Abstract Hidden Markov Models

a monadic account of quantitative information flow

Research output: Contribution to journalArticle

25 Downloads (Pure)


Hidden Markov Models, HMM's, are mathematical models of Markov processes with state that is hidden, but from which information can leak. They are typically represented as 3-way joint-probability distributions. We use HMM's as denotations of probabilistic hidden-state sequential programs: for that, we recast them as "abstract" HMM's, computations in the Giry monad D, and we equip them with a partial order of increasing security. However to encode the monadic type with hiding over some state X we use DX→D2X rather than the conventional X→DX that suffices for Markov models whose state is not hidden. We illustrate the DX→D2X construction with a small Haskell prototype. We then present uncertainty measures as a generalisation of the extant diversity of probabilistic entropies, with characteristic analytic properties for them, and show how the new entropies interact with the order of increasing security. Furthermore, we give a "backwards" uncertainty-transformer semantics for HMM's that is dual to the "forwards" abstract HMM's - it is an analogue of the duality between forwards, relational semantics and backwards, predicate-transformer semantics for imperative programs with demonic choice. Finally, we argue that, from this new denotational-semantic viewpoint, one can see that the Dalenius desideratum for statistical databases is actually an issue in compositionality. We propose a means for taking it into account.
Original languageEnglish
Article number36
Pages (from-to)1-50
Number of pages50
JournalLogical Methods in Computer Science
Issue number1
Publication statusPublished - 29 Mar 2019

Bibliographical note

Copyright the Author(s). Version archived for private and non-commercial use with the permission of the author/s and according to publisher conditions. For further rights please contact the publisher.


  • Abstract Hidden Markov Models
  • Giry Monad
  • Quantitative Information Flow

Fingerprint Dive into the research topics of 'Abstract Hidden Markov Models: a monadic account of quantitative information flow'. Together they form a unique fingerprint.

  • Cite this