Descriptions may be stale. Consult the linked Lean sources.
Wasserstein Smooth
Composite Minimization
We minimize over a set of absolutely continuous probability
measures on with
finite second moments, where
is an integer. The potential is convex
with -Lipschitz gradient, , and is convex along generalized Wasserstein
geodesics in . A given is within Wasserstein distance
of a minimizer . The oracle returns
gradient pushforwards, proximal minimizers of , and optimal transport maps. For , the goal is with ,
almost surely under the joint law of algorithm and oracle
randomness.
Write for the Borel
probability measures with finite second moments, and use the Euclidean
norm on . For ,
let be their
probability couplings and define All these coupling costs are finite.
For a measurable map , the
notation denotes its
pushforward. The objective and optimization problem are where .
The algorithm receives , and ; the functions , the set , and the minimizer are accessed only
through the specified oracle.
Our assumptions:
Assumption 1 (Convex smooth potential). The function
is differentiable everywhere, and
for all and
,
Assumption 2 (Generalized geodesic convexity). Every
is absolutely continuous
with respect to Lebesgue measure. For every base measure and
every , there exists a
probability measure on , with respective
marginals , such that
For this same , every gives a measure
that satisfies
Assumption 3 (Minimizer and initial distance). The
given initial measure and a minimizer satisfy and
The oracle supports three kinds of queries. A forward query at any
and returns
the distribution A proximal query at and returns
satisfying There is no correctness requirement for
proximal queries with .
A transport query at
and any returns
a measurable map such that
, the function is integrable under
, and There is no correctness requirement
for transport queries whose first argument lies outside . Each query of any of the three kinds
counts toward a single oracle-call budget. These reply conditions hold
for every query history and every realization of oracle randomness;
different valid proximal minimizers or transport maps may be
returned.
After the allotted calls, the output distribution must satisfy where is the joint law of the algorithm’s
randomness and the oracle’s randomness, which are independent.