Subadditivity and Portfolio Diversification

Subadditivity and the Meaning of Diversification

A portfolio combines losses from multiple risks. The combined loss is the sum of the component losses, but the risk assigned to that sum need not equal the sum of the stand-alone risk assessments. Subadditivity says the portfolio assessment cannot exceed that sum. It is one mathematical expression of the idea that combining risks should not make the assessed risk larger than assessing each risk separately and adding the results.

This condition does not promise that combining risks will reduce assessed risk, and it does not say that every possible portfolio has less risk than each component. It compares the risk of the combined loss with the sum of the risks assessed separately. To make that comparison, the component losses must be defined on a common set of possible outcomes, so their sum represents the portfolio loss in each outcome.

Definition: Let $X$ and $Y$ be loss random variables on the same probability space, and let $\rho$ be a risk measure that assigns a number to each loss. The risk measure is subadditive if $$ \rho(X+Y) \leq \rho(X)+\rho(Y) $$ for every eligible pair $X,Y$. Here $X+Y$ is the aggregate loss in each outcome. The two sides have the same units, such as dollars or thousands of dollars.

The amount $\rho(X)+\rho(Y)-\rho(X+Y)$ is the subadditivity margin for this particular pair of losses. A positive margin means the assessed portfolio risk is below the sum of the stand-alone assessments. A zero margin means equality. A negative margin means the inequality fails for this pair, so the risk measure is not subadditive over a class that includes these losses.

How Subadditivity Fits the Coherence Axioms

A coherent risk measure for losses is required to satisfy four properties. The earlier material on monotonicity, translation invariance, and positive homogeneity gives their interpretation; here, the key distinction is that each property imposes a different test.

Property	Condition for losses	What it tests
Monotonicity	If $X \leq Y$ in every outcome, then $\rho(X) \leq \rho(Y)$.	Whether a loss that is never larger receives no greater risk assessment.
Translation invariance	For a constant $c$, $\rho(X+c)=\rho(X)+c$.	Whether adding a certain amount to every outcome increases assessed risk by that amount.
Positive homogeneity	For a constant $a \geq 0$, $\rho(aX)=a\rho(X)$.	Whether scaling every loss scales assessed risk in the same way.
Subadditivity	$\rho(X+Y) \leq \rho(X)+\rho(Y)$.	Whether combining losses is assessed at no more than the sum of stand-alone assessments.

Subadditivity is not another name for monotonicity. Monotonicity compares two losses when one is no larger than the other outcome by outcome. Subadditivity compares the risk assigned to a sum with the sum of two separate risk assignments. Nor does translation invariance imply subadditivity: the first concerns adding a fixed amount to every outcome, while the second concerns combining potentially variable losses.

Likewise, positive homogeneity concerns scaling one loss, not combining two different losses. A risk measure must satisfy each required axiom separately to be coherent. Passing a subadditivity check for one pair $X,Y$ is not enough to establish the property for every eligible pair.

Why the Maximum-Scenario Risk Is Subadditive

A short derivation makes the portfolio comparison concrete. Suppose a finite set of scenarios is specified, and define the risk assessment of a loss as its largest loss among those scenarios. This is a hypothetical scenario-based measure; it is not a general recipe for estimating risk from observed data.

Definition: For a finite scenario set $\Omega$, define $\rho_{\max}(X)=\max_{\omega\in\Omega}X(\omega)$. This measure assigns each loss its maximum scenario value.

For any particular scenario $\omega$, the loss $X(\omega)$ is at most the largest value of $X$ over all scenarios, and $Y(\omega)$ is at most the largest value of $Y$. Adding these two inequalities gives an upper bound on the portfolio loss in that scenario. Taking the largest portfolio loss over scenarios preserves the bound:

\begin{aligned} X(\omega)+Y(\omega) &\leq \max_{\omega'\in\Omega}X(\omega') +\max_{\omega'\in\Omega}Y(\omega') \quad \text{for every }\omega\in\Omega,\\ \max_{\omega\in\Omega}\bigl(X(\omega)+Y(\omega)\bigr) &\leq \max_{\omega\in\Omega}X(\omega) +\max_{\omega\in\Omega}Y(\omega). \end{aligned}

The final line is precisely $\rho_{\max}(X+Y)\leq\rho_{\max}(X)+\rho_{\max}(Y)$. The derivation does not require the losses to be independent: it uses their values in the same scenarios. The result illustrates subadditivity for this particular measure and setup; it is not a proof that every risk measure is subadditive.

Worked Example: Combining Two Scenario Losses

A hypothetical insurer assesses two lines of business using the largest loss across three specified scenarios. Losses are in thousands of dollars. For lines $A$ and $B$, the scenario losses are $A=(0,100,20)$ and $B=(80,0,30)$, with entries listed in the same scenario order. Calculate the stand-alone assessments and the portfolio assessment, then test subadditivity.

Method. Apply the defined maximum-scenario measure to each line and to the scenario-by-scenario sum. The common mistake to avoid is adding losses from different scenarios instead of matching the entries by scenario.

Stand-alone risks. The largest loss for $A$ is $\max(0,100,20)=100$ thousand dollars. The largest loss for $B$ is $\max(80,0,30)=80$ thousand dollars. Their sum is $100+80=180$ thousand dollars.

Portfolio scenarios. Add the two lines within each scenario: $0+80=80$, $100+0=100$, and $20+30=50$. Thus the portfolio losses are $(80,100,50)$, and the portfolio assessment is $\max(80,100,50)=100$ thousand dollars.

Check. The inequality is $100\leq180$, which holds. The subadditivity margin is $180-100=80$ thousand dollars. The portfolio has a lower maximum-scenario assessment than the sum of the stand-alone assessments, but this conclusion is specific to the stated measure and scenarios.

Equality Is Possible: Expected Loss

Subadditivity allows equality. In fact, expected loss is additive for integrable losses: $E[X+Y]=E[X]+E[Y]$, regardless of dependence. Therefore expected loss satisfies subadditivity with equality, but this equality does not show a reduction in expected loss from combining risks. Other risk assessments may reflect features beyond the mean, but whether they satisfy subadditivity must be checked for the measure in question.

Worked Example: Expected Loss Gives Equality

Two hypothetical claims are modeled over four equally likely scenarios. The losses, in thousands of dollars, are $X=(0,100,0,100)$ and $Y=(0,0,100,100)$. Use expected loss as the risk measure to calculate both stand-alone assessments and the assessment of their aggregate loss.

Method. Calculate each arithmetic mean using the four equally likely outcomes, then add the losses scenario by scenario. The common mistake to avoid is interpreting subadditivity as a requirement for a strictly smaller portfolio assessment.

Risk of $X$. Its expected loss is $(0+100+0+100)/4=200/4=50$ thousand dollars.

Risk of $Y$. Its expected loss is $(0+0+100+100)/4=200/4=50$ thousand dollars. The sum of the stand-alone assessments is $50+50=100$ thousand dollars.

Aggregate loss and assessment. Adding by scenario gives $X+Y=(0,100,100,200)$. Its expected loss is $(0+100+100+200)/4=400/4=100$ thousand dollars.

Check. The portfolio assessment is $100$ thousand dollars, equal to the sum of the stand-alone assessments. Thus $100\leq100$ holds with zero subadditivity margin. The result is consistent with expected loss being additive.

Testing a Risk Assessment Without Confusing the Axioms

In an application, a risk measure may be given by a model or specified assessment method rather than a simple maximum or expectation. To test subadditivity for two particular portfolio components, keep the loss definitions and units consistent, obtain three assessments, and compare the portfolio value with the sum of the stand-alone values. This checks the inequality for that pair; it does not establish subadditivity universally.

Specify the losses.
Identify $X$, $Y$, and the outcomes in which both are defined. The aggregate loss must be $X+Y$ in those same outcomes.

Obtain the three assessments.
Calculate or use the stated values of $\rho(X)$, $\rho(Y)$, and $\rho(X+Y)$ under the same risk measure and conventions.

Compare and interpret.
Check whether $\rho(X+Y)\leq\rho(X)+\rho(Y)$. If it holds, report the margin or equality; if it fails, the pair is a counterexample to subadditivity.

Worked Example: Auditing Three Model Assessments

A hypothetical portfolio model reports risk assessments in thousands of dollars: $\rho(X)=72$, $\rho(Y)=55$, and $\rho(X+Y)=119$. Determine whether the reported combined assessment is subadditive for this pair, and quantify the margin. The values are supplied as outputs of the same model; no probability distribution is needed to perform this audit.

Method. Add the two stand-alone assessments and compare the result with the combined assessment. The common mistake to avoid is comparing the combined value with only one stand-alone value, rather than with their sum.

Stand-alone total. $\rho(X)+\rho(Y)=72+55=127$ thousand dollars.

Subadditivity test. The combined assessment is $119$ thousand dollars, and $119\leq127$, so the condition holds for this pair.

Margin and check. The margin is $127-119=8$ thousand dollars. Equivalently, $119+8=127$, confirming the calculation. This single comparison does not establish that the model is subadditive for every possible pair of losses, nor does it test its other coherence properties.

Common Interpretations and Exam Traps

Subadditivity is a precise inequality, not a guarantee about every informal use of the word “diversification.” The following distinctions help prevent overclaiming:

It does not require a strict reduction. Equality is permitted, as the expected-loss example shows.
It does not compare the portfolio with each component separately. The right side is $\rho(X)+\rho(Y)$, not $\rho(X)$ or $\rho(Y)$ alone.
It is a property of a measure over a specified domain. One successful numerical check establishes the inequality only for that pair.
It does not assume independence. The definition uses the joint losses $X$ and $Y$; independence is neither part of the inequality nor needed for the scenario-based derivation above.
It is distinct from the other coherence axioms. A calculation of $\rho(X+Y)$ versus $\rho(X)+\rho(Y)$ does not by itself test monotonicity, translation invariance, or positive homogeneity.
Keep units and scope aligned. Do not compare assessments stated in dollars with amounts stated in thousands, or use different model conventions for stand-alone and aggregate risks.

Key takeaway: For losses $X$ and $Y$, subadditivity is the condition $\rho(X+Y)\leq\rho(X)+\rho(Y)$. It formalizes a diversification-related limit on assessed aggregate risk, permits equality, and must be checked separately from monotonicity, translation invariance, and positive homogeneity.

Check Your Understanding

For each problem, use the stated loss convention and risk measure. Check the comparison carefully before interpreting it.

A scenario-based measure assigns the maximum loss over two scenarios. Two hypothetical components have losses $A=(30,0)$ and $B=(0,45)$, in thousands of dollars. Find each stand-alone assessment and the portfolio assessment, then state the subadditivity margin. (Answer: $\rho(A)=30$, $\rho(B)=45$, $\rho(A+B)=45$, and the margin is $30$ thousand dollars; subadditivity holds.)
For a risk measure, $\rho(X)=84$, $\rho(Y)=39$, and $\rho(X+Y)=126$, all in thousands of dollars. Does this pair satisfy subadditivity, and by how much? (Answer: No; $126>84+39=123$, so the inequality fails by $3$ thousand dollars.)
Two losses have equally likely scenario values $X=(20,60,40)$ and $Y=(10,0,50)$, in thousands of dollars. Using expected loss, find the two stand-alone assessments, the aggregate assessment, and the margin. (Answer: $E[X]=40$, $E[Y]=20$, $E[X+Y]=60$, all in thousands of dollars; the margin is $0$, so equality holds.)
A model gives $\rho(X)=5.2$ million dollars, $\rho(Y)=3.7$ million dollars, and $\rho(X+Y)=8.4$ million dollars. Calculate the margin and state exactly what this one check establishes. (Answer: The margin is $5.2+3.7-8.4=0.5$ million dollars, so subadditivity holds for this pair only; this check alone does not establish the property for all pairs or test the other coherence axioms.)
For a particular pair, suppose $\rho(X)=70$, $\rho(Y)=50$, and $\rho(X+Y)=120$, all in the same units. A candidate says subadditivity fails because the portfolio value is not below $\rho(X)$. Identify the error and give the correct conclusion. (Answer: The candidate compared the portfolio value with one component instead of the sum of both. Since $120\leq70+50=120$, subadditivity holds with equality.)

Property	Condition for losses	What it tests
Monotonicity	If \(X \leq Y\) in every outcome, then \(\rho(X) \leq \rho(Y)\).	Whether a loss that is never larger receives no greater risk assessment.
Translation invariance	For a constant \(c\), \(\rho(X+c)=\rho(X)+c\).	Whether adding a certain amount to every outcome increases assessed risk by that amount.
Positive homogeneity	For a constant \(a \geq 0\), \(\rho(aX)=a\rho(X)\).	Whether scaling every loss scales assessed risk in the same way.
Subadditivity	\(\rho(X+Y) \leq \rho(X)+\rho(Y)\).	Whether combining losses is assessed at no more than the sum of stand-alone assessments.

What You'll Learn