White Noise Does Not Automatically Mean Independence
A white-noise sequence has no linear association between observations at different times: its autocovariance is zero at every nonzero lag. That is a statement about means, variances, and covariances. It does not, by itself, rule out every form of dependence. In particular, zero covariance does not generally imply independence.
This distinction matters when assessing a time-series model. A residual sequence may satisfy the moment conditions for white noise while still having a dependence pattern that those conditions cannot detect. To interpret a white-noise assumption correctly, separate the conditions on covariance from the stronger claim that observations are independent.
Independence is a different and stronger condition. Two random variables are independent if their joint probabilities factor into the product of their marginal probabilities. A sequence is mutually independent if this factorization holds for every finite collection of distinct observations. When each observation also has the same distribution, the sequence is independent and identically distributed (iid).
If observations are mutually independent, have a common finite mean and variance, and are otherwise described by the conditions above, then distinct observations have covariance zero. The reverse implication does not generally hold: zero covariance only checks a particular measure of linear association, not the whole joint distribution.
What Independence Adds
For random variables \(X\) and \(Y\) with finite second moments, covariance is \(\operatorname{Cov}(X,Y)=E[(X-E[X])(Y-E[Y])]\). If \(X\) and \(Y\) are independent, expectations of suitable products factor, giving \[ \operatorname{Cov}(X,Y)=E[XY]-E[X]E[Y]=0. \] The converse can fail because positive and negative contributions to the covariance may cancel, or because a dependence is expressed through a nonlinear relationship rather than a linear one.
Independence has consequences beyond covariance. For example, it determines joint probabilities from marginal probabilities. A covariance calculation alone does not establish that factorization. The following examples make the distinction explicit.
Worked Example: Independent Bernoulli Observations Are White Noise
Let \(\{X_t\}\) be mutually independent Bernoulli observations with \(P(X_t=1)=0.4\) and \(P(X_t=0)=0.6\) at every time. Verify their mean, variance, and covariance at distinct times, and state whether they are white noise.
Step 1: Find the mean. For a Bernoulli random variable with success probability \(p\), \(E[X_t]=p\). Thus \(E[X_t]=0.4\).
Step 2: Find the variance. Since \(X_t^2=X_t\), \(E[X_t^2]=0.4\), so \[ \operatorname{Var}(X_t)=E[X_t^2]-(E[X_t])^2=0.4-(0.4)^2=0.24. \]
Step 3: Find covariance at distinct times. For \(t\ne s\), independence gives \(E[X_tX_s]=E[X_t]E[X_s]=(0.4)(0.4)=0.16\). Therefore, \[ \operatorname{Cov}(X_t,X_s)=0.16-(0.4)(0.4)=0. \]
Conclusion and check. The observations have the same mean, \(0.4\), the same finite positive variance, \(0.24\), and zero covariance at distinct times. They satisfy the moment-based white-noise conditions. They are also mutually independent by the model assumption, a fact stronger than those conditions. The common mistake to avoid is treating the covariance calculation as proof of independence; here independence was specified separately.
A Nonlinear Relationship Can Have Zero Covariance
A simple way to see why zero covariance does not imply independence is to make one variable a nonlinear function of another. Symmetry can make the covariance zero even though the relationship is exact.
Worked Example: A Dependent Pair with Zero Covariance
Let \(Z\) have the distribution \(P(Z=-2)=0.25\), \(P(Z=0)=0.50\), and \(P(Z=2)=0.25\). Define \(W=Z^2\). Compute the means, variances, and covariance of \(Z\) and \(W\), then decide whether they are independent.
Step 1: Compute the moments of \(Z\). By symmetry, \[ E[Z]=(-2)(0.25)+(0)(0.50)+(2)(0.25)=0. \] Also, \[ E[Z^2]=4(0.25)+0(0.50)+4(0.25)=2, \] so \(\operatorname{Var}(Z)=E[Z^2]-(E[Z])^2=2\).
Step 2: Compute the moments of \(W\). The values of \(W\) are 4 when \(Z=-2\) or \(Z=2\), and 0 when \(Z=0\). Hence \(E[W]=4(0.25)+0(0.50)+4(0.25)=2\). Since \(W^2=Z^4\), \[ E[W^2]=16(0.25)+0(0.50)+16(0.25)=8, \] and \(\operatorname{Var}(W)=8-2^2=4\).
Step 3: Compute the covariance. Because \(ZW=Z^3\), \[ E[ZW]=E[Z^3]=(-8)(0.25)+0(0.50)+(8)(0.25)=0. \] It follows that \(\operatorname{Cov}(Z,W)=E[ZW]-E[Z]E[W]=0-(0)(2)=0\).
Step 4: Check independence. If \(Z=0\), then \(W=0\), so \(P(Z=0,W=0)=0.50\). But \(P(Z=0)P(W=0)=(0.50)(0.50)=0.25\). These probabilities differ, so \(Z\) and \(W\) are not independent. In fact, \(W\) is completely determined by \(Z\).
Check and trap. Both variances are positive and the covariance is exactly zero, yet the pair is dependent. Concluding independence from zero covariance would be the error; checking a joint probability exposes it.
Why the Covariance Formula Gives Zero for Independent Variables
The one-way implication from independence to zero covariance follows directly from the definition, provided the variables have finite second moments so that the expectations are defined. Write covariance in its equivalent form: \[ \operatorname{Cov}(X,Y)=E[XY]-E[X]E[Y]. \] For independent \(X\) and \(Y\), the expectation of their product factors as \(E[XY]=E[X]E[Y]\). Substitution gives \[ \operatorname{Cov}(X,Y)=E[X]E[Y]-E[X]E[Y]=0. \]
This argument proves only that independence implies zero covariance. It cannot be reversed: the equality \(E[XY]=E[X]E[Y]\) is a single moment condition, whereas independence requires factorization of the joint distribution. The nonlinear example above satisfies that one moment condition without satisfying independence.
White Noise Can Hide Dependence Across a Group of Observations
There is a further distinction between observations being pairwise independent and an entire sequence being mutually independent. A sequence can have zero covariance for every distinct pair—even pairwise independent observations—while some three or more observations are dependent.
Worked Example: A White-Noise Sequence That Is Not Mutually Independent
For each integer block \(k\), let \(U_k\) and \(V_k\) be independent Rademacher variables: each takes values \(-1\) and \(1\) with probability \(0.5\) each. Assume all variables from different blocks are independent. Define three observations in block \(k\) by \[ X_{3k-2}=U_k,\qquad X_{3k-1}=V_k,\qquad X_{3k}=U_kV_k. \] Verify the moment-based white-noise conditions, then test mutual independence.
Step 1: Check means and variances. Each of \(U_k\), \(V_k\), and \(U_kV_k\) takes values \(-1\) and \(1\) equally often. For the product, this follows because, among the four equally likely pairs \((U_k,V_k)\), two have product \(1\) and two have product \(-1\). Thus every observation has mean 0 and variance \(1\).
Step 2: Check covariances within a block. For the first and second observations, \[ E[U_kV_k]=E[U_k]E[V_k]=0. \] For the first and third, \[ E[U_k(U_kV_k)]=E[U_k^2V_k]=E[V_k]=0, \] because \(U_k^2=1\). Similarly, \[ E[V_k(U_kV_k)]=E[U_kV_k^2]=E[U_k]=0. \] All three means are zero, so these product expectations are the corresponding covariances, and every within-block covariance is zero.
Step 3: Check different blocks. Observations in different blocks are functions of independent block variables. They are therefore independent, so their covariances are zero. Every distinct pair in the sequence consequently has covariance zero.
Step 4: Check mutual independence. In each block, \(X_{3k}=X_{3k-2}X_{3k-1}\). In particular, \[ P(X_{3k-2}=1,\ X_{3k-1}=1,\ X_{3k}=-1)=0. \] If these three observations were mutually independent, the probability would instead be \[ P(X_{3k-2}=1)P(X_{3k-1}=1)P(X_{3k}=-1) =(0.5)(0.5)(0.5)=0.125. \] The probabilities disagree, so the sequence is not mutually independent.
Conclusion and check. The sequence has constant mean 0, constant variance 1, and zero covariance between every distinct pair. It is therefore white noise in the moment-based sense, but it is not mutually independent. The common mistake is to infer independence of the whole sequence from pairwise covariance checks; the three-observation probability shows why that conclusion fails.
How to Interpret a White-Noise Claim
When a problem states that a sequence is white noise, first identify which definition is being used. Under the moment-based definition used here, the claim supplies a common mean and variance and zero covariance at nonzero lags. It does not automatically supply independence, identical full distributions, or independence of every finite collection.
Some contexts use “white noise” as shorthand for iid observations with finite variance. That is a stronger convention. If a question explicitly says iid white noise, or explicitly assumes independence, use that extra information. If it states only the moment conditions, do not silently add independence.
A practical check is to ask what the proposed conclusion needs. A covariance or autocorrelation calculation uses second-moment information. A claim about joint probabilities, conditional distributions, or a nonlinear relationship may require more than uncorrelatedness. State the assumption actually given before applying an independence rule.
Record whether the problem gives zero covariance, pairwise independence, mutual independence, or an iid assumption. These are not interchangeable.
For white noise in the moment-based sense, check the common finite mean and variance and zero covariance at distinct times.
Use an explicit independence assumption or verify joint-distribution factorization; a zero-covariance result alone is insufficient.
Check Your Understanding
For each problem, distinguish the conclusions supported by the stated assumptions.
- A sequence consists of mutually independent Bernoulli observations with success probability \(0.4\). Find the common mean and variance and the covariance at distinct times. (Answer: Mean \(0.4\), variance \(0.24\), and covariance \(0\) at distinct times; it is white noise and mutually independent.)
- Let \(Z\) be uniform on \(\{-1,0,1\}\), and set \(W=Z^2\). Find \(\operatorname{Cov}(Z,W)\), then compare \(P(Z=0,W=0)\) with \(P(Z=0)P(W=0)\). (Answer: Covariance \(0\); the joint probability is \(1/3\), while the product is \(1/9\), so the variables are dependent.)
- Suppose every observation in a sequence has mean 0 and variance 4, and every pair of distinct observations has covariance 0. What can be concluded under the moment-based definition, and what cannot be concluded from these facts alone? (Answer: The sequence is white noise in the moment-based sense; mutual independence cannot be concluded.)
- Let \(U,V\) be independent Rademacher variables and let \(A=U\), \(B=V\), and \(C=UV\). Find the three pairwise covariances and decide whether \(A,B,C\) are mutually independent. (Answer: All three pairwise covariances are \(0\); they are not mutually independent because \(C=AB\).)
- An analyst observes zero sample autocorrelations at several nonzero lags and reports that all observations are independent. Is that conclusion justified by the white-noise moment conditions? State the missing requirement. (Answer: No. Zero autocorrelation supports zero linear association at the examined lags, not independence; mutual independence requires a separate assumption or joint-distribution evidence.)