drorobust-optimizationambiguity-setrisk-measureoptimization

Distributionally Robust Optimization (分布鲁棒优化)

Definition. Distributionally Robust Optimization (DRO, 分布鲁棒优化) is a modelling framework in which the decision-maker minimises a worst-case expectation of a cost functional over an ambiguity set \mathcal P of probability distributions compatible with the available statistical information, rather than committing to a single nominal distribution.

The general DRO primal

Given a decision x\in X\subseteq\mathbb R^n and random parameters \tilde\xi\in\Omega\subseteq\mathbb R^d on a measurable space (\Xi,\mathcal F), the canonical DRO problem (Rahimian & Mehrotra 2022, eq. 1.5) is

\min_{x\in X}\ \sup_{P\in\mathcal P}\ R_P\!\left[h(x,\tilde\xi)\right]\quad\text{s.t.}\quad \sup_{P\in\mathcal P}\ R_P\!\left[g_j(x,\tilde\xi)\right]\le 0,\ j=1,\dots,m,

where R_P[Z] is a (componentwise) real-valued functional of the random variable Z under probability measure P. Choosing R_P=\mathbb E_P and the ambiguity set \mathcal P = \{P_{\text{true}}\} recovers classical stochastic optimisation; choosing \mathcal P = \{\text{all distributions on the support}\} recovers classical robust optimisation. DRO is the continuous bridge between the two.

flowchart LR
    A["Stochastic Optimization<br/>(known P)"] -- partial info --> B["DRO<br/>(ambiguity set P)"]
    B -- trivial P = all on support --> C["Robust Optimization<br/>(worst-case on support)"]
    A -- no info --> C

Why DRO, in three lines

  1. The true distribution is unknown. Sample-based estimates of mean and covariance are noisy; a "plug-in" expected-cost optimisation that uses the empirical moments is fragile under regime shift.
  2. Pure RO is too pessimistic. Minimising worst-case cost over the support ignores all distributional information we do have (moments, divergence to baseline, structural shape).
  3. DRO is a systematic hedge. \mathcal P encodes "the set of distributions the truth could plausibly be", and the inner \sup quantifies the worst case over that set.

The four equivalence theorems (Rahimian & Mehrotra §3)

These are the load-bearing results that make DRO a unified theory rather than a bag of techniques.

Theorem (paper §) Statement Quant-finance reading
DRO = RO (§3.2, finite \Xi) When the sample space is finite, the DRO supremum can be written as a finite-dimensional LP whose dual is a classical RO problem. DRO with scenario ambiguity is structurally equivalent to RO with budgeted uncertainty.
DRO = coherent risk (§3.3.1) If \mathcal P is convex and compact, \rho_{\mathcal P}(Z):=\sup_{P\in\mathcal P}\mathbb E_P[Z] is a coherent risk measure (Artzner et al. 1999). The "worst-case" view is a coherent risk view — the same axioms (monotonicity, translation invariance, sub-additivity, positive homogeneity) apply.
DRO = law-invariant risk (§3.3.1) If \mathcal P is law-invariant, \rho_{\mathcal P} is law-invariant; for convex \mathcal P this is equivalent to the optimal certainty equivalent representation of Ben-Tal & Teboulle. Worst-case expected loss under all distributions sharing the same law = a particular coherent risk measure indexed by the family.
DRO = regularised ERM (§3.4) Wasserstein-ball DRO is a regularised empirical-risk minimisation problem; \phi-divergence DRO with \epsilon=\chi^2_{1,N,1-\alpha} admits an asymptotic expansion that adds a variance penalty. LASSO, SVM, regularised logistic regression are implicit Wasserstein-DRO solutions.

Construction of the ambiguity set

The two dominant constructions in quant-finance work are:

  1. Wasserstein ball around the empirical distribution — see concepts/Wasserstein Ambiguity Set.
  2. Moment-based set — typically mean and covariance in a confidence region, sometimes higher moments or the support. This is the "Chebyshev / Delage-Ye" family.

Other constructions (φ-divergence, Prohorov, \zeta-structure, kernel / mean-embedding, shape-preserving) are catalogued in summaries/201908-distributionally-robust-optimization-review/201908-distributionally-robust-optimization-review.

Why this is useful for portfolio construction

A typical equity-portfolio DRO workflow:

  1. Empirical moments from a rolling window of returns → mean \hat\mu and covariance \hat\Sigma.
  2. Choose a radius \epsilon (e.g., from Mohajerin Esfahani & Kuhn 2018 finite-sample bound).
  3. Solve \min_x\ \sup_{P:\ W_p(P,\hat P_N)\le\epsilon}\ \mathbb E_P[-\tilde r^\top x] subject to a return floor and a leverage cap.
  4. The result is a robustified Markowitz allocation whose worst-case expected loss is finite-sample-bounded.

In practice, the Wasserstein construction is preferred for portfolio work because the radius has a clean finite-sample statistical interpretation, while moment-based sets tend to produce overly conservative solutions (worst-case distributions are often two-point masses at the boundary of the moment region — see Scarf 1958 for the newsvendor case).

DRO 与鲁棒优化的区别：不确定的是分布还是参数

DRO 与鲁棒优化 (RO) 经常被混用，但两者不确定的对象不同：

不确定的是 决策变量对抗的是什么 内层问题
RO 参数 / 扰动：只知道 \xi 落在不确定集 \mathcal U 里，不知道它长什么样 某个具体的坏参数实现 \sup_{\xi\in\mathcal U} h(x,\xi)
DRO 分布：知道样本或矩信息，但不知道真实分布 P 某个具体的坏分布 \sup_{P\in\mathcal P} \mathbb E_P[h(x,\tilde\xi)]

两点实际差别：

  1. 建模对象不同。RO 的 \mathcal U 建立在收益/损失的取值空间上（区间、盒、椭球、多面体），是"参数能取哪些值"的问题；DRO 的 \mathcal P 建立在概率测度的空间上，是"分布能是哪些"的问题。前者回答"最坏的那次冲击长什么样"，后者回答"最坏的那个世界长什么样"。
  2. 信息使用量不同。RO 只用支撑信息（min/max、上下界），把不确定集内所有实现视为同等可能；DRO 用分布信息（样本、矩、散度距离），因此通常更不保守。这也是两者可以嵌套的原因：当 \mathcal P 包含支撑集上的所有分布（"没有信息"的情形）时，DRO 与 RO 重合。

在组合语境里这对应两个不同的问题：前者要回答「如果某只资产的收益比预期低 10%，我的组合会怎样」，后者要回答「如果收益的分布不是我以为的那个，我的组合会怎样」。鲁棒组合优化 一页把「优化端（把误差写成扰动集）」与「估计端（用稳健统计把协方差估准）」的分工一并列出。

模糊集半径该取多大：让置信水平决定，而不是反过来。 DRO 最常见的批评是模糊集半径 \epsilon 是个人工超参，调大则保守、调小则冒进。Duchi, Glynn & Namkoong 2018 给了一个原则性的答案：用广义经验似然为最优值 T_{\text{opt}}(P)=\inf_{x\in X}\mathbb E_P[\ell(x;\xi)] 构造置信区间，由区间覆盖要求倒推半径 \rho(\alpha)=\inf\{\rho' : P(\chi^2_{d_0}\le\rho')\ge 1-\alpha\}，这样得到的模糊集自带渐近精确的覆盖率保证。论文进一步证明：分布鲁棒解与样本平均近似（SAA）的解满足本质相同的一致性，所以鲁棒化没有引入额外的估计偏差；而且鲁棒形式的渐近展开中第二项恰好是 SAA 问题的正则项，度量的是目标函数的方差——即「鲁棒化 = 按方差正则化」。这条结论把 DRO 从「额外的保险」改写成有统计依据的方差控制。（d_0 是协方差的秩而非维数，细节见 广义经验似然与鲁棒优化统计。）

一致性线索：DRO 的最坏期望算子在 \mathcal P 凸紧时本身就是一致风险测度（见一致风险测度）。这个结论的源头是 Artzner 等 1999 的情景法表示定理——他们证明相干风险测度 \rho 满足 \rho(X)=\sup_{P}\mathbb E_P[-X/r]，即风险可以写成一族概率测度下的最坏期望。DRO 做的是把这一族概率测度换成凸紧模糊集 \mathcal P，并加上数据驱动的构造方式。两条线在"最坏期望 = 相干风险"这一点上会合。

When NOT to use DRO

  • The decision is small and the cost function is convex. A simpler ERM with \ell_1 or \ell_2 regularisation may suffice.
  • Sample size is very small (say N<50 in d=50): the Wasserstein radius required for a non-trivial confidence level is so large that the worst case is essentially "all distributions on the support" — i.e., you fall back to RO anyway.
  • The objective is highly non-convex (e.g., discrete allocation with combinatorial structure). The cutting-surface / exchange method works but is slow; if the inner sup is tractable, prefer RO with a budgeted uncertainty set.

Cross-references

See also

  • Artzner et al. 1999 摘要 —— 情景法表示定理，DRO 的公理源头
  • Duchi-Glynn-Namkoong 2018 摘要 —— 广义经验似然给出模糊集半径的统计依据（置信水平决定半径）
  • 广义经验似然与鲁棒优化统计 —— 构造细节与「鲁棒化 = 按方差正则化」结论
  • 鲁棒组合优化 —— 另一条路线：不确定集设在参数/扰动上
  • Mohajerin Esfahani, P. & Kuhn, D. (2018) "Data-driven distributionally robust optimization using the Wasserstein metric" — the finite-sample consistency result that makes data-driven Wasserstein DRO deployable.
  • Delage, E. & Ye, Y. (2010) "Distributionally robust optimization under moment uncertainty" — the canonical mean-covariance DRO with confidence regions.
  • Scarf, H. (1958) "A min-max solution of an inventory problem" — the first DRO problem; newsvendor with mean+covariance.