rockafellar2000optimization

Optimization of Conditional Value-at-Risk

Authors: R. Tyrrell Rockafellar, Stanislav Uryasev · Year: 2000 · Venue: Journal of Risk, 2(3):21–41

Raw full-text conversion pending — the LAN/marker cluster is down (2026-07-04). This page was built from the full text of the author-hosted PDF read via firecrawl (sites.math.washington.edu/~rtr/papers/rtr179-CVaR1.pdf), not a marker conversion. File the PDF to Docs/raw/pdf/rockafellar2000optimization.pdf and run marker when the cluster returns.

Summary

This is the foundational paper that turns Conditional Value-at-Risk (CVaR) from a descriptive tail statistic into a tractable optimization objective. Rockafellar and Uryasev introduce an auxiliary function $F_{\beta }(\mathbf{x},\alpha )$ whose minimization over the extra scalar $\alpha$ recovers the $\beta$-CVaR of a loss, and whose joint minimization over the decision $\mathbf{x}$ and $\alpha$ is equivalent to minimizing CVaR directly. Because $F_{\beta }$ is convex (and, under sampling, piecewise-linear), CVaR minimization reduces to convex programming — an elementary linear program for scenario-sampled linear losses — while the $\beta$-VaR is obtained for free as a by-product. The construction requires no assumption of normality and works for any loss distribution given as samples.

Key Claims

• CVaR minimization needs no prior VaR computation. The $\beta$-CVaR ${\varphi }_{\beta }(\mathbf{x})$ equals ${\min }_{\alpha }{F}_{\beta }(\mathbf{x},\alpha )$ (Theorem 1), where the awkward, quantile-dependent definition of CVaR is replaced by minimizing a smooth convex function of one extra variable; the minimizing $\alpha$ is the $\beta$-VaR.
• Joint convex program. Minimizing $\beta$-CVaR over $\mathbf{x}\in X$ is equivalent to minimizing $F_{\beta }(\mathbf{x},\alpha )$ jointly over $(\mathbf{x},\alpha )\in X\times \mathbb{R}$ (Theorem 2). $F_{\beta }$ is jointly convex in $(\mathbf{x},\alpha )$ whenever the loss $f(\mathbf{x},\mathbf{y})$ is convex in $\mathbf{x}$; then, for convex $X$, the whole thing is convex programming.
• Sampling ⇒ linear programming. Approximating the expectation by $q$ Monte-Carlo (or quasi-random) samples makes ${\tilde{F}}_{\beta }$ convex and piecewise-linear in $\alpha$; introducing one auxiliary variable $u_k$ per scenario reduces the minimization to an LP — independent of the sample distribution (no normality needed).
• CVaR dominates VaR and is coherent. By construction ${\varphi }_{\beta }(\mathbf{x})\ge {\alpha }_{\beta }(\mathbf{x})$, so low CVaR forces low VaR; unlike VaR, CVaR is subadditive/convex (citing Pflug 2000; Artzner et al. 1999), so it avoids VaR’s multiple local extrema that make VaR hard to optimize.
• Numerically validated. On a 3-instrument portfolio under a normal model, the LP-based min-CVaR solution matches the Markowitz min-variance benchmark (a Proposition they prove coincides for $\beta \ge 0.5$ under normality with an active return constraint); on a non-normal NIKKEI options hedge, min-CVaR and min-variance diverge and CVaR captures tail risk that VaR misses.

Method

Let $f(\mathbf{x},\mathbf{y})$ be the loss for decision $\mathbf{x}\in X\subseteq {\mathbb{R}}^n$ and random vector $\mathbf{y}\in {\mathbb{R}}^m$ with density $p(\mathbf{y})$. The loss CDF is $\Psi (\mathbf{x},\alpha )={\int }_{f(\mathbf{x},\mathbf{y})\le \alpha }p(\mathbf{y})d\mathbf{y}$. In the paper’s notation $\beta \in (0,1)$ is the confidence level (typically 0.90/0.95/0.99), so the risk lives in the upper $(1-\beta )$ tail; $\beta$-VaR is ${\alpha }_{\beta }(\mathbf{x})$ and $\beta$-CVaR is ${\varphi }_{\beta }(\mathbf{x})$, the conditional expectation of loss at or above the VaR. The proof (Appendix) rests on a Lemma (from Shapiro & Wardi 1994): with $G(\alpha )=\mathbb{E}[f(\mathbf{x},\mathbf{y})-\alpha {]}^{+}$, $G$ is convex, $C^1$, with $G^{\prime }(\alpha )=\Psi (\mathbf{x},\alpha )-1$; hence ${\partial }_{\alpha }{F}_{\beta }=(1-\beta {)}^{-1}[\Psi (\mathbf{x},\alpha )-\beta ]$, whose zero set is exactly the VaR interval. The continuity/no-jump assumption on $\Psi$ is stated as a simplification (the general/atomic case is deferred — later handled in the 2002 follow-up, CVaR for General Loss Distributions, J. Banking & Finance 26:1443–1471).

Regime. This is a finance/stochastic-programming paper; it is regime-agnostic (no manipulator dynamics, no free-flying vs free-floating distinction). Its relevance is entirely at the risk/optimization layer — it supplies the machinery by which any planner’s scalar loss can be turned into a convex CVaR objective.

Convention clash with notation.md / conditional_value_at_risk

R&U use $\beta$ for the confidence level and place the tail at $1-\beta$, writing $\beta$-CVaR $={\varphi }_{\beta }$. The wiki’s conditional_value_at_risk page uses $\alpha$ for the confidence level and writes ${\mathrm{CVaR}}_{1-\alpha }$ (subscript = tail mass). They denote the same quantity under $\beta \leftrightarrow \alpha$. Equations below are transcribed in R&U’s own symbols.

Relevance to thesis

This is the primary provenance for every CVaR computation in the risk layer of the thesis. For risk-aware view scoring on the free-flying manipulator, whatever scalar loss a viewpoint/trajectory emits (e.g. a versine pointing error, a singularity-proximity margin, a collision-distance shortfall), Theorem 2 says we can constrain or minimize its CVaR by adding one scalar variable $\alpha$ and solving a convex program — reducing to an LP once the loss is sampled from the sim. Crucially, the reduction needs no distributional assumption, which matters because our inspection uncertainty (estimation error, thruster/contact disturbance) is not Gaussian. This is the load-bearing “how” behind CVaR; the “why CVaR over a chance constraint” is the tail-severity + coherence argument (majumdar2017how) and the tightest-convex-approximation result (nemirovski2006convex).

Connections

Topics: conditional_value_at_risk · value_at_risk · coherent_risk_measures · chance_constraints Sources: nemirovski2006convex (uses this CVaR form as the tightest convex approximation of a chance constraint) · majumdar2017how · dixit2023risk · ren2022chance (all use the R&U optimization form)

Key Equations / Quotes

$\beta$-VaR and $\beta$-CVaR (Eqs. 2–3):

$${\alpha }_{\beta }(\mathbf{x})=\min \{\alpha \in \mathbb{R}:\Psi (\mathbf{x},\alpha )\ge \beta \}{\varphi }_{\beta }(\mathbf{x})=(1-\beta {)}^{-1}{\int }_{f(\mathbf{x},\mathbf{y})\ge {\alpha }_{\beta }(\mathbf{x})}f(\mathbf{x},\mathbf{y})p(\mathbf{y})d\mathbf{y}$$

Auxiliary function (Eq. 4) and the two theorems (Eqs. 5, 10):

$$F_{\beta }(\mathbf{x},\alpha )=\alpha +(1-\beta {)}^{-1}{\int }_{\mathbf{y}\in {\mathbb{R}}^m}[f(\mathbf{x},\mathbf{y})-\alpha {]}^{+}p(\mathbf{y})d\mathbf{y}$$ $${\varphi }_{\beta }(\mathbf{x})=\underset{\alpha \in \mathbb{R}}{\min }{F}_{\beta }(\mathbf{x},\alpha )\text{(Thm 1)}\underset{\mathbf{x}\in X}{\min }{\varphi }_{\beta }(\mathbf{x})=\underset{(\mathbf{x},\alpha )\in X\times \mathbb{R}}{\min }{F}_{\beta }(\mathbf{x},\alpha )\text{(Thm 2)}$$

Sampled (LP-ready) approximation with $q$ scenarios (Eq. 9):

$${\tilde{F}}_{\beta }(\mathbf{x},\alpha )=\alpha +\frac{1}{q(1-\beta )}\sum _{k=1}^q[f(\mathbf{x},{\mathbf{y}}_k)-\alpha {]}^{+}$$

“The $\beta$-VaR is never more than the $\beta$-CVaR, so portfolios with low CVaR must have low VaR as well.” (Introduction)

“$\beta$-CVaR can be calculated without first having to calculate the $\beta$-VaR on which its definition depends… The $\beta$-VaR may be obtained instead as a byproduct.” (§2, after Thm 1)

Open Questions

• The paper assumes the loss CDF $\Psi (\mathbf{x},\alpha )$ is continuous (no atoms). Sample-based sim losses are discrete — does the atomic-distribution refinement of the 2002 follow-up (weighted VaR) change the LP for our scenario counts?
• Theorem 2’s convexity needs $f(\mathbf{x},\mathbf{y})$ convex in $\mathbf{x}$; a collision/pointing loss routed through the nonlinear circumcentroidal dynamics (ffsm_dynamics) is generally non-convex. What local/linearized surrogate keeps the CVaR reduction convex?
• R&U optimize CVaR of a single-period loss; multi-step inspection guidance raises the time_consistency issue flagged by majumdar2017how — does a per-step nested CVaR still admit the $F_{\beta }$ LP reduction?