Вы находитесь на странице: 1из 61

In Einsteins Footsteps: From Brownian Motion to Fractal

Flights
V. Balakrishnan
Indian Institute of Technology Madras

Stella Maris College, February 12, 2015 p.1/46

Anni mirabili

Truly miraculous yearswatersheds in our understanding


of Nature:
1665-66: Isaac Newton (1642-1727), then in his early
20s, discovered the system of the world, and much else
besides.
1859: Charles Darwin (1809-1882), then 50 years old,
published his Theory of Evolution. He made his voyage
on HMS Beagle when he was 22 years old.
1905: Albert Einstein (1879-1955), then just 26 years old,
published his seminal papers that revolutionized our
understanding of the physical universe in major ways in
three different directions: statistical physics, quantum
physics, relativity.
Here we will focus on aspects of statistical physics.

Stella Maris College, February 12, 2015 p.2/46

Einstein in 1905

Stella Maris College, February 12, 2015 p.3/46

Seminal papers by Einstein on Brownian motion


On the motion required by the molecular kinetic theory of heat of
particles suspended in fluids at rest, Annalen der Physik Vol.
17, pp. 549-560. Received on May 11, 1905.

Einstein completed his doctoral thesis (on Brownian


motion and related topics) on 30 April 1905; it was
submitted to the University of Zrich on 20 July 1905.
Einsteins second paper on Brownian motion was
On the theory of Brownian motion,

received by Annalen der Physik on 19 December 1905


and published in Vol. 19, p. 371 (1906).

Stella Maris College, February 12, 2015 p.4/46

Brownian Motion

Microscopic, micron-sized particles (such as pollen


grains), suspended in a liquid, show erratic, sudden,
irregular movements as though they were being
kicked around in a random fashion.
The phenomenon was known from the early 19th
century. It was initially thought to be due to some living
force or vital force.
The motion was first studied systematically in 1827-28
by the botanist Robert Brown, after which it it became
clear that no vital forces were involved.
By the 1850s the motion was believed to be caused
either by internal motions in the fluid, or by collisions
with fluid molecules from different directions.
Stella Maris College, February 12, 2015 p.5/46

A random curve

A smoothed-out, very schematic version of Brownian


motion. True Brownian motion is much more jagged than
this curve.
Stella Maris College, February 12, 2015 p.6/46

Einstein recognised that the random, fluctuating force due


to collisions by the molecules of the liquid could cause
such irregular motion.
But he was not convinced that the precision of the
experimental observations available at the time was
good enough to assert that this was so.
This is essentially why the phrase Brownian motion
does not appear in the title of his first paper on the
subject, which was On the motion required by the molecular
kinetic theory of heat of particles suspended in fluids at rest.

But in the text of that paper, he says:


It is possible that the motions to be discussed here are
identical with so-called Brownian molecular motion ...
Stella Maris College, February 12, 2015 p.7/46

Determination of Avogadros number

Einsteins aim was far more fundamental: to show that,


if the predictions of the theory could be experimentally
verified, then
... an exact determination of actual atomic sizes
becomes possible.
The determination of atomic sizes and of Avogadros
number NA were strongly motivating and recurring and
themes in Einsteins early work on statistical physics.
Einstein returned to the determination of NA again and
again, proposing several independent methods to
estimate this fundamental quantity.
One of his motivations was to establish beyond all
doubt the atomic nature of matter.
Stella Maris College, February 12, 2015 p.8/46

Identification as a diffusion process

Einstein gave an ingenious analysis of Brownian motion


based on physical arguments involving thermodynamic
equilibrium, osmotic pressure, etc. Using these, he
carefully avoided errors that could have arisen from
what we now recognise as mathematical subtleties in
the behaviour of certain random processes.
Markov processes were introduced in the theory of
probability in 1906, by the mathematician A. A. Markov.
A year before that, Einstein had essentially recognized
that Brownian motion was a special kind of Markov
process, now called a diffusion process.
Even earlier, in 1900, Louis Bachelier had studied such
processes in connection with financial fluctuations!
Stella Maris College, February 12, 2015 p.9/46

The diffusion equation

Einstein correctly identified the distinct time scales in


the problem of a micron-sized object being buffeted
incessantly and randomly by much smaller molecules.
This helped him write down the equation governing the
probability distribution of the position of the larger
particle.
This is the famous diffusion equation,
2 p(x , t)
p(x , t)
=D
.
2
t
x

Here D is the diffusion coefficient.


This is a p.d.e. of 2nd order in x, but 1st order in time,
reflecting the irreversibility of diffusion.
Stella Maris College, February 12, 2015 p.10/46

The fundamental solution

Einstein also gave the fundamental Gaussian solution to


the diffusion equation.
If the particle starts from the origin x = 0 at t = 0, then
1

p(x , t) =
e
4Dt

x2 /4Dt

for any

t > 0.

Crucial characteristic feature of diffusion: the average


value of the square of the distance travelled in any given
direction by a Brownian particle in a time interval t is
proportional to t, rather than t2 :

2
x (t) = 2Dt = rms displacement t
Stella Maris College, February 12, 2015 p.11/46

How the Gaussian solution evolves in time

p(x, t)
0.4

0.3

0.2

0.1

x
Stella Maris College, February 12, 2015 p.12/46

How the Gaussian solution evolves in time

p(x, t)
0.4

0.3

0.2

0.1

x
Stella Maris College, February 12, 2015 p.12/46

How the Gaussian solution evolves in time

p(x, t)
0.4

0.3

0.2

0.1

x
Stella Maris College, February 12, 2015 p.12/46

How the Gaussian solution evolves in time

p(x, t)
0.4

0.3

0.2

0.1

x
Stella Maris College, February 12, 2015 p.12/46

How the Gaussian solution evolves in time

p(x, t)
0.4

0.3

0.2

0.1

The width of the Gaussian increases linearly with time.


Stella Maris College, February 12, 2015 p.12/46

Einsteins deep insight: concentrate on the mean squared


displacement, rather than the instantaneous velocity of
the particle, as the quantity to be studied and
measured.
Modern generalizations of diffusion use the same
criterion: If x2 (t) t , then
< 1 = sub-diffusive
= 1 = diffusive
1 < < 2 = super-diffusive
= 2 = ballistic
> 2 = super-ballistic

Physical examples of all the cases above exist.


Stella Maris College, February 12, 2015 p.13/46

The Einstein-Sutherland formula

Using an extremely ingenious argument that


combined thermodynamics with dynamics, Einstein
derived the formula
RT
D=
,
6NA a

where D is the diffusion coefficient, T is the absolute


temperature of the liquid, is its viscosity, R is the gas
constant, and a is the radius of the Brownian particle
(assumed to be spherical).
(Coincidentally, this formula was also obtained
independently by Sutherland (Australia) in March 1905!)
Historically, this is the first example of a
fluctuation-dissipation relationship.
Stella Maris College, February 12, 2015 p.14/46

Experimental verification: Perrin

RT
x (t) =
t.
3NA a
2

Hence NA can be determined by observing Brownian


motion and measuring the mean squared displacement
of a Brownian particle over different intervals of time!
The predictions of Einsteins theory were checked by
Jean Perrin and his students in a series of experiments
from 1908 to 1914, and they were all confirmed with an
until then unmatched precision.

Stella Maris College, February 12, 2015 p.15/46

With this successful explanation of Brownian motion,


resistance to the reality of atoms essentially ended.
Marian von Smoluchowski (Poland) and Paul Langevin
(France) did pioneering and extremely significant work
on the problem of Brownian motion and related matters
concerning deep issues such as macroscopic
irreversibility, around the same time as Einstein.
Langevins approach is the basis on which
non-equilibrium statistical physics and the applications
of random (or stochastic) processes to physical
problems rest. We will return to this aspect
subsequently.

Stella Maris College, February 12, 2015 p.16/46

Random walk model

Brownian motion can be regarded as the limiting case


of a random walk in which the step length and the time
between steps become vanishingly small.
The simplest random walks are on regular lattices, with
the walker jumping from any site to a nearest-neighbour
site at each time step.
A little more generally, the random walker takes a step
in any random direction, but each step has the same
length.
Note that the argument that follows is actually valid in
any spatial dimension.

Stella Maris College, February 12, 2015 p.17/46

A simple random walk

R1

Stella Maris College, February 12, 2015 p.18/46

A simple random walk

R1

R2

Stella Maris College, February 12, 2015 p.18/46

A simple random walk

R3
R1

R2

Stella Maris College, February 12, 2015 p.18/46

A simple random walk

R3

rN

R1

R2

Stella Maris College, February 12, 2015 p.18/46

A simple random walk

R3

rN

R1

R2

What is mean squared end-to-end distance after N random


steps?
There exists a remarkably simple answer to this question!

Stella Maris College, February 12, 2015 p.18/46

The solution

One has

1 + R
2 + R
N .
rN = R

2
rN

= rN rN =

N

i=1

Ri2

j .
i R
R

i=j

Stella Maris College, February 12, 2015 p.19/46

The solution

One has

1 + R
2 + R
N .
rN = R

2
rN

= rN rN =

N

i=1

Ri2

j .
i R
R

i=j

But if each segment has constant length l, each Ri2 = l2 .



2
2
2
cos ij 
rN  = N l + l
Rj
i=j

ij
Ri

Stella Maris College, February 12, 2015 p.19/46

The solution

One has

1 + R
2 + R
N .
rN = R

2
rN

= rN rN =

N


Ri2

i=1

j .
i R
R

i=j

But if each segment has constant length l, each Ri2 = l2 .



2
2
2
Rj
cos ij 
rN  = N l + l
i=j

Another possible configuration with


 = cos R
cos ij
j
ij
= cos ij  = 0 , =

2 1/2
rN


ij
ij

Ri

= l N.

Perhaps the most important square root in the universe!


Stella Maris College, February 12, 2015 p.19/46

The continuum limit of a random walk

The steps do not have to be of the same size, nor do


the steps have to be taken at regular intervals.
Normal Brownian motion is the continuum limit of a
random walk in which the random time interval between
steps is distributed exponentially, and the individual
steps can themselves be random in both size and
direction, but the variance of the step size is finite.
E.g., the simulation of Brownian motion that follows
1/L3.8 for a step of length
corresponds to a probability

L. (Hence the variance
dL L2 /L3.8 , which
converges to a finite value.)

Stella Maris College, February 12, 2015 p.20/46

Random walk simulations

Stella Maris College, February 12, 2015 p.21/46

Random walk simulations

Stella Maris College, February 12, 2015 p.21/46

Brownian motion as a stochastic process

Standard Brownian motion is also called a Wiener


process.
If w(t) is a Wiener process, then dw(t) = (t) dt, where
(t) is a Gaussian white noise: it is a stationary,
Gaussian, -correlated Markov process, satisfying
(t) = 0,

(t) (t ) = (t t ).

Brownian motion is therefore the integral of white


noise. It is a non-stationary Gaussian Markov process
with stationary increments.
Generalization: a diffusion process is one that is driven
by white noise, i.e., the solution of the stochastic
differential equation dx(t) = f (x, t) dt + g(x, t) dw(t).
Stella Maris College, February 12, 2015 p.22/46

Some peculiarities of a Brownian trail

The sample path of a particle undergoing Brownian


motion in the strict mathematical sense (i.e., a Wiener
process) is not an ordinary, smooth curve.
Its trajectory is a continuous, but extremely jagged,
curve: it is a random fractal.
Its instantaneous velocity is unbounded. (Recall
Einsteins emphasis on the mean square displacement
rather than the instantaneous velocity.)
Almost all Brownian paths are nowhere differentiable.
It is statistically self-similar: its degree of jaggedness
remains unchanged under arbitrarily large magnification
of any portion of the curve.

Stella Maris College, February 12, 2015 p.23/46

A sample of Brownian motion in one dimension


x(t)

Stella Maris College, February 12, 2015 p.24/46

How far can Brownian motion really go?

Given that x(0) = 0, at any t > 0 (however small), x(t)


can arbitrarily large, in principle.
But there exists a rigorous result, the Law of the Iterated
Logarithm: Let h(t) = [2t log log (1/t)]1/2 . Then, as t 0,


Prob

lim sup w(t)/h(t) = 1 = 1.




Similarly, as t , Prob lim sup w(t)/H(t) = 1 = 1
where H(t) = [2t log log t]1/2 .
There is also the interesting time-reversal property
t w(1/t) = w(t)

(t > 0)

in distribution.
Stella Maris College, February 12, 2015 p.25/46

Illustrating the law of the iterated logarithm


h(t)
(1+ ) h(t)

11111111111111
00000000000000
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
(1 ) h(t)
00000000000000
11111111111111
00000000000000
11111111111111

t
11111111111111
00000000000000
00000000000000
11111111111111
(1 ) h(t)
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
00000000000000
11111111111111
(1+ ) h(t)

The brownian particle is located infinitely often in the


shaded region. Hence it crosses the origin infinitely
often.
The instants of zero crossings form a random fractal
set, of fractal dimensionality 12 .
Stella Maris College, February 12, 2015 p.26/46

Brownian trails

The curve is space-filling, in the following sense:


If the Brownian motion is restricted to an infinite line or
an infinite plane, then every point of the line or plane is
sure to be visited infinitely often by the particle as
t . (All roads lead to Rome in 1D and 2D!)
However, the mean time between successive visits is
formally infinite!
If the Brownian motion occurs in three-dimensional
space, the so-called fractal dimension of its trail is 2, and
not 1 as would be expected of an ordinary regular
curve.
This remains true for Brownian motion in any spatial
dimension D 2.
Stella Maris College, February 12, 2015 p.27/46

Anomalous diffusion, Levy walks, fractal flights

A long-tailed waiting time distribution between steps


can lead to anomalous or sub-diffusive behaviour.
When the variance of the step size diverges, the random
walk can become super-diffusive.
In either case, the Gaussian solution is no longer valid,
and the probability distribution of the random walker is
generally a Levy distribution.
The paths of such flights are also statistically self-similar
random fractals, but with different fractal dimensions.
For instance, the simulations that follow represent a
Levy flight with a step-size distribution 1/L2.2 . (Hence

dL L2 /L2.2 , which diverges.)


the variance
Stella Maris College, February 12, 2015 p.28/46

Levy walk simulations

Stella Maris College, February 12, 2015 p.29/46

Levy walk simulations

Stella Maris College, February 12, 2015 p.29/46

Langevin equation for diffusion

The diffusion equation used earlier assumes that the


velocity of a Brownian particle is essentially a white noise.
This is unphysical on two counts: The velocity is certainly
finite at all times, and it cannot be completely uncorrelated
(or random) it must have some memory.
A better model is given by the equations of motion
dx(t)
= v(t)
dt

and

dv(t)
m
= mv(t) + (t),
dt

where is the friction coefficient (related to the viscosity of


the fluid) and is a white noise. The latter satisfies



(t) = 0 , (t) (t ) = 2mkB T (t t  ) .
Stella Maris College, February 12, 2015 p.30/46

The forces mv and (t) represent the systematic


and truly random parts of the force acting on the
diffusing particle as a result of its collisions with the
other molecules of the fluid, which is in thermal
equilibrium at absolute temperature T .
The strength of the white noise, 2mkB T , is directly
proportional to the friction coefficient . This is a
fundamental example of a fluctuation-dissipation relation.
Physical significance: the very same random molecular
collisions that give rise to the fluctuations in the velocity
of the diffusing particle are also responsible for
maintaining thermal equilibrium by damping out any
random excursions away from the average values.
Stella Maris College, February 12, 2015 p.31/46

The Fokker-Planck equation

The Langevin equation implies a certain partial differential


equation for the probability density function (x, v, t) of the
diffusing particle in phase space. This master equation is

kB T 2

.
+v
=
(v) +
2
t
x
v
m v

The fundamental solution to this equation is a Gaussian in


the two variables x and v .
If we integrate over all values of the position x, we get the
probability density function of the velocity of the diffusing
particle. This satisfies the Rayleigh equation

kB T 2 p
p
= (vp) +
.
2
t
v
m v
Stella Maris College, February 12, 2015 p.32/46

Suppose the initial velocity is some given value v0 . Then


The initial condition on p is
p(v, 0) = (v v0 ).

Using this, we can calculate the mean velocity and the


mean squared velocity as functions of the time t:
v(t) = v0 et = lim v(t) = 0
t

for any v0 .

This shows how thermal equilibrium is maintained with the


help of the friction in the system: the dissipation that is always
present in a system helps damp out the fluctuations and restore the
system to equilibrium.

Stella Maris College, February 12, 2015 p.33/46

Similarly, the mean squared velocity is found to be




kB T
kB T
2
2
v (t) =
e2t .
+ v0
m
m
This shows exactly how v 2 (t) approaches the equilibrium
value of the mean squared velocity as t .
The variance of the velocity is therefore

2 k T

B
2t
2
v (t) v(t) =
1e
.
m

Stella Maris College, February 12, 2015 p.34/46

The Ornstein-Uhlenbeck distribution

Given the mean and the variance, we can write down the
explicit form of the probability density itself, because the
latter is a Gaussian.
This is the Ornstein-Uhlenbeck distribution



12
2
m
m(v v0 et )
p(v , t | v0 ) =
exp
.
2t
2t
2kB T (1 e
)
2kB T (1 e
)
As t , this tends to the equilibrium or Maxwellian
distribution

12


2
m
mv
p(v, ) =
exp
.
2kB T
2kB T
Stella Maris College, February 12, 2015 p.35/46

The time evolution of the velocity distribution

p(v,t)
(vv0)

1.2

0.8

0.4

v0

v
Stella Maris College, February 12, 2015 p.36/46

The time evolution of the velocity distribution

p(v,t)
(vv0)

1.2

0.8

0.4

v0

v
Stella Maris College, February 12, 2015 p.36/46

The time evolution of the velocity distribution

p(v,t)
(vv0)

1.2

0.8

0.4

v0

v
Stella Maris College, February 12, 2015 p.36/46

The time evolution of the velocity distribution

p(v,t)
(vv0)

1.2

0.8

p(v,

)
0.4

v0

v
Stella Maris College, February 12, 2015 p.36/46

Velocity autocorrelation function

How rapidly does the velocity of a diffusing particle lose the


memory of its initial value?
This is given by the velocity autocorrelation function,
C(t) = v(0)v(t) = (kB T /m)et
C(t)
1

0.8

0.6

0.4

0.2

t
Stella Maris College, February 12, 2015 p.37/46

Mean squared displacement of a diffusing particle

We
 2 saw
 that the diffusion equation predicted that
x (t) = 2Dt for all t 0. The Langevin model refines this
result. A more accurate formula is


 2D

t
x (t) =
t 1 + e
,

where

kB T
.
D=
m

Thus the diffusion approximation is valid for times t


1 ,
which is called the velocity correlation time. This is
typically of the order of microseconds in liquids under
normal conditions.
Note basic relationship between the transport and
dissipation coefficients D and : this is another form of
the fluctuation-dissipation relation.
Stella Maris College, February 12, 2015 p.38/46

Variation of mean squared displacement with time


2

x (t)
0.8

0.6

0.4

0.2

slope = 2D
0

Stella Maris College, February 12, 2015 p.39/46

A fundamental model for randomness

Brownian motion and its generalizations have become a


paradigm for random motion, with a staggering variety of
applications for instance, in
polymer dynamics, colloidal suspensions, neutrons in a
nuclear reactor, stock market fluctuations, resistor
networks, dynamic friction in star clusters in galaxies,
chemical reaction dynamics, the dynamics of
sand-piles, chromatography, animal herds, swarming
behaviour, cloud shapes, Monte Carlo techniques in
computer simulation, ....
to name just a few!

Stella Maris College, February 12, 2015 p.40/46

Vast ramifications

The ramifications of Brownian motion in diverse areas


of both mathematics and physics are just as
unexpected and astounding. For instance:
The theory of stochastic diferential equations and
stochastic calculus have their origins in the early work
on Brownian motion and the Wiener process.
The Gaussian solution written down above leads, via
the so-called Wiener measure, to generalizations of the
heat kernel to arbitrary Riemannian manifolds, to the
Feynman-Kac formula, the path integral formulation of
quantum mechanics, and then on to the modern method
of quantization in quantum field theory.
Stella Maris College, February 12, 2015 p.41/46

Bibliography on Einstein in 1905

A. Pais, Subtle is the Lord... The Science and the Life of Albert
Einstein, Oxford University Press, Oxford, 1982.
J. Stachel, Einsteins Miraculous Year: Five Papers that Changed
the Face of Physics, Princeton University Press, Princeton,
1998.
J. S. Rigden, Einstein 1905: The Standard of Greatness,
Harvard University Press, Cambridge, Mass., 2005.
Numerous websites (of varied historical and technical
accuracy !) also exist.

Stella Maris College, February 12, 2015 p.42/46

Lessons from Einsteins achievements

What can we students of physics learn on general grounds


from Einsteins revolutionary achievements?
We learn that
it is important to ask questions, no matter how peculiar
or silly they may appear to be at first sight, to get the
fundamentals as clear as possible;
it is most important to acquire a very clear
understanding of basic principles and issues.

Stella Maris College, February 12, 2015 p.43/46

Lessons...

We also learn that


it is important to appreciate the vast range of physical
parameters on which natural phenomena occur, in
contrast with the extremely restricted range in which our
everyday macroscopic phenomena occur; and, as a
consequence,
it is necessary to overcome pre-conceived notions that
are not substantiated by the facts; and to avoid
confusing such notions with physical intuition or
understanding.

Stella Maris College, February 12, 2015 p.44/46

Nature operates on an incredibly vast scale!

Mass in kilograms:
Human perception: 104 to 104
Nature: 1030 to > 1040
Length in metres:
Human perception: 104 to 104
Nature: 1035 to > 1026
Time in seconds:
Human perception: 101 to 107
Nature: 1042 to > 1017

Stella Maris College, February 12, 2015 p.45/46

What should be ones motivation?

Physics is a relatively old science. Its modern version is


nearly 400 years old. Much thought has been given to
the underlying basic principles and phenomena, and a
fairly cogent structure has emerged.
Therefore the motivation should not be to make a
revolutionary discovery or to overthrow Newton,
Einstein, ... . Nor should it be to get the Nobel Prize
or other high awards and prizes.
To quote Landau, in his Plain talk to students of physics:
One must never work for the sake of ulterior motives,
for fame, with an aim at making a great discovery
nothing will come of it anyway.
The ultimate motive should be to understand with honesty
the way nature works.
Stella Maris College, February 12, 2015 p.46/46

Вам также может понравиться