I am looking for any remotely related reference for the following problem, for which I have not the least clue what techniques would be useful.

Consider a discrete probability space $\Omega = \{x, y, z\}$, and the space of all probability measures on it, $\mathcal{P}(\Omega)$, which can be identified with the standard 2-simplex $\Sigma_2 \subset \mathbb{R}^3_+$.

Let $H: \mathcal{P}(\Omega) \times (\Gamma, \gamma) \to \Omega$ be a map satisfying $\gamma[H(\mu, \cdot) = w] = \mu(w)$ for all $\mu \in \mathcal{P}(\Omega)$ and $w \in \Omega$, where $(\Gamma, \gamma)$ is some probability space. $H$ can be thought of as a joint/simultaneous coupling of all probability measures on $\Omega$, with the underlying probability space given by $\Gamma$.

One example of $H$ is the jointly independent coupling, where $\Gamma$ is given by the direct product $\Sigma_2^{\Sigma_2}$, with sigma algebra consisting of cylindrical sets that are nontrivial only on finitely many coordinates.

A more interesting one is given as follows. First identify $\mathcal{P}(\Omega)$ with $\Sigma_2$ via $\mu' \mapsto \sigma' := (\mu'(x), \mu'(y), \mu'(z))$. Let $\Gamma = \Sigma_2$ and define $$H_2(\mu, \sigma') = \arg\max_{w \in \Omega} [\mu(w) - \mu'(w)],$$

which is well-defined up to a $\Gamma$-set of measure $0$. Then it is easy to check that $(H_2, \Sigma_2, \lambda)$, where $\lambda$ is the normalized Lebesgue measure restricted to $\Sigma_2$, satisfies the simultaneous coupling condition listed in the beginning. I would like to show that $(H_2, \Sigma_2, \lambda)$ is optimal in the sense that it achieves the infimum value of the following functional of triples $(H, \Gamma, \gamma)$:

$$\Phi((H, \Gamma, \gamma)) := \int_{\mu \in \mathcal{P}(\Omega)} \int_{\nu \in \mathcal{P}(\Omega)} \gamma(\{\sigma \in \Gamma: H(\mu, \sigma) = H(\nu, \sigma)\}) \lambda(d\mu) \lambda(d\nu),$$

where $\lambda$ is the Lebesgue measure restricted to $\Gamma_2 \cong \mathcal{P}(\Omega)$, and normalized to be a probability measure.

Alternatively, show that $H_2$ is Pareto optimal in the sense that for any other admissible triple $(H, \Gamma, \gamma)$, the condition $$\gamma(\{\sigma \in \Gamma: H(\mu, \sigma) = H(\nu, \sigma)\}) \ge \lambda(\{\sigma \in \Sigma_2: H_2(\mu, \sigma) = H_2(\nu, \sigma)\}), ~~~ \forall ~ \mu, \nu \in \mathcal{P}(\Omega)$$ implies that there is a measure preserving isomorphism mod 0, $\psi: (\Sigma_2, \lambda) \to (\Gamma, \gamma)$, such that $H_2(\mu, \sigma) = H(\mu, \psi(\sigma))$, for all $\mu \in \mathcal{P}(\Omega)$ and $\sigma \in \Sigma_2$.


This has been solved in a joint paper with a colleague that is forthcoming. The key is to consider an interpolating probability measure $\rho$ of $\mu$ and $\nu$ defined by $\rho_i = \mu_i \nu_i / \sum_j \max\{\mu_j, \nu_j\}$.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.