Why Zero Correlation Does Not Imply Independence

Why a zero measure can hide dependence

Covariance summarizes how two losses vary together through the expected product of their centered values. A zero covariance says that this particular average is zero. It does not say that the losses have no relationship: a nonlinear relationship can produce positive and negative centered products that cancel. Since correlation is a standardized covariance, zero correlation has the same limitation.

This distinction matters when assessing a portfolio of risks. A zero correlation estimate is not evidence that the losses behave independently, and it does not rule out a predictable relationship between them. Independence is a stronger condition about the joint distribution, not just one summary of it.

Definition: For losses $L_1$ and $L_2$ with finite means, their covariance is $$ \operatorname{Cov}(L_1,L_2) =E[(L_1-E[L_1])(L_2-E[L_2])] =E[L_1L_2]-E[L_1]E[L_2]. $$ If both losses have finite, strictly positive variances, their correlation is $$ \rho_{L_1,L_2} =\frac{\operatorname{Cov}(L_1,L_2)} {\sqrt{\operatorname{Var}(L_1)\operatorname{Var}(L_2)}}. $$ Zero covariance then implies zero correlation, and vice versa. If either variance is zero, correlation is undefined.

Independence does imply zero covariance when the expectations and variances needed in the definition are finite: independence gives $E[L_1L_2]=E[L_1]E[L_2]$, so the covariance formula is zero. The converse does not follow. Covariance is one number, whereas independence requires the joint probabilities or distribution to factor appropriately across events.

A discrete loss pair with zero covariance

Consider a hypothetical portfolio with one of three equally likely claim scenarios. Let $L_1$ and $L_2$ be the two losses, measured in dollars. The joint outcomes are:

Scenario	$L_1$	$L_2$
1	180	60
2	200	30
3	220	60

Worked Example: Zero Covariance in a Three-Scenario Portfolio

Each row in the table has probability $1/3$. Calculate the means, variances, covariance, and correlation, then determine whether the losses are independent. The method is to center each loss at its mean and average the products of the centered values. A common mistake to avoid is treating a zero average product as proof that each outcome is unrelated.

Step 1: Calculate the means.

E[L_1]=\frac{180+200+220}{3}=\frac{600}{3}=200, \qquad E[L_2]=\frac{60+30+60}{3}=\frac{150}{3}=50.

Step 2: Center the losses and calculate the variances. The deviations of $L_1$ from its mean are $-20,0,20$. Those of $L_2$ are $10,-20,10$.

\operatorname{Var}(L_1) =\frac{(-20)^2+0^2+20^2}{3} =\frac{400+0+400}{3} =\frac{800}{3} \approx 266.7, $$ $$ \operatorname{Var}(L_2) =\frac{10^2+(-20)^2+10^2}{3} =\frac{100+400+100}{3} =200.

The variances are reported to one decimal place where needed. In particular, $800/3$ is not $800$: the three squared deviations must be averaged.

Step 3: Calculate covariance and correlation. The products of the paired deviations are $(-20)(10)=-200$, $(0)(-20)=0$, and $(20)(10)=200$. Therefore,

\operatorname{Cov}(L_1,L_2) =\frac{-200+0+200}{3}=0, \qquad \rho_{L_1,L_2} =\frac{0}{\sqrt{(800/3)(200)}}=0.

Step 4: Test independence. In scenario 1, $L_1=180$ guarantees $L_2=60$, so $P(L_2=60\mid L_1=180)=1$. Unconditionally, $P(L_2=60)=2/3$. These probabilities differ, so the losses are dependent despite their zero covariance and correlation.

Check. The three covariance products sum to zero, while the conditional probability differs from the unconditional probability. Both results agree with the conclusion: zero correlation does not establish independence.

Why symmetry can cancel covariance

A useful continuous construction makes the cancellation explicit. Suppose $U$ is uniformly distributed on $(-1,1)$, and define two hypothetical losses $L_1=100+20U$ and $L_2=50+30U^2$, in dollars. The first loss changes linearly with $U$; the second depends on the square of $U$. Knowing $L_1$ determines $U$, and therefore determines $L_2$, so the losses have an exact nonlinear relationship.

Justification: For $U\sim\operatorname{Uniform}(-1,1)$, symmetry about zero gives $E[U]=E[U^3]=0$, while $E[U^2]=1/3$. Thus $$ \operatorname{Cov}(L_1,L_2) =600E\left[U\left(U^2-\frac13\right)\right] =600\left(E[U^3]-\frac13E[U]\right)=0. $$ The integrand $u(u^2-1/3)$ is an odd function, so its integral over the symmetric interval $(-1,1)$ is zero. This cancellation explains the zero covariance; it does not remove the dependence of $L_2$ on $U^2$.

Worked Example: A Continuous Loss Determined by a Square

Let $U\sim\operatorname{Uniform}(-1,1)$, and let $L_1=100+20U$ and $L_2=50+30U^2$, in dollars. Find each mean, each variance, the covariance, and the correlation. Then exhibit events that contradict independence. Use the uniform moments obtained by integration; the common trap is to conclude independence from the zero covariance alone.

Step 1: State the distributional moments. For a uniform variable on $(-1,1)$, integrating its density $1/2$ gives $E[U]=0$, $E[U^2]=1/3$, $E[U^3]=0$, and $E[U^4]=1/5$.

Step 2: Find the means.

E[L_1]=100+20E[U]=100, \qquad E[L_2]=50+30E[U^2]=50+10=60.

Step 3: Find the variances. Since adding a constant does not change variance and multiplying by a constant multiplies variance by its square,

\operatorname{Var}(L_1) =20^2\operatorname{Var}(U) =400\left(\frac13\right) =\frac{400}{3}\approx 133.3.

For the second loss, $\operatorname{Var}(U^2)=E[U^4]-(E[U^2])^2=1/5-1/9=4/45$. Hence

\operatorname{Var}(L_2) =30^2\operatorname{Var}(U^2) =900\left(\frac{4}{45}\right) =80.

Step 4: Calculate covariance and correlation. Centering gives $L_1-E[L_1]=20U$ and $L_2-E[L_2]=30(U^2-1/3)$. Therefore,

\operatorname{Cov}(L_1,L_2) =600E[U(U^2-1/3)] =600(0-0)=0, \qquad \rho_{L_1,L_2} =\frac{0}{\sqrt{(400/3)(80)}}=0.

Step 5: Show dependence directly. Define event $A$ by $L_1>110$ or $L_1<90$, and event $B$ by $L_2>57.5$. Both are equivalent to $|U|>0.5$: for $A$, the thresholds imply $U>0.5$ or $U<-0.5$; for $B$, $50+30U^2>57.5$ implies $U^2>1/4$. Thus $P(A)=P(B)=1/2$, and $P(A\cap B)=1/2$, not $P(A)P(B)=1/4$. The events are dependent.

Check. The denominator in the correlation is positive because both variances are positive. The correlation is therefore defined and equals zero, while the event calculation confirms dependence.

Another discrete pattern: dependence without a linear trend

The cancellation is not limited to a loss that is a square of a continuous variable. It can also arise when a finite set of scenarios pairs low and high outcomes in a nonlinear pattern. In the next example, a conditional probability makes the failure of independence especially clear.

Worked Example: A Four-Scenario Pairing

A hypothetical risk register assigns equal probability to four joint loss outcomes, in thousands of dollars: $(L_1,L_2)=(10,40),(20,30),(30,30),(40,40)$. Calculate the means and covariance, then test independence. The method is to use the joint outcomes with their probabilities; a trap is to infer independence merely because the centered products cancel.

Step 1: Calculate the means.

E[L_1]=\frac{10+20+30+40}{4}=25, \qquad E[L_2]=\frac{40+30+30+40}{4}=35.

Step 2: Calculate the covariance. The centered values of $L_1$ are $-15,-5,5,15$; those of $L_2$ are $5,-5,-5,5$. Their paired products are $-75,25,-25,75$.

\operatorname{Cov}(L_1,L_2) =\frac{-75+25-25+75}{4}=0.

Step 3: Test independence. When $L_1=10$, $L_2=40$ with probability one. But $P(L_2=40)=2/4=1/2$. Since $P(L_2=40\mid L_1=10)=1\ne 1/2$, the losses are dependent.

Check. The four product terms sum to zero, but the conditional and unconditional probabilities differ. The zero covariance is compatible with this clear dependence.

Interpretation and exam traps

Covariance and correlation describe linear co-movement. When centered high values of one variable tend to pair with centered high values of another, covariance tends to be positive; when high values tend to pair with low values, it tends to be negative. But a joint pattern may be structured without producing either average trend. In the examples, a symmetric loss pattern creates positive and negative products that balance.

For risk analysis, the practical conclusion is precise: a zero correlation can support a statement about the linear association measured by correlation, but not a statement that the losses are independent. Dependence may still affect the chance of joint events or the distribution of aggregate losses. Independence requires checking the joint distribution or suitable event probabilities, not just the covariance.

Wrong: “The covariance is zero, so the losses are independent.”
Correction: “The covariance is zero, so the average product of centered losses is zero. Independence must be assessed separately.”
Wrong: Divide by the variances rather than their square roots when calculating correlation.
Correction: Divide covariance by the product of the standard deviations; both variances must be finite and positive.
Wrong: Forget to average squared deviations or centered products in an equally likely discrete model.
Correction: Weight each term by its probability. For equally likely outcomes, sum the terms and divide by the number of outcomes.

Key takeaway: Independence implies zero covariance when the relevant moments are finite, but zero covariance does not imply independence. A nonlinear or symmetric relationship can be strong even when covariance and correlation are both zero.

Check Your Understanding

For each problem, calculate the requested quantities from the stated model and distinguish zero covariance from independence.

$Z$ takes values $-1,0,1$ with equal probabilities. Let $X=12+3Z$ and $Y=8+6Z^2$. Find $E[X]$, $E[Y]$, and $\operatorname{Cov}(X,Y)$; decide whether $X$ and $Y$ are independent. (Answer: $E[X]=12$, $E[Y]=12$, covariance $=0$, and they are dependent.)
$W\sim\operatorname{Uniform}(-2,2)$. Let $A=80+5W$ and $B=25+3W^2$. Find their means and covariance. (Answer: $E[A]=80$, $E[B]=29$, and $\operatorname{Cov}(A,B)=0$.)
Four equally likely outcomes for $(X,Y)$ are $(1,7),(2,5),(3,5),(4,7)$. Find the covariance and correlation, and use a conditional probability to determine independence. (Answer: covariance $=0$, correlation $=0$, and not independent because $P(Y=7\mid X=1)=1\ne P(Y=7)=1/2$.)
Let $W\sim\operatorname{Uniform}(-1,1)$, $X=50+10W$, and $Y=20+12W^2$. Let $A$ be the event $X>55$ or $X<45$, and $B$ be the event $Y>23$. Find $P(A)$, $P(B)$, and $P(A\cap B)$; do the events act independently? (Answer: $P(A)=1/2$, $P(B)=1/2$, $P(A\cap B)=1/2$, so they are not independent.)
$U$ takes values $-2,-1,1,2$ with probabilities $0.2,0.3,0.3,0.2$, respectively. Let $X=100+10U$ and $Y=40+5U^2$. Find $E[Y]$, both variances, and the correlation. (Answer: $E[Y]=51$, $\operatorname{Var}(X)=220$, $\operatorname{Var}(Y)=54$, and correlation $=0$; the losses are dependent because $Y$ is determined by $U^2$.)

What You'll Learn