Nonlinear Dependence Hidden by Correlation

Why Zero Correlation Can Hide a Relationship

Correlation summarizes linear association: it measures whether large values of one variable tend to accompany large or small values of the other in a linear pattern. A relationship can be strong and still fail to produce a nonzero correlation. In particular, symmetry can make positive and negative contributions to covariance cancel, even when one variable is a nonlinear function of the other.

This distinction matters when interpreting a joint model. A correlation of zero does not establish independence. To assess a nonlinear pattern, inspect the model itself and, where useful, examine events or transformed variables in addition to the ordinary covariance.

Definition: For random variables $X$ and $Y$ with finite, positive variances, their Pearson correlation is $$ \operatorname{Corr}(X,Y)=\frac{\operatorname{Cov}(X,Y)}{\sqrt{\operatorname{Var}(X)\operatorname{Var}(Y)}}. $$ A correlation of zero is equivalent to zero covariance under these variance conditions. It describes the absence of linear association as measured by this standardized covariance; it does not, by itself, imply independence.

How Symmetry Cancels Covariance

Suppose $X$ has a distribution symmetric about zero, so that $X$ and $-X$ have the same distribution. Let $g$ be an even function, meaning $g(-x)=g(x)$, and suppose the relevant expectations are finite. If $Y=g(X)$, then $Y$ is unchanged when the sign of $X$ changes. Positive and negative values of $X$ can therefore contribute equally sized, opposite-signed terms to the covariance.

Theorem: If $X$ is symmetric about zero, $Y=g(X)$ for an even function $g$, and $E[X]$, $E[g(X)]$, and $E[|Xg(X)|]$ are finite, then $$ \operatorname{Cov}(X,Y)=0. $$ If $X$ and $Y$ also have positive finite variances, their correlation is zero. The theorem does not say that $X$ and $Y$ are independent.

The reason can be seen directly. Symmetry gives $E[X]=0$. The product $xg(x)$ is odd, because replacing $x$ by $-x$ changes its sign while leaving $g(x)$ unchanged. Its expectation is therefore zero. Hence

\operatorname{Cov}(X,Y) =E[Xg(X)]-E[X]E[g(X)] =0-0\cdot E[g(X)] =0.

For a continuous symmetric density $f_X$, the cancellation in the first expectation is explicit:

E[Xg(X)] =\int_{-\infty}^{\infty}xg(x)f_X(x)\,dx =\int_0^\infty \left[xg(x)f_X(x)+(-x)g(-x)f_X(-x)\right]\,dx =0.

The two terms inside the brackets cancel because $g(-x)=g(x)$ and $f_X(-x)=f_X(x)$. For a symmetric discrete distribution, the same cancellation occurs between the contributions from $x$ and $-x$. This argument explains why zero covariance can coexist with an exact nonlinear rule such as $Y=X^2$.

Worked Example: A Continuous U-Shaped Relationship

A hypothetical model has $X\sim\operatorname{Uniform}(-2,2)$ and $Y=X^2$. Calculate the means, variances, covariance, and correlation, then verify that the variables are dependent. Method: use uniform moments and symmetry. The common mistake to avoid is treating a zero covariance as proof of independence.

The density of $X$ is $1/4$ on $[-2,2]$. By symmetry, $E[X]=0$, and

E[X^2]=\frac{1}{4}\int_{-2}^{2}x^2\,dx =\frac{1}{4}\left[\frac{x^3}{3}\right]_{-2}^{2} =\frac{4}{3}.

Thus $E[Y]=E[X^2]=4/3$. Also,

E[X^4]=\frac{1}{4}\int_{-2}^{2}x^4\,dx =\frac{1}{4}\left[\frac{x^5}{5}\right]_{-2}^{2} =\frac{16}{5}.

The variances are

\operatorname{Var}(X)=E[X^2]-(E[X])^2=\frac{4}{3}, \qquad \operatorname{Var}(Y)=E[X^4]-(E[X^2])^2 =\frac{16}{5}-\frac{16}{9} =\frac{64}{45}.

For covariance, $E[XY]=E[X^3]=0$, since the distribution is symmetric and $x^3$ is odd. Therefore

\operatorname{Cov}(X,Y) =E[XY]-E[X]E[Y] =0-0\cdot\frac{4}{3} =0, \qquad \operatorname{Corr}(X,Y)=\frac{0}{\sqrt{(4/3)(64/45)}}=0.

To check dependence, define $A=\{|X|\leq 1\}$ and $B=\{Y\leq 1\}$. Since $Y=X^2$, these are the same event. The uniform model gives $P(A)=P(B)=2/4=1/2$, and $P(A\cap B)=1/2$. If $X$ and $Y$ were independent, this intersection would have probability $P(A)P(B)=1/4$, not $1/2$. They are dependent despite their zero correlation.

Reading the Joint Model, Not Just the Correlation

In the continuous example, $Y$ records the magnitude of $X$ squared. Values of $X$ near either end of the interval produce large $Y$, while values near zero produce small $Y$. That U-shaped pattern is clear from the equation $Y=X^2$, but its contributions to linear covariance cancel across the two sides of zero.

A reliable interpretation therefore uses both the numerical summary and the model. If the model specifies $Y=g(X)$, inspect the form of $g$; if a joint probability table or density is given, look for how the conditional distribution of one variable changes with the other. Independence requires that the conditional distribution of $Y$ given $X$ be the same as the unconditional distribution of $Y$, for all relevant conditioning values. A changing conditional distribution signals dependence even when covariance is zero.

Worked Example: A Discrete Symmetric Model

A hypothetical discrete model assigns probability $1/4$ to each of $X=-3,-1,1,3$, and sets $Y=X^2$. Find the covariance and correlation, then establish dependence from an event probability. Method: calculate the required moments from the four equally likely outcomes. The common mistake to avoid is assuming that a symmetric probability table makes the variables independent.

The mean of $X$ is

E[X]=\frac{-3-1+1+3}{4}=0.

The values of $Y$ are $9,1,1,9$, so

E[Y]=\frac{9+1+1+9}{4}=5, \qquad E[X^2]=5, \qquad E[Y^2]=\frac{81+1+1+81}{4}=41.

Since $Y=X^2$, $E[XY]=E[X^3]$. The positive and negative cubes cancel:

E[XY]=\frac{-27-1+1+27}{4}=0.

It follows that

\operatorname{Var}(X)=5,\qquad \operatorname{Var}(Y)=41-5^2=16, $$ $$ \operatorname{Cov}(X,Y)=0-(0)(5)=0,\qquad \operatorname{Corr}(X,Y)=\frac{0}{\sqrt{5\cdot16}}=0.

For a dependence check, let $A=\{X=1\}$ and $B=\{Y=1\}$. Then $P(A)=1/4$, $P(B)=1/2$, and $P(A\cap B)=1/4$. Independence would require $P(A\cap B)=(1/4)(1/2)=1/8$. The actual probability differs, so $X$ and $Y$ are dependent.

Nonlinear Structure Can Persist with Added Noise

A zero correlation need not mean that one variable is an exact function of the other. The same symmetry can hide a nonlinear pattern in a conditional mean. For instance, suppose $Y=X^2+Z$, where $Z$ is independent noise with mean zero. The linear covariance between $X$ and $Y$ can still vanish, while $E[Y\mid X=x]$ changes with $x^2$. A covariance between transformed variables can help expose that pattern.

Worked Example: A Quadratic Conditional Mean with Noise

Let $X\sim\operatorname{Uniform}(-1,1)$. Independently, let $Z$ equal $1$ or $-1$, each with probability $1/2$, and set $Y=X^2+Z$. Calculate $\operatorname{Corr}(X,Y)$ and $\operatorname{Cov}(X^2,Y)$. Method: use independence to separate moments, and symmetry to eliminate odd moments. The common mistake to avoid is concluding there is no dependence just because the ordinary covariance is zero.

The needed moments of $X$ are

E[X]=0,\quad E[X^2]=\frac{1}{3},\quad E[X^3]=0,\quad E[X^4]=\frac{1}{5}.

Also $E[Z]=0$ and $\operatorname{Var}(Z)=1$. Thus $E[Y]=E[X^2]+E[Z]=1/3$. The variance of $X^2$ is

\operatorname{Var}(X^2)=E[X^4]-(E[X^2])^2 =\frac{1}{5}-\frac{1}{9} =\frac{4}{45}.

Independence gives $\operatorname{Cov}(X^2,Z)=0$, so

\operatorname{Var}(Y)=\operatorname{Var}(X^2)+\operatorname{Var}(Z) =\frac{4}{45}+1=\frac{49}{45}.

For the ordinary covariance, independence gives $\operatorname{Cov}(X,Z)=0$, while symmetry gives $\operatorname{Cov}(X,X^2)=E[X^3]-E[X]E[X^2]=0$. Therefore $\operatorname{Cov}(X,Y)=0$, and both variances are positive, so $\operatorname{Corr}(X,Y)=0$.

Now examine the transformed variable $X^2$:

\operatorname{Cov}(X^2,Y) =\operatorname{Cov}(X^2,X^2+Z) =\operatorname{Var}(X^2)+\operatorname{Cov}(X^2,Z) =\frac{4}{45}.

This nonzero covariance reveals a relationship involving $X^2$, although the correlation between $X$ and $Y$ is zero. The conditional mean gives the same interpretation: independence of $Z$ from $X$ yields $E[Y\mid X=x]=x^2$, which varies with $x$ and is not the constant $E[Y]=1/3$. Thus the variables are dependent.

Exam Traps and Practical Checks

The most common error is to interpret zero correlation as independence. The definition of correlation only standardizes covariance; it does not test whether the full joint distribution factors into the product of its marginal distributions. A second error is to overlook a zero variance. If $Y$ is constant, then $\operatorname{Var}(Y)=0$ and correlation is undefined, even though covariance is zero. In that case, do not report a correlation of zero.

When a model has symmetry or an explicit nonlinear rule, a useful sequence is to calculate the covariance, check that both variances are positive before calculating correlation, and then test dependence separately. Event probabilities can demonstrate dependence directly. Conditional distributions or transformed covariances can reveal structure that ordinary correlation does not capture.

Identify the model and its assumptions.
Record the joint rule or distribution, any symmetry, and the moment conditions needed for the calculations.

Calculate covariance and check variances.
Use the centered-product definition or the equivalent moment formula. Correlation is defined only if both variances are finite and positive.

Assess dependence separately.
Use an event-probability check, conditional behavior, or a covariance involving a nonlinear transformation.

Key takeaway: Symmetry can force positive and negative contributions to linear covariance to cancel. A zero correlation therefore rules out only linear association as measured by Pearson correlation; it does not rule out a deterministic nonlinear relationship or other dependence.

Check Your Understanding

For each model, calculate what is requested and distinguish zero correlation from independence.

Let $X$ take each value in $\{-2,-1,1,2\}$ with probability $1/4$, and let $Y=X^2$. Find the covariance and correlation, and use events to show whether the variables are independent. (Answer: $\operatorname{Cov}(X,Y)=0$, $\operatorname{Corr}(X,Y)=0$; they are dependent, since $P(X=1,Y=1)=1/4\ne P(X=1)P(Y=1)=1/8$.)
Let $X\sim\operatorname{Uniform}(-3,3)$ and $Y=|X|$. Find the means, variances, covariance, and correlation. (Answer: $E[X]=0$, $E[Y]=3/2$, $\operatorname{Var}(X)=3$, $\operatorname{Var}(Y)=3/4$, $\operatorname{Cov}(X,Y)=0$, and $\operatorname{Corr}(X,Y)=0$; the variables are dependent because $Y=|X|$.)
Let $X\sim\operatorname{Uniform}(-1,1)$ and $Y=4X^2-2$. Find $\operatorname{Cov}(X,Y)$, $\operatorname{Var}(Y)$, and the correlation. (Answer: $\operatorname{Cov}(X,Y)=0$, $\operatorname{Var}(Y)=64/45$, and $\operatorname{Corr}(X,Y)=0$; the variables are dependent because $Y$ is determined by $X^2$.)
Let $X\sim\operatorname{Uniform}(-1,1)$, and independently let $Z$ equal $1$ or $-1$ with probability $1/2$ each. Set $Y=X^2+2Z$. Find $\operatorname{Cov}(X,Y)$, $\operatorname{Var}(Y)$, and $\operatorname{Cov}(X^2,Y)$. (Answer: $\operatorname{Cov}(X,Y)=0$, $\operatorname{Var}(Y)=184/45$, and $\operatorname{Cov}(X^2,Y)=4/45$.)
Let $X$ equal $-2,0,2$ with probabilities $0.4,0.2,0.4$, respectively, and let $Y=X^4$. Find the means and variances, then determine whether correlation is zero and whether the variables are independent. (Answer: $E[X]=0$, $E[Y]=12.8$, $\operatorname{Var}(X)=3.2$, $\operatorname{Var}(Y)=40.96$, and $\operatorname{Corr}(X,Y)=0$; the variables are dependent because $Y=X^4$.)

What You'll Learn