Equivalent Tests for Subadditivity and Diversification

From an Axiom to a Portfolio Test

Subadditivity is often stated as an inequality between two loss variables. Its practical meaning is a portfolio comparison: measure the risks together, then compare that result with the risks measured separately. If the measure is subadditive, combining risks cannot produce a larger risk amount than the sum of the separate amounts. The gap is a measure of the diversification benefit under that risk measure.

For this discussion, let $X$ and $Y$ be losses on the same valuation basis, and let $\rho$ be a risk measure that assigns a real-valued risk amount to each loss in its domain. Assume the sums and weighted sums considered below remain in that domain. Losses are positive amounts of adverse outcome; the results do not depend on a particular currency unit.

Definition: A risk measure $\rho$ is subadditive if, for every pair of losses $X,Y$ in its domain, $$ \rho(X+Y)\leq \rho(X)+\rho(Y). $$ For a portfolio with component losses $X_1,\ldots,X_n$ and nonnegative exposures $w_1,\ldots,w_n$, the corresponding weighted portfolio loss is $\sum_{j=1}^n w_jX_j$. Subadditivity implies that its risk measure is no greater than the sum of the appropriately scaled standalone risk measures.

The word “appropriately” matters. The portfolio loss and the standalone losses must be put on the same exposure scale. For example, the risk measure of $2X$ is compared with the risk of two units of $X$, not with the risk of one unit. Positive homogeneity makes this scaling exact: $\rho(cX)=c\rho(X)$ for $c\geq0$. Without that property, a comparison using normalized portfolio weights does not automatically test subadditivity.

Equivalent Forms Under Positive Homogeneity

Two familiar properties give equivalent ways to test subadditivity. A risk measure is convex if, for $0\leq\lambda\leq1$, $$ \rho\bigl(\lambda X+(1-\lambda)Y\bigr) \leq \lambda\rho(X)+(1-\lambda)\rho(Y). $$ Convexity tests mixtures whose weights sum to one. Subadditivity tests a sum without that normalization. If the measure is positively homogeneous, scaling connects the two.

Theorem: On a domain closed under nonnegative scaling and addition, a positively homogeneous risk measure is subadditive if and only if it is convex. Consequently, for nonnegative weights $w_j$, subadditivity implies $$ \rho\left(\sum_{j=1}^n w_jX_j\right) \leq \sum_{j=1}^n w_j\rho(X_j). $$ If the weights sum to one, this is the convex-combination test.

Here is the full two-loss justification. Suppose first that $\rho$ is subadditive and positively homogeneous. Apply subadditivity to the two scaled losses, then use positive homogeneity:

\begin{aligned} \rho\bigl(\lambda X+(1-\lambda)Y\bigr) &\leq \rho(\lambda X)+\rho((1-\lambda)Y)\\ &=\lambda\rho(X)+(1-\lambda)\rho(Y). \end{aligned}

Thus $\rho$ is convex. Conversely, suppose $\rho$ is convex and positively homogeneous. Convexity at $\lambda=\tfrac12$ gives $\rho((X+Y)/2)\leq(\rho(X)+\rho(Y))/2$. Multiply both sides by two, and use positive homogeneity on the left:

\rho(X+Y) =2\rho\left(\frac{X+Y}{2}\right) \leq \rho(X)+\rho(Y).

This proves subadditivity. The portfolio version follows by repeatedly applying the two-loss inequality to a finite sum, then using positive homogeneity on each component: $\rho(w_jX_j)=w_j\rho(X_j)$. Alternatively, when weights sum to one, the finite-weight inequality follows by repeated application of convexity. This distinction lets an analyst choose a test suited to the data: use normalized weights to compare mixtures, or arbitrary nonnegative weights to test the full scaled portfolio inequality.

Worked Portfolio Comparisons

Worked Example: Calculating a Diversification Benefit

A hypothetical insurer evaluates three lines of business under three equally weighted annual scenarios. Losses are stated in thousands of dollars. The risk measure is the maximum loss across the listed scenarios. Calculate each standalone risk amount, the combined portfolio risk, and the diversification benefit.

Scenario	Line 1 loss	Line 2 loss	Line 3 loss
1	2	8	1
2	7	2	4
3	3	1	6

Method. Under this finite-scenario model, $\rho(X)$ is the largest scenario loss for $X$. First take the maximum separately for each line: line 1 has maximum $7$, line 2 has maximum $8$, and line 3 has maximum $6$, all in thousands. The sum of standalone risk amounts is $7+8+6=21$ thousand dollars.

Next add the three line losses within each scenario. Scenario 1 totals $2+8+1=11$; scenario 2 totals $7+2+4=13$; scenario 3 totals $3+1+6=10$. The portfolio maximum is therefore $13$ thousand dollars. The standalone maxima occur in scenarios 2, 1, and 3 for lines 1, 2, and 3, respectively; they do not all occur in the same scenario.

The diversification benefit, defined here as the sum of standalone risk amounts less the portfolio risk amount, is $21-13=8$ thousand dollars. The check is $13\leq21$, as required by subadditivity of the maximum-over-scenarios measure. The common mistake this method avoids is adding the separate scenario maxima and treating that sum as the portfolio’s actual maximum; those maxima need not occur together.

Worked Example: Applying Unequal Portfolio Weights

A hypothetical reinsurer considers three exposures with scenario losses, again in thousands of dollars: $X_1=(4,7,3)$, $X_2=(8,2,1)$, and $X_3=(1,4,6)$. The risk measure is the maximum across the three scenarios. The proposed portfolio uses weights $0.5$, $0.3$, and $0.2$, respectively. Calculate the weighted portfolio risk and the weighted standalone bound.

Method. First calculate the standalone maxima: $\rho(X_1)=7$, $\rho(X_2)=8$, and $\rho(X_3)=6$. The weighted standalone amount is $0.5(7)+0.3(8)+0.2(6)=3.5+2.4+1.2=7.1$ thousand dollars.

Now form each weighted scenario total. Scenario 1 is $0.5(4)+0.3(8)+0.2(1)=2+2.4+0.2=4.6$. Scenario 2 is $0.5(7)+0.3(2)+0.2(4)=3.5+0.6+0.8=4.9$. Scenario 3 is $0.5(3)+0.3(1)+0.2(6)=1.5+0.3+1.2=3.0$. Therefore the weighted portfolio risk is $4.9$ thousand dollars, and the diversification benefit relative to the weighted standalone amount is $7.1-4.9=2.2$ thousand dollars.

The check is $4.9\leq7.1$. The weights sum to one, so this is a convex-combination test as well as a scaled portfolio comparison. The common mistake is to compare the weighted portfolio with the unweighted standalone sum $7+8+6=21$, which answers a different exposure question and exaggerates the apparent benefit.

Why Weight Normalization Can Mislead

A normalized-weight test is equivalent to the general subadditivity test only when positive homogeneity is available. Without it, passing a convex-combination comparison does not establish subadditivity. The following artificial risk functional makes the limitation explicit; it is an algebraic illustration, not a proposed capital standard.

Worked Example: A Normalized Test That Misses a Failure

Consider losses measured in abstract risk units, and define $\rho(X)=E[X]+(E[X])^2$ for the bounded losses in this example. Let $X=2$ with certainty, so $E[X]=2$. Test both the equal-weight mixture of two identical positions and the unnormalized sum.

Method. The standalone value is $\rho(X)=2+2^2=6$. The equal-weight mixture is $\tfrac12X+\tfrac12X=X$, so its risk is $6$. The corresponding weighted standalone amount is $\tfrac12(6)+\tfrac12(6)=6$. The normalized test passes with equality.

For the unnormalized sum, $X+X=4$ with certainty. Thus $\rho(X+X)=4+4^2=20$, while $\rho(X)+\rho(X)=6+6=12$. The subadditivity inequality would require $20\leq12$, which is false.

The source of the discrepancy is the failed scaling condition: $\rho(2X)=20$, whereas $2\rho(X)=12$. The normalized mixture did not test that scaling. The common mistake is to infer subadditivity merely because a convex-combination comparison passed. The inference requires positive homogeneity; this example’s functional does not have it.

A Practical Testing Procedure

When examining a risk measure or a reported portfolio result, keep the logical test separate from the numerical calculation. State the property being assumed, define the loss units and weights, and then compare like with like. A failed inequality for even one admissible pair disproves subadditivity on that domain; a passing finite set of examples is evidence about those cases, not a proof of the axiom for every possible loss.

Specify the domain and scale.
Identify each component loss, its exposure, and the units of the risk amounts. Confirm that the combined loss is in the risk measure’s domain.

Choose the appropriate inequality.
For a direct axiom check, compare $\rho(X+Y)$ with $\rho(X)+\rho(Y)$. For normalized weights, compare the risk of the mixture with the weighted standalone risks, and state the assumptions that justify this form.

Compute the portfolio before interpreting it.
Combine the losses scenario by scenario or according to the specified model, then apply the risk measure. Do not add separate risk amounts and label the result the portfolio risk.

Check the conclusion and its limits.
A nonnegative difference between weighted standalone risk and portfolio risk is a diversification benefit for that measure and portfolio. It does not by itself prove that the measure is subadditive in every case.

Subadditivity is important because it prevents the risk measure from penalizing a combined portfolio merely for holding several risks together. The size of any calculated benefit depends on the risk measure, the joint loss outcomes, and the chosen exposures. It is not a guaranteed reduction in realized losses, nor does it imply that every pair of components diversifies equally.

A reliable exam approach is to distinguish three claims: the axiom itself, a consequence of the axiom for a particular weighted portfolio, and an empirical calculation under a particular scenario model. Convexity and positive homogeneity connect the first two; the scenario calculation supplies the third. These tests are aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Key takeaway: Subadditivity means the risk of a sum is no greater than the sum of the component risks. With positive homogeneity, it is equivalent to convexity and supports weighted portfolio tests. A normalized-weight comparison alone is not a general subadditivity test unless the required scaling property holds.

Check Your Understanding

For each problem, keep the portfolio weights and risk amounts on the same scale.

A risk measure is subadditive. For two losses with risk amounts $12$ and $9$, what upper bound does the axiom give for the risk of their sum? If the measured portfolio risk is $17$, what is the diversification benefit? (Answer: Upper bound $21$; benefit $4$.)
A subadditive, positively homogeneous risk measure assigns standalone risk amounts $10$ and $20$. A portfolio holds $0.4$ units of the first loss and $0.6$ units of the second. What is the weighted standalone upper bound for portfolio risk? (Answer: $0.4(10)+0.6(20)=16$.)
Suppose $\rho$ is both convex and positively homogeneous. Use the convexity inequality at weight one-half and the scaling property to obtain the subadditivity bound for $\rho(X+Y)$. (Answer: $\rho(X+Y)=2\rho((X+Y)/2)\leq\rho(X)+\rho(Y)$.)
Under the maximum-over-scenarios measure, two hypothetical losses have scenario values $A=(3,9,4)$ and $B=(8,2,5)$, in thousands of dollars. Find the sum of their standalone risk amounts, the portfolio risk, and the diversification benefit. (Answer: Standalone sum $9+8=17$; portfolio scenarios $(11,11,9)$, so risk $11$; benefit $6$ thousand dollars.)
For deterministic $X=2$, use the artificial functional $\rho(Z)=E[Z]+(E[Z])^2$. Does the equal-weight mixture test for two copies of $X$ detect a subadditivity failure? Give the risk of the mixture, the weighted standalone amount, and the risk of the unnormalized sum. (Answer: No; mixture risk $6$, weighted standalone amount $6$, but $\rho(X+X)=20>12=\rho(X)+\rho(X)$.)

What You'll Learn