Symmetry Arguments for Conditional Means

Why Conditional Symmetry Matters

A distribution that balances positive and negative values often has mean zero. The same cancellation can hold after conditioning, but only when the conditional distribution itself retains that balance. This is useful when the conditioning information records a magnitude, a symmetric region, or some other feature that does not distinguish a value from its negative.

The central idea is sign-pairing: under symmetry, the conditional contributions from $x$ and $-x$ match in probability. An odd function takes opposite values at those two points, so its conditional average cancels. We state the idea precisely, prove it with integrability conditions, and work through discrete, continuous, and event-based examples.

Conditional Symmetry and Cancellation

Definition: Let $X$ be real-valued and let $\mathcal G$ be a sigma-field. The conditional distribution of $X$ given $\mathcal G$ is symmetric about zero if, for every bounded Borel function $\varphi:\mathbb R\to\mathbb R$, $$ E[\varphi(X)\mid\mathcal G]=E[\varphi(-X)\mid\mathcal G]\quad\text{almost surely}. $$ Equivalently, a version of the conditional distribution assigns the same conditional probability to a set and its reflection through zero, whenever such a conditional distribution is being used.

This definition concerns the distribution after the information in $\mathcal G$ is known. It is stronger than saying that the unconditional distribution of $X$ is symmetric. The distinction matters: conditioning can reveal which sign occurred, destroying the balance.

Theorem: Suppose the conditional distribution of $X$ given $\mathcal G$ is symmetric about zero. If $h:\mathbb R\to\mathbb R$ is Borel measurable and $E|h(X)|<\infty$, then $h(-X)$ is integrable and $$ E[h(X)\mid\mathcal G]=E[h(-X)\mid\mathcal G]\quad\text{almost surely}. $$ In particular, if $h$ is odd, meaning $h(-x)=-h(x)$ for every $x$, then $$ E[h(X)\mid\mathcal G]=0\quad\text{almost surely}. $$

Proof. The definition directly gives the claimed conditional-expectation equality for bounded Borel functions. To extend it to an integrable $h(X)$, first apply the definition to the bounded functions $\varphi_n(x)=\min\{|h(x)|,n\}$. Taking expectations gives $E[\min\{|h(X)|,n\}]=E[\min\{|h(-X)|,n\}]$. Letting $n$ increase to infinity and applying the monotone convergence theorem shows that $E|h(-X)|=E|h(X)|<\infty$.

Now define the bounded Borel truncation $h_n(x)=\max\{-n,\min\{h(x),n\}\}$. Conditional symmetry gives $E[h_n(X)\mid\mathcal G]=E[h_n(-X)\mid\mathcal G]$. Since $h_n(X)\to h(X)$ and $h_n(-X)\to h(-X)$ in $L^1$, the $L^1$ contraction property of conditional expectation implies convergence of both conditional expectations in $L^1$. Their limits are therefore equal almost surely. If $h$ is odd, then $h(-X)=-h(X)$, so the equality says that $E[h(X)\mid\mathcal G]$ equals its own negative. It must be zero. $\square$

Taking $h(x)=x$ yields the basic zero-mean result whenever $E|X|<\infty$. Other choices provide further cancellations: for example, $h(x)=x^3$ gives a zero conditional third moment when it is integrable. Even functions behave differently: their values at $x$ and $-x$ agree, so symmetry does not force their conditional means to vanish.

Worked Examples

Worked Example: Conditioning on a Discrete Absolute Value

Let $X$ have probabilities $P(X=-2)=0.1$, $P(X=-1)=0.4$, $P(X=1)=0.4$, and $P(X=2)=0.1$. Find $E[X\mid |X|]$.

Write $R=|X|$. The event $R=1$ has probability $0.4+0.4=0.8$, and $$ P(X=1\mid R=1)=\frac{0.4}{0.8}=\frac12,\qquad P(X=-1\mid R=1)=\frac12. $$ Thus $E[X\mid R=1]=1(1/2)+(-1)(1/2)=0$.

Likewise, $P(R=2)=0.1+0.1=0.2$, and each of $X=2$ and $X=-2$ has conditional probability $1/2$ given $R=2$. Therefore $E[X\mid R=2]=2(1/2)+(-2)(1/2)=0$. Since these are the only possible values of $R$, the conditional mean is zero almost surely: $E[X\mid |X|]=0$.

Worked Example: A Normal Variable Given Its Magnitude

Let $Z\sim N(0,1)$ and $R=|Z|$. Determine $E[Z\mid R]$.

For $r>0$, the two possible signs correspond to $Z=r$ and $Z=-r$. The standard normal density, denoted by $\phi$, satisfies $\phi(r)=\phi(-r)$. The density of $R$ at $r>0$ is $\phi(r)+\phi(-r)=2\phi(r)$, so the conditional probabilities of the two signs given $R=r$ are $$ P(Z=r\mid R=r)=\frac{\phi(r)}{2\phi(r)}=\frac12,\qquad P(Z=-r\mid R=r)=\frac{\phi(-r)}{2\phi(r)}=\frac12. $$ Consequently, the conditional mean at magnitude $r$ is $r(1/2)+(-r)(1/2)=0$. Since $P(R=0)=0$, the value assigned to a conditional-mean version at $R=0$ does not affect the result. Hence $E[Z\mid |Z|]=0$ almost surely.

Worked Example: A Symmetric Event Region

Let $(X,Y)$ be uniformly distributed on the square $[-1,1]\times[-1,1]$. Let $A=\{Y>0,\ |X|<Y\}$. Find $E[X\mid A]$.

The joint density is $1/4$ on the square. The region $A$ is a triangle whose horizontal slice at height $y$, for $0<y<1$, runs from $x=-y$ to $x=y$. Its area is $$ \int_0^1\int_{-y}^{y}1\,dx\,dy =\int_0^1 2y\,dy=1. $$ The whole square has area $4$, so $P(A)=1/4>0$. The conditional mean is therefore defined by $E[X\mid A]=E[X\mathbf 1_A]/P(A)$. Its numerator is $$ E[X\mathbf 1_A] =\frac14\int_0^1\int_{-y}^{y}x\,dx\,dy =\frac14\int_0^1\left[\frac{x^2}{2}\right]_{-y}^{y}dy =0. $$ Thus $E[X\mid A]=0/(1/4)=0$. The cancellation occurs because each point $(x,y)$ in the region is paired with $(-x,y)$, also in the region, and the two values of $X$ have opposite signs.

Symmetry About a Conditional Center

The same argument applies when the balance point is not zero. The center may itself depend on the information being conditioned on, but it must be integrable for the usual conditional expectation of $X$ to be defined.

Corollary: Let $m$ be a finite, $\mathcal G$-measurable random variable with $E|m|<\infty$. Suppose $E|X-m|<\infty$ and the conditional distribution of $X-m$ given $\mathcal G$ is symmetric about zero. Then $X$ is integrable and $$ E[X\mid\mathcal G]=m\quad\text{almost surely}. $$

Indeed, $|X|\leq |X-m|+|m|$, so $E|X|<\infty$. The theorem applied to the odd function $h(z)=z$ gives $E[X-m\mid\mathcal G]=0$. Since $m$ is integrable and $\mathcal G$-measurable, linearity of conditional expectation gives $E[X\mid\mathcal G]=E[X-m\mid\mathcal G]+E[m\mid\mathcal G]=m$.

Worked Example: A Random Conditional Center

Let $Y$ take values $-2$ and $1$ with probabilities $0.3$ and $0.7$. Independently, let $\varepsilon$ take values $-1$ and $1$, each with probability $1/2$, and set $X=Y+\varepsilon$. Find $E[X\mid Y]$.

Given $Y=y$, the variable $X-y=\varepsilon$ is equally likely to be $-1$ or $1$, so its conditional distribution is symmetric about zero. Also, $E|Y|=0.3(2)+0.7(1)=1.3<\infty$, and $E|\varepsilon|=1$. The corollary applies with $\mathcal G=\sigma(Y)$ and $m=Y$, giving $E[X\mid Y]=Y$.

Directly, when $Y=-2$, $X$ is $-3$ or $-1$ with equal probabilities and has mean $-2$. When $Y=1$, $X$ is $0$ or $2$ with equal probabilities and has mean $1$. These conditional means verify the result at both possible values of $Y$.

Symmetric Events and the Limits of Unconditional Symmetry

An event-based version is often convenient. Suppose a transformation of the sample space preserves probabilities, leaves an event $A$ unchanged, and changes $X$ to $-X$. If $X$ is integrable and $P(A)>0$, then the transformation pairs every contribution to $E[X\mathbf 1_A]$ with an equal-probability contribution of opposite sign. Therefore $E[X\mathbf 1_A]=-E[X\mathbf 1_A]$, so $E[X\mid A]=0$. The triangle calculation above is an integral form of this argument.

The condition that the event be preserved is essential. An event that selects positive values does not retain the sign balance, even when the original distribution is symmetric. More generally, unconditional symmetry alone does not imply a zero conditional mean.

Worked Example: Unconditional Symmetry Does Not Guarantee Conditional Cancellation

Let $Y$ equal $-1$ or $1$ with probability $1/2$ each, and set $X=Y$. The unconditional distribution of $X$ is symmetric, and $E[X]=(-1)(1/2)+(1)(1/2)=0$. But given $Y$, the value of $X$ is known exactly. Thus $E[X\mid Y]=Y$, which is not zero: it equals $-1$ when $Y=-1$ and $1$ when $Y=1$.

The conditioning information here reveals the sign rather than just the magnitude. The conditional distribution is concentrated at one point, so it is not symmetric about zero except on a probability-zero event. This is why a zero unconditional mean cannot replace a check of conditional symmetry.

Common Pitfall

Using unconditional symmetry after conditioning. A symmetric marginal distribution only guarantees cancellation before conditioning. Check that the conditional distribution, or the event being conditioned on, preserves the sign-pairing.
Forgetting integrability. The usual $E[X\mid\mathcal G]$ is defined for integrable $X$. A symmetry statement about $X-m$ alone does not establish that $X$ is integrable; in the shifted result, $E|m|<\infty$ and $E|X-m|<\infty$ together ensure it.
Assuming every function cancels. Symmetry forces the conditional mean of an integrable odd function to be zero. An even function, such as $x^2$, generally does not cancel because its paired values have the same sign.

Key takeaway: Conditional sign-pairing makes integrable odd functions average to zero. Identify the conditioning information, verify that it preserves the symmetry, and check integrability before using the cancellation.

Check Your Understanding

For each question, identify the relevant symmetry and state any integrability condition needed.

If the conditional distribution of $X$ given $\mathcal G$ is symmetric about zero and $E|X|<\infty$, what is $E[X\mid\mathcal G]$?
In the discrete absolute-value example, why are the conditional probabilities of $X=r$ and $X=-r$ equal given $|X|=r$?
For a symmetric conditional distribution, what can be concluded about $E[X^3\mid\mathcal G]$ if $E|X|^3<\infty$?
Why does conditioning on $Y$ in the example $X=Y$, with $Y$ equally likely to be $-1$ or $1$, fail to give a zero conditional mean?
If $X-m$ is conditionally symmetric about zero given $\mathcal G$, which integrability conditions in the corollary ensure that $E[X\mid\mathcal G]=m$ is meaningful?