Measure and Index in Robust Optimization

Definition

Satisficing Measure [4]
Let $(\Omega ,\mathcal{F},\mathbb{P})$ be a measure space and let $\mathcal{X}$ be a set of random variables on $\Omega$, i.e., a set of functions $X:\Omega \to \mathbb{R}$. Each $X\in \mathcal{X}$ represents the payoff (or return) of a different, risky position. Consider the situation in which the investor has an aspiration level $\tau$, which she hopes to achieve via these positions. We assume $\tau$ a random variable on $\Omega$ as well. Given an uncertain payoff $X\in \mathcal{X}$, we define the target premium $V$ to be $V=X-\tau \in \mathcal{X}$.

Definition. A function $\rho :\mathcal{X}\to [0,\bar{\rho }]$, where $\bar{\rho }\in [1,\infty ]$, is a satisficing measure defined on the target premium if it satisfies the following axioms for all $X,Y\in \mathcal{X}$:

1. Attainment Content: If $X\ge 0$, then $\rho (X)=\bar{\rho }$.
2. Nonattainment apathy: If $X<0$, then $\rho (X)<0$.
3. Monotonicity: If $X\ge Y$, then $\rho (X)\ge \rho (Y)$.
4. Gain Continuity: ${\lim }_{a\downarrow 0}\rho (X+a)=\rho (X)$

Definition. A function $\rho :\mathcal{X}\to [0,\bar{\rho }]$, where $\bar{\rho }\in [1,\infty ]$, is a quasiconcave satisficing measure (QSM) defined on the target premium if, in addition to the conditions of satisficing measure, it satisfies the following axioms for all $X,Y\in \mathcal{X}$:
1… Quasiconcavity: If $\lambda \in [0,1]$, ρ ( λ X + ( 1 − λ ) Y ) ≥ min ⁡ { ρ ( X ) , ρ ( Y ) } \rho(\lambda X+(1-\lambda)Y)\ge\min\{\rho(X),\rho(Y)\} ρ(λX+(1−λ)Y)≥min{ρ(X),ρ(Y)}.
If in addition, $\rho$ satisfies the following condition, we say it a coherent satisficing measure (CSM)
2. Scale Invariance: If $k>0$, then $\rho (kX)=\rho (X)$.

Risk Measure [4]
Definition. A function $\mu :\mathcal{X}\to \mathbb{R}$ is a risk measure if it satisfies the following for all $X,Y\in \mathcal{X}$:

1. Monotonicity: If $X\ge Y$, then $\mu (X)\le \mu (Y)$.
2. Translation invariance: For any $a\in \mathbb{R}$, $\mu (X+a)=\mu (X)+a$.

Definition. A function $\mu :\mathcal{X}\to \mathbb{R}$ is a convex risk measure if, in addition to the conditions of risk measure, it satisfies the following for all $X,Y\in \mathcal{X}$:

1. Convexity: For any $\lambda \in [0,1]$, $\mu (\lambda X+(1-\lambda )Y)\le \lambda \mu (X)+(1-\lambda )\mu (Y)$.
   If in addition, $\mu$ satisfies the following condition, we say it a coherent risk measure :
2. Positive Homogeneity: If $\lambda \ge 0$, then $\mu (\lambda X)=\lambda \mu (X)$.

It is well known that every coherent risk measure may be written in the form $$\mu (X)=\underset{\mathbb{Q}\in \mathcal{Q}}{\sup }{\mathbb{E}}_{\mathbb{Q}}(-X)$$ for a family of generating measures $\mathcal{Q}$.

Riskiness Index [3]
Definition. Given a random delay denoted by the random variables $\tilde{\xi }\in \mathcal{V}$ with probability distribution $\mathbb{P}$, the riskness index ${\rho }_R:\mathcal{V}\to \mathbb{R}$ is defined as $${\rho }_R(\tilde{\xi })=\inf \left\{\alpha >0|C_{\alpha }(\tilde{\xi })\le 0\right\}$$
where $\alpha >0$ and $C_{\alpha }(\tilde{\xi })$ is the certainty equivalent of the $\tilde{\xi }$ under exponential disutility given by $$C_{\alpha }(\tilde{\xi })=\alpha \ln {\mathbb{E}}_{\mathbb{P}}\left(\exp \left(\frac{\tilde{\xi }}{\alpha }\right)\right).$$

Essential Riskiness Index [2]
Definition. Given a random delay denoted by the random variables $\tilde{\xi }\in \mathcal{V}$ with probability distribution $\mathbb{P}$, the essential riskness index ${\rho }_E:\mathcal{V}\to [0,\infty ]$ is defined as $${\rho }_E(\tilde{\xi })=\inf \left\{\alpha \ge 0|{\mathbb{E}}_{\mathbb{P}}\left(\max \left(\tilde{\xi },-\alpha \right)\right)\le 0\right\}.$$

Proposition For all $\tilde{\xi },{\tilde{\xi }}_1,{\tilde{\xi }}_2\in \mathcal{V}$

1. Satisficing: ${\rho }_E(\tilde{\xi })=0$ if and only if $\mathbb{P}(\tilde{\xi }\le 0)=1$.
2. Infeasiblity: If ${\mathbb{E}}_{\mathbb{P}}(\tilde{\xi })>0$, then ${\rho }_E(\tilde{\xi })=\infty$.
3. Convexity: For any $\lambda \in [0,1]$, ${\rho }_E(\lambda {\tilde{\xi }}_1+(1-\lambda ){\tilde{\xi }}_2)\le \lambda {\rho }_E({\tilde{\xi }}_1)+(1-\lambda ){\rho }_E({\tilde{\xi }}_2)$.
4. Delay bounds: $\mathbb{P}\left(\tilde{\xi }>{\rho }_E(\tilde{\xi })\theta \right)\le \frac{1}{1+\theta },\forall \theta >0$.

Adversarial Impact Measure [1]
The function $\rho :\mathcal{L}\to [0,+\infty ]$ is an adversarial impact measure if it has the following representation
$$\rho (\tilde{v})=\min k$$ $$\text{s.t. }{\mathbb{E}}_{\mathbb{P}}[\tilde{v}]\le k\Delta (\mathbb{P},\hat{\mathbb{P}}),\forall \mathbb{P}\in {\mathcal{P}}_0$$ $$k\ge 0$$ for some probability distance function $\Delta$.

THEOREM 1 (Properties of Adversarial Impact Measure [1]) The function $\rho :\mathcal{L}\to [0,+\infty ]$ is an adversarial impact measure if and only if it is lower semi-continuous and satisfies the following properties:
5. Monotonicity: If ${\tilde{v}}_1\ge {\tilde{v}}_2$ , then $\rho ({\tilde{v}}_1)\ge \rho ({\tilde{v}}_2)$.
6. Infeasible Baseline: If ${\mathbb{E}}_{\mathbb{P}}[\tilde{v}]>0$, then $\rho (\tilde{v})=\infty$ .
7. Full Robustness: If $\tilde{v}\le 0$, then $\rho (\tilde{v})=0$.
8. Positive Homogeneity: For any $\lambda \ge 0$, we have $\rho (\lambda \tilde{v})=\lambda \rho (\tilde{v})$.
9. Convexity: For any $\lambda \in [0,1]$, we have $\rho (\lambda {\tilde{v}}_1+(1-\lambda ){\tilde{v}}_2)\le \lambda \rho ({\tilde{v}}_1)+(1-\lambda )\rho ({\tilde{v}}_2)$.

Moreover, the underlying probability distance $\Delta$ is such that Δ ( P , P ^ ) = sup ⁡ v ~ ∈ L { E P [ v ~ ] ∣ ρ ( v ~ ) ≤ 1 } \Delta(\mathbb{P},\hat{\mathbb{P}})=\mathop{\sup}\limits_{\tilde{v}\in\mathcal{L}}\{\mathbb{E}_{\mathbb{P}}[\tilde{v}] | \rho(\tilde{v})\leq 1\} Δ(P,P^)=v~∈Lsup​{EP​[v~]∣ρ(v~)≤1}.

Normalized Convex Risk Measure [1]
A normalized convex risk measure is a lower semi-continuous functional $\mu :\mathcal{L}\to \mathbb{R}$, that satisfies the following properties:

1. Monotonicity: If ${\tilde{v}}_1\ge {\tilde{v}}_2$ , then $\mu ({\tilde{v}}_1)\ge \mu ({\tilde{v}}_2)$.
2. Translation invariance: For any $a\in \mathbb{R}$, $\mu (\tilde{v}+a)=\mu (\tilde{v})+a$.
3. Convexity: For any $\lambda \in [0,1]$, $\mu (\lambda {\tilde{v}}_1+(1-\lambda ){\tilde{v}}_2)\le \lambda \mu ({\tilde{v}}_1)+(1-\lambda )\mu ({\tilde{v}}_2)$.
4. Normalized: $\mu (0)=0$.

Proposition (Risk-based Representation [1]). The functional $\rho :\mathcal{L}\to [0,+\infty ]$ is an adversarial impact measure if and only if there exists some convex risk measure $\mu :\mathcal{L}\to \mathbb{R}$, such that $$\rho (\tilde{v})=\inf \left\{k>0|k\mu \left(\frac{\tilde{v}}{k}\right)\le 0\right\}$$ for which $\mu (\tilde{v})$ is strictly positive whenever ${\mathbb{E}}_{\mathbb{P}}[\tilde{v}]>0$.

证明： 定义
μ ( v ~ ) : = inf ⁡ { a : ρ ( v ~ − a ) ≤ 1 } \mu(\tilde{v}) := \inf\{a:\rho(\tilde{v}-a)\leq 1\} μ(v~):=inf{a:ρ(v~−a)≤1}
那么
$$\mu \left(\frac{\tilde{v}}{k}\right)\le 0$$ $$⟺{\mathbb{E}}_{\mathbb{P}}\left(\frac{\tilde{v}}{k}\right)\le \Delta \left(\mathbb{P},\hat{\mathbb{P}}\right)$$ $$⟺\rho \left(\frac{\tilde{v}}{k}\right)=1$$ $$⟺\rho (\tilde{v})=k$$

Literature

[1] The Dao of Robustness
[2] Routing optimization with time windows under uncertainty
[3] Aumann, R.J., Serrano, R.: An economic index of riskiness. J. Polit. Econ. 116(5), 810–836 (2008)
[4] Brown and Sim: Satisficing Measures for Analysis of Risky Positions, Management Science(2019) 55(1), 71-84