Stop-Loss Characterization of Convex Order

From Convex Functions to Threshold Losses

Convex order compares dispersion while holding the mean fixed. Its definition ranges over every convex function, which can make the comparison seem to require many unrelated calculations. The stop-loss characterization reduces that task to a single family of functions: the positive-part functions \(x\mapsto (x-t)_+\), one for each real threshold \(t\).

For an integrable random variable \(X\), the quantity \(E[(X-t)_+]\) is the expected amount by which \(X\) exceeds \(t\). In insurance, this is the expected payment under a deductible \(t\), often called the stop-loss premium. The characterization says that, among variables with the same mean, one has larger convex dispersion exactly when all of these expected excesses are larger.

Definition: For real-valued integrable random variables \(X\) and \(Y\), write \(X\le_{\mathrm{cx}}Y\) (say, \(X\) is smaller than \(Y\) in convex order) when \(E[X]=E[Y]\) and \(E[\phi(X)]\le E[\phi(Y)]\) for every finite convex function \(\phi:\mathbb{R}\to\mathbb{R}\). These expectations are understood as extended expectations when necessary; a finite convex function has an affine lower bound, so its negative part has finite expectation under an integrable variable. Define the stop-loss transform by \(S_X(t)=E[(X-t)_+]\), where \(u_+=\max(u,0)\).

Theorem (Stop-Loss Characterization of Convex Order): If \(X\) and \(Y\) are integrable and \(E[X]=E[Y]\), then \(X\le_{\mathrm{cx}}Y\) if and only if \(S_X(t)\le S_Y(t)\) for every \(t\in\mathbb{R}\).

The Direct Implication

For any fixed threshold \(t\), the function \(\phi_t(x)=(x-t)_+\) is convex: it is the maximum of the two affine functions \(x-t\) and \(0\). Therefore, if \(X\le_{\mathrm{cx}}Y\), the defining inequality for convex functions applies to \(\phi_t\), giving \(E[(X-t)_+]\le E[(Y-t)_+]\). This holds for every real \(t\), proving the forward implication.

Worked Example: A Constant Compared With a Symmetric Spread

Let \(X=0\) with probability one, and let \(Y\) take the values \(-1\) and \(1\), each with probability \(1/2\). Both variables have mean zero. We verify the stop-loss inequalities over the full range of thresholds.

If \(t<-1\), both values of \(Y\) exceed \(t\), so \(S_Y(t)=E[Y-t]=-t\). Also \(S_X(t)=-t\). If \(-1\le t<0\), only the value \(1\) contributes for \(Y\), giving \(S_Y(t)=(1-t)/2\), while \(S_X(t)=-t\). Their difference is \((1-t)/2-(-t)=(1+t)/2\ge0\). If \(0\le t<1\), then \(S_X(t)=0\) and \(S_Y(t)=(1-t)/2\ge0\). For \(t\ge1\), both transforms are zero.

Thus \(S_X(t)\le S_Y(t)\) for every \(t\), and the theorem gives \(X\le_{\mathrm{cx}}Y\). The calculation also shows where the two transforms agree and where the greater spread of \(Y\) produces a strict inequality.

Representing Convex Functions by Stop-Loss Functions

The converse rests on a representation of any finite convex function as an affine function plus a nonnegative mixture of positive-part functions. The nonnegative weights encode the function’s curvature. To make the representation valid on both sides of zero, it is convenient to subtract an affine term and use a nonnegative kernel.

Definition: For \(x,t\in\mathbb{R}\), define \(K(x,t)=(x-t)_+-(-t)_+-x\,\mathbf{1}_{\{t<0\}}\), where \(\mathbf{1}_{\{t<0\}}\) equals 1 when \(t<0\) and 0 otherwise. This kernel is nonnegative. When \(t\ge0\), it is \((x-t)_+\); when \(t<0\), it is \((t-x)_+\).

A finite convex function \(\phi\) has left and right derivatives everywhere, with nondecreasing slopes. Its distributional second derivative is a nonnegative, locally finite measure \(\mu\); the mass \(\mu((a,b])\) records the increase in slope across an interval. With \(a=\phi'_-(0)\), the function has the representation

\phi(x)=\phi(0)+ax+\int_{\mathbb{R}}K(x,t)\,\mu(dt).

Here \(\phi'_-(0)\) denotes the left derivative at zero. The integral is finite for each fixed \(x\), since it equals the difference between finite values of \(\phi\) and an affine function. The representation includes a possible corner at zero: its curvature appears as an atom of \(\mu\) at \(t=0\).

Proof of the representation: For \(s>0\), the slope satisfies \(\phi'_+(s)=a+\mu([0,s])\). Integrating from 0 to \(x\ge0\) gives \(\phi(x)-\phi(0)=ax+\int_{[0,x]}(x-t)\,\mu(dt)\), by interchanging the nonnegative integrals. This is the stated formula because \(K(x,t)=(x-t)_+\) for \(t\ge0\), and \(K(x,t)=0\) for \(t<0\) when \(x\ge0\). For \(s<0\), the left slope satisfies \(\phi'_-(s)=a-\mu([s,0))\). Integrating from \(x<0\) to 0 gives \(\phi(x)-\phi(0)=ax+\int_{(x,0)}(t-x)\,\mu(dt)\). This is the stated formula because for \(t<0\), \(K(x,t)=(t-x)_+\), while for \(t\ge0\) it is zero when \(x<0\). The slope identities follow from the measure of slope increases, and convex functions are locally absolutely continuous, so these integrations recover their function values. \(\square\)

Proof of the Converse

Suppose now that \(E[X]=E[Y]\) and \(S_X(t)\le S_Y(t)\) for every real \(t\). The kernel was chosen so that its expectations are ordered by exactly these assumptions. For \(t\ge0\), \(K(X,t)=(X-t)_+\), so the required inequality is the assumed stop-loss inequality. For \(t<0\), use the definition of \(K\) and equal means:

E[K(X,t)] =S_X(t)-(-t)_+-E[X]\mathbf{1}_{\{t<0\}} \le S_Y(t)-(-t)_+-E[Y]\mathbf{1}_{\{t<0\}} =E[K(Y,t)].

The kernel is nonnegative, so Tonelli’s theorem applies to its integral against the positive measure \(\mu\), even if an expectation of a convex function is infinite. Taking expectations in the convex-function representation therefore gives

E[\phi(X)] =\phi(0)+aE[X]+\int_{\mathbb{R}}E[K(X,t)]\,\mu(dt) \le \phi(0)+aE[Y]+\int_{\mathbb{R}}E[K(Y,t)]\,\mu(dt) =E[\phi(Y)].

The affine terms agree because the means agree, and the integral inequality follows from the pointwise-in-\(t\) expectation comparison. This holds for every finite convex \(\phi\), proving \(X\le_{\mathrm{cx}}Y\) and completing the theorem. \(\square\)

Worked Example: Comparing Two Discrete Spreads

Let \(X\) take \(-1\) and \(1\) with probability \(1/2\) each. Let \(Y\) take \(-2,0,2\) with probabilities \(1/4,1/2,1/4\), respectively. Both means are zero. We calculate their stop-loss transforms and compare them on every interval determined by their support points.

For \(X\), \(S_X(t)=-t\) when \(t<-1\), \(S_X(t)=(1-t)/2\) when \(-1\le t<1\), and \(S_X(t)=0\) when \(t\ge1\). For \(Y\), \(S_Y(t)=-t\) when \(t<-2\). When \(-2\le t<0\), the values 0 and 2 contribute, so \(S_Y(t)=\tfrac12(0-t)+\tfrac14(2-t)=\tfrac12-\tfrac34t\). When \(0\le t<2\), only the value 2 contributes, giving \(S_Y(t)=\tfrac14(2-t)=\tfrac12-\tfrac14t\). For \(t\ge2\), \(S_Y(t)=0\).

For \(t<-2\), the transforms agree. On \([-2,-1)\), their difference is \((\tfrac12-\tfrac34t)-(-t)=\tfrac12+\tfrac14t\ge0\). On \([-1,0)\), the difference is \((\tfrac12-\tfrac34t)-(\tfrac12-\tfrac12t)=-\tfrac14t\ge0\). On \([0,1)\), it is \((\tfrac12-\tfrac14t)-(\tfrac12-\tfrac12t)=\tfrac14t\ge0\). On \([1,2)\), \(S_X(t)=0\le S_Y(t)\), and for \(t\ge2\) both are zero. Hence \(X\le_{\mathrm{cx}}Y\).

Why Equal Means Matter

The equal-mean condition is not optional. Stop-loss functions use thresholds in the increasing direction, so comparing them alone can also reflect a shift in location rather than a pure increase in dispersion. Equal means remove that shift and ensure the affine parts of every convex function have the same expectation.

Worked Example: Stop-Loss Inequalities Without Convex Order

Take \(X=0\) and \(Y=1\), both constant. For every real \(t\), \((X-t)_+\le(Y-t)_+\), since \(x\mapsto(x-t)_+\) is increasing. Thus \(S_X(t)\le S_Y(t)\) at every threshold. But \(E[X]=0\ne1=E[Y]\), and \(X\not\le_{\mathrm{cx}}Y\) under the definition requiring equal means.

Indeed, the function \(\phi(x)=-x\) is convex because it is affine, but \(E[\phi(X)]=0\) and \(E[\phi(Y)]=-1\), so the convex-function inequality in the direction \(X\le_{\mathrm{cx}}Y\) fails. Stop-loss comparison without equal means does not characterize convex order.

Interpretation and Common Pitfall

The theorem is useful because threshold losses are often easier to calculate or estimate than expectations of arbitrary convex functions. It also makes the comparison operational: at every threshold, the expected excess above that threshold is no larger for \(X\) than for \(Y\). The requirement that every real threshold be checked matters; a comparison at only a few selected thresholds is not the theorem’s condition.

Pitfall: forgetting the mean condition. The incorrect reasoning is, “All stop-loss expectations for \(X\) are below those for \(Y\), therefore \(X\le_{\mathrm{cx}}Y\).” The constant example shows why this fails: the means differ. Check equality of means as well as the full family of threshold inequalities.
Pitfall: checking only nonnegative thresholds. The stop-loss criterion requires all \(t\in\mathbb{R}\). Negative thresholds capture behavior below zero; the equal-mean condition does not make those comparisons redundant.
Pitfall: treating one threshold as decisive. A single inequality verifies only one convex test function, not all convex functions. The representation argument is what justifies passing from every threshold to an arbitrary convex function.

Key takeaway: For integrable variables with equal means, convex order is equivalent to comparing \(E[(X-t)_+]\) at every real threshold. The forward direction tests convex order with positive-part functions; the converse follows by representing any convex function as an affine term plus a nonnegative mixture of threshold kernels.

Check Your Understanding

Use the definitions and proof above to answer these questions.

Why is \(x\mapsto(x-t)_+\) a valid convex test function for every real \(t\)?
If \(X\) and \(Y\) have equal means and \(S_X(t)\le S_Y(t)\) for all \(t\), what does the stop-loss characterization imply?
For \(t<0\), identify the term in \(K(x,t)\) whose expectation comparison uses equality of means.
Why does nonnegativity of \(K\) permit Tonelli’s theorem in the converse proof?
What condition fails in the comparison \(X=0\), \(Y=1\) when their stop-loss transforms are ordered?