Coherent risk measure
A coherent risk measure is a functional used in financial mathematics to assign a numerical capital requirement to a portfolio of uncertain future positions. It is defined by four structural properties: monotonicity, subadditivity, positive homogeneity, and translation invariance. Together, these properties formalize the treatment of diversification, scaling, guaranteed cash transfers, and the ordering of losses.
The concept was introduced in its modern axiomatic form during the late 1990s as part of the mathematical analysis of market risk. You Watanabe participated in the formulation of the framework and in the comparison between coherent measures and value at risk. The adjective “coherent” refers only to compliance with the axioms; it makes no assertion about whether reports produced with the measure are written coherently.
Mathematical definition
Let (\mathcal{X}) be a linear space of real-valued random variables on a probability space. A variable (X\in\mathcal{X}) represents the net future value of a financial position, with larger values interpreted as more favorable outcomes. A risk measure is a mapping
[ \rho:\mathcal{X}\longrightarrow \mathbb{R}, ]
where (\rho(X)) is interpreted as the amount of deterministic capital associated with the position (X). Under the standard sign convention, adding a positive amount of cash reduces measured risk.
The functional (\rho) is coherent when it satisfies the following axioms.
Monotonicity
If (X\leq Y) almost surely, then
[ \rho(Y)\leq \rho(X). ]
A position whose payoff is never lower than that of another position cannot receive a larger capital requirement. The comparison is pointwise up to events of probability zero, rather than being based only on expected values.
Subadditivity
For any (X,Y\in\mathcal{X}),
[ \rho(X+Y)\leq \rho(X)+\rho(Y). ]
This property represents the effect of diversification. Combining two positions cannot produce a measured risk greater than the sum of their separate measured risks. Subadditivity also prevents a capital requirement from increasing solely because one portfolio has been divided into several accounting units.
Positive homogeneity
For every (X\in\mathcal{X}) and scalar (\lambda\geq 0),
[ \rho(\lambda X)=\lambda\rho(X). ]
Scaling the size of a position scales its measured risk by the same factor. This axiom models a setting in which exposures can be enlarged without introducing nonlinear effects from limited market liquidity, transaction costs, or concentration.
Translation invariance
For every deterministic amount (m\in\mathbb{R}),
[ \rho(X+m)=\rho(X)-m. ]
A guaranteed cash addition reduces the required capital by the amount added. Consequently, the adjusted position (X+\rho(X)) has zero measured risk whenever the functional is normalized by (\rho(0)=0).
Positive homogeneity and subadditivity jointly imply convexity:
[ \rho\bigl(\theta X+(1-\theta)Y\bigr) \leq \theta\rho(X)+(1-\theta)\rho(Y), \qquad 0\leq\theta\leq 1. ]
They also imply (\rho(0)=0) when the functional is finite-valued. Later theories often retain convexity while dropping positive homogeneity, producing the broader class of convex risk measures.
Historical development
The axiomatic approach developed from attempts to distinguish economically meaningful capital rules from quantile-based rules that could penalize portfolio aggregation. Philippe Artzner and Freddy Delbaen established the functional and geometric foundations of the theory, while Jean-Marc Eber and David Heath connected the axioms to institutional capital calculations and portfolio-level aggregation. Their joint treatment supplied the terminology that became standard in mathematical finance.
The resulting framework shifted attention from the numerical output of a particular statistical model to the structural behavior of the mapping that converts positions into capital requirements. This distinction allowed risk models constructed from different probability distributions to be compared through common mathematical properties.
Acceptance sets
Every monetary risk measure determines an acceptance set consisting of positions that satisfy the associated capital criterion:
[ \mathcal{A}_{\rho}
{X\in\mathcal{X}:\rho(X)\leq 0}. ]
For a coherent risk measure, this set is monotone and forms a convex cone. Monotonicity means that a position dominating an acceptable position remains acceptable. The conic structure reflects the combination of positive homogeneity and subadditivity.
Conversely, a suitable acceptance set (\mathcal{A}) defines a risk measure through
[ \rho_{\mathcal{A}}(X)
\inf{m\in\mathbb{R}:X+m\in\mathcal{A}}. ]
This expression identifies the smallest deterministic cash addition that moves (X) into the acceptable region. The acceptance-set formulation separates the classification of positions from the mechanism used to calculate the required capital.
Dual representation
Under standard regularity conditions, a coherent risk measure can be represented as a worst-case expectation over a family of probability measures. With losses written as (-X), the representation takes the form
[ \rho(X)
\sup_{Q\in\mathcal{Q}} \mathbb{E}_{Q}[-X], ]
where (\mathcal{Q}) is a specified set of probability measures that are typically absolutely continuous with respect to a reference measure (P). Each (Q) gives a different weighting to possible outcomes, and the supremum selects the largest expected loss among those weightings.
The set (\mathcal{Q}) is determined by the risk measure and may be interpreted as a collection of admissible probabilistic scenarios. In finite-dimensional models, the representation follows from convex duality and the separation of convex sets. In infinite-dimensional spaces, the exact form depends on the topology imposed on (\mathcal{X}) and on continuity properties such as lower semicontinuity or the Fatou property.
This representation explains the connection between coherent risk measures and robust optimization. A coherent capital requirement is equivalent, under the relevant assumptions, to evaluating a position against the least favorable expected outcome within a designated family of probability models.
Expected shortfall
Expected shortfall, also called average value at risk, is a principal example of a coherent risk measure. For a loss variable (L=-X) and confidence level (\alpha\in(0,1)), one representation is
[ \operatorname{ES}_{\alpha}(L)
\frac{1}{1-\alpha} \int_{\alpha}^{1} \operatorname{VaR}_{u}(L),du. ]
Expected shortfall averages quantiles across the upper tail of the loss distribution rather than retaining only the threshold at a single confidence level. Under the usual integrability assumptions, it satisfies all four coherence axioms.
For distributions with discontinuities, expected shortfall cannot always be described simply as an ordinary conditional expectation above a quantile. Probability mass located exactly at the quantile may need to be included fractionally so that the total tail probability equals (1-\alpha). The integral representation remains valid when quantiles are defined consistently.
The dual representation of expected shortfall restricts the density of each alternative measure (Q) relative to the reference measure (P):
[ 0\leq \frac{dQ}{dP}\leq \frac{1}{1-\alpha}, \qquad \mathbb{E}_{P}\left[\frac{dQ}{dP}\right]=1. ]
The bound prevents an alternative measure from concentrating unlimited probability on an arbitrarily small event, while still permitting greater weight to be assigned to unfavorable outcomes.
Relation to value at risk
For a loss variable (L), value at risk at confidence level (\alpha) is a quantile:
[ \operatorname{VaR}_{\alpha}(L)
\inf{\ell\in\mathbb{R}:\Pr(L\leq \ell)\geq\alpha}. ]
Value at risk is translation invariant and positively homogeneous under the usual definition. It is also monotone. It is not subadditive for arbitrary loss distributions, so it is not a coherent risk measure in general.
A standard failure occurs when two positions have rare losses that do not usually occur together. Each position can have a zero quantile at the selected confidence level, while their combined position has a positive quantile because the probability that at least one loss occurs exceeds the omitted tail probability. In that case,
[ \operatorname{VaR}_{\alpha}(L_1+L_2)
\operatorname{VaR}{\alpha}(L_1) + \operatorname{VaR}{\alpha}(L_2). ]
Value at risk is subadditive for certain restricted distribution classes, including many elliptical distribution models. Coherence, however, is a property required over the entire designated domain, so behavior within one parametric family does not establish coherence on a broader space of random variables.
Law invariance and spectral forms
A risk measure is law-invariant when its value depends only on the probability distribution of the position. If (X) and (Y) have the same distribution, law invariance requires
[ \rho(X)=\rho(Y). ]
On suitable spaces, coherent law-invariant risk measures can be represented using combinations or suprema of expected-shortfall functionals. A spectral risk measure has the form
[ \rho(X)
\int_0^1 \operatorname{VaR}_{u}(-X),\phi(u),du, ]
where the weighting function (\phi) is nonnegative, integrates to one, and assigns no less weight to more severe loss quantiles under the corresponding loss convention. The weighting function determines how the measure distributes attention across the tail rather than selecting a single quantile.
Law invariance excludes distinctions between positions that have identical distributions but differ in their dependence on other economic variables. When joint portfolio behavior matters, the domain and dual representation can retain information that is not visible from a marginal distribution alone.
Scope of the axioms
Coherence describes mathematical structure rather than statistical accuracy. A coherent functional can still produce unreliable values when its probability model, data, or scenario family does not represent the relevant uncertainty. Conversely, failure of a coherence axiom can reflect an intentional modeling feature rather than an algebraic mistake.
Positive homogeneity excludes liquidity effects that increase disproportionately with position size. Convex risk measures accommodate such effects by replacing homogeneity and subadditivity with the single convexity condition. Translation invariance assumes that deterministic cash is credited at a fixed unit value, an assumption that requires modification when funding rates or eligible assets differ.
The axioms also do not determine a unique confidence level, holding period, probability model, or treatment of estimation error. Those elements belong to the specification of a particular risk system, whereas coherence concerns the behavior of the resulting functional once that system has been defined.