Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Coherent and Convex Risk Measures

Abstract. Coherent and convex risk measures specify mathematical conditions under which a financial position, insurance liability, or portfolio of uncertain outcomes can be assigned a monetary requirement. This monograph reconstructs the foundational axioms, distinguishes payoff and loss conventions, derives consequences for capital, diversification, scaling and acceptance sets, and examines dual representations in finite and general probability spaces. It analyzes the original coherent-risk framework of Artzner, Delbaen, Eber and Heath and the convex extensions associated with Föllmer and Schied and Frittelli and Rosazza Gianin. Original, explicitly constructed examples establish the non-subadditivity of quantile-based value at risk; the coherence of expected shortfall with correct treatment of atoms; the dependence of entropic risk on scale and dependence; and the way a penalized family of alternative probability measures changes the effective worst-case scenario. The article develops a rigorous separation between the mathematical properties of a mapping, the estimation procedure used to implement it, and the prudential capital calculation prescribed by a specific regulatory regime. It examines law invariance, continuity, robust probability representations, model ambiguity, time consistency, and the interpretation of penalties. An international insurance analysis relates the axioms to claim liabilities, assets, reinsurance, solvency, concentration, reporting and governance, with precise reference to the European Solvency II framework and the IAIS Insurance Capital Standard. The central finding is that coherence and convexity give distinct, verifiable kinds of consistency and admissible diversification; their practical adequacy depends on the selected position, horizon, cash convention, state model and supervisory purpose. The monograph offers independent calculations, classification, and analytical synthesis within established risk-measure science. It institutes no new theory, doctrine, or named model.

Keywords: Coherent risk measure; convex risk measure; monetary risk measure; monotonicity; cash invariance; translation invariance; subadditivity; positive homogeneity; diversification; acceptance set; capital adequacy; expected shortfall; value at risk; entropic risk; robust representation; convex duality; penalty function; law invariance; Fatou property; risk aggregation; model risk; insurance solvency; reinsurance; risk governance.

Introduction

A risk measure converts uncertain financial outcomes into a number relevant to a decision. An insurer may need a monetary buffer for claims that exceed its ordinary expectation, a reinsurer may evaluate the risk of a layer shared across cedents, or a supervisor may require a prescribed solvency calculation. The word “risk” alone does not specify which random quantity is measured, which amount counts as capital, how losses are aggregated, or whether equal amounts of cash have equal effects in every state. Coherent and convex measures of risk answer part of that problem by imposing explicit mathematical relations on a mapping from positions to monetary assessments.

The starting point is the four-property account given by Artzner et al. (1999) in Coherent Measures of Risk. This account makes the treatment of improved positions, certain cash additions, combined exposures, and scaled exposures open to proof and criticism. Föllmer and Schied (2002), in their study of convex measures and trading constraints, extend the program to a more general relation between diversification and nonproportional financing costs. Frittelli and Rosazza Gianin (2002) independently develop convex risk-measure axioms and duality. Together these works establish the mathematical setting of the present study.

This Research Dossier addresses the axioms themselves, their implications, representation theorems, and boundaries of interpretation. The principal associated Science is Risk Theory; additional approved associations are Actuarial Science, Solvency and Capital, Enterprise Risk Management in Insurance, and Insurance Finance and Asset-Liability Management. Detailed operational comparisons of competing VaR/ES methods and stand-alone portfolio capital-allocation algorithms belong to separate dossiers. They appear here only to demonstrate a theorem, explain a counterexample, or connect mathematical properties to insurance practice.

The research method combines source-critical reconstruction, explicit definitions, deductive derivations, independently constructed finite-state examples, and comparison with authoritative supervisory materials. Sources used for the historical origins and theorems are original journal articles or their identifiable publisher and author records. Where a publisher supplies only an abstract or a bibliographic landing page, the argument rests on the verified general contribution and the derivations carried out in this text, rather than on an assertion that inaccessible pages were read in full. Institutional evidence is confined to stated standards and dates. Numerical cases are didactic mathematical constructions; no simulated number is represented as insurer data.

Six questions structure the inquiry. What is measured and with which sign? What follows from monetary monotonicity and translation invariance? Which distinct relations are imposed by subadditivity, positive homogeneity, and convexity? When does a representation by alternative probabilities or penalties exist? What changes under discontinuous distributions, estimated probabilities, and multiple periods? Finally, what can insurance practitioners and supervisors legitimately infer from these mathematical properties? Each question can be answered while preserving the separation between established theory, the author's demonstration, and positive regulatory law.

Risk Measure Foundations

Mathematical object, domain, and sign convention

Let ((\Omega,\mathcal F,P)) be a probability space, and let (L^\infty) denote the essentially bounded real-valued random variables. Throughout the main mathematical exposition (X\in L^\infty) is a discounted terminal net financial position: larger (X) means a more favorable payoff, whereas a loss is (L=-X). A monetary risk measure is a map (\rho:L^\infty\to\mathbb R), with a larger value indicating a larger assessed capital requirement. This is the payoff convention of the foundational coherent-risk literature. For an alternative loss convention, define (R(L)=\rho(-L)), and reverse the corresponding order and cash signs consistently.

Boundedness keeps the first results elementary and ensures that the numerical quantities in the examples are finite. It is not a universal condition on insurance claims. Severe property losses, liability claims, and certain financial obligations motivate extensions to unbounded spaces and more demanding integrability conditions. A theorem proved on (L^\infty) cannot simply be transplanted to (L^1), a class of heavy-tailed loss distributions, or a path-dependent process without checking its assumptions. The comprehensive discussion in Föllmer and Schied's (2016) Stochastic Finance, fourth edition locates monetary risk measures alongside contingent-claim pricing, hedging, and model uncertainty.

The number (\rho(X)) is interpreted as a monetary assessment at the chosen valuation date, in the units and numeraire of the position. Its sign is meaningful. For a normalized measure, a risk-free terminal amount (m) has (\rho(m)=-m); thus (\rho(X)<0) can indicate a cash surplus relative to the selected acceptance standard. A nonnegative reporting convention may instead quote (\max{\rho(X),0}), but that truncation changes some mathematical properties and must be explicitly distinguished from the underlying measure.

A risk measure does not determine the probability model by itself. Even a law-invariant functional depends on a specified law; a scenario-based functional depends on the chosen scenario set; and a robust functional depends on which alternative measures are admissible and how they are penalized. The mapping, its domain, the model used to evaluate it, and the interpretation attached to its output are separate objects.

The development of the axiomatic tradition

The coherent program begins with a question about the internal consistency of regulatory and institutional capital calculations. Artzner et al. (1999) name monotonicity, translation invariance, subadditivity, and positive homogeneity as desirable requirements. Their original paper investigates scenario-based constructions and the difficulties of quantile-based methods. This is a formal framework rather than an empirical assertion that every risk must be priced or regulated by the same functional.

The convex extension takes seriously the fact that a combined portfolio or a weighted position may generate nonlinear costs, illiquidity, imperfect hedges, or changing feasible trading sets. In the Föllmer–Schied convex-risk analysis, convexity replaces the joint requirement of subadditivity and positive homogeneity as the principal diversification-related condition. The Frittelli–Rosazza Gianin formulation also connects convex risk measures with pricing rules through duality. Convexity leaves room for scale-dependent capital effects while maintaining a precise Jensen-type inequality.

Subsequent lines of research deepen the framework. Kusuoka (2001) studies law-invariant coherent measures; Acerbi (2002) develops spectral risk-measure representations; and Ben-Tal and Teboulle (2007) connect convex monetary assessments to optimized certainty equivalents. These are contributions with different assumptions and objects. Their results become especially valuable when one examines whether a supposedly intuitive statement about “risk” is in fact a theorem.

Monetary and normalized risk measures

The two fundamental monetary axioms in the payoff convention are

X≤YP-a.s. ⟹ ρ(X)≥ρ(Y)(monotonicity),X\leq Y\quad P\text{-a.s.}\ \Longrightarrow\ \rho(X)\geq\rho(Y) \qquad\text{(monotonicity),}

and

ρ(X+m)=ρ(X)−m,m∈R(cash invariance).\rho(X+m)=\rho(X)-m,\qquad m\in\mathbb R \qquad\text{(cash invariance).}

The phrase “translation invariance” in the classic literature refers here to translation by deterministic cash. It must not be confused with statistical invariance under a shift of the data distribution. Cash invariance assumes cash is eligible, immediately transferable in the model, and risk-free in the chosen numeraire over the horizon. Discounting permits the same reasoning for a deterministic terminal cash amount while avoiding a mismatch between current and terminal currencies.

Normalization is the additional condition (\rho(0)=0). The first two properties alone do not force it: (\rho(X)=-E_P[X]+c) is monetary for every constant (c). A normalized coherent risk measure satisfies all four Artzner axioms and consequently has (\rho(0)=0) from positive homogeneity. A convex monetary risk measure is often normalized by convention, but convexity and monetary properties alone need not impose that convention.

These distinctions matter when comparing measures across institutions. Two calculations can share all response properties and differ by a deterministic buffer, target, or risk margin. Their outputs should not be interpreted as interchangeable without reconciling that offset.

Acceptance sets and economic interpretation

Define the acceptance set associated with (\rho) by

Aρ={X∈L∞:ρ(X)≤0}.\mathcal A_\rho=\{X\in L^\infty:\rho(X)\leq0\}.

The set is upward closed under monotonicity: if (X\in\mathcal A_\rho) and (Y\geq X), then (Y\in\mathcal A_\rho). Cash invariance gives the reconstruction

ρ(X)=inf⁡{m∈R:X+m∈Aρ},\rho(X)=\inf\{m\in\mathbb R:X+m\in\mathcal A_\rho\},

whenever the displayed monetary measure is finite-valued. Indeed (X+m) is acceptable precisely when (\rho(X)-m\leq0), so the admissible (m)'s form the interval ([\rho(X),\infty)). This deduction is exact, not a loose analogy between an algebraic number and a regulatory capital account.

If (\rho) is convex, (\mathcal A_\rho) is convex: any weighted average of two acceptable positions remains acceptable. If (\rho) is coherent, the acceptance set is a convex cone and is closed under addition and multiplication by nonnegative constants. Conversely, a suitably monotone cash-compatible convex acceptance set can induce a convex monetary risk measure. These geometric facts explain why acceptance sets, not only risk numbers, occupy an important place in the duality literature.

The insurer's actual eligible own funds, legal solvency requirement, risk margin, collateral, and management buffer must remain separately defined. A mathematical acceptance set can express a chosen tolerance or performance test; a supervisory regime determines which tests carry legal effect.

Monotonicity and Translation

Order consistency

Monotonicity imposes a comparison between positions whose payoffs can be ordered state by state. Suppose (X) is the net position after a contractual claim and (Y) results from an unconditional recovery of ten units in every state. Then (Y=X+10) and its measured risk is ten units lower under cash invariance. If the recovery varies by state but is always nonnegative, monotonicity ensures (\rho(Y)\leq\rho(X)), although the size of the difference depends on the risk measure.

This order is stronger than a comparison of expected values. A claim portfolio with a greater expected profit can have a much worse catastrophe outcome and need not dominate another portfolio state by state. Monotonicity therefore cannot determine which portfolio is safer from a comparison of averages alone. It is a coherent rule about genuinely ordered consequences, not a substitute for specifying risk preferences or the joint distribution.

The order convention also clarifies a common sign error. Under the loss convention, if (L_1\leq L_2) state by state, (R(L_1)\leq R(L_2)). Writing the payoff inequality while verbally interpreting (X) as a loss reverses the economic meaning of the formula. Every derivation below retains (X) as payoff and (L=-X) as loss.

Cash invariance and capital interpretation

The cash law (\rho(X+m)=\rho(X)-m) specifies a one-for-one effect of guaranteed cash. If (\rho(X)=15), adding 15 of eligible terminal cash makes (X+15) acceptable at the boundary. Adding 20 yields a risk assessment of (-5). If the reporting system uses a floor at zero, it may report no additional requirement while still needing to preserve the underlying value for analysis.

The law implies (\rho(m)=\rho(0)-m). With normalization, (\rho(m)=-m). This result is an immediate consistency check: a formula that assigns positive additional risk to a sure cash payment merely because the payment is larger cannot be cash invariant under the stated payoff convention.

In insurance applications, future recoveries from reinsurers are generally contingent on claims, contract terms, counterparty performance, and settlement timing. They are not deterministic cash (m). Treating expected recoveries as a sure translation would silently remove counterparty and basis risk. An unconditional cash contribution and a contingent reinsurance asset have different mathematical structures even if they share an expected nominal amount.

Bounds and stability

Monotonicity and cash invariance impose a useful uniform bound. For a normalized monetary risk measure and a bounded payoff (X),

−ess sup⁡X≤ρ(X)≤−ess inf⁡X.-\operatorname{ess\,sup}X\leq\rho(X)\leq-\operatorname{ess\,inf}X.

The proof follows from the statewise inequalities (\operatorname{ess,inf}X\leq X\leq\operatorname{ess,sup}X), monotonicity, and the known values of deterministic cash. If (X) is a net payoff between (-20) and (10), its normalized monetary risk assessment lies between (-10) and (20), regardless of the probabilities assigned inside the interval.

A second deduction is the one-Lipschitz property in the essential-supremum norm. If (|X-Y|_\infty\leq\varepsilon), then (Y-\varepsilon\leq X\leq Y+\varepsilon). Monotonicity and cash invariance give

∣ρ(X)−ρ(Y)∣≤∥X−Y∥∞.|\rho(X)-\rho(Y)|\leq\|X-Y\|_\infty.

This provides a strong stability statement for uniformly small, statewise errors. It says nothing, on its own, about sensitivity to small changes in probabilities, tail estimation, an omitted scenario, or sampling from an unbounded distribution. Those are different notions of robustness. The distinction between a functional and the procedure used to estimate it is central to Cont, Deguest, and Scandolo's (2010) robustness analysis of risk measurement.

Eligible assets, discounting, and balance sheets

The cash axiom assumes an eligible instrument that pays a specified sure amount over the assessment horizon. A multicurrency balance sheet, illiquid asset portfolio, or capital requirement stated in current rather than discounted units may violate that assumption. In such a setting, the first task is to define the discounted numeraire and the precise payoff attributed to eligible cash.

Suppose an insurer has terminal asset value (A), outstanding claims (C), expenses (E), and reinsurance recoveries (H). One plausible simplified position is (X=A+H-C-E). Its risk assessment depends on the joint distribution of all four components. A guaranteed addition (m) to assets reduces (\rho(X)) by (m); an investment with expected value (m) does not generally do so. This distinction is critical when a theoretical capital expression becomes a dashboard number or a solvency management instruction.

Monetary risk measures are consequently a language for exact statements about financial positions. Their operational meaning requires a disciplined mapping from contracts and accounts to random cash flows.

Subadditivity and Homogeneity

Diversification and scale

Subadditivity requires

ρ(X+Y)≤ρ(X)+ρ(Y).\rho(X+Y)\leq\rho(X)+\rho(Y).

It expresses the principle that assessing two positions jointly should not demand more risk capital than the sum of their separate assessments under the same measure and convention. The inequality is not a promise that every merger creates an empirical diversification benefit. For comoving positions equality can occur, and a portfolio can still have severe common losses. Rather, the inequality prohibits an axiomatic surcharge from aggregation alone under a coherent functional.

Positive homogeneity requires

ρ(λX)=λρ(X),λ≥0.\rho(\lambda X)=\lambda\rho(X),\qquad \lambda\geq0.

Doubling a position doubles its measured risk. The axiom makes sense for an idealized proportional replication of exposures with identical feasible trading and risk conditions. It becomes a substantive modeling assumption in the presence of nonlinear premiums, market impact, capacity limits, operational thresholds, or scale-dependent uncertainty.

Together, the two properties entail convexity. For (0\leq\lambda\leq1),

ρ(λX+(1−λ)Y)≤ρ(λX)+ρ((1−λ)Y)=λρ(X)+(1−λ)ρ(Y).\rho(\lambda X+(1-\lambda)Y) \leq\rho(\lambda X)+\rho((1-\lambda)Y) =\lambda\rho(X)+(1-\lambda)\rho(Y).

This is a one-line proof of the inclusion of coherent risk measures within convex risk measures. The converse is false without further assumptions. In particular, convexity does not imply positive homogeneity and does not, by itself, imply unrestricted subadditivity.

Coherence and convexity

A coherent monetary risk measure combines monotonicity, cash invariance, subadditivity and positive homogeneity. The Artzner et al. axioms constrain a functional as a whole. A convex monetary risk measure retains monotonicity and cash invariance while replacing the pair of scaling and aggregation requirements with the convexity inequality. Normalization can be added to either class for comparison.

For a normalized convex (\rho), convexity yields (\rho(\lambda X)\leq\lambda\rho(X)) when (0\leq\lambda\leq1), because (X) is mixed with zero. For (\lambda\geq1), the same inequality applied to (X=\lambda^{-1}(\lambda X)+(1-\lambda^{-1})0) gives (\rho(\lambda X)\geq\lambda\rho(X)). Larger positions can therefore carry more-than-proportional measured risk, consistent with a penalty or scale effect. Neither conclusion asserts that every convex measure strictly penalizes scale; a coherent measure attains equality.

A normalized convex measure can be subadditive in some special cases, such as pairs of independent exposures under an entropic measure, while failing it for other pairs. Hence “convex” and “subadditive” should not be used as interchangeable descriptive words.

Value-at-risk counterexample

A simple finite probability model demonstrates the difficulty of general subadditivity for a quantile. Let events (A) and (B) be disjoint, each of probability (0.04), with a residual event of probability (0.92). Define losses (L_A=100\mathbf 1_A) and (L_B=100\mathbf 1_B). At confidence (\alpha=0.95), define the lower quantile

VaR⁡α(L)=inf⁡{x:P(L≤x)≥α}.\operatorname{VaR}_{\alpha}(L)=\inf\{x:P(L\leq x)\geq\alpha\}.

Each loss individually has probability (0.96) of being zero, so (\operatorname{VaR}{0.95}(L_A)=\operatorname{VaR}{0.95}(L_B)=0). Their sum is 100 with probability (0.08), and zero with probability (0.92). Therefore

VaR⁡0.95(LA+LB)=100>0+0.\operatorname{VaR}_{0.95}(L_A+L_B)=100>0+0.

This violates subadditivity. No empirical estimate or independence assumption is involved; the joint distribution was fully stated. The counterexample concerns a quantile risk measure, not the sum of losses themselves, and it does not imply that VaR violates subadditivity for every distribution or portfolio. For some elliptical families and relevant aggregation settings, useful subadditivity relations do hold.

Artzner et al. (1999) discuss quantile limitations in their foundational treatment. The present disjoint-event construction supplies an independent, directly checkable instance. Its decision implication is narrow but important: two positions that each escape the chosen tail quantile can jointly cross it, so a numerical capital test should not silently infer diversification credit from that quantile.

Expected shortfall and finite-tail accounting

Define expected shortfall for an integrable loss (L) by the quantile integral

ES⁡α(L)=11−α∫α1VaR⁡u(L) du,0<α<1.\operatorname{ES}_{\alpha}(L) =\frac{1}{1-\alpha}\int_\alpha^1 \operatorname{VaR}_{u}(L)\,du, \qquad 0<\alpha<1.

This definition handles atoms and discrete distributions without relying on the possibly misleading expression (E[L\mid L>\operatorname{VaR}_\alpha(L)]). Acerbi and Tasche (2002) explain why careful definitions matter in their coherence study of expected shortfall. Rockafellar and Uryasev (2000) supply the CVaR optimization formulation, equivalent under the usual integrability conventions to

ES⁡α(L)=min⁡t∈R[t+11−αE[(L−t)+]].\operatorname{ES}_{\alpha}(L) =\min_{t\in\mathbb R} \left[t+\frac{1}{1-\alpha}E[(L-t)_+]\right].

The infimum version covers general cases where a minimizer requires conditions; for the finite examples here a minimizer exists. These formulas make expected shortfall a coherent loss-based monetary risk measure after translating back to payoffs. Its dual representation, for bounded (L), maximizes (E_Q[L]) over probabilities (Q\ll P) whose density satisfies (0\leq dQ/dP\leq(1-\alpha)^{-1}). The dual constraint captures concentration of attention on a limited tail mass.

Return to the disjoint-event example. Each individual loss has expected shortfall (80) at 95% confidence: the worst 5% of outcomes consists of 4% with loss 100 and 1% with loss zero. The sum has expected shortfall 100, because the worst 5% outcomes all have loss 100. Thus (100\leq80+80), in line with subadditivity. This is a comparison of axioms, not a demonstration that expected shortfall is always easy to estimate.

For a separate three-state loss model, let (P(L=0)=0.80), (P(L=20)=0.15), and (P(L=100)=0.05). The 90% quantile is 20, but the worst 10% comprises 5% of outcomes with loss 100 and 5% with loss 20. Hence (\operatorname{ES}_{0.90}(L)=(0.05\times100+0.05\times20)/0.10=60). This exact calculation illustrates how the tail-averaging convention retains information that a single quantile leaves aside.

Dependence and portfolio aggregation

Diversification operates through joint distributions. If (L_1) and (L_2) have the same marginal distributions, their sum can nevertheless have different tail behavior under independence, positive dependence, or disjoint extremes. Coherence guarantees (\rho(X+Y)\leq\rho(X)+\rho(Y)) for a fixed valid functional over the admissible joint model; it cannot determine the correct dependence model from the two marginals.

A common insurance mistake is to equate the subadditivity inequality with unlimited capital relief from combining lines. If two portfolios are exposed to the same hurricane season, economic downturn, interest-rate change, or cloud-service failure, joint tail losses can be substantial. Coherence supplies a consistency bound; credible dependence assessment determines whether the measured capital benefit is large, small, or absent.

The question of capital allocation adds further constraints about marginal contributions, coalitional incentives, and accounting. Tasche (2002) examines expected shortfall and capital allocation. The present monograph does not derive a general capital-allocation algorithm; it establishes the aggregation properties on which such algorithms may depend.

Convex Risk Measures

Convexity and capital under mixtures

Convexity requires

ρ(λX+(1−λ)Y)≤λρ(X)+(1−λ)ρ(Y),0≤λ≤1.\rho(\lambda X+(1-\lambda)Y) \leq\lambda\rho(X)+(1-\lambda)\rho(Y), \qquad0\leq\lambda\leq1.

The weighted position may be interpreted as a combination of financial exposures, a proportional sharing of contracts, or a mixture in a common state space. It is not the same mathematical operation as randomizing between (X) and (Y) by tossing an independent coin. The first operation forms the statewise payoff (\lambda X+(1-\lambda)Y); the second creates a new random mechanism whose law may be different.

The geometric meaning is that the risk functional lies below its secants. From the financial perspective, the measure can recognize pooling while allowing marginal risk to increase as a position is enlarged. This accommodates convex funding costs, market impact, ambiguity penalties, and nonlinear constraints under particular models. The 2002 convex risk literature shows how such measures connect with trading constraints and robust shortfall criteria.

Convexity is a precise but partial behavioral statement. It does not establish that combining every pair of insurance businesses reduces actual insolvency probability, nor that a price charged by a market participant must follow a convex mapping. Those are separate economic and institutional questions.

Entropic risk and nonlinear scaling

For a fixed reference probability (P), a bounded payoff (X), and (\gamma>0), define the entropic risk measure

ργ(X)=1γlog⁡EP[e−γX].\rho_\gamma(X)=\frac{1}{\gamma} \log E_P[e^{-\gamma X}].

It is normalized, monotone, cash invariant, and convex. Monotonicity follows from (X\leq Y\Rightarrow e^{-\gamma X}\geq e^{-\gamma Y}); cash invariance follows by extracting (e^{-\gamma m}) from the expectation; convexity follows from Hölder's inequality. It is generally not positively homogeneous. The exponential and logarithmic transformation can be interpreted through exponential certainty equivalents, while the exact risk-measure construction is an established part of the convex-risk literature and the optimized certainty equivalent analysis.

Take a loss (L) equal to 1 with probability 0.10 and to 0 otherwise. Put (X=-L) and (\gamma=2). Then

ρ2(X)=12log⁡(0.9+0.1e2)≈0.2470.\rho_2(X)=\tfrac12\log(0.9+0.1e^2)\approx0.2470.

For twice the same position, meaning the identical event causes loss 2, the result becomes

ρ2(2X)=12log⁡(0.9+0.1e4)≈0.9250,\rho_2(2X)=\tfrac12\log(0.9+0.1e^4)\approx0.9250,

which exceeds (2\rho_2(X)\approx0.4940). These figures are rounded evaluations of the displayed expressions. This independent counterexample proves failure of positive homogeneity and, for two perfectly dependent copies of the position, failure of subadditivity. It is fully compatible with convexity.

If (X) and (Y) are independent under (P), the exponential moment factorizes and (\rho_\gamma(X+Y)=\rho_\gamma(X)+\rho_\gamma(Y)), provided the moments exist. The same functional is additive for independent exposures and superadditive for the constructed perfectly dependent copies. Its response cannot be summarized by a universal slogan about diversification without specifying dependence.

The entropic measure has the robust representation

ργ(X)=sup⁡Q≪P{EQ[−X]−1γDKL(Q∥P)},\rho_\gamma(X)=\sup_{Q\ll P} \left\{E_Q[-X]-\frac{1}{\gamma}D_{\mathrm{KL}}(Q\|P)\right\},

where (D_{\mathrm{KL}}(Q|P)=E_Q[\log(dQ/dP)]) when defined. The equality follows from the Gibbs variational principle. The relative-entropy term penalizes departures from the reference model; its magnitude is governed by (\gamma). It is a theoretical measure of model distortion, not a claim that an insurer can assign an observable market price to every alternative distribution.

Penalized scenarios and finite-state calculations

Consider three outcomes and two probability vectors (Q_A=(0.50,0.25,0.25)) and (Q_B=(0.25,0.50,0.25)). Let (L=-X) be the loss vector, and define

ρ(X)=max⁡{EQA[L], EQB[L]−1.5}.\rho(X)=\max\{E_{Q_A}[L],\,E_{Q_B}[L]-1.5\}.

This is the maximum of two affine monetary risk functions. It is monotone, cash invariant, and convex. It is normalized because the least penalty is zero. For loss vector (L=(0,10,0)), the two terms are (2.5) and (5-1.5=3.5), so (\rho(X)=3.5). For twice the loss vector, the terms are (5) and (10-1.5=8.5); therefore (\rho(2X)=8.5>7=2\rho(X)).

The fixed 1.5 penalty does not double when the position is doubled. The second probability model is relatively less disadvantaged at greater scale, and the maximizing scenario changes its economic influence. A coherent counterpart would remove the positive penalty, giving (\max{E_{Q_A}[L],E_{Q_B}[L]}); it would scale linearly. This three-state example demonstrates the mathematical role of a penalty without treating the arbitrary number 1.5 as calibrated empirical evidence.

Convex risk measures therefore allow a distinction between uncertainty that is assigned zero penalty, uncertainty that is admitted at an explicit cost, and uncertainty excluded through an infinite penalty. The difference is encoded in a model set and a functional, not in a novel axiom.

Shortfall and optimized certainty equivalents

A family of convex measures can be constructed by specifying a loss or utility function and a tolerance condition. For a strictly increasing loss function (\ell) and a threshold (c), one standard shortfall construction uses

ρℓ,c(X)=inf⁡{m∈R:E[ℓ(−X−m)]≤c}.\rho_{\ell,c}(X)=\inf\{m\in\mathbb R:E[\ell(-X-m)]\leq c\}.

With suitable monotonicity, convexity, integrability and nondegeneracy conditions, this produces a convex monetary risk measure. The formula is an implementation of an acceptance criterion: additional cash shifts the assessed loss until its transformed expected value crosses the threshold. Choice of (\ell) determines which outcomes receive progressively larger weight.

Ben-Tal and Teboulle (2007) investigate the optimized certainty equivalent and its relation to convex risk measures, relative entropy, and special cases such as conditional value at risk. Their result gives a disciplined route from specified decision preferences to a monetary functional. It does not establish one universally correct loss function for insurers or regulators.

The mathematical choice becomes operationally meaningful only after its domain, calibration and intended decision have been stated. A steep convex loss transform can magnify rare severe outcomes, yet practical use may be dominated by uncertainty in the estimated tail. The selection of (\ell), the confidence or tolerance parameter, and the reference probability model should therefore be reported as separate assumptions.

Decisions about penalties and admissible positions

The interpretation of convex penalties matters for comparative insurance analyses. A penalty may represent aversion to an alternative probability model, a constraint on plausible parameter shifts, transaction costs, or a dual price of accepting certain exposures. These interpretations need not coincide. In a particular application the penalty can be estimated, stipulated as policy, or chosen as part of a stress framework; its provenance governs the strength of the conclusions.

A numerical comparison between a coherent and a convex measure is informative only if both act on comparable positions, horizons, currency units, and models. A coherent risk measure can be overly conservative under an excessively broad zero-penalty scenario set; a convex measure can understate risk if its penalties dismiss plausible adverse states. The axioms establish mathematical consistency within the adopted setup, not the correctness of that setup.

Representations and Counterexamples

Finite-state duality and acceptance-set geometry

On a finite probability space with (n) states, identify payoffs with vectors (x\in\mathbb R^n). A finite convex, monotone, cash-invariant risk functional admits a representation by probability vectors (q) with nonnegative coordinates summing to one:

ρ(x)=sup⁡q∈Δn{−q⋅x−α(q)},\rho(x)=\sup_{q\in\Delta_n} \left\{-q\cdot x-\alpha(q)\right\},

where a minimal convex penalty can be defined by

α(q)=sup⁡x∈Rn{−q⋅x−ρ(x)}.\alpha(q)=\sup_{x\in\mathbb R^n} \{-q\cdot x-\rho(x)\}.

This is a standard application of finite-dimensional Fenchel duality to a finite convex functional. The probability restrictions arise from monotonicity and cash invariance, and the penalty describes how much the risk measure discounts a particular alternative probability. The convexity and lower semicontinuity needed for the usual duality are automatic for a finite convex function on all of (\mathbb R^n). The Frittelli–Rosazza Gianin duality framework and the Föllmer–Schied representation establish related general principles.

For a normalized coherent measure, the minimal penalty is zero on a nonempty closed convex set of admissible probability vectors and infinite outside it. The representation then becomes

ρ(x)=max⁡q∈QEq[−X].\rho(x)=\max_{q\in\mathcal Q}E_q[-X].

The supremum form extends to relevant noncompact settings. A coherent measure can thus be interpreted as the greatest expected loss among a specified family of probabilistic scenarios, subject to the mathematical assumptions of the representation. This does not mean that every scenario set used in practice is well founded or that every worst-case probability vector is physically plausible.

The acceptance set and dual representation are two descriptions of the same functional: one describes which positions are admitted, while the other describes their supporting linear expectations and any penalties. In a finite model the relation can be checked directly by computing (\rho(x)) for constructed payoff vectors and comparing against the support functions.

General-space regularity and dual measures

Moving to an infinite state space requires more care. Convexity and monetary properties on (L^\infty) by themselves lead naturally to dual objects that can include finitely additive charges. To obtain a representation in terms of countably additive probabilities absolutely continuous with respect to (P), the relevant lower semicontinuity or Fatou-type conditions must be imposed. A common Fatou property reads

ρ(X)≤lim inf⁡k→∞ρ(Xk)\rho(X)\leq\liminf_{k\to\infty}\rho(X_k)

for almost-surely convergent uniformly bounded sequences (X_k\to X). In the appropriate setting this supports a dual representation over countably additive (Q\ll P). Distinctions among the norm topology, weak-star topology, and convergence in probability cannot be ignored.

Föllmer and Schied's (2016) monetary risk-measure treatment explains the dependence of representation results on the chosen space and continuity conditions. Jouini, Schachermayer, and Touzi (2006) analyze the Fatou property under law invariance, including structural results under assumptions on the underlying space. The appropriate theorem in a specific application must name both the function space and the regularity hypothesis.

For insurance loss distributions, this caveat is substantive. Catastrophe claims can have unbounded support, and the existence of a quantile does not guarantee that all relevant expected losses or entropic moments are finite. When (E[e^{\gamma L}]) diverges, the entropic measure can be infinite. When (E[|L|]=\infty), ordinary expected-shortfall calculations may also be infinite. No elegant dual representation permits unverified integrability assumptions to be waived.

Law invariance and spectral constructions

A risk measure is law invariant with respect to (P) if (\rho(X)=\rho(Y)) whenever (X) and (Y) have the same distribution under (P). Expected shortfall and entropic risk under a fixed (P) possess this property. A functional defined by arbitrary state-specific scenario restrictions generally need not be law invariant, even if two positions happen to have identical (P)-distributions.

Kusuoka (2001) establishes a characterization of law-invariant coherent risk measures in an appropriate atomless setting with the required continuity conditions. Broadly, such representations can be expressed through suprema over mixtures of expected-shortfall levels. Their exact normalization and admissible mixing measures depend on the stated theorem. Acerbi's (2002) spectral risk-measure study analyzes coherent risk measures obtained by weighting tail quantiles with suitably ordered, nonnegative weights. The mathematical consistency of a spectral measure rests on those weight conditions rather than on an arbitrary preference for severe losses.

Law invariance is a useful property when only the loss distribution matters for a question. Yet insurance applications often depend on who suffers the loss, when claims are paid, which jurisdiction governs the contract, and which other exposures share the same event. Two positions with the same marginal law can have different joint effects on an insurer's total balance sheet. Consequently, law invariance at a stand-alone level must not be used to erase dependence or cash-flow structure at a portfolio level.

Dynamic consistency and risk-measure procedures

Single-period monetary measures assign risk to terminal positions. Insurance obligations develop through time: claims are reported and settled at different dates, asset portfolios are rebalanced, recoveries may be delayed, and information arrives sequentially. A dynamic risk measure must specify conditional assessments indexed by time and a rule connecting them. Merely applying the same one-period formula repeatedly does not establish consistent multi-period decisions.

Cheridito, Delbaen, and Kupper (2006) study dynamic monetary risk measures for bounded discrete-time processes and analyze time consistency, acceptance-set decomposition, and dual representations. In general, time consistency imposes stability or recursion requirements on the underlying structure; it is an additional property beyond static coherence or convexity.

A second operational distinction concerns the estimator. A theoretical map (\rho) acts on a distribution or on random variables in a specified mathematical space. A risk measurement procedure first estimates an unknown distribution from data and then evaluates or approximates (\rho). Cont, Deguest, and Scandolo (2010) show in their study of risk-estimator robustness that the same risk measure can have different empirical sensitivities under different estimation procedures, and that the desire for subadditivity can interact with robustness. A theorem about coherence therefore cannot replace validation of data quality, tail extrapolation, stress scenarios, or the stability of an estimator.

Statistical backtesting introduces yet another question. A measure may be mathematically coherent while being difficult to elicit or score with an isolated observation and a single standard loss function. Gneiting (2011) discusses forecast functionals and scoring rules, including the significance of elicitability for point forecasting. This concerns evaluation of forecasts and does not invalidate the mathematical properties of expected shortfall.

Sensitivity, empirical validity, and methodological limitations

The results developed here apply under exact assumptions. The bounded-state calculations prove particular properties and counterexamples. The finite-state dual representation depends on a properly defined finite convex functional. The expected-shortfall equations use the quantile-integral definition, including discrete atoms. The entropic results assume a fixed (\gamma>0) and finite exponential moments. The insurance interpretation assumes the position includes all selected asset, liability, and contingent components.

None of these calculations constitutes a historical estimate of insurer capital, a forecast of future catastrophe losses, or proof that a particular regulatory calibration is optimal. The key distinction is between three layers: mathematical validity of (\rho), validity of the probability or scenario model used with (\rho), and institutional suitability of applying that model to a particular decision. Each layer has different evidence requirements.

A practical source-to-claim audit must also separate original publications from later summaries. Article titles, dates, volumes, and DOIs are bibliographic identities; broad theorems supported by an accessible original or a reliable publisher record are attributed to the correct authors; the precise finite-state numbers and algebra in this study are the author's own calculations. Claims about the existing EU or global insurance framework derive from the relevant supervising institution or legislative text and retain their temporal limits.

Insurance Implications

Insurance liabilities and financial position

An insurer's financial position is exposed to underwriting losses, investment prices, interest rates, credit, expenses, and operational events. Choosing (X=A+H-C-E) as a simplified net terminal payoff allows risk measures to assess the combined uncertainty. The representation is intentionally stylized: real balance sheets involve valuation bases, technical provisions, discount curves, contract boundaries, regulatory eligibility of assets, and group relationships that require separate modeling.

Monotonicity establishes that an unequivocal improvement in the insurer's terminal net position cannot increase the measured risk. Cash invariance makes the effect of an additional certain monetary buffer explicit. Subadditivity or convexity clarifies which diversification relations the chosen functional recognizes. These are valuable design tests for a capital model, but they do not determine the legal amount of required own funds.

Expected shortfall may be appropriate for examining average losses in a tail region, while a quantile can be appropriate for a specified probability threshold. The choice depends on the objective, the source model, the horizon and the applicable rules. A numerical tail measure is one part of the decision architecture and cannot substitute for claim-settlement liquidity, policyholder-protection arrangements, or governance.

Reinsurance, diversification and concentration

Reinsurance changes the distribution of retained losses and creates recoverables whose value depends on counterparty credit and contract terms. Let (L_g) be the gross loss and (H) the contingent contractual recovery, so retained loss is (L_n=L_g-H). If (0\leq H\leq L_g) in every state, a monotone loss-based risk measure gives (R(L_n)\leq R(L_g)). That conclusion is conditional on including the recovery as actually payable. If a reinsurer defaults in the catastrophe state, the realized (H) may be smaller than the promised amount, and the statewise comparison has to use the realized contingent payment.

The coherent subadditivity inequality can help examine aggregate retained positions across policies, lines or entities. It does not imply that identical insureds have independent losses. Climate hazards, cyber infrastructure dependencies, supply chains, geopolitical events and correlated legal developments can generate accumulation. The actual joint distribution remains central.

Convex measures add a way to express increasing marginal exposure costs. Where scaling a reinsurance layer requires progressively greater capital or produces worse liquidity terms, the strict equality demanded by positive homogeneity may be unsuitable as a descriptive assumption. A convex capital proxy can represent the scale effect, provided the chosen functional and penalty structure are justified. It does not establish an actuarially fair premium or a market-clearing price on its own.

Prudential regimes and capital measurement

Positive regulatory law must be read independently of mathematical axioms. The EU's Solvency II Article 101 specifies an overall Solvency Capital Requirement calibrated to a 99.5% one-year value-at-risk of basic own funds under the directive's assumptions. The aggregation provisions in Article 104 require risk modules and prescribed correlation treatment under the standard formula. A stand-alone quantile's possible failure of subadditivity does not, by itself, prove that the Solvency II standard formula breaks every desired aggregation property; the regulatory construction has its own detailed rules.

At the research cutoff of November 14, 2026, the EU's Solvency II review overview explains that amendments introduced by Directive (EU) 2025/2 are scheduled to take effect on January 30, 2027. A comparative statement about the current regime must distinguish rules already effective in 2026 from those approaching implementation. The present work is a mathematical and comparative assessment, not a country-specific legal opinion.

At global level, the IAIS reports that its Insurance Capital Standard was adopted in December 2024 as a risk-based quantitative component of ComFrame for internationally active insurance groups. The ICS addresses group capital adequacy through its own specified architecture. It should not be labeled simply “the coherent measure” or “the convex measure” without a precise mathematical derivation from the applicable standard. The institutional objective includes international comparability, supervisory intervention, and policyholder protection.

A cross-sector reference is instructive: the Basel Committee's 2019 market-risk capital standard adopted an expected-shortfall-oriented internal-model framework for banks, replacing earlier design elements that emphasized VaR. This shows an institutional role for tail-averaging considerations. It does not establish that bank market-risk rules automatically apply to insurers, that every component of Basel capital is coherent, or that expected shortfall is an adequate one-number representation of an insurer's total risk.

Model governance, backtesting and disclosure

A theoretically sound risk measure can produce a misleading number if exposures are omitted, correlations are misestimated, recovery values are idealized, or tail observations are scarce. Model governance should document the random variable, numeraire, time horizon, probability assumptions, group perimeter, eligible capital definition, stress scenarios, and treatment of contingent claims. It should also state the difference between internal economic-capital results and prescribed supervisory calculations.

For a finite scenario model, validation can begin with basic exact checks: monotonicity on ordered vectors, the cash-shift identity, convexity under weighted positions, and, where coherence is claimed, subadditivity and scale behavior. For an entropic measure, verify the exponential-moment domain and parameter sensitivity. For a robust scenario measure, inspect the alternative probabilities and penalties. For estimated ES, compare estimators, inspect extreme-event sensitivity, and evaluate relevant forecast tests. These checks complement one another because they detect different types of error.

The risk measurement procedure literature justifies separating functional-level properties from statistical implementation. The point-forecast evaluation literature cautions that the performance measure used to judge a forecast must fit the functional being forecast. A process that passes a formal axiom test can still fail empirical validation or produce an unreliable capital signal.

A scientific report should therefore disclose not only a final capital number but also the domain on which it is defined, its scenario provenance, main concentrations, material estimation uncertainty, and policy status. This form of transparency follows from ordinary methodological accountability; it is not presented as a newly named governance theory.

Research agenda and cross-Science synthesis

Several continuing research questions arise within established theory. First, the interaction of convex risk measures with nonstationary insurance portfolios calls for explicit treatment of evolving hazards and changing dependence. Second, links between finite-scenario validation and infinite-state continuity require study when losses have very heavy tails. Third, model ambiguity and statistical estimation interact: an apparently conservative robust set can be fragile if the set itself is estimated from a biased sample.

Fourth, multi-period insurance valuation requires careful separation of time consistency, cash-flow timing and liquidity. Fifth, cross-border groups must reconcile mathematical aggregation with local capital eligibility and fungibility constraints. Sixth, any comparison of internal and regulatory measures should document why the institution chose that objective and horizon rather than imply that a single axiom determines policy.

The joint relevance to five Science areas is specific. Risk Theory supplies the axiomatic mathematical core; Actuarial Science supplies loss models and statistical procedures; Solvency and Capital supplies prudential purpose and required capital concepts; Enterprise Risk Management in Insurance supplies governance and aggregation decisions; Insurance Finance and Asset-Liability Management supplies the full financial position, discounting and liquidity connections. This shared membership gives the dossier one research identity while supporting several domains of application.

The analytical synthesis is that a risk measure should be evaluated simultaneously as an ordered monetary functional, a choice of diversification and scale behavior, a scenario or probability representation, a statistical estimation procedure, and an instrument used under specific institutional rules. The parts interact through explicitly defined assumptions. Nothing in the literature reviewed warrants collapsing them into a single universal formula for every insurer, portfolio, horizon and regulator.

Conclusion

Coherent and convex risk measures provide a precise mathematical framework for assessing uncertain monetary positions. Monotonicity makes the assessment respect statewise improvement, and cash invariance gives an exact unit-for-unit response to eligible certain cash. Subadditivity and positive homogeneity jointly establish coherence and imply convexity. Convexity preserves a weighted-diversification inequality while admitting nonlinear scale effects. These distinctions determine different classes of mathematical objects and have explicit consequences for acceptance sets, bounds, dual representations and financial interpretation.

The original calculations in this monograph verify the principal structural claims. A finite disjoint-event construction shows that a quantile can violate subadditivity. Expected shortfall, defined through a tail-quantile integral that correctly handles atoms, satisfies coherent tail aggregation in the illustrated cases. Entropic risk and a penalized two-scenario measure demonstrate how convexity can coexist with nonlinear scaling and non-subadditive behavior. The derivations also establish acceptance-set reconstruction, cash-based stability, and the interpretation of convex dual penalties under stated assumptions. Each calculation is reproducible from the numbers and formulas supplied.

Representation theory reveals the importance of assumptions that are often hidden in informal discussions. Finite-state convex duality, countably additive general-space representations, law invariance, Fatou continuity, and dynamic consistency concern different mathematical conditions. The statistical properties of an estimated risk measurement procedure are additional. A coherent formula can be implemented poorly; a convex formula can be calibrated with implausible penalties; either can be applied to an incomplete financial position. The scientific assessment must therefore examine its object, model and intended decision.

In insurance, the axioms clarify relationships among liabilities, assets, reinsurance recoveries, cash buffers and portfolio concentration. Legal solvency calculations are specified by the applicable regime, including the current Solvency II directive and the internationally adopted IAIS Insurance Capital Standard. Mathematical consistency informs model selection and governance while prudential rules retain their distinct legal structure and effective dates.

The fundamental conclusion is that coherence and convexity establish verifiable conditions on monetary risk assessments. Their analytical value becomes operational when the position, horizon, numeraire, probability structure, approximation method and purpose are identified together. This synthesis extends practical clarity within established risk-measure mathematics, without claiming a new scientific theory or an unverified novelty.

References

  • Acerbi, Carlo. 2002. "Spectral Measures of Risk: A Coherent Representation of Subjective Risk Aversion." Journal of Banking & Finance 26 (7): 1505–1518. Publisher record. DOI: 10.1016/S0378-4266(02)00281-9.

  • Acerbi, Carlo, and Dirk Tasche. 2002. "On the Coherence of Expected Shortfall." Journal of Banking & Finance 26 (7): 1487–1503. Publisher record. DOI: 10.1016/S0378-4266(02)00283-2.

  • Artzner, Philippe, Freddy Delbaen, Jean-Marc Eber, and David Heath. 1999. "Coherent Measures of Risk." Mathematical Finance 9 (3): 203–228. DOI.

  • Basel Committee on Banking Supervision. 2019. Minimum Capital Requirements for Market Risk. Revised January 2019; consolidated in Basel Framework. Bank for International Settlements. Official standard.

  • Ben-Tal, Aharon, and Marc Teboulle. 2007. "An Old-New Concept of Convex Risk Measures: The Optimized Certainty Equivalent." Mathematical Finance 17 (3): 449–476. DOI.

  • Cheridito, Patrick, Freddy Delbaen, and Michael Kupper. 2006. "Dynamic Monetary Risk Measures for Bounded Discrete-Time Processes." Electronic Journal of Probability 11: 57–106. Author preprint. DOI: 10.1214/EJP.v11-302.

  • Cont, Rama, Romain Deguest, and Giacomo Scandolo. 2010. "Robustness and Sensitivity Analysis of Risk Measurement Procedures." Quantitative Finance 10 (6): 593–606. DOI.

  • European Insurance and Occupational Pensions Authority. 2026. Solvency II: Regulation and Policy Overview. European Union. Consulted November 14, 2026. Official overview.

  • European Insurance and Occupational Pensions Authority. n.d. Solvency II Single Rulebook: Article 104, Design of the Basic Solvency Capital Requirement. Consulted November 14, 2026. Official rulebook.

  • European Parliament and Council of the European Union. 2009, consolidated 2024. Directive 2009/138/EC on the Taking-Up and Pursuit of the Business of Insurance and Reinsurance (Solvency II), Article 101. EUR-Lex consolidated text.

  • Föllmer, Hans, and Alexander Schied. 2002. "Convex Measures of Risk and Trading Constraints." Finance and Stochastics 6 (4): 429–447. DOI.

  • Föllmer, Hans, and Alexander Schied. 2016. Stochastic Finance: An Introduction in Discrete Time. 4th revised ed. Berlin and Boston: De Gruyter. Publisher book record.

  • Frittelli, Marco, and Emanuela Rosazza Gianin. 2002. "Putting Order in Risk Measures." Journal of Banking & Finance 26 (7): 1473–1486. Publisher record. DOI: 10.1016/S0378-4266(02)00270-4.

  • Gneiting, Tilmann. 2011. "Making and Evaluating Point Forecasts." Journal of the American Statistical Association 106 (494): 746–762. DOI.

  • International Association of Insurance Supervisors. 2024. Insurance Capital Standard. Adopted December 5, 2024, with related Level 1 and Level 2 texts. Official ICS portal.

  • Jouini, Elyes, Walter Schachermayer, and Nizar Touzi. 2006. "Law Invariant Risk Measures Have the Fatou Property." Advances in Mathematical Economics 9: 49–71. DOI.

  • Kusuoka, Shigeo. 2001. "On Law Invariant Coherent Risk Measures." Advances in Mathematical Economics 3: 83–95. DOI.

  • Rockafellar, R. Tyrrell, and Stanislav Uryasev. 2000. "Optimization of Conditional Value-at-Risk." The Journal of Risk 2 (3): 21–41. DOI.

  • Tasche, Dirk. 2002. "Expected Shortfall and Beyond." Journal of Banking & Finance 26 (7): 1519–1533. Publisher record. DOI: 10.1016/S0378-4266(02)00272-8.