Type-infity Wasserstein Ball
type- ∞ \infty ∞ Wasserstein distance
给定两个分布 P 1 , P 2 \mathbb{P}_1, \mathbb{P}_2 P1,P2，它们之间的 type- ∞ \infty ∞ Wasserstein distance 定义为：
d ∞ ( P 1 , P 2 ) : = inf Q ∈ P ( P 1 , P 2 , W ) Q -ess sup ∥ w ~ 1 − w ~ 2 ∥ d_{\infty}\left(\mathbb{P}_1, \mathbb{P}_2\right) := \mathop{\inf}\limits_{\mathbb{Q}\in\mathcal{P}(\mathbb{P}_1, \mathbb{P}_2,\mathcal{W})} \mathbb{Q}\text{-ess sup}\lVert \tilde{\boldsymbol{w}}_1 - \tilde{\boldsymbol{w}}_2 \rVert d∞(P1,P2):=Q∈P(P1,P2,W)infQ-ess sup∥w~1−w~2∥ 其中 w ~ 1 ∼ P 1 , w ~ 2 ∼ P 2 \tilde{\boldsymbol{w}}_1\sim\mathbb{P}_1, \tilde{\boldsymbol{w}}_2\sim\mathbb{P}_2 w~1∼P1,w~2∼P2, P ( P 1 , P 2 , W ) \mathcal{P}(\mathbb{P}_1, \mathbb{P}_2,\mathcal{W}) P(P1,P2,W) 表示 W × W \mathcal{W}\times\mathcal{W} W×W 上所有以 P 1 , P 2 \mathbb{P}_1, \mathbb{P}_2 P1,P2 为边缘的分布。
Q -ess sup ∥ ⋅ ∥ \mathbb{Q}\text{-ess sup} \lVert\cdot\rVert Q-ess sup∥⋅∥ 表示相对于联合分布 Q \mathbb{Q} Q 的范数 ∥ ⋅ ∥ \lVert\cdot\rVert ∥⋅∥ 的本质上界（essential supremum）：
Q -ess sup ∥ w ~ 1 − w ~ 2 ∥ : = inf { t : Q [ ∥ w ~ 1 − w ~ 2 ∥ > t ] = 0 } \mathbb{Q}\text{-ess sup}\lVert \tilde{\boldsymbol{w}}_1 - \tilde{\boldsymbol{w}}_2 \rVert := \inf \left\{ t : \mathbb{Q}\left[ \lVert \tilde{\boldsymbol{w}}_1 - \tilde{\boldsymbol{w}}_2 \rVert > t \right]=0 \right\} Q-ess sup∥w~1−w~2∥:=inf{t:Q[∥w~1−w~2∥>t]=0}
type- ∞ \infty ∞ Wasserstein ball
F ∞ ( θ ) : = { P ∈ P ( W ) ∣ d ∞ ( P , P ^ ) ≤ θ } \mathcal{F}_{\infty}(\theta) := \left\{ \mathbb{P}\in\mathcal{P}(\mathcal{W}) \big| d_{\infty}(\mathbb{P},\hat{\mathbb{P}})\leq\theta \right\} F∞(θ):={P∈P(W)∣∣d∞(P,P^)≤θ} 其中 P ^ = 1 N ∑ i ∈ [ N ] δ w ^ i \hat{\mathbb{P}} =\frac{1}{N}\mathop{\sum}\limits_{i\in[N]}\delta_{\hat{\boldsymbol{w}}_i} P^=N1i∈[N]∑δw^i。
一个等价形式
Bertsimas et al. (2018) 证明：
F ∞ ( θ ) = { P ∈ P ( W ) ∣ P = 1 N ∑ i ∈ [ N ] P i , P i ∈ F i ( θ ) , ∀ i ∈ [ N ] } \mathcal{F}_{\infty}(\theta) = \left\{ \mathbb{P}\in\mathcal{P}(\mathcal{W}) \big| \mathbb{P}= \frac{1}{N}\mathop{\sum}\limits_{i\in[N]} \mathbb{P}_i, \mathbb{P}_i\in\mathcal{F}_i(\theta), \forall i\in[N] \right\} F∞(θ)=⎩⎨⎧P∈P(W)∣∣P=N1i∈[N]∑Pi,Pi∈Fi(θ),∀i∈[N]⎭⎬⎫ 其中 F i ( θ ) = { P ∈ P ( W ) ∣ P [ ∥ w ~ − w ^ i ∥ ≤ θ ] = 1 } \mathcal{F}_i(\theta) = \left\{ \mathbb{P}\in\mathcal{P}(\mathcal{W}) \big| \mathbb{P}[\lVert\tilde{\boldsymbol{w}}-\hat{\boldsymbol{w}}_i\rVert\leq\theta]=1 \right\} Fi(θ)={P∈P(W)∣∣P[∥w~−w^i∥≤θ]=1}
type- ρ \rho ρ Wasserstein distance
对任何 ρ ∈ [ 1 , ∞ ) \rho\in[1,\infty) ρ∈[1,∞)，type- ρ \rho ρ Wasserstein distance 定义为
d ρ ( P 1 , P 2 ) : = inf Q ∈ P ( P 1 , P 2 , W ) E Q [ ∥ w ~ 1 − w ~ 2 ∥ ρ ] ρ d_{\rho}\left(\mathbb{P}_1, \mathbb{P}_2\right) := \mathop{\inf}\limits_{\mathbb{Q}\in\mathcal{P}(\mathbb{P}_1, \mathbb{P}_2,\mathcal{W})} \sqrt[\rho]{ \mathbb{E}_{\mathbb{Q}} \left[ \lVert \tilde{\boldsymbol{w}}_1 - \tilde{\boldsymbol{w}}_2 \rVert^{\rho} \right]} dρ(P1,P2):=Q∈P(P1,P2,W)infρEQ[∥w~1−w~2∥ρ]
Givens et al. 1984 证明：当 ρ \rho ρ 趋于无穷时，type- ρ \rho ρ Wasserstein distance 收敛到 type- ∞ \infty ∞ Wasserstein distance.
参考文献
[1] Bertsimas, Dimitris, Shtern Shimrit, Sturt Bradley. 2018. A data-driven approach for multi-stage linear optimization. Available at Optimization Online.
[2] Givens, Clark R, Rae Michael Shortt, et al. 1984. A class of wasserstein metrics for probability distributions. The Michigan Mathematical Journal 31(2) 231–240.
[3] Zhi Chen, Weijun Xie. Regret in the Newsvendor Model Revisited: A Data-Driven Perspective.