Reparameterized Gaussian VI — Gaussian Variational Inference
Descriptions may be stale. Consult the linked Lean sources.
Gaussian Variational Inference · Lean proof
Gaussian
Variational Inference
We minimize the Gaussian variational objective over and lower triangular
with
positive diagonal, where . The
differentiable potential has an -Lipschitz, -strongly monotone gradient, with
. A given feasible
pair is within
Euclidean–Frobenius distance
of a minimizer .
The oracle returns at
. For , the goal is a
measurable feasible output with integrable gap satisfying .
Let be an integer, equip
with its Euclidean
inner product and norm, and write for the Frobenius
norm. Define
The coordinates of are
independent standard normal random variables. With the objective defined above, the problem is The algorithm receives , where
, , , and .
Our assumptions:
Assumption 1 (Smoothness and strong monotonicity).
The potential is differentiable
everywhere. For every ,
Assumption 2 (Minimizer and initial distance). There
exists
such that
Queries may be made at any . The oracle returns
at . The Gaussian expectation defining
is over ; output expectations are under , the law of the algorithm’s randomness.
After the allotted oracle calls, the output must be measurable and
satisfy for
every realization. Its objective gap must be integrable under , and the required accuracy is
Proved result
OracleTemplate.Convergence LongTails.Setups.GaussianVariationalInference.spec
OracleTemplate.GaussianVariational.ReparameterizedGaussianVI.budget
Relevant previous results
No qualifying prior reference verified; this is not a proof of novelty.