rockafellar2002conditional

Conditional Value-at-Risk for General Loss Distributions

Authors: R. Tyrrell Rockafellar, Stanislav Uryasev · Year: 2002 · Venue: Journal of Banking and Finance, 26(7):1443–1471

Full-text via firecrawl, ligatures repaired — this page was built from the full text of the author-hosted preprint read via firecrawl (sites.math.washington.edu/~rtr/papers/rtr187-CVaR2.pdf), not a marker conversion. The raw capture at rockafellar2002conditional.md drops the confidence-level/auxiliary-variable Greek glyphs ($\alpha ,\zeta ,\Psi ,\lambda$) in most inline prose and several boxed theorem statements — a font-encoding artifact, not missing content. The equations below were reconstructed from the surrounding definition/theorem numbering and cross-checked against the roughly one-third of display equations that survived the extraction as fully-formed LaTeX (Theorem 10's proof, Proposition 8's atomic-CVaR formulas, Theorem 16's risk-shaping formula) — those are the notation anchor. Re-run through marker/LightOnOCR when the cluster is free for a fully clean capture.

Summary

This is the direct successor to rockafellar2000optimization: it drops the earlier paper’s simplifying assumption that the loss CDF $\Psi (\mathbf{x},\zeta )$ is continuous, and re-derives CVaR, its minimization formula, and its coherence for general loss distributions that may have probability atoms — exactly the case of scenario-based/finite-sample optimization, where the loss distribution is a step function by construction. Because a distribution with an atom at the VaR threshold makes “the expected loss beyond VaR” ambiguous (three candidate values — VaR itself, and two conditional expectations either side of the jump — can all differ), the paper first disentangles $\mathrm{VaR}$, upper/lower ${\mathrm{CVaR}}^{+/-}$, and the single well-defined $\mathrm{CVaR}$ (via a rescaled “$\alpha$-tail distribution” that splits the atom), then proves the rockafellar2000optimization minimization formula and LP reduction survive unchanged in this general setting.

Key Claims

• VaR/CVaR are genuinely different objects when the loss has an atom at the threshold. For a continuous loss distribution $\mathrm{VaR}={\mathrm{CVaR}}^{-}={\mathrm{CVaR}}^{+}$ (Prop. 5, eq. 13), but with a jump at ${\zeta }_{\alpha }(\mathbf{x})$ the three values strictly separate (eq. 17); ${\mathrm{CVaR}}^{+}$ (“mean shortfall”) and ${\mathrm{CVaR}}^{-}$ (“tail VaR” in Artzner et al.) are each individually not coherent once atoms are present, only the paper’s $\mathrm{CVaR}$ (Def. 3) is.
• CVaR is a weighted average of VaR and upper-CVaR — the atom is split, not ignored. ${\varphi }_{\alpha }(\mathbf{x})={\lambda }_{\alpha }(\mathbf{x}){\zeta }_{\alpha }(\mathbf{x})+[1-{\lambda }_{\alpha }(\mathbf{x})]{\varphi }_{\alpha }^{+}(\mathbf{x})$ (Prop. 6, eq. 21), where ${\lambda }_{\alpha }(\mathbf{x})\in [0,1]$ is exactly the fraction of the VaR-level probability atom that gets folded into the tail. This is the mechanism, not a side remark: it is what makes $\mathrm{CVaR}$ well-behaved (coherent, continuous in $\alpha$) where $\mathrm{VaR}$ and ${\mathrm{CVaR}}^{\pm }$ are not.
• Closed-form CVaR for finite scenario sets (the SAA/atomic case). For $N$ scenarios with loss values $z_1<\cdots <z_N$ and probabilities $p_k$, with $k_{\alpha }$ the index where the cumulative probability first reaches $\alpha$ (eq. 23), ${\zeta }_{\alpha }(\mathbf{x})=z_{k_{\alpha }}$ exactly (eq. 24) and
  ${\varphi }_{\alpha }(\mathbf{x})=\frac{1}{1-\alpha }\left[\right({\sum }_{k=1}^{k_{\alpha }}{p}_k-\alpha )z_{k_{\alpha }}+{\sum }_{k=k_{\alpha }+1}^N{p}_k{z}_k]\text{(Prop. 8, eq. 25)}$
  — a direct, closed-form CVaR for exactly the discrete/atomic distributions that finite-sample (SAA) risk estimation produces.
• The 2000 minimization formula and LP reduction need no continuity assumption. Theorem 10 re-derives ${\varphi }_{\alpha }(\mathbf{x})={\min }_{\zeta }{F}_{\alpha }(\mathbf{x},\zeta )$ and Theorem 14 re-derives the joint minimization ${\min }_{\mathbf{x}\in X}{\varphi }_{\alpha }(\mathbf{x})={\min }_{(\mathbf{x},\zeta )\in X\times \mathbb{R}}{F}_{\alpha }(\mathbf{x},\zeta )$ without assuming $\Psi (\mathbf{x},\cdot )$ continuous — the auxiliary function $F_{\alpha }$ and its convexity/attainment properties (Thm. 10) hold for atomic distributions exactly as they did for continuous ones in rockafellar2000optimization. This directly answers that paper’s own open question about whether discreteness breaks the LP reduction: it does not.
• CVaR alone (of VaR, CVaR${}^{+}$, CVaR${}^{-}$) is coherent for general distributions. Building on Pflug’s argument, Cor. 12 confirms sublinearity + monotonicity + translation-equivariance for ${\varphi }_{\alpha }$ under $f$ linear in $\mathbf{x}$, matching Artzner–Delbaen–Eber–Heath coherence, with no continuity assumption on the loss distribution.

Method

Notation in this paper (⚠ differs from rockafellar2000optimization — see warning box below): $\alpha \in (0,1)$ is the confidence level (they no longer use $\beta$ for this), the loss CDF is $\Psi (\mathbf{x},\zeta )=P\{\mathbf{y}\mid f(\mathbf{x},\mathbf{y})\le \zeta \}$ with left limit $\Psi (\mathbf{x},{\zeta }^{-})$, and $\zeta$ (not $\alpha$) is now the auxiliary scalar variable. $\alpha$-VaR is ${\zeta }_{\alpha }(\mathbf{x})=\min \{\zeta :\Psi (\mathbf{x},\zeta )\ge \alpha \}$ (Def. 1, eq. 4); a probability atom is present at $\zeta$ whenever $\Psi (\mathbf{x},\zeta )-\Psi (\mathbf{x},{\zeta }^{-})>0$.

CVaR (Def. 3) is defined as the mean of the $\alpha$-tail distribution of the loss, obtained by rescaling $\Psi$ above ${\zeta }_{\alpha }(\mathbf{x})$:

$${\Psi }_{\alpha }(\mathbf{x},\zeta )=\{\begin{array}{ll}0& \zeta <{\zeta }_{\alpha }(\mathbf{x})\\ \frac{\Psi (\mathbf{x},\zeta )-\alpha }{1-\alpha }& \zeta \ge {\zeta }_{\alpha }(\mathbf{x})\end{array}\text{(eq. 8)}$$

This rescaling — spanning the graph of $\Psi$ between height $\alpha$ and $1$ back out to $[0,1]$ — is precisely what “splits the atom” at ${\zeta }_{\alpha }(\mathbf{x})$ if one is present, rather than assigning it wholesale to either the tail or the body.

Upper/lower CVaR (Def. 4) are the two naive candidates that CVaR interpolates between:

$${\varphi }_{\alpha }^{+}(\mathbf{x})=\mathbb{E}\{f(\mathbf{x},\mathbf{y})\mid f(\mathbf{x},\mathbf{y})>{\zeta }_{\alpha }(\mathbf{x})\},{\varphi }_{\alpha }^{-}(\mathbf{x})=\mathbb{E}\{f(\mathbf{x},\mathbf{y})\mid f(\mathbf{x},\mathbf{y})\ge {\zeta }_{\alpha }(\mathbf{x})\}\text{(eqs. 10–11)}$$

Fundamental minimization formula (Thm. 10), same auxiliary function as in rockafellar2000optimization but with the general Def. 3 CVaR on the left:

$$F_{\alpha }(\mathbf{x},\zeta )=\zeta +\frac{1}{1-\alpha }\mathbb{E}\{[f(\mathbf{x},\mathbf{y})-\zeta {]}^{+}\}\text{(eq. 27)}$$ $${\varphi }_{\alpha }(\mathbf{x})=\underset{\zeta \in \mathbb{R}}{\min }{F}_{\alpha }(\mathbf{x},\zeta )\text{(eq. 28)}$$

with ${\zeta }_{\alpha }(\mathbf{x})$ the lower endpoint and ${\zeta }_{\alpha }^{+}(\mathbf{x})$ the upper endpoint of ${\mathrm{argmin}}_{\zeta }{F}_{\alpha }(\mathbf{x},\zeta )$ (eq. 29) — VaR is again a free byproduct of the CVaR minimization.

Joint minimization / optimization shortcut (Thm. 14) — the inf-formula proper:

$$\underset{\mathbf{x}\in X}{\min }{\varphi }_{\alpha }(\mathbf{x})=\underset{(\mathbf{x},\zeta )\in X\times \mathbb{R}}{\min }{F}_{\alpha }(\mathbf{x},\zeta )\text{(eq. 43)}$$

Scenario/atomic closed form (Prop. 8), for $N$ scenarios ordered $z_1<\cdots <z_N$ with probabilities $p_k>0$ and $k_{\alpha }$ defined by ${\sum }_{k=1}^{k_{\alpha }}{p}_k\ge \alpha >{\sum }_{k=1}^{k_{\alpha }-1}{p}_k$ (eq. 23):

$${\zeta }_{\alpha }(\mathbf{x})=z_{k_{\alpha }}\text{(eq. 24)},{\varphi }_{\alpha }(\mathbf{x})=\frac{1}{1-\alpha }\left[\right(\sum _{k=1}^{k_{\alpha }}{p}_k-\alpha )z_{k_{\alpha }}+\sum _{k=k_{\alpha }+1}^N{p}_k{z}_k]\text{(eq. 25)}$$

with atom-splitting weight ${\lambda }_{\alpha }(\mathbf{x})=\frac{1}{1-\alpha }\left({\sum }_{k=1}^{k_{\alpha }}{p}_k-\alpha \right)\le \frac{p_{k_{\alpha }}}{p_{k_{\alpha }}+\cdots +p_N}$ (eq. 26). Corollary 9: if $p_N>1-\alpha$ (the confidence level is inside the top scenario’s mass), ${\varphi }_{\alpha }(\mathbf{x})={\zeta }_{\alpha }(\mathbf{x})=z_N$ — CVaR collapses to worst-case/max-loss.

Sampled LP form (Thm. 16), for scenario weights $p_k$ over $N$ sample points ${\mathbf{y}}_k$ and possibly several confidence levels ${\alpha }_i$ with tolerances ${\omega }_i$:

$${\varphi }_{{\alpha }_i}(\mathbf{x})=\underset{{\zeta }_i\in \mathbb{R}}{\min }{F}_{{\alpha }_i}(\mathbf{x},{\zeta }_i),F_{{\alpha }_i}(\mathbf{x},{\zeta }_i)={\zeta }_i+\frac{1}{1-{\alpha }_i}\sum _{k=1}^N{p}_k[f(\mathbf{x},{\mathbf{y}}_k)-{\zeta }_i{]}^{+},F_{{\alpha }_i}(\mathbf{x},{\zeta }_i)\le {\omega }_i$$

identical in structure to the 2000 paper’s sampled ${\tilde{F}}_{\beta }$, but now proved valid for the general (possibly tied/atomic) scenario weights $p_k$ that finite-sample simulation actually produces, rather than assuming i.i.d. continuous scenarios.

Regime. Same as rockafellar2000optimization: regime-agnostic finance/stochastic-programming paper, no manipulator dynamics. Relevance is entirely at the risk/optimization layer.

Notation swap between the two Rockafellar–Uryasev papers

rockafellar2000optimization uses $\beta$ for the confidence level and $\alpha$ for the auxiliary minimization variable. This 2002 paper swaps both: $\alpha$ is the confidence level, $\zeta$ is the auxiliary variable. Same authors, same $F_{(\cdot )}(\mathbf{x},\cdot )$ construction, different letters — do not mix symbols across the two source pages. The wiki’s Conditional Value at Risk page uses yet a third convention ($\alpha$ = confidence level, ${\mathrm{CVaR}}_{1-\alpha }$ = subscript is tail mass); this page’s equations are transcribed in this paper’s own $\alpha ,\zeta$ symbols.

Relevance to thesis

This paper is the provenance for the atomic-tail case of the SAA (sample-average-approximation) CVaR obligations A1/A5: our risk-aware planner’s loss (pointing-error versine, collision-distance shortfall, singularity-proximity margin) is only ever observed through a finite set of simulated scenarios, so the empirical loss distribution is a step function with probability atoms by construction — exactly the setting rockafellar2000optimization excludes by assumption and this paper covers. Two results carry directly into the proof obligations:

1. The inf/min-formula (Thm. 10, 14) needs no continuity assumption, so the LP reduction used for the sampled CVaR objective is valid as-is for our scenario-weighted, possibly-tied simulated losses — this closes the open question left on rockafellar2000optimization about whether discreteness breaks the reduction.
2. The closed-form atomic CVaR (Prop. 8, eq. 25) gives an exact, sortable expression for ${\varphi }_{\alpha }$ over $N$ weighted scenarios — useful both as a closed-form check on the LP solution and for stating the SAA estimator’s bias/consistency properties without appealing to continuity.

Connections

Topics: Conditional Value at Risk · Value at Risk · Coherent Risk Measures · Chance Constraints Sources: rockafellar2000optimization (predecessor; this paper generalizes its minimization formula to atomic/discrete loss distributions) · nemirovski2006convex · majumdar2017how · dixit2023risk · ren2022chance

Key Equations / Quotes

$\alpha$-tail distribution — the atom-splitting construction (Def. 3, eq. 8):

$${\Psi }_{\alpha }(\mathbf{x},\zeta )=\{\begin{array}{ll}0& \zeta <{\zeta }_{\alpha }(\mathbf{x})\\ \frac{\Psi (\mathbf{x},\zeta )-\alpha }{1-\alpha }& \zeta \ge {\zeta }_{\alpha }(\mathbf{x})\end{array}$$

CVaR as a weighted average of VaR and upper-CVaR (Prop. 6, eqs. 20–21):

$${\lambda }_{\alpha }(\mathbf{x})=\frac{\Psi (\mathbf{x},{\zeta }_{\alpha }(\mathbf{x}))-\alpha }{1-\alpha }\in [0,1],{\varphi }_{\alpha }(\mathbf{x})={\lambda }_{\alpha }(\mathbf{x}){\zeta }_{\alpha }(\mathbf{x})+[1-{\lambda }_{\alpha }(\mathbf{x})]{\varphi }_{\alpha }^{+}(\mathbf{x})$$

The inf/min-formula (Thm. 14, eq. 43):

$$\underset{\mathbf{x}\in X}{\min }{\varphi }_{\alpha }(\mathbf{x})=\underset{(\mathbf{x},\zeta )\in X\times \mathbb{R}}{\min }{F}_{\alpha }(\mathbf{x},\zeta )$$

“Generally ${\mathrm{CVaR}}^{-}\le \mathrm{CVaR}\le {\mathrm{CVaR}}^{+}$, with equality holding when the loss distribution function does not have a jump at the VaR threshold; but when a jump does occur, which for scenario models is always the situation, both inequalities can be strict.” (Introduction)

“[CVaR] can be viewed as a weighted average of VaR and CVaR${}^{+}$ … This seems surprising, in the face of neither VaR nor CVaR${}^{+}$ being coherent. The weights arise from the particular way that CVaR ‘splits the atom’ of probability at the VaR value, when one exists.” (Introduction)

Open Questions

• Prop. 8’s closed-form ${\varphi }_{\alpha }$ (eq. 25) is stated for a fixed $\mathbf{x}$ with the scenario ordering $z_1<\cdots <z_N$ depending on $\mathbf{x}$ (the paper flags this itself). For a gradient-based planner sweeping $\mathbf{x}$ continuously, does the ordering/$k_{\alpha }$ switching introduce non-smoothness in ${\varphi }_{\alpha }(\mathbf{x})$ beyond what Theorem 10’s $F_{\alpha }$-based reduction already handles?
• Theorem 16’s multi-threshold risk-shaping form is the natural fit for constraining CVaR at several confidence levels simultaneously (e.g. a 0.95 and a 0.99 pointing-error bound) — has this been checked against our current single-$\alpha$ CVaR gate in the risk-aware planner?
• As in rockafellar2000optimization: multi-step inspection guidance still raises the Time Consistency issue; this paper’s atomic-case results are single-period, same as its predecessor.