Bregman Loss Regret and Conditional Means

Why the Conditional Mean Appears

Squared-error loss is not the only loss for which the conditional mean is the optimal prediction. A broader family comes from differentiable convex functions called potentials. Their associated Bregman losses measure the gap between a function and its tangent approximation. With the outcome in the first argument and the action in the second, the conditional regret of an action has a particularly simple form: it is the Bregman divergence from the conditional mean to that action.

The exact algebraic identity and the claim of optimality are related but distinct. Differentiability and a well-defined conditional mean give the identity; convexity makes its right-hand side nonnegative; strict convexity makes it positive for every action other than the mean. Keeping those steps separate prevents a common error: treating an algebraic identity as if it alone proved that the mean minimizes risk.

Potential, Divergence, and Conditional Risk

Let \(C\) be an open convex subset of \(\mathbb{R}^d\), and let \(\phi:C\to\mathbb{R}\) be differentiable. The gradient \(\nabla\phi(a)\) is the vector of first derivatives at \(a\). Define the Bregman divergence from \(a\) to \(y\) by the difference between \(\phi(y)\) and its first-order approximation at \(a\).

The order of the arguments matters: the gradient is evaluated at the action \(a\), not the outcome \(y\). In general \(D_\phi(y,a)\neq D_\phi(a,y)\). The potential need not be quadratic, and the divergence need not be symmetric.

For the conditional statement, let \(\mathcal{G}\) represent the information available when choosing an action. Suppose \(Y\) takes values in \(C\), \(E[\|Y\|\mid\mathcal{G}]<\infty\), and its conditional mean \(M=E[Y\mid\mathcal{G}]\) belongs to \(C\) almost surely. Actions \(A\) under consideration are \(\mathcal{G}\)-measurable and take values in \(C\). Assume the terms in the conditional risks below are integrable; in particular, the products involving \(\nabla\phi(A)\) and \(Y-A\), and those involving \(\nabla\phi(M)\) and \(Y-M\), must have finite conditional expectations. These assumptions ensure the algebra uses well-defined quantities. They can be checked for the particular potential, outcome distribution, and actions in an application.

The Conditional Regret Identity

Write \(R(A\mid\mathcal{G})=E[D_\phi(Y,A)\mid\mathcal{G}]\) for the conditional risk of an available action, and define \(R(M\mid\mathcal{G})\) in the same way. The identity below does not require \(\phi\) to be convex. Convexity enters only when interpreting the sign of the resulting difference.

Proof of the identity. Since \(A\) is \(\mathcal{G}\)-measurable, \(\phi(A)\) and \(\nabla\phi(A)\) are known given \(\mathcal{G}\). Expand the divergence and use \(E[Y\mid\mathcal{G}]=M\):

Subtract the second expression from the first. The terms \(E[\phi(Y)\mid\mathcal{G}]\) cancel, leaving \(\phi(M)-\phi(A)-\nabla\phi(A)^\mathsf{T}(M-A)=D_\phi(M,A)\), as claimed. This calculation used differentiability to define the divergence, and the conditional-mean property to replace \(E[Y\mid\mathcal{G}]\) by \(M\); it did not use convexity. \(\square\)

Why convexity gives minimization. For differentiable convex \(\phi\), the tangent-plane inequality says \(\phi(u)\geq\phi(v)+\nabla\phi(v)^\mathsf{T}(u-v)\) for \(u,v\in C\). Taking \(u=M\) and \(v=A\) gives \(D_\phi(M,A)\geq0\). The identity therefore implies \(R(A\mid\mathcal{G})\geq R(M\mid\mathcal{G})\). If \(\phi\) is strictly convex, the tangent-plane inequality is strict whenever \(M\neq A\); hence the risk difference is strictly positive on that event. Any minimizing action must equal \(M\) almost surely. \(\square\)

Worked Examples

What the Assumptions Do

The formula is useful because it gives more than an optimizer: it expresses the exact cost of using an alternative action. Under convexity, this cost is nonnegative and determined by the gap between the action and the conditional mean through the chosen potential. For a quadratic potential it is a squared distance; for another potential, the geometry of the regret can differ. The conditional-mean conclusion is therefore not a claim that all Bregman losses are squared errors. It follows from the specific orientation of the divergence and its tangent-plane inequality.

The assumptions also matter in practice. The mean must exist and be an available action in the domain. If it falls outside the allowed action set, the theorem as stated does not show that it is an admissible minimizer. Conditional risks and the terms used to expand them must be well-defined as well; otherwise subtracting them or invoking conditional linearity may not be justified.

Common Pitfall: Convexity and Argument Order

A tempting but incorrect argument is: “The regret identity holds for a differentiable potential, so the conditional mean must minimize risk.” The identity alone says the risk difference is \(D_\phi(M,A)\); it does not say that this quantity is nonnegative. For example, with \(\phi(t)=-t^2/2\), the identity still holds, but \(D_\phi(y,a)=-(y-a)^2/2\). If \(Y\) is equally likely to be \(-1\) or \(1\), its mean is zero. The risks at actions \(0\) and \(1\) are respectively \(-1/2\) and \(-1\), so action \(1\) has lower risk. Convexity is what ensures nonnegative regret; strict convexity ensures uniqueness.

A second error is to reverse the arguments and use \(D_\phi(A,Y)\) as the loss while applying the same identity. The proof specifically expands \(D_\phi(Y,A)\), whose gradient is evaluated at the chosen action. Reversing the arguments changes the loss and does not, in general, preserve the conditional-mean result. Always check which argument is the outcome and which is the action before using the formula.