B j= RCF. All that remains is to prove that for each n, IH(n) implies IH(n + 1). We rst make a claim: Claim. If E j= ORF f n g and p is a polynomial of degree at most n + 1 with coe cients from E and a < d are in E such that p(a) < 0 < p(d) then there is a model F such that E F, F j= ORF and there is b 2 such that a < b < d and p(b) = 0. Let us rst see how this claim helps us to prove that IH(n) implies IH(n+1). Let E j= ORF we will use the claim to build a model F such that E F and F j= n+1 . We rst construct a chain of models of ORF E = E0 E1 : : : Em Em+1 such that for each m and each polynomial p of degree at most n +1 with coe cients from Em and each pair of a, d of elements of Em such that p(a) < 0 < p(d) there is a b 2 Em+1 such that a < b < d and p(b) = 0. Suppose Em has been constructed we construct Em+1 as follows: let m be the set of all existential sentences of LEm of the form (9x)(ca < x ^ x < cd ^ p(x) = 0) where p is a polynomial of degree at most n + 1 and such that ca , cd and the coe cients of the polynomial p are constant symbols from LEm and (Em )Em j= p(ca ) < 0 ^ 0 < p(cd) We claim that ORF 4Em m is satis able.

5. THE SEVENTEENTH PROBLEM

f 1

43

Using the Compactness Theorem, it su ces to nd, for each nite subset : : : k g of m , a model C such that Em C and C j= ORF f 1 : : : k g: By IH(n), obtain a model F1 such that Em F1 and F1 j= ORF f n g. By the claim, obtain a model F2 such that F1 F2 and F2 j= ORF f 1 g. Again by IH(n), obtain F3 such that F2 F3 and F3 j= ORF f n g. Again by the claim, obtain F4 such that F3 F4 and F4 j= ORF f 2 g. Continue in this manner, getting models of ORF Em F1 : : : F2k with each F2j j= j . Since each j is existential, we get that F2k is a model of each j (see Exercise 18). Let D j= ORF 4Em m and then use the Diagram Lemma to get Em+1 such that Em Em+1 , Em+1 j= ORF and Em+1 j= m , thus satisfying the required property concerning polynomials from Em . Let F be the union of the chain. Since ORF is a universal-existential theory, F j= ORF (see Exercise 22) and F j= n+1 by construction. So IH(n + 1) is proved. We now nish the entire proof by proving the claim. Proof of Claim. If p(x) = 0 for some x in E such that a < x < d then we can let F = E. Otherwise, introduce a new element b to E where the place of b in the ordering is given by: b = supremumft 2 E : t < d and p(t) < 0g: Note that continuity-style considerations show that b 6= d. We now show that since E j= n the polynomial p cannot be factored in E. Suppose p(x) = q(x) s(x). The de nition of b allows us to nd a1 and d1 such that a1 < b < d1 and p(a1 ) < 0 < p(d1 ) and such that, other than possibly b, q and s have no roots in the interval of E between a1 and d1 . Now since p(a1 ) p(d1 ) < 0 either q(a1 ) q(d1 ) < 0 or s(a1 ) s(d1 ) < 0 and q and s each have degree n, forcing b to be an element of E. The fact that p is irreducible over E means that we can extend hE + 0 1i by quotients of polynomials in b of degree n to form a eld hF + 0 1i in the usual way. We leave the details to the reader. Note that the construction cannot force q(b) = 0 for any polynomial q(x) with coe cients from E of degree n. This is because we could take such a q(x) of lowest degree and divide p(x) by it to get p(x) = q(x) s(x) + r(x) where degree of r < degree of q. This means that r(x) = 0 constantly and so p could have been factored over E. Now we must expand hF + 0 1 i to an ordered eld F while preserving the order of E. We are aided in this by the fact that if q is a polynomial of degree at most n with coe cients from E then there are a1 and a2 in E such that a1 < b < a2 and q doesn't change sign between a1 and a2 this comes from the fact that E j= n .

5. THE SEVENTEENTH PROBLEM

44

Proof of the Corollary. Using Lemma 11 we see that it su ces to prove the corollary for a polynomial p(x1 : : : xn ) such that p(a1 : : : an ) 0 for all a1 : : : an 2 R. Let B = hR(x1 : : : xn ) + 0 1i be the eld of \rational functions". Note that B contains the reduct of R to f+ 0 1g as a sub eld, where R is de ned as in Example 3 as the usual real numbers. By Lemma 12, if p is not the sum of squares in B, then we can nd an ordering <B on B, extending the ordering on the reals, such that the expansion B0 of B is an ordered eld and p(x1 : : : xn ) <B 0. We now use Lemma 13 to embed B0 as a submodel of a real closed eld M, 0 M. B Let '(v1 : : : vn ) be the quanti er free formula which we informally write as p(v1 : : : vn ) < 0 where ' involves constant symbols cri for the real coe cients ri of p. Let be the formula of the language of eld theory, obtained from ' by substituting a new variable ui for each cri . We have B0 j= 9v1 : : : 9vn r1 : : : rk ] and so M j= 9v1 : : : 9vn r1 : : : rk ] Since RCF is model complete and R B0 M, Theorem 14 gives R M and so R j= 9v1 : : : 9vn r1 : : : rk ] i.e. there exist a1 : : : an in R such that p(a1 : : : an ) < 0.

Hilbert also asked: If the coe cients of a positive de nite rational function are rational numbers (i.e. it is an element of Q (x1 : : : xn )) is it in fact the sum of squares of elements of Q (x1 : : : xn )? The answer is \yes" and the proof is very similar. Let Q = hQ + : < 0 1i be the ordered eld of rationals as in Example 3. Lemma 12 holds for A = Q and B = hQ (x1 : : : xn ) + 0 1 i by Lemma 11 every positive rational number is the sum of squares since every positive integer is the sum of squares n = 1 + 1 + + 1. Exercise 23. Finish the answer to Hilbert's question by making any appropriate changes to the proof of the corollary. Hint: create an ordering on R(x1 : : : xn ) so that B0 and R are each submodels.

CHAPTER 6

Submodel Completeness
Definition 30. A theory T is said to admit elimination of quanti ers in L whenever for each formula '(v0 : : : vp ) of L there is a quanti er free formula (v0 : : : vp ) such that: T j= (8v0 : : : 8vp )('(v0 : : : vp ) $ (v0 : : : vp )) Remark. There is a ne point with regard to the above de nition. If ' is actually a sentence of L there are no free variables v0 : : : vp . So T j= ' $ for some quanti er free formula with no free variables. But if L has no constant symbols, there are no quanti er free formulas with no free variables. For this reason we assume that L has at least one constant symbol, or we restrict to those formulas ' with at least one free variable. This will become relevant in the proof of Theorem 17 for (2) ) (3). Exercise 24. If T admits elimination of quanti ers in L and L has no constant symbols, show that for each sentence of L there is a quanti er free formula (v0 ) such that T j= $ 8v0 $ 9v0 Definition 31. A theory T is said to be submodel complete whenever T 4A is complete in LA for each submodel A of a model of T . Exercise 25. Use Theorem 14 and the following theorem to nd four proofs that every submodel complete theory is model complete. Theorem 17. Let T be a theory of a language L. The following are equivalent: (1) T is submodel complete (2) If B and C are models of T and A is a submodel of both B and C, then every existential sentence which holds in BA also holds in CA . (3) T admits elimination of quanti ers (4) whenever A B, A C, B j= T and C j= T there is a model D such that both BA and CA are elementarily embedded in DA . Proof. (1) ) (2) Let B j= T and C j= T with A B and A C. Then BA j= T 4A and CA j= T 4A. So (1) and Lemma 6 give BA CA . Thus (2) is in fact proved for all sentences, not just existential ones. (2) ) (3) Lemma 4 shows that it su ces to prove (3) for formulas in prenex normal form. We do this by induction on the prenex rank of '. This claim is the rst step. Claim. For each existential formula '(v0 : : : vp ) of L there is a quanti er free formula '(v0 : : : vp ) such that T j= (8v0 : : : 8vp )(' $ )
45

6. SUBMODEL COMPLETENESS

46

= L fc0 : : : cp g and to form a sentence ' of L obtained by replacing each instance of vi in ' with the corresponding ci ' is an existential sentence. It su ces to prove that there is a quanti er free sentence of L such that T j= ' $ :
L

Proof of Claim. Add new constant symbols c0 : : : cp to L to form

Let S = f quanti er free sentences of L : T j= ' ! g: It su ces to nd some in S such that T j= ! ' . Since a nite conjunction of sentences of S is also in S , it su ces to nd 1 : : : n in S such that T j= 1 ^ ^ n!' : If no such nite subset f 1 : : : n g of S exists, then each
T

would be satis able. So it su ces to prove that T S j= ' . Let C j= T S with the intent of proving that C j= ' . Let A be the least submodel of C in the sense of the language L . That is, every element of A is the interpretation of a constant symbol from L fc0 : : : cp g or built from these using the functions of C. We can ensure that L LA . Let P = f of 4A : is a sentence of L g: We wish to show that T f' g P is satis able. By compactness, it su ces to consider T f' g where is a sentence in P . If this set is not satis able then T j= ' ! : so that by de nition of S we have : 2 S and hence C j= : . But this is impossible since A C means that CA j= 4A. Let B0 j= T f' g P . The interpretations of fc0 : : : cp g generate a submodel of B0 isomorphic to A. So there is a model B for L such that B = B0 and A B. In order to invoke (2) we use the restrictions AjL, BjL and CjL of A, B and C to the language L. We have BjL j= T , CjL j= T , AjL BjL and AjL CjL. ' is an existential sentence of L LA and since B0 j= ' we have (BjL)A j= ' . So by (2), (CjL)A j= ' and nally C j= ' which completes the proof of the claim. We now do the general cases for the proof of the induction on prenex rank. There are two cases, corresponding to the two methods available for increasing the number of alternations of quanti ers: (a) the addition of universal quanti ers (b) the addition of existential quanti ers. For case (a), suppose '(v0 : : : vp ) is 8w0 : : : 8wm (v0 : : : vp w0 : : : wm ) and has prenex rank lower than '. Then : also has prenex rank lower than ' and we can use the inductive hypothesis on : to obtain a quanti er free formula 1 (v0 : : : vp w0 : : : wm ) such that T j= (8v0 : : : 8vp )(8w0 : : : 8wm )(: $ 1 ) So T j= (8v0 : : : 8vp )(9w0 : : : 9wm : $ 9w0 : : : 9wm 1 )

f 1

: : : ng

f:' g

6. SUBMODEL COMPLETENESS

47

By the claim there is a quanti er free formula 2 (v0 : : : vp ) such that T j= (8v0 : : : 8vp )(9w0 : : : 9wm 1 $ 2 ) So T j= (8v0 : : : 8vp )(9w0 : : : 9wm : $ 2 ) So T j= (8v0 : : : 8vp )(8w0 : : : 8wm $ : 2 ) and so : 2 is the quanti er free formula equivalent to '. For case (b), suppose '(v0 : : : vp ) is 9w0 : : : 9wm (v0 : : : vp w0 : : : wm ) and has prenex rank lower than '. We use the inductive hypothesis on to obtain a quanti er free formula 1 (v0 : : : vp w0 : : : wm ) such that T j= (8v0 : : : 8vp )(8w0 : : : 8wm )( $ 1 ) So T j= (8v0 : : : 8vp )(9w0 : : : 9wm $ 9w0 : : : 9wm 1 ) By the claim there is a quanti er free formula 2 (v0 : : : vp ) such that T j= (8v0 : : : 8vp )(9w0 : : : 9wm 1 $ 2 ) So T j= (8v0 : : : 8vp )(9w0 : : : 9wm $ 2 ) and so 2 is the quanti er free formula equivalent to '. This completes the proof. (3) ) (4) Let A B, A C, B j= T and C j= T . Using the Elementary Diagram Lemma it will su ce to show that Th(BB ) Th(CC ) is satis able. Without loss of generosity, we can ensure that LB \ LC = LA . By the Robinson Consistency Theorem, it su ces to show that there is no sentence of LA such that both: Th(BB ) j= and Th(CC ) j= : Suppose is such a sentence, Th(BB ) j= . Let fca0 : : : cap g be the set of constant symbols from LA n L appearing in . Let '(v0 : : : vp ) be obtained from by exchanging each cai for a new variable ui . Let (v0 : : : vp ) be the quanti er free formula from (3): T j= (8v0 : : : 8vp )(' $ ) Let be the result of substituting cai for each ui in . is also quanti er free. Since BB j= , B j= ' a0 : : : ap ]. Since B j= T , B j= a0 : : : ap ] and so since BA j= . Since is quanti er free and AA BA we have AA j= AA CA we then get that CA j= . Hence C j= a0 : : : ap ] and then since C j= T we then get that C j= ' a0 : : : ap ]. But then this means that CA j= and so CC j= so is in Th(CC ) and we are done. (4) ) (1) Let B j= T and A B we show that T 4A is complete. Noting that BA j= T 4A , we see that it su ces by Lemma 6 to show that BA C0 for each 0 j= T 4A . C For each such C0 , by the Diagram Lemma, there is a model C for L such that A C and CA = C0 . Then C j= T so by (4) there is a D into which both BA and CA are elementarily embedded. In particular BA DA CA so we are done.
Example 11. (Chang and Keisler) Let T be the theory in the language L = fU V W R S g where U , V and W are

6. SUBMODEL COMPLETENESS

48

unary relation symbols and R and S are binary relation symbols having axioms which state that there are in nitely many things, that U V W is everything, that U , V and W are pairwise disjoint, that R is a one-to-one function from U onto V and that S is a one-to-one function from U V onto W . Exercise 26. Show that T above is complete and model complete but not submodel complete. Hints: For completeness, use the Los-Vaught test and for model completeness use Lindstrom's test. For submodel completeness use (2) of the theorem with B j= T and A B where a 2 A = fb 2 B : B j= W (v0 ) b]g along with the sentence (9v0 )(U (v0 ) ^ S (v0 ca )): Remark. We will prove that each of the following theories admits elimination of quanti ers: 1. dense linear orders with no end points (DLO) 2. algebraically closed elds (ACF) 3. real closed ordered elds (RCF) C. H. Langford proved elimination of quanti ers for DLO in 1924. The cases of ACF and RCF were more di cult and were done by A. Tarski. Thus, by Exercise 25, we will have model completeness of RCF which was promised at the beginning of Chapter 5. Exercise 27. Use part (4) of the previous theorem and the fact that RCF admits elimination of quanti ers to prove that RCF is complete another result originally due to A. Tarski. Hint: Show that Q of Example 3 can be isomorphically embedded into any real closed eld and then use (4) from Theorem 17. Exercise 28. Let T be the theory DLO in the language L = f< c1 c2 g where c1 and c2 are constant symbols. Use the fact that DLO admits elimination of quanti ers in its own language f<g to show that T is submodel complete. But, show also, that T is not complete. As an application of quanti er elimination of ACF we have the following: Corollary 7. (Tarski) The truth value of any algebraic statement about the complex numbers can be determined algebraically in a nite number of steps. Proof. Let C be the complex numbers in the language of eld theory L let be a sentence of LC . Then let A be the nite subset of C consisting of these elements of C (other than 0 or 1 ) which are mentioned in . Let ' be the formula of L formed by exchanging each ca for a new variable. Then ACF j= 8v0 : : : 8vp(' $ ) for some quanti er free . Hence C j= i C j= ' a0 : : : ap ] i C j= a0 : : : ap ] but checking this last statement amounts to evaluating nitely many polynomials in a0 : : : ap .
Remark. In fact Tarski's original proof actually gave an explicit method for nding the quanti er free formulas and this led, via the corollary above, to an effective decision proceedure for determining the truth of elementary algebraic statements about the reals or the complex numbers.

6. SUBMODEL COMPLETENESS

49

As an application of quanti er elimination of RCF we have: Corollary 8. (The Tarski-Seidenberg Theorem) The projection of a semi-algebraic set in Rn to Rm for m < n is also semialgebraic. The semi-algebraic sets of Rn are de ned to be all those subsets of Rn which can be obtained by repeatedly taking unions and intersections of sets of the form fhx1 : : : xn i 2 Rn : p(x1 : : : xn ) = 0g and fhx1 : : : xn i 2 Rn : q(x1 : : : xn ) < 0g where p and q are polynomials with real coe cients. Proof. We rst need two simple results which we state as exercises. Let L be the language of RCF augumented with a constant symbol for each element of the reals R. Let T be RCF considered as a theory in this language. Exercise 29. T admits elimination of quanti ers as a theory in the language L . Let R = hR + < 0 1i be the usual model of the reals then RR is a model for the language L . Exercise 30. A set X Rn is semi-algebraic i there is a quanti er free formula '(v1 : : : vn ) of L such that X = fhx1 : : : xn i : RR j= ' x1 : : : xn ]g: Now, in order to prove the corollary, let X Rn be semi-algebraic and let ' be its associated quanti er free formula. The projection Y of X into Rm is Y = fhx1 : : : xm i : 9wm+1 : : : 9wn hx1 : : : xm xm+1 : : : xn i 2 X g = fhx1 : : : xm i : RR j= 9vm+1 : : : 9vn ' x1 : : : xm ]g Since T admits elimination of quanti ers, there is a quanti er free formula of L such that T j= (8v1 : : : 8vm )(9vm+1 : : : 9vn ' $ ) So RR j= 9vm+1 : : : 9vn ' x1 : : : xm ] i RR j= x1 : : : xm ]: So Y = fhx1 : : : xm i : RR j= x1 : : : xm ]g and by the exercise, Y is semi-algebraic.

CHAPTER 7

Model Completions
Closely related to the notions of model completeness and submodel completeness is the idea of a model completion. Definition 32. Let T T be two theories in a language L. T is said to be a model completion of T whenever T 4A is satis able and complete in LA for each model A of T . Lemma 14. Let T be a theory in a language L. (1) If T is a model completion of T , then for each A j= T there is a B j= T such that A B. (2) If T is a model completion of T , then T is model complete. (3) If T is model complete, then it is a model completion of itself. (4) If T1 and T2 are both model completions of T , then T1 j= T2 and T2 j= T1 . Proof. (1) Easy. Just use the diagram lemma and the word \satis able" in the de nition of model completion. (2) Easier. (3) Easiest. (4) This needs a proof. Let A j= T2 . It will su ce to prove that A j= T1 . Let A0 = A. since A0 j= T and T1 is a model completion of T we obtain, from (1), a model A1 j= T1 such that A0 A1. Similarly, since A1 j= T and T2 is a model completion of T we obtain A2 j= T2 such that A1 A2 . Continuing in this manner we obtain a chain: A0 A1 A2 : : : An An+1 Let B be the union of the chain, fAn : n 2 N g. Each even A2n j= T2 . By (2) and (4) of Theorem 14 we get that for each n, A2n A2n+2 and by the Elementary Chain Theorem A0 B. Similarly A1 B. so A0 A1 and hence A j= T1 .
Remark. Part (4) of the above lemma shows that model completions are essentially unique. That is, if model completions T1 and T2 of T are closed theories in the sense of De nition 12 then T1 = T2 . Since there is no loss in assuming that model completions are closed theories, we speak of the model completion of a theory T . Lemma 15. If T T are theories for a language L such that for each A j= T there is a B j= T such that A B, then the following are equivalent: (1) T is the model completion of T . (2) For each A j= T , B j= T and C j= T such that A B and A C we have a model D such that both BA and CA are elementarily embedded into DA .
50

7. MODEL COMPLETIONS

51

(3) For each A j= T , B j= T and C j= T such that A B and A C we have a model D such that BA DA and CA is elementarily embedded into DA . (4) For each A j= T , B j= T and C j= T such that A B and A C we have a model D such that BA is isomorphically embedded into DA and C D. Proof. (1) ) (2) By the Elementary Diagram Lemma it su ces to prove that the union of the elementary diagrams of BA and CA , is satis able. By the Robinson Consistency Theorem it su ces to show that there is no sentence of LA such that ThBA j= and ThCA j= : . Indeed, if ThBA j= then BA j= . Now T 4A is a complete theory in LA and both BA j= T 4A and CA j= T 4A so BA CA . Therefore CA j= and hence is in ThCA . (2) ) (3) and (3) ) (4) easily follow from the de nitions. (4) ) (1) We rst show that T is model complete using Theorem 14 we show that T is existentially complete. Let A j= T we show that A is existentially closed. Let B j= T such that A B and let be an existential sentence of LA with BA j= our aim is to prove that AA j= . We invoke (4) with C = A to get a model D such that A D and an isomorphic embedding f : BA ! DA . Since is existential it is of the form 9v0 : : : 9vp ' for some quanti er free formula '(v0 : : : vp ) of LA . BA j= 9v0 : : : 9vp ' So for some b0 : : : bp in B we have BA j= ' b0 : : : bp ]: By Exercises 8 and 18 we have DA j= ' f (b0 ) : : : f (bp )] and so DA j= 9v0 : : : 9vp ': Now A D implies that AA DA so AA j= 9v0 : : : 9vp ': Hence AA j= and T is model complete. We now show that T is the model completion of T . Let A j= T by the hypothesis on T and T we have that T 4A is satis able. We show that T 4A is complete in LA by showing that for each B j= T and C j= T with A B and A C we have BA CA . Letting B and C be as above, we invoke (4) to obtain a model D such that BA is isomorphically embedded into DA and C D. C D gives that D j= T . The isomorphic embedding gives us a model E such that B E and DA = EA . So E j= T . Using model completeness of T and Theorem 14 we can conclude that B E. We have: BA EA DA CA and we are done. Let's compare the de nitions of model completion and submodel complete. Let T be the model completion of T . Then T will be submodel complete provided

7. MODEL COMPLETIONS

52

that every submodel of a model of T is a model of T . Unfortunately this is not always the case. However this seems like a promising approach to show submodel completeness (and hence elimination of quanti ers) of some theories T | we just need to show that T is the model completion of some theory T and T has the property that every submodel of a model of T is also a model of T . Since T T , it would be enough to show that every submodel of a model of T is again a model of T . And this is indeed the case whenever T is a universal theory, that is, whenever T has a set of axioms consisting of universal sentences. Our ultimate aim is to show that DLO, ACF and RCF are submodel complete. We will in fact show that these theories are the model completions of LOR, FEI and ORF respectively. See Example 5 to recall the axioms for these theories. Now LOR is a universal theory but FEI and ORF are not. The culprits are the existence axioms for inverses: 8x9y (x + y = 0) and 8x9y (y x = 1) In fact, a submodel A of a eld B is only a commutative ring, not necessarily a sub eld. Nevertheless, A generates a sub eld of B in a unique way. This motivates the following de nition. Definition 33. A theory T is said to be almost universal whenever A B, B j= T and A C, C j= T imply there are models D and E such that D j= T , A D B and E j= T , A E C and DA = EA . Example 12. LOR is almost universal since any universal theory T is almost universal | just let D = E = A and note A j= T . Example 13. FEI is almost universal | just let D and E be the sub elds of B and C, respectively, generated by A. The isomorphism DA = EA is the natural one obtained from the identity map on A. Example 14. ORF is almost universal | again just let D and E be the ordered sub elds of B and C, respectively, generated by A. The extension of the identity map on A to the isomorphism DA = EA is aided by the fact that the order placement of the inverse of an element a is completely determined by the order placement of a. Theorem 18. Let T and T be theories of the language L such that T is almost universal and T is the model completion of T . Then T is submodel complete. Proof. We show that condition (2) of Theorem 17 is satis ed. Let B and C be models of T with A a submodel of both B and C. We will show that, in fact, BA CA . Now T T so B j= T and C j= T . Since T is almost universal there are models D and E of T such that A D B, A E C and DA = EA . So BD j= T 4D and CE j= T 4E. Now BD is a model for the language LD whereas CE is a model for LE . We wish to obtain a model C0 for LD which \looks exactly like" CE . We just let C0 be C and in fact let C0 jLA = CE jLA . The interpretation of a constant symbol cd 2 LD n LA is the interpretation of ce 2 LE n LA in CE where the isomorphism DA = EA takes d to e.

7. MODEL COMPLETIONS

53

Now D j= T and since T is the model completion of T , T 4D is complete. The isomorphism DA = EA ensures that C0 j= T 4D. So BD C0 . Hence BD jLA C0 jLA that is, BA CA . The way to show that DLO, ACF and RCF admit elimination of quanti ers is now clear. We will use Theorem 17 and 18. This reduces to showing that DLO, ACF and RCF are the model completions of LOR, FEI and ORF respectively. To do this we will use Lemma 15, so we rst need to show that each pair of these theories satisfy the general hypothesis of Lemma 15: if A j= T then there is a B j= T such that A B. For the case T = LOR and T = DLO is easy every linear order can be enlarged to a dense linear order without endpoints by judiciously placing copies of the rationals into the linear order. The case T = FEI and T = ACF is just the well known fact that every eld has an algebraic closure. The case T = ORF and T = RCF is just Lemma 13. So all that remains of the quest to prove elimination of quanti ers for DLO, ACF and RCF is to verify condition (4) of Lemma 15 in each of these cases. We rephrase this condition slightly as: For each A j= T , B j= T and C j= T with A B and A C there is a D such that C D and an isomorphic embedding f : B ,! D such that f A is the identity on A. At this point the reader may already be able to verify this condition for one or more of the pairs T = LOR and T = DLO, T = FEI and T = ACF, or T = ORF and T = RCF. However the remainder of this chapter is devoted to a uniform method. Definition 34. Let L be a language and (v0 ) a set of formulas of L in the free variable v0 . A model A for L is said to realize (v0 ) whenever there is some a 2 A such that A j= ' a] for each '(v0 ) in (v0 ). Definition 35. The set of formulas (v0 ) in the free variable v0 , is said to be a type of the model A whenever (1) every nite subset of (v0 ) is realized by A (2) (v0 ) is maximal with respect to (1). Remark. Every set of formulas (v0 ) having property (1) of the de nition of type can be enlarged to also have property (2). Lemma 16. Suppose A is an in nite model for a language L. Let X A and let (v0 ) be a type of AX in the language LX . Then there is a B such that A B and BX realizes (v0 ). Proof. Let T = ThAA (c) where c is a new constant symbol and (c) = f'(c) : ' 2 (v0 )g and of course '(c) is '(v0 ) with c replacing v0 . By de nition of type, for each nite T 0 T , there is an expansion A0 of A such that A0 j= T 0 . The Elementary Diagram Lemma and the Compactness Theorem will complete the proof.

7. MODEL COMPLETIONS

54

Lemma 17. Suppose A is an in nite model for a language L. There is a model B for L such that A B and BA realizes each type of AA in the language LA . Proof. Let f (v0 ) : 2 I g enumerate all types of AA in the language LA . For each 2 I introduce a new constant symbol c and let (c ) = f'(c ) : ' 2 (v0 )g: Let = f (c ) : 2 I g. Let 0 be any nite subset. Claim. 0 ThAA is satis able for the language LA fc : 2 I g. Proof of Claim. Let 1 (v0 ) : : : n (v0 ) be nitely many types such that 0 1 (c0 ) 2 (c1 ) n (c n ): By Lemma 16 there is a model A1 such that A A1 and (A1 )A realizes 1 (v0 ). Using Lemma 16 repeatedly, we can obtain A A1 A2 An such that each (Aj )A realizes j (v0 ). Now A An so (An )A j= ThAA . It is easy to check that since each Aj An, An realizes each j (v0 ) and furthermore so does (An )A . So we can expand (An )A to the language LA fc 1 : : : c n g to satisfy 0 ThAA .

By the claim and the Compactness Theorem, there is a model C j= ThAA . By the Elementary Diagram Lemma, A is elementarily embedded into CjL, the restriction of C to the language L. Therefore there is a model B for L such that A B and BA = CjLA . It is now straightforward to check that BA realizes each type (v0 ).
Definition 36. A model A for L is said to be -saturated whenever we have that for each X A with jX j < , AX realizes each type of AX . Recall that + is de ned to be the cardinal number just larger than . So a model A will be + -saturated whenever we have that for each X A with jX j , AX realizes each type of AX . In particular, if B is any in nite set, A will be jB j+ saturated whenever we have that for each X A with jX j jB j, AX realizes each type of AX . Remark. A model A is said to be saturated whenever it is jAj-saturated, where jAj is the size of the universe of A. For example, hQ < i is saturated to prove this let X be a nite subset of Q and let (v0 ) be a type of hQ < iX . By Lemma 16 and the Downward Lowenheim-Skolem Theorem get a countable B such that hQ < iX BX and B realizes (v0 ). Use the hint for Exercise 11 to show that hQ < iX = BX and then note that this means that (v0 ) is realized in hQ < iX . Lemma 18. (R. Vaught) Suppose C is an in nite model for L and B is an in nite set. There is a jBj+ saturated model D such that C D.

7. MODEL COMPLETIONS

55

C = C 0 C1 C 2 Cn n2N such that for each n 2 N (Cn+1 )Cn realizes each type of (Cn )Cn . This comes immediately by repeatedly applying Lemma 17. Let D be the union of the chain the Elementary Chain Theorem assures us that C D and indeed each Cn D. Let X D with jX j jBj and let (v0 ) be a type of DX . If X Cn for some n, then (v0 ) is a type of (Cn )X since (Cn )X DX . Now (v0 ) can be enlarged to a type of (Cn )Cn which is realized in (Cn+1 )Cn and so (v0 ) is realized in (Cn+1 )Cn . Since (Cn+1 )Cn DCn , we can easily check that (v0 ) is realized in DCn . Since (v0 ) involves only constant symbols associated with X , we have that DX realizes (v0 ). We have almost proved that D is jBj+ -saturated, but not quite, because there is no guarantee that if X D = fCn : n 2 N g and jX j jBj, then X Cn for some n. However, there would be no problem if X was nite. The problem with in nite X is that the elementary chain is not long enough to catch X . The solution is to upgrade the notion of an elementary chain to include chains which are indexed by any well ordered set, not just the natural numbers. We sketch the appropriate generalization of the above argument from the case of hN < i to the case of an arbitrary well ordered set hI <i with least element 0. We construct an elementary chain of models C = C0 C ::: 2 I recursively as follows. At stage , suppose we have already constructed C for each 2 I with < . The union of the chain up to E = fC : 2 I and < g falls under the scope of an upgraded Elementary Chain Theorem (which is proved exactly as Theorem 4) and so C E for each 2 I with < . We now use Lemma 17 as before to get C such that E C and (C )E realizes each type of EE . As before, let D = fC : 2 I g be the union of the entire chain and by the upgraded Elementary Chain Theorem C D. Also as before, DX realizes each type of DX for each X D such that X C for some 2 I . But now we can complete the proof of the lemma by choosing a well ordered set hI <i large enough so that if X D = fC : 2 I g and jX j jBj then there is some 2 I such that X C . Such a well ordered set is well known to exist | for example, any ordinal with co nality > jBj.
1. A B and 2. there is some b 2 B such that no properly smaller submodel of B contains A fbg.
Definition 37. We say that B is a simple extension of A whenever

Proof. We build an elementary chain

7. MODEL COMPLETIONS

56

Theorem 19. (Blum's Test) Suppose T T are theories of a language L. Suppose further that: (1) T is an almost universal theory (2) every model of T can be extended to a model of T and (3) for each A j= T , B a submodel of a model of T , C j= T such that C is jBj+ saturated, A C and B is a simple extension of A, there is an isomorphic embedding f : B ! C such that f A is the identity on A. Then: (4) for each A j= T , B j= T and C j= T with A B and A C there is a model D such that C D and an isomorphic embedding f : B ! D such that f A is the identity on A, (5) T is the model completion of T and (6) T admits elimination of quanti ers. Proof. Because of Lemma 15, Theorem 17 and Theorem 18, (5) and (6) follow from (1), (2) and (4). We will therefore only need to prove (4). Let A, B and C be as in (4). Using Lemma 18 we obtain a jBj+ -saturated model D such that C D. We wish to nd an isomorphic embedding f : B ,! D which is the identity on A. Let f be a maximal element of fg : for some B0 B g is an isomorphic embedding from B0 into D such that g A is the identity on Ag in the sense that no other such g properly extends f . We will prove that this f satis es condition (4). The proof is by contradiction suppose f : B0 ,! D and B0 6= B. Claim. There is a model G such that B0 G B, G j= T and a model H such that G H and an isomorphism j : H ! D such that j B0 = f . Proof of Claim. From the isomorphic embedding f : B0 ,! D we get a model E such that B0 E and an isomorphism e : E ! D such that e B0 = f . We have B0 E and B0 B with both E j= T and B j= T . By condition (1) there are models F and G of T such that B0 F E, B0 G B and FB0 = GB0 . This gives an isomorphic embedding g : G ,! E such that g B0 is the identity map. From g we get a model H such that G H and an isomorphism h : H ! E such that h G = g. Now let j = e h. Thus j is an isomorphism from H to D such that j B0 = e h B0 = e g B0 = e B0 = f which nishes the proof of the claim.

Once we have the claim, there are two cases: G 6= B and G = B. For the case G 6= B, there must be some b 2 B n G and we use b to form the simple extension B00 of G by b. Now use condition (3) on G, B00 and H to obtain an embedding k : B00 ,! H which is the identity map on G. Now j k extends f , which is a contradiction. For the case G = B, the function j G extends f and gives the contradiction.

7. MODEL COMPLETIONS

57

The following lemma completes the proofs that each of the theories DLO, ACF and RCF admit elimination of quanti ers. Lemma 19. Each of the following three pairs of theories T and T satisfy condition (3) of Blum's Test. (1) T = LOR theory of linear orderings. T = DLO, theory of dense linear orderings without endpoints. (2) T = FEI, theory of elds. T = ACF, theory of algebraically closed elds. (3) T = ORF, theory of ordered elds. T = RCF, theory of real closed elds. Proof of (1). Let A and B be linear orders, with B = A fbg and A B. Let C be a jBj+ -saturated dense linear order without endpoints with A C. We wish to nd an isomorphic embedding f : B ! C which is the identity on A. Consider a type of CA containing the following formulas: ca < v0 for each a 2 A such that a < b v0 < ca for each a 2 A such that b < a Since C is a dense linear order without endpoints each nite subset of the type can be realized in CA . Saturation now gives some t 2 C realizing this type. We set f (b) = t and we are nished. Proof of (2). Let A be a eld and B a simple extension of A witnessed by b such that B is a submodel of a eld (a commutative ring). Let C be a jBj+ -saturated algebraically closed eld such that A C. We wish to nd an isomorphic embedding f : B ! C which is the identity on A. There are two cases: (I) b is algebraic over A, (II) b is transcedental over A. Case(I). Let p be a polynomial with coe cients from A such that p(b) = 0 but b is not the root of any such polynomial of lower degree. Since C is algebraically closed there is a t 2 C such that p(t) = 0. We extend the identity map f on A to make f (b) = t. We extend f to the rest of B by letting f (r(b)) = r(t) for any polynomial r with coe cients from A. It is straightforward to show that f is still a well-de ned isomorphic embedding. las:
Case (II). Let us consider a type of CA containing the following set of formuf:(p(v0 ) = 0)g

where p is a polynomial with coe cients in fca : a 2 Ag. Since C is algebraically closed, it is in nite and hence each nite subset is realized in CA . Saturation will now give some t 2 C such that t realizes the type. We set f (b) = t. Since t is transcedental over A, the extension of f to all of B comes easily from the fact that every element of B n A is the value at b of some polynomial function with coe cients from A.

7. MODEL COMPLETIONS

58

Proof of (3). Let A be an ordered eld and B be a simple extension of A witnessed by b such that B is a submodel of an ordered eld (an ordered commutative ring). Let C be a jBj+ -saturated real closed eld such that A C. We wish to nd an isomorphic embedding f : B ! C which is the identity on A. There are two cases: (I) b is algebraic over A. (II) b is transcedental over A. Case (I). Since b is algebraic over A we have a polynomial p with coe cients in A such that p(b) = 0. All other elements of the universe of the simple extension B are of the form q(b) where q is a polynomial with coe cients in A. Before beginning the main part of the proof we need some algebraic facts. Claim. Let D be a real closed ordered eld and q(x) be a polynomial over D of degree n. Then for any e 2 D we have: n X q (m) (e) m q(x) = m! (x ; e) m=0

where q(m) stands for the polynomial which is the m-th derivative of q. Proof of Claim. This is Taylor's Theorem from Calculus unfortunately we cannot use Calculus to prove it because we are in D, not necessarily the reals R. However the reader can check that the Binomial Theorem gives the identity for the special cases of q(x) = xn and that these special cases readily give the full result.

Claim. Let D be a real closed ordered eld and q(x) a polynomial over D with e 2 D and q(e) = 0. If there is an a < e such that q(x) > 0 for all a < x < e then q0 (e) 0. If there is an a > e such that q(x) > 0 for all e < x < a then q0 (e) 0. Here q0 is the rst derivative of q. Proof of Claim. From the previous claim we get ! n q(x) ; q(e) = q0 (e) + (x ; e) X q(m) (e) (x ; e)m;2 x;e m=2 m!

for any x 6= e in D. By choosing x close enough to e we can ensure that the entire right hand side has the same sign as q0 (e). A proof by contradiction now follows readily.
Claim. Let D be a real closed ordered eld and q(x) be a polynomial over D with e 2 D and q(e) = 0. If w and z are in D such that w < e < z and q(w) q(z ) > 0 then there is a d in D such that w < d < z and q0 (d) = 0.

7. MODEL COMPLETIONS

59

Proof of Claim. Without loss of generosity q(w) > 0 and q(z ) > 0. Since q has only nitely many roots, we can pick d1 to be the least x such that w < x e and q(x) = 0. Since q(x) 6= 0 for all w < x < d1 , the Intermediate Value Property of Real Closed Ordered Fields shows that q cannot change sign here and so q(x) > 0 for all w < x < d1 . By the previous claim, q0 (d1 ) 0. A similar argument with z shows that there is a d2 such that e d2 < z and q0 (d2 ) 0. If d1 = e = d2 take d = e. If d1 < d2 the Intermediate Value Property gives a d with the required properties. Claim. Let D be a real closed ordered eld with an ordered eld E D. Let f : E ! C be an isomorphic embedding into a real closed ordered eld. Let q be a polynomial with coe cients in E such that fx 2 D : q0 (x) = 0g E. Let d 2 D n E be such that q(d) = 0 but d is not a root of a polynomial with coe cients from E which has lower degree. Then f can be extended over the sub eld of D generated by E fdg. Proof of Claim. Since the nitely many roots of q0 from D actually lie in E, we can get e1 and e2 in E such that e1 < d < e2 and q0(x) 6= 0 for all x in D such that e1 < x < e2 . Furthermore for all x in E we have q(x) 6= 0. We can now apply the previous claim to get that q(w) q(z ) < 0 for all w and z in E such that e1 < w < d < z < e2 . We now move to the real closed ordered eld C and the isomorphic embedding f . For each w and z in E such that e1 < w < d < z < e2 we have f (w) < f (z ) and q(f (w)) q(f (z )) < 0. By the Intermediate Value property of C we get, for each such w and z , a y 2 C such that f (w) < y < f (z ) and q(y) = 0. Since q has only nitely many roots there is some t 2 C such that q(t) = 0, f (w) < t for all e1 < w < d and t < f (z ) for all d < z < e2 . We now extend f by letting f (d) = t and f (r(d)) = r(t) for any polynomial r with coe cients from E. It is straightforward to check that the extension is a well-de ned isomorphic embedding of the simple extension of E by d into C. We

use the fact that ORF is almost universal to extend the isomorphic embedding to all of the sub eld of D generated by E fdg, since we can rephrase the de nition of almost universal as follows: Whenever C j= T , D j= T , E0 D and f : E0 ,! C is an isomorphic embedding, then there is a model E00 j= T such that E0 E00 D and f extends over E00 .

It is now time for the main part of the proof of this case. Using Lemma 13, let D be a real closed ordered eld with B D. We have a polynomial p with coe cients from A such that p(b) = 0. By induction on the degree of p, we can show that there is a sequence of elements d0 : : : dm = b of elements of D, a sequence of sub elds of D: A = E0 E1 : : : Em+1 with each dj 2 Ej+1 n Ej and corresponding isomorphic embeddings fj : Ej ! C

7. MODEL COMPLETIONS

60

coming from the previous claim and having the property that f0 is the identity and fj+1 extends fj . In this way we extend the identity map f0 : A0 ,! C until we reach fm+1 : Em+1 ,! C. We then note that since b 2 Em+1 we have B Em+1 and we are nished.
Case (II). Let us consider a type of CA containing the following formulas: ca < v0 for all a 2 A with a < b

CA .

v0 < ca for all a 2 A with b < a :(p(v0 ) = 0) for all polynomials p with coe cients in fca : a 2 Ag Since each interval of C is in nite, each nite subset of this type is realized by

Saturation now gives t 2 C which realizes this type. We put f (b) = t. We can now extend f on the rest of B n A, since each such element is the value at b of a polynomial function with coe cients from A.

Question. Suppose K is the reduct of a real closed ordered eld to the language p of eld theory. Can you show that K ;1] is algebraically closed using Model Theory and the Fundamental Theorem of Algebra?

Bibliography
1. J. Barwise (ed.), Handbook of Mathematical Logic, North Holland, 1997. 2. J. L. Bell and A. B. Slomson, Models and Ultraproducts, North Holland, 1969. 3. Felix E. Browder (ed.), Mathematical developments arising from Hilbert's problems, Proceedings of Symposia in Pure Mathematics, vol. XXVIII - Parts 1 and II, AMS, 1976. 4. S. Buechler, Essential Stability Theory, Perspectives in Mathematical Logic, Springer, 1996. 5. C. C. Chang and H. J. Keisler, Model Theory, second ed., North Holland, 1977. 6. G. Cherlin, Model Theoretic Algebra, Selected Topics, LNM521, Springer-Verlag, 1976. 7. N. Jacobson, Basic Algebra I, W. H. Freeman, 1985. 8. H. J. Keisler, Foundations of In nitesimal Calculus, Prindle, Weber and Schmidt, 1976. 9. M. D. Morley (ed.), Studies in Model Theory, MAA Studies in Mathematics, MAA, 1973. 10. A. Robinson, Introduction to Model Theory and the Metamathematics of Algebra, second ed., North-Holland, 1965. 11. , Non Standard Analysis, North Holland, 1966. 12. G. E. Sacks, Saturated Model Theory, W. A. Benjamin Inc., 1972. 13. D. H. Saracino and V. B. Weispfenning (eds.), Model theory and algebra: A Memorial Tribute to Abraham Robinson, LNM498, Springer Verlag, 1975. 14. S. Shelah, Classi cation Theory and the Numbers of Nonisomorphic models, second ed., North Holland, 1990. 15. J. R. Shoen eld, Mathematical Logic, Addison-Wesley, 1967. 16. A. Tarski, A Descision Method for Elementary Algebra and Geometry, second ed., The Rand Corporation, 1957.

61

Index
A B, 18 A = fAn : n 2 Ng, 19 ThA, 14 ThAA , 27 j=, 14 , 18 A jL , 17 , 17 t x0 : : : xq ], 7 hC + 0 1i, 13 hQ < + 0 1i, 13 < hR < + 0 1i, 13 hN + < 0 1i, 18 A. Robinson, 35 ACF, 14 submodel complete, 57 algebraically closed elds axioms,theory of, 14 almost universal, 57, 64 axioms, 14
0 0

Blum's Test, 60 bound variable, 6 categorical -categorical theory, 24 chain of models, 19 elementary, 19 Compactness Theorem, 17 complete theory, 24 Completeness Theorem, 17 complex, 13 Craig Interpolation Theorem, 33 dense linear orders without endpoints axioms,theory of, 14 diagram lemmas, 27 DLO, 14 submodel complete, 57 elementarily embedded model, 27 elementarily equivalent models, 18 Elementary Chain Theorem, 19 elementary diagram, 27 elementary extension, 18 elementary submodel, 18
62

elimination of quanti ers, 49 existentially closed, 35 expansion language, 17 model, 17 FEI, 14 almost universal, 57 elds axioms,theory of, 14 formula, 5 free variable, 6 isomorphic models, 18 isomorphically embedded model, 27 language, 6 Leibniz Principle, 29 Lindstrom's Test, 38 linear orders axioms,theory of, 14 LOR, 14 almost universal, 57 Los-Vaught Test, 24 Lowenheim-Skolem Theorems Downward, 23 Upward, 23 model, 6 satis es, 7 model complete theory, 35 model completion, 55 submodel complete, 57 Number Theory, 18 number theory non-standard models, 18 ordered eld, 45 ordered elds axioms,theory of, 14 ORF, 14 almost universal, 57 prenex normal form, 10 rational numbers, 13 RCF, 14 submodel complete, 57

INDEX

63

real closed ordered eld, 45 real closed ordered elds axioms, theory, 15 axioms,theory of, 14 Intermediate Value Property, 15 real numbers, 13 realize, 58 reduction language, 17 model, A jL , 17 Robinson Consistency Theorem, 30 satisfaction A j= , 14 saturated -saturated model, 59 sentence, 10 simple extension, 60 subformula, 5 submodel, 18 submodel complete, 49 submodel complete theory, 49 T. Skolem, 18 Tarski's Elementary Chain Theorem, 19 Tarski-Vaught Condition, 21 term, 5 theory, 14 almost universal, 57 model completion, 55 theory of A, 14 type, 58 variable, 5
0 0

Greens functions for planarly layered media (continued)

Massachusetts Institute of Technology 6.635 lecture notes

We shall here continue the treatment of multilayered media Greens functions, starting from the TE/TM decomposition we have presented in the previous document. For the sake of illustration, let us consider a one layer medium with a reection and transmission region, as shown in Fig. 1.
z r J( ) #0: ( 0 , 0 ) x #1: ( 1 , 1 ) d #2: ( 0 , 0 )

z
PSfrag replacements

Figure 1: Geometry of the problem.

. We have shown before that the Greens functions in various layers are expressed as:
G
where = 0, z < z :
r r Ke (k0z ) = e(k0z ) eiK + RT E e(k0z ) eik ,

i 8 2

dk

1 r r Ke (k z )(k0z ) eiK + Kh (k z )h(k0z ) eiK , e k0z

(1)

(2a) , (2b)

Kh (k0z ) = h(k0z ) e =1 :

r iK

+R

TM

h(k0z ) e

r ik

r r Ke (k1z ) = A e(k1z ) eik1 + B e(k1z ) eiK1 , r r Kh (k1z ) = C h(k1z ) eik1 + D h(k1z ) eiK1 ,

(2c) (2d)

=2 :

r Ke (k2z ) = T T E e(k0z ) eiK ,

(2e) . (2f)

Kh (k2z ) = T T M h(k0z ) e

r iK

Section 1. Transmission line analogy for multilayered media

By satisfying the boundary conditions at z = 0 and z = d, we obtain the following system (for TE waves, TM waves can be solved similarly):

1 + RT E =A + B k1 k0z (1 + RT E ) = z (A B) , 0 1 A eik1z d + Beik1z d = T T E eik0z d , k1z k0 (Aeik1z d Beik1z d ) = z T T E eik0z d . 1 0 Upon solving, we obtain:

(3a) (3b) (3c) (3d)

RT E = TTE =

1 e2ik1z d RT E , TE TE 1 + R01 R10 e2ik1z d 01 (1 + pT E ) (1 01 4ei(k1z k0z )d . TE TE + pT E ) (1 + R01 R10 e2ik1z d ) 10

(4a) (4b)

RT M = TTM =

1 e2ik1z d RT M , TM TM 1 + R01 R10 e2ik1z d 01 4ei(k1z k0z )d . TM TM (1 + pT M ) (1 + pT M ) (1 + R01 R10 e2ik1z d ) 01 10

(5a) (5b)

A= B= C=

TE TE 1 pT E 1 + R01 R10 10 e2ik1z d , TE TE 2 1 + R01 R10 e2ik1z d TE 1 + R01 , TE TE 1 + R01 R10 e2ik1z d

(6a) (6b) (6c) (6d)

TM 1 k 0 2R10 e2ik1z d , TM TM 0 k1 (1 + pT M ) (1 + R01 R10 e2ik1z d ) 01 1 k 0 2 D= . TM TM 0 k1 (1 + pT M ) (1 + R01 R10 e2ik1z d ) 01

Transmission line analogy for multilayered media

Let us consider a plane wave incident from region 0, with its plane of incidence parallel to the (xz) plane. The medium it is incident upon is multilayered. All elds vectors are independent on y, so that y = 0 in Maxwells equations. Thus, we can decompose the electromagnetic eld into TE/TM components. We get, in region (for TE waves):

E H H

y x

=[A eikz z + B eikz z ] eikx x , kz = [A eikz z B eikz z ] eikx x , kx [A eikz z B eikz z ] eikx x . =

(7a) (7b) (7c)

Following standard notation in transmission line theory, we shall use here the j notation!! From 6.630, we know that a transmission line is characterized by its length d, characteristic
(p) PSfrag replacements wavenumber kz , as dened in Fig. 2. impedance Zc and
(p)

I1

I2

(p)

V1

(p)

[Zc , kz ]

(p)

V2

(p)

Figure 2: Transmission line circuit.

In a source-free region, the transmission line equations are written as (p) (p) V = jkz Zc I (p) , z (p) 1 I = jkz (p) V (p) , z Zc where p refers to the polarization (TE or TM), and
T Zc E = T Zc M

(8a) (8b)

, kz kz . =

(9a) (9b)

The solution to Eqs. (8) is:

V (p) =A(p) eikz z + B (p) eikz z , 1 I (p) = (p) [A(p) eikz z B (p) eikz z ] , Zc

(10a) (10b)

As it can be seen, there is a direct analogy between the voltage/current in a transmission line and the components of the electric and magnetic elds. For example, referring back to Eq. (7), we write:

Section 1. Transmission line analogy for multilayered media

E H

y x

=V T E (z)eikx x , = I T E (z)eikx x .

(11a) (11b)

Therefore, the problem of computing the elds in multilayered media in the spectral domain comes down to determining the voltage/current in equivalent transmission line network. The analogy is illustrated in Fig. 3.
$

Figure 3: Transmission line analogy for horizontal electric source and vertical magnetic source. Other cases are obtained by duality.

The treatment of the source will not be demonstrated here and we shall just state the nal results (details can be found in the literature). Thus, depending on the source type and orientation, dierent generators will have to be placed in the transmission line network. The various cases are (magnetic sources can be obtained by duality): Horizontal electric source Vertical electric source The algorithm is therefore as follows: 1. Write the eld components in terms of voltage and current. Locate the source and observation point. 2. Compute the equivalent transmission line network. Locate the source (type and position) and observation. 3. Starting from the source replace all the layers above and below the source by equivalent impedances (see Fig. 4). To do this, start with the extreme boundary conditions and propagate back to the source using: current generator voltage generator value: 1/2 value: 1/2

 #

 #

 #

 

 

%& %      !"      

Z up

PSfrag replacements Z down

Figure 4: Equivalent upper and lower impedance.

(p) Zin (p) = Zc

ZL + jZc tan(kz d) Zc + jZL tan(kz d)

(p) (p)

(p)

(p)

(12)

4. Using standard circuit theory, compute V (p) and I (p) at the upper and lower limits. 5. Propagate V (p) and I (p) until the observation point using V2 I2 = cos(kz d)
(p) j/Zc sin(kz d)

jZc sin(kz d) cos(kz d)

(p)

V1 I1

(13)

6. Get the elds in the spectral domain.

Coming back to the space domain: Sommerfeld integral

In the rest of these notes we come back to the i notation!!

To come back to the space domain, we need to evaluate the inverse Fourier transform. A typical integral we have to perform is: f () = r 1 (2)2 dkx dky f (kx , ky ) eikx x eiky y . (14)

By symmetry of the problem (x and y axis are equivalent), we can make a change of variables and integrate one integral analytically. The proper change of variables is the following: kx =k cos k , ky =k sin k , We can transform the exponential part as: eikx x eiky y = ei[k cos k cos +k sin k sin ] = eik cos(k ) , so that f () = r 1 (2)2
2

x = cos , y = sin .

(15) (16)

(17)

dk
0 0

dk k f (k , k ) eik cos(k ) .

(18)

Section 2. Coming back to the space domain: Sommerfeld integral

By rotational symmetry, f (k , k ) = f (k ). In addition, we can expand the exponential part using the following identity:

ik cos(k )

=e

i cos

=
m=

(i)m Jm () eim ,

(19)

(where and have just been dened to simplify the notation in the identity and have no connection to physical parameters). As we can see, the exponential function is the only term depending on k . Performing the integration, we get: 0 2 if m = 0 , (20) eim = 2 if m = 0 . Therefore, we end up with
1 dk k f (k ) J0 (k ) , 2 0 which is known as a Sommerfeld integral.

f () = r

(21)

Note however that in Eq. (7b), the kernel is function of kx also, which adds a k cos k term in the integral, so that the integral in k cannot be performed exactly as shown above. However, the generalization to this case is straightforward. Upon performing the same expansion of the Bessel function, we see that this time the non-vanishing contribution will come from the m = 1 term (or 1 depending on how the integral is written). Without further details, we generalize the denition of Sommerfeld integral to the n th order as: n+1 (22) dk Jn (k )k f (k ) , S n [f ] =
0

and the transformation from spectral to spatial can be summarized as follows (where A is a function of k only): spectral domain G=A G = ikx A G = iky A
2 G = kx A 2 G = ky A

space domain G = S0 [A] G = cos S1 [A] G = sin S1 [A] G= cos 2 2 S1 [A] cos2 S0 [k A] cos 2 2 S1 [A] sin2 S0 [k A]

G= G=

G = kx ky A

sin 2 1 2 S1 [A] sin 2 S0 [k A] 2

Numerical evaluation of Sommerfeld integrals

Sommerfeld integrals are dicult (but not impossible) to evaluate for two reasons: 1. The spectral kernel can present poles (and in general does). 2. They have an oscillatory tail. Fortunately these two problems appear in essentially two distinct regions.

3.1

Poles of the Greens function

It can be shown (out of the scope here) that the poles of the Greens functions are associated with propagating waves (e.g. in single-mode regime, the Greens function has only one pole). Therefore, we must have at least one value of (the index of the region) where k z is real, i.e. k
z

2 2 kl k =

2 2 k0 k

(23)

must be real. Since this needs to happen in at least one layer, it yields the condition: k < k0 max(
l

),

(24)

which puts an upper limit for the location of the poles. Although it can also be shown that poles have to correspond to k > k0 , we do not need this constraint here and we can limit ourselves to the interval [0, k0 maxl ( )]. In the lossless situation, the poles lie on the real axis, which renders the integral impossible to evaluate as is. We therefore need to deform the contour in the complex k -plane: At innity, convergence is ensured by Sommerfelds radiation condition. On [0, k0 maxl ( )], we perform the integration over an ellipse, as shown in Fig. 5.

(k )

a b
PSfrag replacements

(k )

Figure 5: Contour deformation in the complex k -plane.

3.2

Oscillatory tail

3.2

Oscillatory tail

The problem in this case is not the divergence of the integral (it does not diverge) bu the convergence, which is very slow because of the oscillatory behavior of the kernel. Yet, we can apply acceleration techniques to sum the series (these acceleration techniques belong to the family of extrapolation techniques). For the sake of illustration, we can mention Eulers transformation, which is one of the best known acceleration technique: Sn + Sn+1 Sn = , (25) 2 where Sn = ui is the partial sum (ui being the terms of the original series). i=0 By applying the formula repeatedly, we get:
k+1 Sn

Sn + Sn+1 . = 2

(k)

(k)

(26)

A direct improvement of Euler technique is to weight the partial sums and write: Sn = wn Sn + wn+1 Sn+1 . wn + wn+1 (27)

In our specic situation, we need to evaluate integrals of the form:

S=

f (x)dx ,

(28)

where is related to k0 maxl ( ). In order to transform this integral in a series, we apply the approach of integration then summation. Thus, we dene:
xi

ui =
xi1

f (x)dx , ui ,
i=0 n

(29a) (29b) (29c)

S= Sn =
i=0

ui .

The break points xi have to be well chosen, and may for example be chosen based on the asymptotic behavior of f . If we refer back to Sommerfeld integrals, we can take the asymptotic expansion of the Bessel function: J (k ) 2 cos(k ) . k 2 4 (30)

Hence, the easiest/simplest choice of break points will be xn = k n = x 0 + n , (31)

where x0 is the rst break point greater than . We will then approximate S by Sn as
xn

Sn =

f (x)dx .

(32)

The problem of Sommerfelds integrals is that the remainder

rn = S S n =

f (x)dx
xn

(33)

decays slowly, so that we want to accelerate the series from which S n is evaluated (see Eq. (29c)). For Sommerfeld-type integrals, we write the generic form as:

I=

g(k ) f (k ) dk ,
xn

(34)

and the partial integral as I(n) =

g(k ) f (k ) dk .

(35)

The remainder is therefore I I(n) = g(k ) f (k ) dk .

xn

(36)

This integral can be expanded into an innite series of inverse powers of by integration by part. For example, with g(k ) = eik (which is a generic form for Sommerfeld integrals), we can write: i i i I I0 (n) = eian fn + fn + ( )2 fn + . . . , (37) where fn = f (xn ), fn = f (xn ), etc. Note that if the right-hand side term converges (and it does in our case), the dominant term is O(1 ). Yet, if we now construct I1 (n) = fn+1 I0 (n) + fn I0 (n + 1) , fn + fn+1 (38)

it appears to be a better estimate of I since the error is in O(2 ). If, in addition, f (k ) Ck ek , we can approximate fn and fn+1 and write I1 (n) = I0 (n) + 0 I0 (n + 1) , 1 + 0

(39)

(40)

where 0 = [n/(n + 1)] e/ . At higher orders, a better approximation is given by: I2 (n) = I1 (n) + 1 I1 (n + 1) , 1 + 1 (41)

where 1 = [(n 1/2)/(n + 1/2)]2 e/ . In practical applications, the parameters and may have to be adjusted for an optimum convergence. The technique presented briey here is known as the weighted average method, and more details can be found in the literature under this keyword.

Integral Equations in Electromagnetics

Massachusetts Institute of Technology 6.635 lecture notes

Most integral equations do not have a closed form solution. However, they can often be discretized and solved on a digital computer. Proof of the existence of the solution to an integral equation by discretization was rst presented by Fredholm in 1903. In general, integral equations can be divided into two families: 1. When the unknown is in the integral only, the integral equation is called of the rst kind. 2. When the unknown is both inside and outside the integral, the integral equation is called of the second kind. For electromagnetic applications, we can have both scalar and vector integral equations.

Scalar integral equations

Let us consider the situation of Fig. 1: two regions are dened in space, region 2 bounded by the closed surface S and region 1 being all the remaining space, bounded by S and S , in which sources are located.
#1 J #2

n S

Figure 1: Denition of the geometry of the problem.

It is already known that we can write: ( (

2 2 2 + k1 ) 1 () = J() , r r 2 + k2 ) 2 () = 0 , r

(1a) (1b)

Section 1. Scalar integral equations

and similarly for the Greens functions ( (

2 2 2 + k1 ) g1 (, r ) = ( r ) r r

in region 1, in region 2.

(2a) (2b)

2 k2 ) g2 (, r ) r

= ( r ) r

Upon performing (Eq. (1a)g1 (, r )Eq. (2b)1 (, r )), we get: r r [g1 (, r ) r

2

1 () 1 () r r

g1 (, r )] = J()g1 (, r ) + ( r )1 () . r r r r r

(3)

Upon integrating Eq. (3) over the entire volume and using the identity 2 g, we get: dv
V

(g g) = (4)

[g1 (, r )1 () 1 () g1 (, r )] = r r r r

dvJ() g(, r ) + 1 ( ) . r r r
V

By Gauss theorem, we reduce the left-hand side integral to a surface integral. Also dvJ()g1 (, r ) = inc ( ) . r r r (5)

We therefore obtain
S+S

ds n [g1 (, r ) 1 () 1 () g1 (, r )] = inc () + 1 ( ) , r r r r r r

r V .

(6)

By invoking the radiation condition, the integral over S vanishes, leaving (and exchanging r and r so that primed coordinates correspond to sources and unprimed ones to observation): 1 () = inc () r r ds n [g1 (, r ) r 1 ( ) 1 ( ) r r g1 (, r )] r r V1 . (7)

For r V2 , the wave equation has no source and therefore the integration of the delta function yields a zero value. Performing the same steps for this second case, we get the generic relation: () r V 1 r 1 g1 (, r )] = r 0 r V2

inc () r

ds n [g1 (, r ) r

1 ( ) 1 ( ) r r

(8)

This is directly evocative of Huygens principle:

For r V1 : the total eld 1 () is the sum of the incident eld plus the eld due to the r surface currents on the surface S. For r V2 : the surface source on S produces a eld that exactly opposes inc , yielding the extinction theorem. Applying the same reasoning to region 2, we write (where there is no incident eld): 0 r V1

ds n [g2 (, r ) r

2 ( ) 2 ( ) r r

g2 (, r )] = r

2 () r V2 r

(9)

Note that the sign reversal is due to the denition of the normal vector n which has to point outward from the surface. Here, since we use the same n as before, we have to take it as being negative. Eqs. (8) and (9) have four independent unknowns: 1 , 2 ; n 1 , n 2 , (10) g(, r ) are the kernel r

which can be related by the boundary conditions. Here, g(, r ) and n r of the integral equation.

Vector integral equation

For the sake of completeness, we shall write the vector wave equation as well, although we will not use it here directly. Considering the same situation as above, we know that the elds have to satisfy: r r r E1 () 2 1 1 E1 () =i1 J() , r r E2 () 2 2 2 E2 () =0 ,

and the Greens functions:

(11a) (11b)

G2 (, r ) 2 2 2 G2 (, r ) = I( r ) . r r r
V

r r r G1 (, r ) 2 1 1 G1 (, r ) = I( r ) , r dv [Eq. (11a) G1 (, r ) Eq. (12a) E1 ()]dv) r G1 (, r ) = i1 r

V

(12a) (12b)

By the same technique as before ( dv

V

r r r E1 () G1 (, r ) E1 ()

r r dv J() G1 (, r ) E1 () . r

(13) The left-hand side can be transformed into a surface integral (left as exercise) and the right-hand side written in terms of incident eld, yielding: r r E1 ( ) = Einc ( )+
S+S

r r ds n [ E1 ()]G1 (, r )+ E1 () G1 (, r ) , r r

r V1 . (14)

By reciprocity of the Greens functions, we can transform n[ r E1 ()] G1 (, r ) = n [ r r E1 ()] G1 (, r ) r (15)

r = i1 G1 (, r ) n H1 () . r

In addition, for an unbounded homogeneous medium ( n E1 () r

G(, r ) is reciprocal): r G1 (, r ) r (16)

r G1 (, r ) = n E1 () r = [

r G1 (, r )] n E1 () . r

Section 3. Problem with the internal resonance

For Greens functions that satisfy the radiation condition, Eq. (14) becomes: r r E1 ( ) = Einc ( ) +
S

r ds n i1 G1 (, r ) n H1 () [ r

r G1 (, r )] n E1 () . r

(17)

We perform the same steps for the other region to eventually obtain (and again interchanging r and r ): E () r V 1 r 1 = 0 r V2 = 0 E2 () r V2 r r V1 (18b)

r Einc () +
S

ds

r r i1 G1 (, r ) n H1 ( ) [

r G1 (, r )] n E1 ( ) r

(18a)

ds
S

r r i2 G2 (, r ) n H2 ( ) [

r G2 (, r )] n E2 ( ) r

Together with the boundary conditions r r n H1 () = n H2 () , r r n E1 () = n E2 () , (19a) (19b)

r r this system can be solved for n E1 () and n H1 (). Note also that Eqs. (18) can be written in terms of electric and magnetic currents, and magnetic Greens functions: r r Jeq ( ) = n H1 ( ) ; so that for example: r Einc () +
S

r r Meq ( ) = n E1 ( ) ,

(20a)

ds

r r i1 Ge1 (, r ) J( ) + Gm1 (, r ) M ( ) r r

E () r V 1 r 1 = 0 r V2

(21)

Problem with the internal resonance

A question arises: is the integral equation equivalent to Maxwells equations? Or asked dierently, if we solve the integral equation and Maxwells equations, do we get the same solution? The answer is actually no, that is they are not always equivalent to each-other. The problem comes from spurious solutions at frequencies corresponding to the eigenfrequencies of the cavity enclosed by the surface S. This problem is generally referred to as the internal resonance of the integral equation. However, this lack of complete equivalence between the physical problem and its dening integral equation is rather minor and infrequent phenomenon, and is therefore often tolerated in practice.

Scattering by a rough surface

Let us consider the 2D problem (for which we shall use a scalar integral equation) depicted in Fig. 2.
z

PSfrag replacements

f (x)

x
1

Figure 2: Rough surface S separating two media.

The integral equation in scalar form is given by: inc () + r

S

ds n [( ) g(, r ) g(, r ) ( )] = r r r r

() r 0

r V0 r V1

(22)

4.1

Dirichlet boundary conditions: EFIE

For a TE wave (E = E y ) and PEC surface, the boundary condition is () = 0 , r for r S . () r 0 (23)

The integral equation becomes, for r S , r S: inc () r

S

ds g(, r ) n r

( ) = r

(24)

This equation, in which represent the electric eld, is referred to as the electric eld integral equation (EFIE). Note that as r gets closer to the surface, () 0 (from the boundary condition) so that r we do not need to distinguish between approaching the surface from one side or the other. In fact, we can unify the equations and write: inc () r ds g(, r ) n r ( ) = 0 , r r S ,r S . (25)

In addition, g(, r ) has an integrable singularity as r r . Let us consider the surface r depicted in Fig. 2, with z = f (x): ds = dx2 + dz 2 = dx 1+ df dx
2

(26)

such that the integral equation becomes: inc () = r

x

dx

1+

df dx

g(x, f (x), x , f (x )) ( n

( )) , r

at z = f (x ) ,

(27)

4.2

Neumann boundary conditions: MFIE

where we can limit x to [L/2, L/2]. By letting 1+ df dx

2

( n

( ))|z =f (x ) = U (x ) , r

(28a) (28b) (28c)

inc (x, f (x)) = b(x) , K(x, x ) = g(x, f (x), x , f (x )) , we can rewrite the integral equation as
L/2

dx K(x, x ) U (x ) = b(x) ,
L/2

(29)

which has to be solved numerically. Before doing this, we shall write the integral equation with Neumann boundary conditions.

4.2

Neumann boundary conditions: MFIE

This corresponds to a TM case with H = H y with a PEC surface. In this case, the boundary condition is n () = 0 , r (30) and the integral equation becomes inc () + r
S

ds () n r

In this case, it makes a dierence if we approach the surface from the top or from the bottom. In fact inc (+ ) + r
S

() r V r 0 g(, r ) = r 0 r V1

(31)

ds ( ) n r ds ( ) n r

g(+ , r ) = (+ ) , r r g( , r ) = 0 . r

(32a) (32b)

inc ( ) + r
S

These two equations seems inconsistent with one another. We shall show that in fact, they are actually consistent with each-other, due to the singularity of the Greens functions. We shall examine what happens when we let r approach the surface. Fig. 3 is an illustration of the situation at the immediate vicinity of the surface. The integral part of the equation can be written as:
S

ds ( ) n r

g(, r ) = P V r
S

ds ( ) n r

g(, r ) + r
piece

ds ( ) n r

g(, r ) , r

(33)

where P V denotes the principal value and piece refers to the integration over the local domain shown in Fig. 3. For this integral, we use the local coordinates: ds = dX , n = Z , etc (34)

PSfrag replacements

Z a r = (0, Z) a X
Figure 3: Zoom on the rough surface.

The integral becomes:

a piece

ds ( ) n r

g(, r ) = lim lim () r r

a0 |Z|0 a

dX

g(, r ) . r Z

(35)

Over the small piece, we have: | r | = r g= X 2 + (Z Z )2 , (36a) (36b) (36c)

1 i (1) | r | r r H0 (| r |) ln( ), 4 2 2 Z g . = Z Z =0 2(X 2 + Z 2 )

Thus, the integral becomes

a

= lim lim () r
piece a0 |Z|0 a

dX

1 Z 2 + Z2 2 X
a a

The two parts of the integral then become: inc () + P V r

S

() r X = lim lim tan1 a0 |Z|0 2 Z 1 () r for Z > 0 = 2 1 () for Z < 0 r

2

(37)

ds ( ) n r ds ( ) n r

inc () + P V r
S

1 g(, r ) + () = () , r r r 2 1 g(, r ) () = 0 , r r 2

(38a) (38b) (38c)

which is cast into one equation: inc () + P V r

S

ds ( ) n r

1 r g(, r ) = () . r 2

(39)

Eq. (39) is called the magnetic eld integral equation (since represents the magnetic eld) and is an integral equation of the second kind.

Section 5. Solving the EFIE

In the same way as before, we can dene the kernel as K(x, x ) = and write the MFIE as

1+

df dx

g(, r )|z=f (x),z =f (x ) , r

(40)

inc () + P V r

dx (x ) K(x, x ) =

1 (x) . 2

(41)

Solving the EFIE

Upon using the notation introduced in Section 4.1, the problem is reduced to solving the following integral equation:
L/2

dx K(x, x ) U (x ) = b(x) .
L/2

(42)

Let us subdivide the integration domain into N small elements, each of length = L/N , and centered at xm (m [1, N ]). Thus, constraining the observation at these discrete locations, the integral equation becomes
L/2

dx K(xm , x ) U (x ) = b(xm ) .
L/2

(43)

Next, if we suppose that U (x) is constant in each interval, we replace the integral by a sum over all segments, excluding the singular term: x
n=1 n=m

K(xm , xn ) U (xn ) + U (xm )

m

dx K(xm , x ) = b(xm ) .

(44)

Note that we have to single out the singularity, i.e. the mth interval because K(xm , x ) is singular at x = xm . This part is known as the self-patch contribution. For a 2D problem, we have: K(xm , x ) = i (1) H (k 4 0 (xm x )2 + (f (xm ) f (x ))2 ) . (45)

Upon approximating f (x ) f (xm ) K(xm , x ) =

(1)

f (xm )(x xm ), we write: i (1) H (k|x xm | 1 + f (xm )2 ) . 4 0 (46) 1.78). (47)

2 For small argument: H0 () 1 + i ln(/2) where is the Euler constant ( Therefore: i 1 k(x xm ) 1 + f 2 (xm ) . K(xm , x ) = 1 + i ln 4 2

Therefore, the integral becomes:

xm +x/2 x/2

dx K(xm , x ) = 2
m xm x/2

dx K(xm , x ) = 2
0

dx K(xm , x + xm )

i 1 kx 1 + f 2 (xm ) dx 1 + i ln 2 0 2 2 k ix 1 + i ln x 1 + f 2 (xm ) , = 4 4e where ln(e) = 1. We can therefore cast the integral equation into a matrix equation of the form:
N

(48)

Amn Un = bm ,
n=1

(49)

where Un = U (xn ) is the unknown, bm = b(xm ) , x K(xm , xn ) Amn = ix 2 4 1 + i ln (50a) (50b) for m = n
k 4e x

1 + f 2 (xm )

for m = n

(50c)

The system can now easily be solved numerically.

The Method of Moments in Electromagnetics

Massachusetts Institute of Technology 6.635 lecture notes

Introduction

In the previous lecture, we wrote the EFIE for an incident TE plane wave on a PEC surface. The solution was then obtained by some types of intuitive arguments, such as dividing the integration domain into small elements and supposing that the unknown does not vary too much over each elementary cell. We shall now see more rigorously what we actually did, and show that it was in fact a simple version of the Method of Moments. R. F. Harrington was the rst to use the method of moments (MoM) in electromagnetics and his book remains a fundamental reference (and very easy to read!): R. F. Harrington, Field Computation by Moment Method (is now available from IEEE Press).

PEC surface with TE incident wave: EFIE

The situation we studied last time is depicted in Fig. 1. The integral equation (EFIE) we
z

f (x)

x
1

Figure 1: Rough surface S separating two media.

eventually obtained was:

L/2

inc () = r
L/2

dx

1+

df dx

g(x, f (x), x , f (x )) [ n

( )]|z =f (x ) , r

(1)

which we rewrote as
L/2

b(x) =
L/2

dx K(x, x ) U (x ) ,

(2)

Section 2. PEC surface with TE incident wave: EFIE

where 1+ df dx
2

[ n

( )]|z =f (x ) = U (x ) , r

(3a) (3b) (3c)

inc (x, f (x)) = b(x) , K(x, x ) = g(x, f (x), x , f (x )) , where U (x ) is the unknown we are solving for.

The important step we did from this integral, although it probably appeared straightforward, was to say that x can only take discrete values on the surface, thus dening N intervals of length x: x {xi } , i = 1, . . . , N, (4) The assumptions were therefore: 1. x {xi } , i = 1, . . . , N.

2. U (x ) is constant on each interval. which eventually yielded the following system of equations (supposing that the problem related to the singularity of the Greens function has been accounted for):
N

Amn Un = bm .
n=1

(5)

This is a matrix equation with the two indices m and n corresponding to: m: observation point unprimed coordinates. n: source point primed coordinates. Physical interpretation: element (m, n) represents the eect of cell n on cell m. element (m, m) represents the self-term. Mathematically: The steps we had to perform to from Eq. (2) to Eq. (5) are 1. Write the unknown as U (x ) =
n

Un (xn )x ,

(6)

stating that now the unknowns become {Un } which are the amplitudes of the function. The integral equation becomes:
L/2

dx K(x, x )U (x ) = x
L/2 n

Un K(x, xn ) = b(x) .

(7)

2. Dot-multiply both sides by (xm ): x

n

Un K(xm , xn ) = b(xm ).

(8)

These two steps are at the basis of the method of moments: 1. First step can actually be decomposed into two steps: (a) Mesh the structure (i.e. choose the intervals over which Un will be dened). (b) Expand the unknown U (x) into basis functions. 2. The second steps concerns the observation: dot-multiply both sides of the equation by a test function (or weighting function). In the previous example: basis functions: pulse basis function. testing functions: we point-patch the integral equation at x = xm (the method is therefore called point matching). This is a very simple, yet very widely used version of the method of moments.

General considerations on MoM

Let us consider the inhomogeneous equation: L(f ) = g , (9)

where L is a linear operator, g is known, and f is to be determined. We shall now perform the two essential steps we have highlighted above. 1. Let f be expanded in a series of functions: f=
n

n fn ,

(10)

where n are constant. The set fn is called expansion function, or basis functions. Note that for an exact solution, the summation should be taken to , but has to be truncated in practice. 2. It is assumed that a suitable inner product has been dened for the problem. Now, we dene a set of weighting functions, or testing functions, w1 , w2 , . . . , wN in the range of L, and take the inner product of the previous equation with wm :
n

n < wm , Lfn >=< wm , g > .

(11)

Section 4. A simple example for electrostatic

The system can now be written in matrix form as: [Amn ] [n ] = [gm ] , where < w1 , Lf1 > < w1 , Lf2 > . . . [Amn ] = < w2 , Lf1 > < w2 , Lf2 > . . . , . . .. . . . . . 1 [n ] = 2 , . . . < w1 , g > [gm ] = < w2 , g > . (13) . . . (12)

If the matrix [Amn ] is not singular, the unknowns n are simply given by: [n ] = [Amn ]1 [gm ] , (14)

and the original function f can be reconstructed using Eq. (10). We can now generalize the following denitions: The basis functions used previously are dened as: 1 if x belongs to the interval n Pulse basis functions: fn = 0 otherwise The testing (or weighting functions): Point matching = taking Dirac functions as testing functions.

(15)

A simple example for electrostatic

The example is taken from the reference mentioned at the beginning of this document. Let us consider a square plate of side 2a lying on the z = 0 plane with its center at the origin (see Fig. 2. Let (x, y) represent the surface density on the plate, assumed to have zero thickness. The electrostatic potential at any point in space is given by
z y 2a

PSfrag replacements

2a

2b 2b

Figure 2: Discretized square plate.

5
a a

(x, y, z) =
a

dx
a

dy (x , y ) g(, r ) , r

(16a) (16b) (16c)

g(, r ) = r R=

1 , 4 R (x x )2 + (y y )2 + z 2 .

The boundary condition is = V = constant on the plate. The integral equation for the problem is therefore
a a

V =
a

dx
a

dy

1 4

(x , y ) (x x )2 + (y y )2 + z 2

(17)

where the unknown to determine is (x , y ). Let us perform the three steps mentioned before: 1. Mesh the structure: divide the plate into N squares of size 2b (see Fig. 2). 2. Basis functions: let us choose
N

(x , y )
n=1

n fn (x , y )

with fn (x , y ) =

1 0

on Sn on Sm , m = n.

(18)

3. Test functions: we choose to satisfy the integral equation at the mid-point (x m , ym ) of each Sm : wm = (x xm )(y ym ). (19) With these three steps, we construct the matrix as (z = 0): [Amn ] =
xn

dx
yn

dy

1 4

1 (xm x )2 + (ym y )2

(20)

It is obvious to see that this integral is singular at (xm , ym ) Sm . In this simple case fortunately, we can perform the integration analytically (this is not always the case), and write:
b b

Ann =
b

dx
b

dy

1 4

1 x2+y2 b2

2b ln(1 + 2) , , m = n.

(21a) (21b)

Amn

Sn = 4 Rmn

(xm x )2 + (ym y )2

To rewrite this with the language of linear space: f (x, y) = (x, y) , g(x, y) = V on the plate (the discretization gives: gm
a a

(22a) V V ) = . . . (22b)

L(f ) =

dx
a a

dy

1 4

f (x , y ) (x x )2 + (y y )2

(22c)

Note: if we add another plate under the existing one, at z = 2d, with a potential V , we build a new problem that can be analyzed in two ways:

Section 5. Vectorial MoM

1. By meshing both plates (i.e. meshing everything, which is of course always possible). This will yield matrix twice as big as the previous one, to solve for twice as many unknowns. 2. By using the image theory, and saying that the new problem is equivalent to the one of a unique plate on top of a ground plane at z = d. In that case, we only have to change the Greens function to take the ground plane into account, and we keep the same number of unknowns as in the initial problem. When possible, this solution is better because computationally less expensive (analytically more expensive). Basically, a general trend is to have a Greens function that represents as much as possible the environment and to mesh only those parts that are external to the environment. This is in fact the reason why people are looking for Greens functions in layered media, periodic media, etc.

Vectorial MoM

We can of course apply the MoM to the vectorial case, like for example the equation: r E() = i
S

r ds Ge (, r ) J( ) . r r n fn ( ) ,
n

(23)

The general expansion of the current will be: r J( ) = yielding r E() = i

n

(24)

n
Sn

r ds Ge (, r ) fn ( ) . r

(25)

The third step is to dot-multiply the equation with a testing function hm and integrate over the cell surface (i.e. perform the inner product):
Sm

r r E() hm () = i

n
n Sn

ds
Sm

r r ds hm () Ge (, r ) fn ( ) . r

(26)

The double double-integral on the right-hand side of this equation is known as the impedance term since we can cast this system of equation into a matrix representation as: [Em ] = [Zmn ] [n ] . (27)

Other basis and testing functions

The advantage of the MoM over purely numerical methods is that there is still a large part that remains analytic (like the Greens functions for example). Yet, it remains a numerical method based on a matrix inversion technique and therefore, convergence issues need to be examined. The convergence of the MoM is closely related to the choice of basis functions and, although to a lesser extend, to the choice of testing functions. There are essentially two families of basis functions:

1. Entire domain basis functions: using these functions to expand the unknowns is analogous to a Fourier expansion or to a modal expansion. These types of functions yield a good convergence of the method but are not versatile since the geometry need be regular in order to have the modes dened. Note that in this case there is no use to mesh the geometry. 2. Sub-domain basis functions: they rely on a proper meshing of the geometry, which can be rectangular, triangular, etc. The choice of basis functions is here very wide, from Dirac (like for the weighting functions shown in this document), pulses (basis functions shown in this document), piecewise linear, etc. Finally, we can mention that point matching, which is easy to grasp and straightforward to implement, may not yield an optimal convergence. In most of the applications, the Galerkin technique is better, which consists in choosing the same testing functions as the basis functions. This applies to both sub-domain and entire domain functions.

Time Domain Method of Moments

Massachusetts Institute of Technology 6.635 lecture notes

Introduction

The Method of Moments (MoM) introduced in the previous lecture is widely used for solving integral equations in the frequency domain. Yet, some attempts have been made recently at the use of the MoM in the time domain. We shall briey expose this approach here.

Time domain equations

The rst step is of course to write Maxwells equation and all other relations (constitutive relations and continuity) in time domain: r B(, t) M (, t) , r t r B(, t) = m(, t) , r r r D(, t) = E(, t) , r E(, t) =
r J(, t) + (, t) = 0 , r t

r D(, t) + J(, t) , r t r D(, t) = (, t) , r r r B(, t) = H(, t) , r H(, t) =

r M (, t) + m(, t) = 0 . r t

(1a) (1b) (1c)

(1d)

For the time-domain MoM, it is easier to work with the potentials, and make use of the well-known retarded potentials theory. In view of doing this, we write the denition: 1 r H(, t) = 0
r A(, t) (2a)
(2b)

r E(, t) = (, t) A(, t) . r r t

Both the vector potential A and the scalar potential satisfy the wave equation which, in time-domain domain, writes:
2

2 r A(, t) = 0 J(, t) , r t2 2 (, t) r 2 (, t) 0 0 2 (, t) = r r . t 0 r A(, t)

0 0

(3a)
(3b)

These potentials are linked by the time-domain Lorentz gauge: r A(, t) +

0 0

(, t) = 0 . r t

(4)

Section 2. Time domain equations

We can dened also a time-domain Greens function which satises the time-domain scalar equation: 1 2 ( 2 2 2 )g(, r , t, t ) = ( r ) (t t ) , r r (5) c t which solution is (in free-space): | | r r 1 c ) t>t, r r (t t (6) g(, r , t, t ) = 4| | r 0 t<t. From this, the solution to the wave equation for A and can be written as:

r A(, t) =0
V

dv

r dt J( , t ) g(, r , t, t ) = 0 r
V

dv

r J( , t R/c) , 4R

(7a)

(, t) = r
V

dv

(, t R/c) r , 4 0 R

(7b)

where R = | r |. These wave equations are known as the time retarded potentials, and r essentially say that the potential (either A or ) can be calculated at a given point in space r and given time t from all previous times. From these equations, we can calculate the space-time electromagnetic elds: 1 r H(, t) = 4 dv r J( , ) , = t R/c , R r (, ) 0 r J( , ) dv dv . R 4 t R (8a) (8b)

1 r E(, t) = 4 0

Let us continue with the electric eld rst: 1 r E(, t) = 4 0 dv

V

1 1 0 ( , ) + ( , ) r r R R 4

dv
V

1 J( , ) . r R t

(9)

At this point, we need to use the following relations: R , R 1 R = 3, R R 1 1R ( , ) = r ( , ) = r R ( , ) = r (, ) . r c c R R= We can therefore continue with the electric eld as: r E(, t) = 1 4 0 1 = 4 0 dv dv 1 R 0 R ( , ) + 3 ( , ) r r 2 cR R 4 R 1 0 1 ( , ) + ( , ) 2 r r c R R 4 1 J( , ) r R 1 J( , ) . r dv R dv (10a) (10b) (10c)

(11)

We can perform the same type of calculations for the magnetic eld using the relation 1R r J( , ) . r J( , ) = c R We get: 1 r H(, t) = 4 dv R 1 R r J( , ) 3 J( , ) . r c R2 R (12)

(13)

Upon gathering the expressions for the electric and magnetic eld, we eventually get: 1 r E(, t) = 4 1 r H(, t) = 4 dv dv 1 1 R 0 J( , ) r ( , ) + ( , ) r r 2 c R R 0R R 1 1 J( , ) + J( , ) 2 . r r c R R (14a) (14b)

Upon using the boundary conditions for the electric and magnetic eld, we construct the integral equations in a standard way: EFIE: n (E i + E scat ) = 0 on PEC surface 1 r n n E i (, t) + 4 MFIE: n (H i + H scat ) = Js . As we have seen before (in a previous class), this integral equation is expressed in terms of the principal value of the integral with a 1/2 additional factor. Thus: 1 1 r J(, t) = n H i (, t) + r n PV 2 4 ds [. . .] (16) ds [. . .] (15)

For the sake of comparison, we can write the MFIE in the frequency domain and in the time domain: r J() = 2 H i () + 2 P V n r n 1 r n PV J(, t) = 2 H i () + n r 2 r ds J( ) ds g(, r ) r r S, (17a) (17b)

1 1 R J( , ) + J( , ) 2 . r r c R R

Note that in the principal value, we essentially exclude the part for which R = 0. Since = t R/c and R = 0, we always have that < t. The time domain equations therefore r state that the current at location r and time t is equal to a known term 2 H i (, t) plus a n term (integral) known from the past history of J. This is the basis for solving the time domain integral equation by iterative methods, the most well-known one being the marching-on-in-time.

Section 3. The marching-on-in-time technique

3
3.1

The marching-on-in-time technique

General equations

The integral equation can often be cast in the following form: r r J(, t) = J i (, t) +
S

dv
0

r dt K(, r , t t ) J( , t ) . r

(18) eq.10

eq.10 r Note that we have a time integral also as in Eq. (18), J( , t ) has not yet been set to satisfy r any causality condition. Hence, we must then impose J i (, t) = 0 for t < 0, r S.

In order to apply the MoM, we discretize the current both in space and in time:
M N

r J( , t ) =
m =1 n =0

Jp (m , n ) Ps ( rm ) Pt (t tn ) , r

(19)

where P denotes the simple pulse function. In addition, we also apply point-matching, which means that we take the following testing functions: Wmn (m , tn ) = ( rm ) (t nt) = ( rm ) (t tn ) , r r r (20)

r where we take t = min{Rmm /c}, Rmm = |m rm |. The method is best illustrated on the following example.

3.2

Example

Let us consider a 1D example governed by the following integral equation:

x0

g(x, t) =
x0

K(x, x ) f (x , )dx ,

x [x0 , x0 ],

= (x, x , t) = t

|x x | . c

(21)

Let us chose the following expansion for f :

N J

f (x , )
i =1 j =1

ai j Pi j (x , ) ,

(22)

Note that we use the denitions: xi = i dx , tj = j dt , and dx = cdt . In order to apply point matching, we take the following testing functions: wij (x, t) = (x xi ) (t tj ) . (24)

where the pulse basis functions are dened as 1 for x [x dx , x + i i 2 Pi j = 0 elsewhere.

dx 2 ]

and t [tj

dt 2 , tj

dt 2]

(23)

Upon expanding and testing, we get:

x0 N J

g(xi , tj ) = gij =
x0 N

K(xi , x )
i =1 j =1
1 (i + 2 )dx

ai j Pi j (x , )(t tj ) (25)

1 i =1 (i 2 )dx

dx
j =1

ai j Pi j (x , )K(idx , x )(t tj ) .

Coming back to the denition of , we write (with the test and the expansion): = tj dx |xi xi | = jdt |i i | = (j |i i |)dt , c c (26)

such that the coecient ai j becomes ai ,j|ii | . The integral equation becomes:
N (i + 1 )dx 2

gij =
i =1

ai ,j|ii |

1 (i 2 )dx

dx K(idx , x ) .

(27)

We can dene the term Zii = and rewrite the previous system as
N

1 (i + 2 )dx 1 (i 2 )dx

dx K(idx , x )

(28)

gij =
i =1

Zii ai ,j|ii | = Zii aij + Zi,i1 ai1,j1 + Zi,i2 ai2,j2 + . . . + Zi1 a1,ji+1 + Zi,i+1 ai+1,j1 + Zi,i+2 ai+2,j2 + . . . + Z1,N aN,jN +i . (29)

In this equation, only the rst term involves time step j, all the others terms being at j 1, j 2, . . . Therefore, we can solve for aij : 1 aij = gij Zii
N

Zii ai ,j|ii | .
i =1 i =i

(30)

The value of all aik are known for k < j, so that aij is completely specied in closed form by those and the present value of gij . This process is known as a 1D march-on-in-time approach. Time-domain MoM is nowadays in its early stage and, although it has been successfully applied to various simple situations, still suers from numerical instabilities. More work is in progress...

Study of EM waves in Periodic Structures

with addenda: Study of EM waves in Periodic Structures (mathematical details)

Massachusetts Institute of Technology 6.635 lecture notes

Introduction

We will study here the distribution of electromagnetic elds in dielectric periodic media. The main dierence with the previous topic comes from the word dielectric. Obviously, even a 2D periodic dielectric medium cannot be studied with the Greens functions presented in a previous lecture, since the Greens function was for periodic metallic structures. In this topic, we will study the EM elds in media where: The material is macroscopic and isotropic. Since the material is constituted of real dielectric, we suppose we work in a small enough frequency band such that we can ignore the frequency dispersive behavior of . The dielectric are lossless, so that is purely real.

2
2.1

Wave equations
Wave equations for H
= (), it is straightforward to r

Starting from Maxwells equations and using a permittivity show that we can write the following equations:
2

1 () r

r H() =

r H() ,

(1a) (1b)

r H() = 0 . In this approach, the strategy is therefore: r 1. Find the modes of H(). r 2. Find those of E() by solving Maxwells equations.

2.2

Wave equations for E

Note that we can rewrite Eq. (1a) as an eigenvalue problem: r H() = where = c
2

r H() ,

(2)

1 . () r

(3)

Upon solving, we get the eigenvectors which correspond to the eld patterns of the harmonic modes, and the eigenvalues which are proportional to the squared frequencies of these modes. Note that is a linear operator and that it is Hermitian. The demonstration of the last property is straightforward: r Dene a scalar product < F , G >= d F () G() and show that (by r r integration by part for example): < F , G >=< F , G >, which is the denition of a Hermitian operator. The consequence of this property is of course that has real eigenvalues.

2.2

Wave equations for E

Another approach to get the elds would be of course to write the wave equation for E instead of for H: 2 r E() = ()E() . r r (4) c However, this system cannot be cast in a simple eigenvalue problem. Although it can still be solved for, it is far more complicated to get accurate results since the operator we would have to dene would not be Hermitian. For this reason, this approach is in general avoided.

2.3

Bloch states

In a periodic medium, we know that the elds can be written as:

r r Hk () = eikuk (), r

uk ( + R) = uk () . r r

(5)

Inserting into Eq. (2) yields: (ik + ) 1 (ik + () r ) uk () = r (k) c

2

uk () , r

(6)

so that the operator becomes: = (ik + ) 1 (ik + () r ) (7)

Note that because uk ( + R) = uk (), the eigenvalue problem can be restricted to a small r r zone in space, which would necessarily imply a discrete spectrum of eigenvalues. Therefore, we expect a set of discrete modes for each k.

Fundamentals of photonic crystals

We shall briey explain some terminology here related essentially to solid state physics, but which is of prime importance for the study of the structures we are dealing with here.

3.1

Direct lattice (some details are given in the mathematical details addenda)

A photonic crystal is a periodic structure (that we will take to be dielectric here) in 1D, 2D or 3D. Any vector r in space can be written as r = r +R, where R is the translational vector in space dened by R = 1 a1 + 2 a2 + 3 a3 , where 1,2,3 {. . . , 2, 1, 0, 1, 2, 3, . . .} and a1 , a2 and a3 are the lattice vectors. From the lattice, we can construct the Wigner-Seitz cell as shown in Fig. 1. (9) (8)

Figure 1: Wigner-Seitz cell for an arbitrary position of points: the cell is constructed by joining the center element to its closest neighbors and drawing perpendicular lines from to the center of these segments. The polygon thus created is the smallest repeatable cell of the periodic lattice, and is dened as the Wigner-Seitz cell.

Note that there exist only one type of lattice for a 1D photonic crystal, ve distinct types for 2D photonic crystals (rectangular, square, hexagonal or triangle, centered rectangular and oblique), and fourteen for 3D photonic crystals.

3.2

Reciprocal lattice (some details are given in the mathematical details addenda)

We will use here the same notation as [Joannopoulos, Meade, and Winn, Photonic Crystals] and write the reciprocal translational vector as G: G = 11 + 22 + 33 , b b b (10)

where 1,2,3 {. . . , 2, 1, 0, 1, 2, 3, . . .} and 1 , 2 and 3 are the lattice vectors in the spectral b b b domain. For the sake of illustration, Tab. 1 gives the denition of vectors a and for square and b triangular lattices.

3.3

Bloch-Floquet theorem

Square lattice Triangular lattice

a1 a2 a1 a2

= a x = a y = a x = ( + 3) x y

1 b 2 b 1 b 2 b

= 2/a x = 2/a y = 2/a ( 3/3 y ) x = 2/a 2 3/3 y

Table 1: Denition of a and vectors for square and triangular (or hexagonal) lattice. b

3.3

Bloch-Floquet theorem

From Bloch-Floquet theorem, we know that we can write the electric and magnetic elds as a summation over reciprocal vectors G (see the mathematical details part of the notes). This means that dierent k do not necessarily correspond to dierent modes and that therefore there is a redundancy in the label k: we can therefore reduce the study to what is called the rst Brillouin zone (which is the Wigner-Seitz cell in the reciprocal lattice). Some examples of Brillouin zones are given in Fig. 2 and Fig. 3.

a2 a1
PSfrag replacements

2 b

1 b

Figure 2: Direct square lattice and corresponding reciprocal lattice with highlighted Brillouin zone.
2 b a2 a1
PSfrag replacements

1 b

Figure 3: Direct triangular (or hexagonal) lattice and corresponding reciprocal lattice with highlighted Brillouin zone.

Bragg-like diraction

The standard Bragg diraction is illustrated in Fig. 4. Here, we will derive another diraction condition, equivalent to Bragg, and shall see that the diraction is entirely governed by the reciprocal vector G.

PSfrag replacements

Figure 4: Schematic representation of Bragg diraction. Maximal diraction occurs at 2a sin = n where is the wavelength of the electromagnetic wave and n is an integer.

Referring to Fig. 5, we can write the scattering amplitude in terms of the reection coecient at position r times a phase factor.

r O k
PSfrag replacements

Figure 5: Diraction from an elementary volume of a periodic medium: k is the is the wavevector of the diracted wave. wavevector of the incident wave whereas k

Upon integrating over the whole volume, we get: F (k, k ) = Since the medium is periodic, we can write: ( + R) = () = r r
G r ()ei(kk )dv . r

(11)

r (G)eiG ,

(12)

such that F (k, k ) =

G r (G)ei(Gk)dv ,

(13)

Section 4. Bragg-like diraction

where k = k k. This amplitude is maximal when G k = 2m or, when m = 0, G = k. (14)

This is an important relation which, again, is a condition for maximal diraction. Upon ex panding back in terms of k and k and rising to the square, we write (noting that |k| = |k | = k): k 2 = k 2 + 2 k G + G2 , or (taking k instead of k): 2k G = G 2 . (15)

(16)

As an exercise, it is interesting to show that this condition is equivalent to the standard Bragg diraction. Upon dividing both terms of Eq. (16) by 4, we eventually write G G k ( ) = ( )2 . 2 2 (17)

This last equation has a nice geometrical interpretation shown in Fig. 6 which shows that the vectors k that satisfy the maximum diraction condition are actually those which lie on the edge of the Brillouin zone.
D

GD /2

O PSfrag replacements

GC /2

Figure 6: Graphical representation of Eq. (17): each vector k (black vector) with its tip on a dashed line (not all represented) will satisfy the equation. Graphically: all those k have the same projection on the generating vector G/2 (red vector).

Therefore: The edge of the Brillouin zone plus its center (G = 0) satisfy the maximum diraction condition. This condition can also be rewritten in terms of group velocity: for those k which tip lie on the edge of the Brillouin zone and k = 0, the component of the group velocity normal to the

Bragg diraction planes tends to zero since the electromagnetic wave tends to be completely reected for these k:
norm vg (k BZ tip) =

k (k)

norm

(k BZ tip) 0 .

(18)

For the symmetry points, the diracted wave is reected in the direction of the incident wave so that for these points, the total group velocity is zero. This can be directly seen on the dispersion curves where, at the symmetry points of the crystal, the tangent to the curve is horizontal (except possibly for those points corresponding to a zero frequency).

Mathematical details

Using all the principles shown before, we can construct the eigenvalue system for H and then solve for E. The detailed mathematical manipulations are given in the annex document Study of EM waves in Periodic Structures (mathematical details). Note that to build the system, we need to evaluate the Fourier coecients of the permittivity (or the inverse of the permittivity, ). We shall show how to get these coecients for the case of innite dielectric rods a of circular cross-section organized in a square lattice, embedded in a background of b . We therefore place ourselves in a 2D situation where the G vector will be to denote that it does not depend on z (and similarly R will be noted R ). written G As a reminder, we write the permittivity as: () =
G (G )eiG ,

1 (G ) =

()eiG ,

(19)

where denotes the surface of the elementary cell. The idea is to write the permittivity as () =
b

+(

b) R

S(Rc | R |) ,

(20)

where again the subscript in R denotes a dependency on x and y only, Rc is the radius of the dielectric rods and S denotes the step function. Merging these two equations, we get: (G ) = = 1 1
b be iG

+(

b) R

S(Rc | R |) eiG d , (
a

d +

b) R

S(Rc | R |)eiG d .

(21)

Let us call the rst integral I1 (G ) and the second integral I2 (G ), and evaluate them separately. Evaluation of I1 ()

Section 5. Mathematical details

The rst integral is easily evaluated as: I 1 ( G ) =

b

if G = 0 elsewhere

(22)

Evaluation of I2 () For the second integral, we can make the change of variable = R . Since spans the whole domain and R is the translational vector, spans the whole space. We can therefore replace the sum of integrals over by a single integral over the whole 2D space. We write then: If G = 0: I 2 ( G ) = where fr =
2 Rc

d S(Rc | R |) = (

b)

2 Rc = fr (

b ),

(23)

is the fractional volume.

If G = 0: 1 I 2 ( G ) = ( =
a a

b) Rc

d S(Rc | R |) eiG 2

a =

d
0 0 Rc

deiG

cos()

2
0

d J0 ( G ),

(24)

where we have used the change of variable (now standard) x = sin , y = cos , Gx = G sin , Gy = G cos , and the well-known identity for the Bessel function. Upon using xJ0 (x)dx = x/ J1 (x), we continue with I 2 ( G ) = J1 (G ) G 2J1 (G Rc ) = fr ( a b ) . G Rc
a b)

2 (

Rc 0

(25)

This lead to the nal result that: f + (1 f ) a r r b (G ) = ( a b )fr 2J1 (G Rc )

G Rc

if G = 0 elsewhere.

(26)

The reconstruction of the permittivity is then straightforward, and examples of two lattice are given in Fig. 7.

(a) Side view.

(b) Top view.

Figure 7: Reconstruction of the permittivity for cylindrical rods in a square lattice a = 10, b = 1, Rc /a = 0.2 (yielding a fractional volume fr = 12.5%).

10

Section 6. Dispersion curves

Dispersion curves

At this point, we have everything to build the eigensystem (34) given in the additional document Study of EM waves in periodic structures (mathematical details). Solving it gives a set of eigenvalues that are directly related to the dispersion curve of the material. An example is given in Fig. 8.
Square lattice: Rc/a = 0.2 (f=12.5664%), a=10, b=1.
0.8

0.7

0.6

Frequency a/2 c

0.5

0.4

0.3

0.2

PSfrag replacements
0.1

TE TM

Figure 8: Dispersion diagram for the photonic crystal of Fig. 7.

Study of EM waves in Periodic Structures (mathematical details)

Massachusetts Institute of Technology 6.635 partial lecture notes

Introduction: periodic media nomenclature

1. The space domain is dened by a basis,(a1 , a2 , a3 ), where any vector can be written as
r = r + R = r + 1 a1 + 2 a2 + 3 a3 ,
where R is the translation vector, with 1 , 2 , 3 integers.
2. The spectral domain is dened by a basis, (b1 , b2 , b3 ), and similarly, the translational vector is written as
G = 1 b1 + 2 b2 + 3 b3 ,
where 1 , 2 , 3 are integers. 3. The two basis are linked since the functions (elds, permittivity) are periodic. For example, if we write the permittivity: Fourier expansion: (r) =
G (G) eiGr
where (G) =
1
Vcell dr3 (r) eiGr . (3)

(1)

(2)

Periodicity:

(r + R) =
G
=
G

(G) eiG(r+R)
(G) eiGr eiGR = (r)
(4)

so that eiGR = 1 and

G R = 2m
where m {. . . , 1, 0, 1, 2, . . .} . (5)

We can see that condition (5) is immediately veried if we impose:

bj ai = 2ij .
(6)

Section 1. Introduction: periodic media nomenclature

4. Bloch-Floquet theorem: Since EM elds are periodic, we can write them as a propagating function times a function with the same periodicity as the medium:
(r) = eik r (r) k k

where

(r + R) = (r) , k k

(7)

and where can represent either the electric or magnetic elds, E or H.

Since (r) is periodic, we can Fourier expand it:
(r) = k
G
so that we shall write:
E (r) = k
G
H (r) = k
G
5. Wave equation in source-free region: From Maxwells equation, we can easily obtain the following wave equations in source-free regions (with = (r)):

eiGr , G

(8)

e ei(k +G)r , G
h ei(k +G)r . G

(9a)
(9b)

1 r (r)

E(r) =
H(r)
=

c
c

r r (r) E(r) ,
2

(10a)
(10b)

r H(r) ,

To make these equations more symmetrical, we shall work with 1/ r (r) instead of directly, so that we dene
r (r) =
1 = r (r)
r (G) eiGr . G

r (r)

(11)

The wave equations are rewritten as:

r (r)

E(r) =
H(r)
=

c
c

r E(r) ,
2

(12a)
(12b)

r (r)

r H(r) .

2
2.1

Treatment of the E eld

Method 1: direct expansion of the permittivity

We want to write Eq. (10a) with the decomposition of Eq. (9a). First, let us compute the rst curl (taking G as the variable for the expansion):
E k (r) =
G
Taking the curl one more time gives E (r) = k
G (k + G ) (k + G ) e
ei(k +G )r .
(14)

e ei(k +G )r = i G
G

(k + G ) e

ei(k +G )r .

(13)

Upon using Eq. (3) but changing the index G into G , we write
r (r)E(r)

=
G G

r (G ) e ei(k +G +G )r . G

(15)

By changing the variables G = G + G :

r (r)E(r)

=
G G

r (G G ) e ei(k +G)r . G

(16)

The wave equation (see Eq. (10a)) can therefore be rewritten as:
2

(k + G ) (k + G ) e ei(k +G )r = G

r
G G

r (G G ) e ei(k +G)r . (17) G

We can simplify by exp (ik r) and multiply by exp (iG r) to get:

ei(G G )r =
c
2

(k + G ) (k + G ) e

r
G G

r (G G ) e ei(GG )r . (18) G

If we integrate this equation over the entire space, we can pull all the terms out of the integral, except ei(G G )r on the left-hand side and. ei(GG )r on the right-hand side. Yet, we have
dr3 ei(GG )r =
V

1 (G G ) , (2)3

(19)

so that Eq. (18) becomes (upon substituting G by G since these are dummy variables):
(k + G) (k + G) e
= c
2

r
G

r (G G ) e , G

G .

(20)

2.2

Method 2: expansion of the inverse of the permittivity

2.2

Method 2: expansion of the inverse of the permittivity

Instead of working with Eq. (10a), we can also use Eq. (12a), which would need the expansion of Eq. (11). Applying the same method (and transforming the index of Eq. (11) from G to G ), we get:
2

G G

r (G )(k + G ) (k + G ) e ei(k +G +G )r = G

r
G

e ei(k +G )r . (21) G

which, upon substituting G = G + G , simplifying by exp (ik r), multiplying by exp (iG r), integrating over the whole space, using Eq. (19) and nally substituting G by G, becomes:

G r (G G )(k + G ) (k + G ) e G
= c
2

r e , G

G .

(22)

Treatment of the H eld

The H eld is treated in an exactly similar way to eventually obtain very similar equations. However, these equations can still be pushed further by using the fact that H (r) = 0. k Upon using this equality, we see from Eq. (9b) that (using G for the expansion of the eld):
(k + G ) h = 0 . G We can therefore dene three vectors (1 , e2 , e3 ) such that e
k + G = |k + G | e3 ,
e1 e 3 = e 2 e 3 = 0 , and (1 , e2 , e3 ) for an orthonormal tryad. In that case, we can decompose e
h
G = h1
G e1 + h 2
G e2 =
=1,2

(23)

(24a) (24b)

e .

(25)

We need now to introduce this expression into Eq. (12b). First, we compute H (r) = i k
G
so that r (r)
H (r) = i k
G G
= i
G G
Taking the next curl, we write:

(k + G ) e ei(k +G )r ,

(26)

r (G ) (k + G ) e ei(k +G +G )r
r (G G ) (k + G ) e ei(k +G)r .
(27)

r (r) H (r) = k

G G

r (GG ) (k+G)[(k+G ) ] ei(k +G)r , (28) e

so that the wave equation (see Eq. (12b)) becomes:

G G

r (GG ) (k+G)[(k+G ) ] ei(k +G)r = e

r
G

ei(k +G )r e .

(29) Always by the same token (multiplying by the proper functions and integrating over whole space), we write:

r (G G ) (k + G) [(k + G ) e ]

G . (30)

We can further simplify this expression by dot-multiplying the equation by e and noting that (using C (A B) = B (C A))
(k + G) [(k + G ) e ] e = (k + G ) e (k + G) e
Therefore, dot-multiplying Eq. (30) by e , we get the nal result: (31)

(k + G ) e (k + G) e

r (G G ) h

r h

(32)

Upon exchanging G and G (transformations: G G , G G, G G ), we obtain

c
2

(k + G) e (k + G ) e

r (G G) h

r h

(33)

which is the relation given in [Joannopoulos et al., 1995, p. 129]. Upon using the same notation, we rewrite Eq. (33) as:
k
G

(G),(G)

(G)

r h

(G)

(34a)

where k
(G),(G)

= r (G G) (k + G) e (k + G ) e .

(34b)

3.1

Matrix form

3.1

Matrix form

We can cast Eq. (33) in matrix form. First, we rewrite the kernel of the operator of Eq. (34b) as: (k + G) e (k + G ) e
= (k + G) (k + G ) [3 e ] [3 e ] . e e
(35)

Remembering that e3 e1 = e2 and e3 e2 = 1 , we can write: e [3 e ] [3 e ] = e e so that we write the operator as: e2 e2 2 e1 e 1 e2 e e1 e 1 , (36)

(G),(G)

= r (G G) (k + G) e (k + G ) e
= r (G G) (k + G) (k + G )
e2 e2 2 e1 e 1 e2 e e1 e 1 , (37)

used in Eq. (34a).

Study of EM waves in Periodic Structures: Photonic Crystals and Negative refraction

Massachusetts Institute of Technology 6.635 lecture notes

Introduction

In the previous class, we have introduced various concepts necessary for the study of EM waves in photonic crystal structures. We shall now use these concepts to explain various results such as: Reconstruction of the permittivity prole. The band diagrams for rectangular and triangular lattices. k-surfaces for various eigenvalues. In particular, we will show an example of how a periodic structure can exhibit k-surfaces typical of a negative refraction material (the concept of k-surface for left-handed materials was rst introduced in the class of February 24, 2003). This class is based on the following two references:: [1] [2] C. Luo, S. G. Johnson, and J. D. Joannopoulos, All-angle negative refraction without negative eective index, Physical Review B, vol. 65, 2002, #201104. J. D. Joannopoulos, R. D. Meade, and J. N. Winn, Photonic Crystals, Molding the Flow of Light, Princeton University Press, 1995.

Retrieving the permittivity

We have presented already the way to calculate the Fourier coecients of the permittivity for cylindrical inclusions, with circular cross-section (see section 5 of March 19 2003 notes). An example was also given in Fig. 7. We shall just briey comment on this. It is known already that a photonic crystal is dened by a lattice (rectangular, triangular, . . . ) and a basis (the shape of the inclusion). Both have to be incorporated in the retrieval of for example. In the treatment presented last time (section 5 of the reference mentioned above) we obtained the Fourier coecients of the permittivity for a specic lattice (square). The result was:

Section 2. Retrieving the permittivity

Yet, the information on the basis is also included in Eq. (1), since both G and fr will depend on it. It is straightforward to show that:  Square lattice: fr = lattice constant). 2 Rc 2  Triangular lattice: fr = . 3 a An illustration for square and triangular lattices is given in Fig. 1 (note that since the permittivity is obtained from its Fourier coecients, a unavoidable Gibbs phenomenon will occur). Rc a
2

f + (1 f ) a r r b (G ) = 2J1 (G Rc ) ( a b )fr
G Rc

if G = 0 elsewhere.

(1)

(where Rc is the radius of the inclusions and a the

(a) Square lattice (fr = 38.48%).

(b) Triangular lattice (fr = 44.43%).

Figure 1: Reconstruction of the permittivity for cylindrical rods of circular crosssection for (a) a square lattice and (b) a rectangular lattice. Other parameters are: a = 1, b = 12, Rc /a = 0.35.

Band diagrams

The purpose of all the mathematical developments presented so far (getting the Fourier coefcients of the permittivity, building the eigensystem, etc) is to eventually obtain the eigenvalues and eigenvectors of the problem. Eigenvalues correspond to dispersion diagrams whereas eigenvectors correspond to the actual eld distributions. We shall limit ourselves here to the consideration of eigenvalues only. Upon solving the eigensystem: r H() = = k 2 H() , r c 1 , () r (2a) (2b)

r (with the condition H() = 0 to reduce the size of the system), we get a set of eigenvalues for each incident k. The band diagrams are then constructed by sweeping all possible k. Because of the periodicity of the medium the all possible k can be reduced to the rst Brillouin zone and, by symmetry, further reduced to the irreducible Brillouin zone. In addition, as we have seen, it is also enough to span the edge of the Brillouin zone since it corresponds to the maximum diraction condition. Again by symmetry, we further reduce the domain to the edge of the irreducile Brillouin zone. This zone will of course depend on the lattice, as is shown in Figs 2 and 3 of the March 19 2003 notes. For both cases, the irreducible Brillouin zone is depicted in Fig. 2.

PSfrag replacements

PSfrag replacements

(a) Square lattice.

(b) Triangular lattice.

Figure 2: Irreducible Brillouin zones (red region) for (a) a square lattice and (b) a triangular lattice. Each region is dened between symmetry points of the crystal.

In order to span the edge of the irreducible Brillouin zone, we therefore need to know the coordinates of the symmetry points of the crystal. It is easy to show that:

Section 3. Band diagrams

Square lattice: (kx = 0, ky = 0) , X (kx = , ky = 0) , a M (kx = , ky = ) . a a Triangular lattice: (kx = 0, ky = 0) , 2 2 K (kx = , ky = ) , 3a 3a 2 M (kx = 0, ky = ) . 3a (4a) (4b) (4c) (3a) (3b) (3c)

Having these limit points, we can sweep k, solve the eigensystem and obtain the eigenvalues. An example is given in Fig. 3.

Square lattice: Rc/a = 0.48 (f=72.3823%), a=1, b=13.

1 TE TM 0.9

0.8

0.7

Frequency a/2 c
PSfrag replacements

0.6

0.5

0.4

0.3

0.2

0.1

(a) Square lattice.

Hexagonal lattice: Rc/a = 0.48 (f=83.5799%), a=1, b=13.

1 TE TM 0.9

0.8

0.7

Frequency a/2 c
PSfrag replacements

0.6

0.5

0.4

0.3

0.2

0.1

(b) Triangular lattice.

Figure 3: Band diagram for TE and TM modes as function of the normalized frequency for (a) a square lattice and (b) a triangular lattice. Notice the absolute band gap for TM modes. Parameters are: a = 1, b = 13, Rc /a = 0.48. Important note: TE modes are here dened as transverse to the axis of the crystal (z axis), therefore with an in-plane electric eld!! Note also that these curves have not yet fully reached convergence.

Section 4. k-surfaces

4
4.1

k-surfaces
Well-chosen example

We can generalize the band diagram of the previous section (which, again, spans only the edge of the Brillouin zone), to the entire zone or even to the entire photonic crystal. The example we shall consider from now on is taken from [1] and the parameters are summarized in Tab. 1. In addition, we shall work with TE modes (remember that in this notation, TE corresponds to an in-plane electric eld). Lattice: Inclusions: Background: Radius : square a =1 b = 12 Rc /a = 0.35

Table 1: Parameters for our example [1].

The band diagram obtained for this case is depicted in Fig. 4.

Square lattice: Rc/a = 0.35 (f=38.4845%), =1, =12.
a b
0.5

0.4

Frequency a/2 c

0.3

0.2

TE 0.1

PSfrag replacements
0 0

Figure 4: Band diagram for TE modes for the structure dened in Tab. 1. The horizontal line is at the frequency corresponding to a change of curvature of the ksurface, and the green line is the light-line shifted to M.

We can also extend the plot from the edge of the irreducible Brillouin zone to the entire structure, for various eigenstates. This is represented in Figs. 5 to 7.

4.2

Negative refraction

From the k-surfaces shown in the gures of this section, we see that there is the potential of negative refraction (the iso-frequency surfaces converge to a point as frequency increases). Yet, we need to make sure that: 1. We can couple to one of these surfaces from free-space. 2. The power is bending on the same side of the normal. Answer to both points is shown in Fig. 4. 1. Light-line: the light-line gives the radius of the k-surface of an EM wave impinging on the crystal from free-space. In order to have a possible coupling, whole (or part) of the free-space k-surface has to be included in one of the k-surfaces of the crystal. The light-line is represented in Fig. 4 by the green line (it is actually the translation of the light-line to M ). Its intersection with say the curve of the rst eigenvalue gives the maximal frequency for which the condition mentioned above (total inclusion of the free-space k-surface into a k-surface of the crystal) is satised. Note also that in order to have a negative refraction, the free-space k-surface has to be included in one of the k-surfaces converging to M and therefore, the actual crystal needs to be rotated by 45 . 2. Power bending: around M , the power is converging to a single point when frequency is increased. Yet, the direction of the power can be deduced from the gradient of the k-curves and therefore, directly from their radius of curvature. Hence, if the radius of curvature is such that the gradient is pointing toward M , the refraction will be negative. The frequency at which the radius of curvature diverge is given by the horizontal line in Fig. 4, which can be obtained by a direct inspection of Fig. 5(c).

As it can be seen, the power can bend on the opposite side of the normal. However, the phase is still propagating forward, which justies the title of the paper: All-angle negative refraction without negative index. We can also examine the second eigenvalue and see if a similar phenomenon can appear. The k-surface has been depicted in Fig. 6(b) and is shown again in Fig. 9 with less curves represented. It is clear from this gure that again, there is a point at which the radius of curvature of the k-curves changes, and energy converges to a single point ( this time) as frequency increases. It is therefore again possible to have a negative index of refraction. Notice that in this case

4.2

Negative refraction

Squ. lattice. Band of eigenvalue nb 1. Rc/a = 0.35 (f=38.4845%), a=1, b=12.

0.25

0.2

0.15

0.1

0.05

0 4000 2000 0 0 2000 4000 4000 2000 2000 4000

(a) Brillouin cell.

(b) All structure.

Squ. lattice. Band of eigenvalue nb 1. Rc/a = 0.35 (f=38.4845%), a=1, b=12. 0.22 160 0.2 140 0.18 120 0.16 0.14 0.12 0.1 60 0.08 40 0.06 20 0.04 0.02 20 40 60 80 100 120 140 160

100

80

(c) 2D k-surface.

Figure 5: k-surfaces for the rst eigenvalue of the system.

Squ. lattice. Band of eigenvalue nb 2. Rc/a = 0.35 (f=38.4845%), a=1, b=12.

0.36

0.34

0.32

0.3

0.28

0.26 4000 2000 0 0 2000 4000 4000 2000 2000 4000

(a) Brillouin cell.

Squ. lattice. Band of eigenvalue nb 2. Rc/a = 0.35 (f=38.4845%), a=1, b=12.

160

0.34

140 0.33 120 0.32 100 0.31 80

60

0.3

40

0.29

20 0.28 20 40 60 80 100 120 140 160

(b) 2D k-surface.

Figure 6: k-surfaces for the second eigenvalue of the system.

however, because of phase matching, power and phase are in directions that make an angle greater than /2. Although it is not necessarily like in a pure left-handed regime, it is still in the regime of left-handed behavior. Fig. 10 can be completed to see this phenomenon.

10

4.2

Negative refraction

Squ. lattice. Band of eigenvalue nb 3. Rc/a = 0.35 (f=38.4845%), =1, =12.

a b

0.45 0.44 0.43 0.42 0.41 0.4 0.39 0.38 4000 2000 0 0 2000 4000 4000 2000 2000 4000

(a) Brillouin cell.

Squ. lattice. Band of eigenvalue nb 3. Rc/a = 0.35 (f=38.4845%), a=1, b=12.

160

0.435

140

0.43

120

0.425

100

0.42

80

0.415

60

0.41

40

0.405

20

0.4

0.395 20 40 60 80 100 120 140 160

(b) 2D k-surface.

Figure 7: k-surfaces for the third eigenvalue of the system.

11

Squ. lattice. Band of eigenvalue nb 2. Rc/a = 0.35 (f=38.4845%), a=1, b=12.

55

0.31

50

45

0.3

40

0.29
35

30

0.28
25

20

0.27

15

10

0.26

10

15

20

25

30

35

40

45

50

55

0.25

Figure 9: k-surface for the second eigenvalue.

12

4.2

Negative refraction

Squ. lattice. Band of eigenvalue nb 2. Rc/a = 0.35 (f=38.4845%), a=1, b=12. 1.2 34 1 32 0.8

30

0.6

0.4 28 0.2 26

24

0.2

0.4 22 0.6 22 24 26 28 30 32 34

Figure 10: k-surface for the second eigenvalue for the specic frequency where the radius of curvature is positive.