Divergence-Ball Distributionally Robust Risk

Why Put a Neighborhood Around a Reference Distribution?

A prediction rule can have low risk under the distribution used to evaluate it and still perform poorly if the evaluation distribution differs. In distributionally robust prediction, we do not select one alternative distribution and treat it as known. Instead, we specify a neighborhood around a reference distribution and ask for the largest risk over that neighborhood. The result is a guarantee against a declared class of distribution shifts, not against every possible shift.

Let $Z$ denote the random observation on which a prediction is evaluated, let $P$ be its reference distribution, and let $h$ be a fixed prediction rule. Write $\ell_h(Z)$ for its measurable loss. For now, assume the loss is bounded: $a\leq \ell_h(Z)\leq b$, $P$-almost surely, for finite constants $a$ and $b$. This assumption ensures the risks below are finite for the distributions considered.

Definition: The reference risk of $h$ is $R_P(h)=E_P[\ell_h(Z)]$. Given a nonnegative divergence $D(Q\|P)$ and radius $\rho\geq 0$, the divergence ball around $P$ is $\mathcal{B}_{D,\rho}(P)=\{Q:Q\ll P,\ D(Q\|P)\leq\rho\}$. The divergence-ball distributionally robust risk is $$ R^{\mathrm{rob}}_{D,\rho}(h;P)=\sup_{Q\in\mathcal{B}_{D,\rho}(P)} E_Q[\ell_h(Z)]. $$ Here $Q\ll P$ means that $Q$ is absolutely continuous with respect to $P$: every event with zero probability under $P$ also has zero probability under $Q$.

The absolute-continuity restriction makes the comparison with density ratios explicit: for every admissible $Q$, a Radon–Nikodym derivative $w=dQ/dP$ exists and $E_P[w]=1$. It also means the ambiguity set cannot assign positive probability to an event that is impossible under $P$. A divergence ball therefore expresses robustness only within its specified support and divergence constraint.

Three Ways to Measure Distributional Discrepancy

For $Q\ll P$, set $w=dQ/dP$. The following three divergences all vanish when $Q=P$, but they constrain changes in different ways. In the formulas, the KL divergence is allowed to be infinite if its integrand is not integrable.

Definitions: The Kullback–Leibler divergence, chi-square divergence, and total-variation distance are $$ D_{\mathrm{KL}}(Q\|P)=E_P[w\log w],\qquad D_{\chi^2}(Q\|P)=E_P[(w-1)^2],\qquad d_{\mathrm{TV}}(Q,P)=\sup_A |Q(A)-P(A)|. $$ We use $0\log 0=0$. For $Q\ll P$, $d_{\mathrm{TV}}(Q,P)=\tfrac12E_P[|w-1|]$.

KL divergence is directional: in general $D_{\mathrm{KL}}(Q\|P)\neq D_{\mathrm{KL}}(P\|Q)$, and reversing the arguments changes what support failures are penalized. The chi-square divergence directly measures the squared deviation of the density ratio from one. Total variation controls the largest probability change over any event. These quantities have different scales, so equal numerical radii do not mean equal neighborhoods.

Worked Example: A Chi-Square Ball Uses Reference Variance

Suppose $Z$ takes three values, with $P=(0.5,0.3,0.2)$, and the corresponding losses are $0,1,3$. Consider a chi-square ball of radius $\rho=0.04$. Find a bound on its worst-case risk.

The reference risk and second moment are $$ R_P(h)=0.5(0)+0.3(1)+0.2(3)=0.9, \qquad E_P[\ell_h^2]=0.5(0^2)+0.3(1^2)+0.2(3^2)=2.1. $$ Thus $\operatorname{Var}_P(\ell_h)=2.1-(0.9)^2=1.29$. The chi-square risk bound proved below gives $$ R^{\mathrm{rob}}_{\chi^2,0.04}(h;P) \leq 0.9+\sqrt{1.29(0.04)} =0.9+\sqrt{0.0516} \approx 1.1272. $$ The loss is also at most $3$, so this bound is below the trivial upper bound of $3$. It is a guarantee, not a claim that some distribution in the ball attains risk $1.1272$.

Basic Upper Bounds on Worst-Case Risk

The loss range and its variability under $P$ determine how strongly a divergence constraint limits risk. For a total-variation ball, the range gives a bound. For a chi-square ball, the reference variance gives a potentially sharper one. A KL ball can be handled through Pinsker’s inequality, which converts KL divergence into a total-variation bound.

Theorem: Suppose $a\leq \ell_h\leq b$, $P$-almost surely, and define $\mu=E_P[\ell_h]$ and $v=\operatorname{Var}_P(\ell_h)$. For the divergence balls defined above, $$ \begin{aligned} R^{\mathrm{rob}}_{\mathrm{TV},\rho}(h;P)&\leq \min\{b,\ \mu+(b-a)\rho\},\\ R^{\mathrm{rob}}_{\chi^2,\rho}(h;P)&\leq \min\{b,\ \mu+\sqrt{v\rho}\},\\ R^{\mathrm{rob}}_{\mathrm{KL},\rho}(h;P)&\leq \min\{b,\ \mu+(b-a)\sqrt{\rho/2}\}. \end{aligned} $$ The total-variation bound applies when $d_{\mathrm{TV}}(Q,P)\leq\rho$; the chi-square bound applies when $D_{\chi^2}(Q\|P)\leq\rho$; and the KL bound applies when $D_{\mathrm{KL}}(Q\|P)\leq\rho$.

Proof. Fix an admissible $Q$ and let $w=dQ/dP$. For total variation, $E_P[w-1]=0$, and the positive and negative parts of $w-1$ have equal integrals. Each integral is $d_{\mathrm{TV}}(Q,P)$. Therefore $$ E_Q[\ell_h]-E_P[\ell_h] =E_P[(w-1)(\ell_h-a)] \leq (b-a)E_P[(w-1)_+] =(b-a)d_{\mathrm{TV}}(Q,P). $$ Here $x_+=\max\{x,0\}$. The inequality uses $0\leq\ell_h-a\leq b-a$: the contribution where $w-1<0$ is nonpositive, and the positive contribution is at most $(b-a)E_P[(w-1)_+]$. The asserted total-variation bound follows.

For chi-square divergence, centering the loss and applying Cauchy–Schwarz gives $$ \begin{aligned} E_Q[\ell_h]-E_P[\ell_h] &=E_P[(w-1)(\ell_h-\mu)]\\ &\leq \sqrt{E_P[(w-1)^2]\,E_P[(\ell_h-\mu)^2]}\\ &=\sqrt{D_{\chi^2}(Q\|P)\,v} \leq \sqrt{\rho v}. \end{aligned} $$ This proves the chi-square bound, including when $v=0$.

For the KL bound, first establish $d_{\mathrm{TV}}(Q,P)\leq\sqrt{D_{\mathrm{KL}}(Q\|P)/2}$. Let $A=\{w\geq1\}$, $p=P(A)$, and $q=Q(A)$. The positive-part identity above implies $q-p=d_{\mathrm{TV}}(Q,P)$. Convexity of $x\log x$, applied separately on $A$ and its complement, yields $$ D_{\mathrm{KL}}(Q\|P)\geq q\log\frac{q}{p}+(1-q)\log\frac{1-q}{1-p}. $$ For $0<p<1$, denote the expression on the right by $g(q)$. It has $g(p)=g'(p)=0$, and for $0<q<1$, $$ g''(q)=\frac{1}{q(1-q)}\geq4. $$ Integrating this second-derivative bound from $p$ to $q$, in either direction, gives $g(q)\geq2(q-p)^2$. Endpoint cases follow by continuity; if $p=0$ or $p=1$, absolute continuity gives $Q=P$ on the corresponding event and the total variation is zero. This proves the stated KL-to-total-variation inequality. Applying the total-variation risk bound then gives the KL risk bound. Finally, every admissible risk is at most $b$, so taking the smaller of $b$ and each derived bound completes the proof. $\square$

Worked Example: A Total-Variation Ball Gives an Exact Bound

Let $Z$ be binary with $P(Z=1)=0.2$, $P(Z=0)=0.8$, and loss $\ell_h(Z)=Z$. For a total-variation ball of radius $0.1$, what is the worst-case risk?

For any admissible $Q$, $E_Q[\ell_h]=Q(Z=1)$. The defining event bound for total variation gives $$ Q(Z=1)\leq P(Z=1)+d_{\mathrm{TV}}(Q,P)\leq0.2+0.1=0.3. $$ This upper bound is attained by $Q(Z=1)=0.3$, $Q(Z=0)=0.7$. This $Q$ is absolutely continuous with respect to $P$, and its total-variation distance is $\lvert0.3-0.2\rvert=0.1$. Hence the worst-case risk is exactly $0.3$. Here the range-based theorem bound is also exact.

Worked Example: Different Divergences for the Same Shift

Let $P=(0.5,0.5)$ and $Q=(0.6,0.4)$ on two outcomes. Take the loss to be $1$ on the first outcome and $0$ on the second. Calculate the risk change and the three discrepancies.

The reference and shifted risks are $0.5$ and $0.6$, so the risk change is $0.1$. The total-variation distance is $$ d_{\mathrm{TV}}(Q,P)=\tfrac12(|0.6-0.5|+|0.4-0.5|)=0.1. $$ The density ratios are $1.2$ and $0.8$, so $$ D_{\chi^2}(Q\|P)=0.5(1.2-1)^2+0.5(0.8-1)^2=0.04. $$ The KL divergence is $$ D_{\mathrm{KL}}(Q\|P)=0.6\log(1.2)+0.4\log(0.8)\approx0.02014. $$ The risk bounds reflect these different scales. For the indicator loss, $b-a=1$, $\mu=0.5$, and $v=0.25$. A chi-square radius of $0.04$ gives the upper bound $0.5+\sqrt{0.25(0.04)}=0.6$. A KL radius of approximately $0.02014$ gives $0.5+\sqrt{0.02014/2}\approx0.6003$, slightly above the actual risk. The divergence values describe the same shift but are not interchangeable radius values.

What the Choice of Ball Means

A divergence ball is a modeling choice about which alternatives count as nearby. Total variation has a direct probability interpretation: no event’s probability may change by more than the radius. Chi-square divergence limits the mean squared density-ratio deviation under $P$, which makes its risk bound especially responsive to reference-loss variance. KL divergence limits an information-theoretic discrepancy and, through Pinsker’s inequality, controls event-probability changes as well. The resulting risk bounds are general guarantees; they need not be sharp for a particular loss or reference distribution.

The common support restriction matters in all three cases here. If an outcome has zero probability under $P$, no admissible $Q$ can assign it positive probability. Thus these balls cannot protect against shifts that introduce entirely new support. And because KL is directional, a constraint on $D_{\mathrm{KL}}(Q\|P)$ should not be read as a constraint on $D_{\mathrm{KL}}(P\|Q)$.

Common Pitfall

Wrong reasoning: “The KL divergence is small, so the worst-case risk must be close to the reference risk by an amount proportional to the KL radius.” This treats the divergence value itself as a risk difference and ignores both the loss scale and the square-root relationship in the general Pinsker bound.

Correction: State the loss range and the divergence convention, then apply a bound with the correct scale. For a loss in $[a,b]$, the KL guarantee established above is $\mu+(b-a)\sqrt{\rho/2}$, capped at $b$. It is an upper bound, not generally an equality. Also verify the direction of the divergence and whether the support restriction matches the shifts of interest.

Key takeaway: Divergence-ball robust risk is the largest expected loss over a specified neighborhood of a reference law. Total variation controls probability changes, chi-square divergence controls squared density-ratio changes, and KL divergence yields a range-based guarantee through Pinsker’s inequality. The loss scale, reference variability, support, and divergence direction all affect what the guarantee says.

Check Your Understanding

Use the definitions and bounds above to answer the following.

Why does the definition require $Q\ll P$, and what kind of shift is excluded by that restriction?
For a loss bounded between $a$ and $b$, which feature of the reference loss appears in the chi-square risk bound?
If $d_{\mathrm{TV}}(Q,P)\leq\rho$, how much can the probability of any single event change?
Explain why a KL radius cannot be compared numerically with an equal chi-square radius without further interpretation.
For a fixed loss in $[0,1]$, what does the general KL risk bound imply when the KL radius is $\rho$?