Ergodic Processes

Tobias Oechtering

KTH Royal Institute of Technology,

EECS, Devision of Information
Science and Engineering
Stockholm, Sweden

Course: EQ1220 Signal Theory

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Ergodicity - Motivation

Ensemble averages (mean, acf) are often needed for the system design (filter,
estimators, etc).
⇒ Estimation of those is an important task, but how?

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Ergodicity - Motivation

Ensemble averages (mean, acf) are often needed for the system design (filter,
estimators, etc).
⇒ Estimation of those is an important task, but how?

Intuitively, the property of ergodicity describes processes where ensemble averages

can be estimated from time averages from almost all realizations of a process.
formal definition requires measure-theoretic arguments
averaging over time should be sufficiently long so that initial states/particularities of
a realization are forgotten, i.e.,
general definition in particular allows some non-stationarity that vanishes

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Ergodicity - Motivation

Ensemble averages (mean, acf) are often needed for the system design (filter,
estimators, etc).
⇒ Estimation of those is an important task, but how?

Intuitively, the property of ergodicity describes processes where ensemble averages

can be estimated from time averages from almost all realizations of a process.
formal definition requires measure-theoretic arguments
averaging over time should be sufficiently long so that initial states/particularities of
a realization are forgotten, i.e.,
general definition in particular allows some non-stationarity that vanishes

Ergodicity is another fundamental property of stochastic processes besides

stationarity, i.e.,
Concept of ergodicity does not require stationarity although in engineering
contexts/textbooks it is often introduced assuming stationarity to simplify exposition!
KTH course: EQ1220 Signal Theory c Tobias Oechtering
Ergodicity - General Definition

τ-invariant
An event F is said to be τ-invariant if x(t) ∈ F implies that x(t + τ) ∈ F.

Example: Discrete-time random process event F consisting of all binary sequences

having a limiting relative frequency of 1s of p.

Ergodicity
A random process is ergodic if for any τ all τ-invariant events have probability one or
zero.
Remark: It can be shown that an iid process is ergodic.

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Ergodicity - General Definition

τ-invariant
An event F is said to be τ-invariant if x(t) ∈ F implies that x(t + τ) ∈ F.

Example: Discrete-time random process event F consisting of all binary sequences

having a limiting relative frequency of 1s of p.

Ergodicity
A random process is ergodic if for any τ all τ-invariant events have probability one or
zero.
Remark: It can be shown that an iid process is ergodic.

Definition of ergodicity is somewhat difficult! If we additionally assume (weakly)

stationary, then we can have a convenient constructive definition based on the
desired operational property (time average ⇒ ensemble average)

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Problem of Interest

Let xi (n) be i-th realization of process X(n).

Sample mean:
N
(i) 1 X
m̂x = xi (n)
2N + 1
n=−N

Ensemble mean: Z ∞
mX (n) = E{X(n)} = x fX(n) (x) dx
−∞

Question
When does the sample mean converge to the ensemble mean?

Think about how to approach the question! E.g., convergence in what sense?

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Partially ergodic wrt the mean

Approach: Evaluate the asymptotic stochastic performance of the estimator!

Define new RV that describes the estimator as follows

N
1 X
M̂X (N) = X(n)
2N + 1
n=−N

For a weakly stationary process X(n) we have E{M̂X (N)} = mX (estimator is unbiased)!

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Partially ergodic wrt the mean

Approach: Evaluate the asymptotic stochastic performance of the estimator!

Define new RV that describes the estimator as follows

N
1 X
M̂X (N) = X(n)
2N + 1
n=−N

For a weakly stationary process X(n) we have E{M̂X (N)} = mX (estimator is unbiased)!

Definition: Partially ergodic with respect to the mean

The weakly stationary stochastic process X(n) is partially ergodic with respect to the
mean mX if n o
Var(M̂X (N)) = E (M̂X (N) − mX )2 → 0 as N → ∞.

Convergence in the mean square sense (quadratic error → 0 as N → ∞).

KTH course: EQ1220 Signal Theory c Tobias Oechtering
 2N
2N+1−|k|
Claim: Var(M̂X (N)) =
P
(2N+1)2
CovX (k) with
k=−2N
CovX (k) = E{(X(n) − mX )(X(n − k) − mX )}

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Example: IID sequence

iid (independent identical distributed) RV X(n, φ) = An (φ)

P{An = +1} = P{An = −1} = 0.5
output independently changes with n

N
RV M̂X (N) = 1 P
2N+1 X(n): Mean:
n=−N

mM̂X (N) = mX = mAn = 0

Variance:
1
Var(M̂X (N)) = E{(M̂X (N) − mX )2 } = ... = →0
2N + 1
as N → ∞ ⇒ Process X(n, φ) is ergodic wrt the mean.

KTH course: EQ1220 Signal Theory c Tobias Oechtering

To be shown: E{(M̂X (N) − mX )2 } = 1
2N+1

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Example: Battery holder
Measure output voltage of battery holder: Y(n, φ) = B(φ)
B(φ) battery polarity: P{B = +1} = P{B = −1} = 0.5
output is constant - does not change with n

N
RV M̂Y (N) = 1 P
2N+1 Y(n):
n=−N

P{M̂Y (N) = +1} = P{M̂Y (N) = −1} = 0.5 ∀N

Mean:
mM̂Y (N) = mY = mB = 0
Variance:
Var(M̂Y (N)) = E{(M̂Y (N) − mY )2 }
= P{M̂Y (N) = +1}(1 − 0)2 + P{M̂Y (N) = −1}(−1 − 0)2 = 1

Var(M̂Y (N)) = 1 6→ 0 ⇒ Process Y(n, φ) is not ergodic wrt the mean.

KTH course: EQ1220 Signal Theory c Tobias Oechtering
Partially ergodic wrt to acf

X(n) weakly stationary stochastic process with realization x(n)

N
1 X
r̂(k, N) = x(n + k)x(n)
2N + 1
n=−N

Question: When does r̂(k, N) converge to rX (k) as N → ∞?

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Partially ergodic wrt to acf

X(n) weakly stationary stochastic process with realization x(n)

N
1 X
r̂(k, N) = x(n + k)x(n)
2N + 1
n=−N

Question: When does r̂(k, N) converge to rX (k) as N → ∞?

Same procedure:
Define RV R̂(k, N).
R̂(k, N) denotes an unbiased estimator, i.e., E{R̂(k, N)} = rX (k)
Definition: WSS X(n) is partially ergodic with respect to the acf if for all k
n o
Var(R̂X (k, N)) = E (R̂X (k, N) − rX (k))2 → 0 as N → ∞.

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Comments

More tests on partial ergodicity in textbook.

Concepts readily transfer to continuous time stochastic processes.

iid random process are ergodic

ergodicity can be seen as a generalization of iid (iid excludes any kind of correlation)

In general it is not possible to determine if a measured source is ergodic

To measure almost all realizations is usually not feasible.
Natural approach is to assume ergodicity if this mathematical model is plausible.

WSS processes ergodic wrt mean and acf are also known as ergodic in the wide
sense.

KTH course: EQ1220 Signal Theory c Tobias Oechtering

Summary lecture 3: Ergodicity

Motivation: How to deduce stochastic process parameters from available data?

Obtain ensemble averages from time averages...

General definition of ergodicity does not require (weak) stationarity, but

non-stationarities are averaged out in the long run
Our simplified definitions require process to be weakly stationary!

Ergodicity wrt mean (or acf) are operationally defined, i.e.,

process is called ergodic wrt mean (or acf) if estimator performance converges in the
mean square sense!

Ergodicity is an abstract concept that cannot be measured, i.e.

ergodicity property is assumed where it is plausible!

KTH course: EQ1220 Signal Theory c Tobias Oechtering