Propagating Conditional Distribution Error to Decision Regret

From Distribution Error to Decision Regret

A conditional distribution estimate is often used not only to describe future outcomes, but also to choose an action. For example, a clinician may choose a treatment using an estimated distribution of a patient’s response, or a planner may choose an inventory level using an estimated distribution of demand. The relevant question is then not just whether the estimated law is close to the true law. It is whether replacing the true law with the estimated one causes a poor decision.

This tutorial gives a direct answer when the action loss changes smoothly with the outcome. If the loss is Lipschitz in the outcome, a Wasserstein bound on conditional distribution error controls the excess conditional risk. The bound depends on the losses at the true and estimated optimal actions; a uniform Lipschitz condition gives a simpler, action-independent bound.

Conditional Risk, Regret, and Wasserstein Distance

Fix a covariate value \(x\). Let \(P_x\) be the true conditional law of an outcome \(Y\) given \(X=x\), and let \(Q_x\) be an estimated conditional law. Outcomes lie in a metric space \((\mathcal{Y},d)\). The available actions form a set \(\mathcal{A}\), and \(\ell(a,y)\) is the loss from taking action \(a\) when the outcome is \(y\). Define the conditional risk under a law \(G\) by \(R_G(a)=\int \ell(a,y)\,dG(y)\), whenever this expectation is finite.

For \(r\geq 1\), the \(r\)-Wasserstein distance is defined, for probability measures with finite \(r\)th moments, by \(W_r(P,Q)=\inf_{\pi\in\Pi(P,Q)}\{\int d(y,y')^r\,d\pi(y,y')\}^{1/r}\). Here \(\Pi(P,Q)\) is the set of joint distributions, or couplings, whose two marginal distributions are \(P\) and \(Q\). The bound below uses \(W_1\); finite first moments suffice. The decision regret from using an action \(a_Q\) chosen under \(Q_x\) is its excess risk under the true law, \(R_{P_x}(a_Q)-\inf_{a\in\mathcal A}R_{P_x}(a)\).

The Basic Wasserstein Risk Bound

The key step is a bound on the risk difference at one fixed action. It follows from the coupling definition and does not require a separate dual representation of Wasserstein distance.

Proof. Take any coupling \((Y,Y')\) with marginal laws \(P\) and \(Q\). The marginal property gives \(R_P(a)-R_Q(a)=E[\ell(a,Y)-\ell(a,Y')]\). Therefore, \(|R_P(a)-R_Q(a)|\leq E|\ell(a,Y)-\ell(a,Y')|\leq L(a)E[d(Y,Y')]\). This holds for every coupling, so taking the infimum of the final expectation over couplings gives the claimed inequality. \(\square\)

Proof. Add and subtract the risks under \(Q\) at the two actions: \[ R_P(a_Q)-R_P(a_P) =\{R_P(a_Q)-R_Q(a_Q)\} +\{R_Q(a_Q)-R_Q(a_P)\} +\{R_Q(a_P)-R_P(a_P)\}. \] The middle term is nonpositive because \(a_Q\) minimizes risk under \(Q\). The first and third terms are bounded above by their absolute values. Applying the lemma at \(a_Q\) and \(a_P\) gives \[ R_P(a_Q)-R_P(a_P) \leq L(a_Q)W_1(P,Q)+L(a_P)W_1(P,Q), \] as required. Under the uniform condition, \(L(a_Q)+L(a_P)\leq 2L\). \(\square\)

This argument explains why the optimal actions matter. Without a uniform Lipschitz constant, it is enough that the loss be Lipschitz at the two actions actually compared, with finite constants there. Conversely, small Wasserstein error does not by itself ensure small regret if those constants are large or unavailable.

Approximate Decisions and Varying Covariates

In practice, an optimization algorithm may return an action that is close to optimal under the estimated law, rather than an exact minimizer. If \(a_Q\) satisfies \(R_Q(a_Q)\leq \inf_{a\in\mathcal A}R_Q(a)+\delta\) for some \(\delta\geq0\), and a true optimal action \(a_P\) exists, the same decomposition gives \[ R_P(a_Q)-R_P(a_P) \leq \{L(a_Q)+L(a_P)\}W_1(P,Q)+\delta. \] Only the middle term changes: it is now at most \(\delta\), rather than nonpositive. Under a uniform \(L\)-Lipschitz condition, the bound is \(2LW_1(P,Q)+\delta\).

For conditional decisions, apply the pointwise result at each covariate value. Write \(a_P(x)\) and \(a_Q(x)\) for measurable choices of optimal actions under \(P_x\) and \(Q_x\), and assume the relevant conditional risks and Wasserstein distances are measurable and integrable. Then the conditional regret satisfies, for almost every \(x\), \[ R_{P_x}(a_Q(x))-R_{P_x}(a_P(x)) \leq \{L(a_Q(x))+L(a_P(x))\}W_1(P_x,Q_x). \] Taking expectation over \(X\) gives the corresponding bound on average regret. If the loss is uniformly \(L\)-Lipschitz, this simplifies to \(2L E[W_1(P_X,Q_X)]\). For approximate conditional minimizers, add \(E[\delta(X)]\), provided it is finite. These statements hold for each fixed estimated conditional law satisfying the assumptions; if that law is random, they can be read conditionally on the data used to construct it, whenever the stated assumptions hold conditionally.

Common Pitfall: Wasserstein Error Is Not a Universal Regret Bound

A tempting but incorrect argument is: “The estimated conditional law is close in Wasserstein distance, so every decision made from it must have small regret.” The missing step is a link between distribution distance and the particular loss. The coupling proof supplies that link only when the loss is Lipschitz in the outcome, with suitable constants at the actions being compared. For example, an indicator loss with a discontinuity on a continuous outcome space need not be Lipschitz under the usual Euclidean metric. A small movement in outcome can then change the loss abruptly, and the lemma does not apply. One must establish a different loss-specific bound or add assumptions that control probability near the discontinuity.

A second mistake is to bound the estimated action’s risk difference and omit the true optimal action’s risk difference. The regret decomposition has two distribution-comparison terms, one at each action. This is why the uniform result contains \(2L\), not \(L\). Also, the bound is an upper bound, not an equality: as the squared-error example shows, actual regret can be much smaller than the Wasserstein-based bound.