Вы находитесь на странице: 1из 12

# 10 - 1 MMSE Estimation

S. Lall, Stanford

2011.02.02.01

S. Lall, Stanford

2011.02.02.01

10 - MMSE Estimation
Estimation given a pdf
Minimizing the mean square error
The minimum mean square error (MMSE) estimator
The MMSE and the mean-variance decomposition
Example: uniform pdf on the triangle
Example: uniform pdf on an L-shaped region
Example: Gaussian
Posterior covariance
Bias
Estimating a linear function of the unknown
MMSE and MAP estimation

10 - 2 MMSE Estimation

Suppose x :

## One can predict or estimate the outcome as follows

Given cost function c : Rn Rn ! R
pick estimate x to minimize E c(x, x)

c(x, x) = kx

xk2

## Then the mean square error (MSE) is

E kx

xk

kx

xk2px(x) dx

10 - 3 MMSE Estimation

S. Lall, Stanford

2011.02.02.01

## Minimizing the MSE

Lets find the minimum mean-square error (MMSE) estimate of x; we need to solve
min E kx
x

xk2

We have
E kx

xk2 = E (x

x)T (x

x)

= E xT x

2
xT x + xT x

= Ekxk2

2
xT E x + xT x

## Dierentiating with respect to x gives the optimal estimate

xmmse = E x

10 - 4 MMSE Estimation

S. Lall, Stanford

## The MMSE estimate

The minimum mean-square error estimate of x is
xmmse = E x

E kx

since E kx

xmmsek2 = E kx

## xmmsek2 = trace cov(x)

E xk2

2011.02.02.01

10 - 5 MMSE Estimation

S. Lall, Stanford

2011.02.02.01

## The mean-variance decomposition

We can interpret this via the MVD. For any random variable z, we have
E kzk2 = E kz

Applying this to z = x

E zk2 + kE zk2

x gives

E kx

xk2 = E kx

x
E xk2 + k

E xk2

## The first term is the variance of x; it is the error we cannot remove

The second term is the bias of x; the best we can do is make this zero.

10 - 6 MMSE Estimation

S. Lall, Stanford

2011.02.02.01

## Suppose x, y are random variables, with joint pdf p(x, y).

We measure y = ymeas.

2
1

0
1
2

: Rm ! R n .

3
2

## which minimizes the cost function

J = E k (y)

xk2

10 - 7 MMSE Estimation

S. Lall, Stanford

2011.02.02.01

S. Lall, Stanford

2011.02.02.01

Notation
Well use the following notation.
py is the marginal or induced pdf of y
p (y) =
y

p(x, y) dx

## p|y is the pdf conditioned on y

p|y (x, y) =

p(x, y)
py (y)

10 - 8 MMSE Estimation

## The MMSE estimator

The mean-square-error conditioned on y is econd(y), given by
econd(y) =

k (y)

## Then the mean square error J is given by

J = E econd(y)

because
J=
=

Z Z
Z

k (y)

xk2 p(x, y) dx dy

py (y) econd(y) dy

10 - 9 MMSE Estimation

S. Lall, Stanford

2011.02.02.01

S. Lall, Stanford

2011.02.02.01

## The MMSE estimator

We can write the MSE conditioned on y as
econd(ymeas) = E k (y)

xk2 | y = ymeas

## For each ymeas, we can pick a value for (ymeas)

So we have an MMSE prediction problem for each ymeas

10 - 10

MMSE Estimation

## The MMSE estimator

Apply the MVD to z = (y)

x conditioned on y = w to give
xk2 | y = w

econd(w) = E k (y)
= E kx

h(w)k2 | y = w + k (w)

## where h(w) is the mean of x conditioned on y = w

h(w) = E(x | y = w)
To minimize econd(w) we therefore pick
(w) = h(w)

h(w)k2

10 - 11

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

S. Lall, Stanford

2011.02.02.01

## The error of the MMSE estimator

With this choice of estimator
econd(w) = E kx

h(w)k2 | y = w

= trace cov(x | y = w)

10 - 12

MMSE Estimation

## Summary: the MMSE estimator

The MMSE estimator is
mmse (ymeas )

= E(x | y = ymeas)

## econd(ymeas) = trace cov(x | y = ymeas)

We often write

xmmse =

mmse (ymeas )

econd(ymeas) = E kx

xmmsek2 y = ymeas

## The estimate only depends on the conditional pdf of x | y = ymeas

The means and covariance are those of the conditional pdf
The above formulae give the MMSE estimate for any pdf on x and y

10 - 13

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

## Example: MMSE estimation

(x, y) are uniformly distributed on the triangle A.
i.e., the pdf is
(
2 if (x, y) 2 A
p(x, y) =
0 otherwise

0.5

0
0

0.5

## the conditional distribution of x | y = ymeas is uniform on [ymeas, 1]

the MMSE estimator of x given y = ymeas is
xmmse = E(x | y = ymeas) =

1 + ymeas
2

## the conditional mean square error of this estimate is

E kx

10 - 14

xmmsek2 y = ymeas =

1
(ymeas
12

1)2

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

1

0.5

## The mean square error is

0

E k

mmse (y)

xk2 = E econd(y)

x=0

x=0

1
24

p(x, y) econd(y) dy dx

y=0

1
(y
y=0 6
x

1)2 dy dx

0.5

10 - 15

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

## Example: MMSE estimation

(x, y) are uniformly distributed on the L-shaped
region A, i.e., the pdf is
(
4
if (x, y) 2 A
p(x, y) = 3
0 otherwise

0.5

0
0

## the MMSE estimator of x given y = ymeas is

xmmse = E(x | y = ymeas) =

(1

## the mean square error of this estimate is

E kx

10 - 16

xmmsek

y = ymeas =

if 0 ymeas

2
3
4

if

(1

12
1
48

1
2

0.5

1
2

< ymeas 1

if 0 ymeas
if

1
2

1
2

< ymeas 1

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

1

0.5

## The mean square error is

0

E k

mmse (y)

xk2 = E econd(y)

0.5

p(x, y) econd(y) dy dx

x=0

1
16

1
2

1
p(x, y) dy dx +
y=0 12

x= 21

1
2

y= 12

1
p(x, y) dy dx
48

10 - 17

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

x
N (, ) where
Suppose
y

x
=
= Tx
y
xy

0.35
0.3
0.25
0.2
0.15

xy

0.1

0.05
0

## We know that the conditional density of

x | y = ymeas is N (1, 1) where
1 = x +
1

(ymeas

xy

xy

1 T
y
xy

3
2
1

y )

0
1
2
3

10 - 18

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

## MMSE estimation for Gaussians

The MMSE estimator when x, y are jointly Gaussian is
mmse (ymeas )

= x +

econd(ymeas) = trace

xy

(ymeas

xy

y )

1 T
xy

## The conditional MSE econd(y) is independent of y; a special property of Gaussians

Hence the optimum MSE achieved is
E k

mmse (y)

xk2 = trace

## The estimate xmmse is an ane function of ymeas

xy

1 T
xy

10 - 19

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

S. Lall, Stanford

2011.02.02.01

Posterior covariance
Lets look at the error z = (y)

x. We have

## We use this to find a confidence region for the estimate

cov(x | y = ymeas is called the posterior covariance

10 - 20

MMSE Estimation

x
Here
N (0, ) with
y

2.8 2.4
=
2.4 3

3
2

We measure

ymeas = 2

0
1

mmse (ymeas )

Let 2 = QF

= 12 221ymeas
= 0.8ymeas
= 1.6

3
3

xy

yx

= 0.88

## Prob |x 1.6| y = ymeas = 0.9

10 - 21

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

S. Lall, Stanford

2011.02.02.01

Bias
The MMSE estimator is unbiased; that is the mean error is zero.
E

10 - 22

mmse (y)

x =0

MMSE Estimation

## Estimating a linear function of the unknown

Suppose
p(x, y) is a pdf on x, y

q = Cx is a random variable
we measure y and would like to estimate q
The optimal estimator is
qmmse = C E x | y = ymeas
Because E q | y = ymeas = C E x | y = ymeas
The optimal estimate of Cx is C multiplied by the optimal estimate of x
Only works for linear functions of x

10 - 23

MMSE Estimation

S. Lall, Stanford

2011.02.02.01

## MMSE and MAP estimation

Suppose x and y are random variables, with joint pdf f (x, y)
The maximum a posterior (MAP) estimate is the x that maximizes
h(x, ymeas) = conditional pdf of x | (y = ymeas)

## The MAP estimate also maximizes the joint pdf

xmap = arg max f (x, ymeas)
x

3
2

## When x, y are jointly Gaussian, then the peak of

the conditional pdf is the conditional mean.

1
0
1

## i.e., for Gaussians, the MMSE estimate is equal

to the MAP estimate.

2
3