Conference item icon

Conference item

Densities of almost surely terminating probabilistic programs are differentiable almost everywhere

Abstract:
We study the differential properties of higher-order statistical probabilistic programs with recursion and conditioning. Our starting point is an open problem posed by Hongseok Yang: what class of statistical probabilistic programs have densities that are differentiable almost everywhere? To formalise the problem, we consider Statistical PCF (SPCF), an extension of call-by-value PCF with real numbers, and constructs for sampling and conditioning. We give SPCF a sampling-style operational semantics à la Borgström et al., and study the associated weight (commonly referred to as the density) function and value function on the set of possible execution traces. Our main result is that almost surely terminating SPCF programs, generated from a set of primitive functions (e.g. the set of analytic functions) satisfying mild closure properties, have weight and value functions that are almost everywhere differentiable. We use a stochastic form of symbolic execution to reason about almost everywhere differentiability. A by-product of this work is that almost surely terminating deterministic (S)PCF programs with real parameters denote functions that are almost everywhere differentiable. Our result is of practical interest, as almost everywhere differentiability of the density function is required to hold for the correctness of major gradient-based inference algorithms.
Publication status:
Published
Peer review status:
Reviewed (other)

Actions


Access Document


Files:
Publisher copy:
10.1007/978-3-030-72019-3_16

Authors


More by this author
Institution:
University of Oxford
Division:
MPLS
Department:
Computer Science
Role:
Author
More by this author
Institution:
University of Oxford
Department:
COMPUTER SCIENCE
Sub department:
Computer Science
Role:
Author
ORCID:
0000-0001-7509-680X
More by this author
Institution:
University of Oxford
Division:
MPLS
Department:
Computer Science
Role:
Author
More by this author
Institution:
University of Oxford
Division:
MPLS
Department:
Computer Science
Role:
Author
ORCID:
0000-0002-2807-8462


Publisher:
Springer
Host title:
Proceedings of the 30th European Symposium on Programming (ESOP 2021)
Journal:
Lecture Notes in Computer Science More from this journal
Volume:
12648
Pages:
432-461
Series:
Lecture Notes in Computer Science
Publication date:
2021-03-23
Acceptance date:
2021-01-28
Event title:
30th European Symposium on Programming, ESOP 2021, held as part of the European Joint Conferences on Theory and Practice of Software, ETAPS 2021
Event website:
https://etaps.org/2021/esop
Event start date:
2021-03-27
Event end date:
2021-04-01
DOI:
ISSN:
0302-9743


Language:
English
Keywords:
Pubs id:
1166367
Local pid:
pubs:1166367
Deposit date:
2021-03-08

Terms of use



Views and Downloads






If you are the owner of this record, you can report an update to it here: Report update to this record

TO TOP