Dilations and information flow axioms in categorical probability

Basic Information

• Paper ID: 2211.02507
• Title: Dilations and information flow axioms in categorical probability
• Authors: Tobias Fritz, Tomáš Gonda, Nicholas Gauguin Houghton-Larsen, Antonio Lorenzin, Paolo Perrone, Dario Stein
• Classification: math.CT cs.IT cs.LO math.IT math.PR
• Publication Date: October 17, 2025 (arXiv version)
• Paper Link: https://arxiv.org/abs/2211.02507

Abstract

This paper investigates the positivity axiom and causality axiom in Markov categories as properties of dilations and information flow, generalizing these concepts to arbitrary semicartesian monad categories. The research demonstrates that being a positive Markov category is merely an additional property (rather than additional structure) of symmetric monad categories. The paper characterizes positivity in representable Markov categories, proves that causality implies positivity (but not conversely), and shows that positivity fails in quasi-Borel spaces, interpreting this failure as a privacy property arising from probabilistic name generation.

Research Background and Motivation

Problem Background

Markov categories provide a categorical-theoretic approach to probability and statistics, achieving success in proving multiple classical theorems including the sufficient statistic theorem, 0/1 laws, statistical experiment comparison, and de Finetti's theorem. However, these developments typically require additional conditions such as existence of conditional probabilities, causality axioms, or positivity axioms.

Core Issues

1. Relationships between information flow axioms: The relationship between causality and positivity axioms was previously an open problem
2. Geometric meaning of axioms: A deeper understanding of the intuitive meaning and geometric interpretation of these axioms is needed
3. Categorical characterization: How to develop probability theory in more general semicartesian categories
4. Practical applications: How to verify these axioms in concrete probabilistic models

Research Motivation

The paper's primary motivations are:

• Clarifying the meaning and mutual relationships of information flow axioms
• Providing tools to determine whether a given Markov category satisfies these axioms
• Exploring the possibility of developing probability theory within the semicartesian category framework
• Demonstrating through concrete examples (such as quasi-Borel spaces) cases where axioms fail and their significance

Core Contributions

1. Proved that causality implies positivity: This is the paper's main result, resolving an open problem, with counterexamples constructed to show the converse does not hold
2. Established equivalent characterizations of positivity: Equivalently characterized positivity as deterministic marginal independence (DMI), providing more intuitive understanding
3. Characterized positivity in representable Markov categories: Provided complete characterization through the concept of strongly affine monads
4. Discovered failure of positivity in quasi-Borel spaces: Connected this to privacy equations, interpreting it as a privacy property arising from probabilistic name generation
5. Developed probability theory in semicartesian categories: Studied probability theory in a more general framework through the concept of dilations
6. Provided categorical characterization of positive Markov categories: Proved that positive Markov category structure can be recovered from semicartesian structure

Methodology Details

Core Concept Definitions

Dilations

For a morphism p : A → X in a semicartesian category D, a dilation is a morphism π : A → X ⊗ E satisfying:

The X-marginal of π = p

Intuitively, a dilation π describes a process consistent with p but potentially leaking information to the "environment" E.

Positivity Axiom

A Markov category C is positive if and only if for morphisms f : X → Y and g : Y → Z, whenever g∘f is deterministic:

The copy of g∘f = the copies of f and g acting respectively

Causality Axiom

A Markov category C is causal if and only if it satisfies the strengthened parametric equality property: if two morphisms are equal under some parameter, they remain equal under any dilation of that parameter.

Technical Framework

Deterministic Marginal Independence (DMI)

The paper proves that positivity is equivalent to DMI: for any deterministic morphism p : A → X, any dilation π : A → X ⊗ E of p exhibits conditional independence of X and E given A.

Dilation Categories

For each morphism p : A → X, define its dilation category Dilations(p), where:

• Objects are dilations of p
• Morphisms are equivalence classes of morphisms satisfying specific commutative diagrams

Initial Dilations

A dilation π is initial if and only if it is an initial object in Dilations(p). The paper proves that positivity is equivalent to the bloom (input-copying dilation) of deterministic morphisms being an initial dilation.

Technical Innovations

1. Dilation equivalence: Introduces the concept of dilation equivalence, generalizing the notion of almost sure equality
2. Categorical methods: Unifies treatment of various probabilistic models through categorical-theoretic approaches
3. Information flow perspective: Interprets probabilistic axioms as constraints on information flow
4. Semiring-valued kernel applications: Constructs counterexamples through semiring-valued random kernels, revealing deep connections between algebraic structures and probabilistic properties

Main Results

Theorem 1: Equivalent Characterizations of Positivity

For a Markov category C, the following are equivalent:

1. C is positive
2. C satisfies deterministic marginal independence
3. For all q : A → X ⊗ E, q is deterministic on X if and only if qX is deterministic

Theorem 2: Causality Implies Positivity

If a Markov category C is causal, then C is positive. The converse does not hold.

Proof Strategy: Constructs special dilations and transforms causality into positivity conditions using strengthened parametric equality.

Counterexample Construction: Uses the ideal lattice of the ideal Z2i to construct a semiring-valued Markov category that is positive but not causal.

Theorem 3: Characterization of Positivity in Representable Markov Categories

Let C be a representable Markov category with associated affine commutative monad P. Then C is positive if and only if P is strongly affine.

Theorem 4: Failure of Positivity in Quasi-Borel Spaces

In the Markov category QBStoch of quasi-Borel spaces:

1. The privacy equation holds: for atomless measures ν, the random singleton distribution equals the Dirac measure on the empty set
2. This causes deterministic marginal independence to fail
3. Therefore, QBStoch is neither positive nor causal

Theorem 5: Semicartesian Characterization

A semicartesian category D can be equipped with copy morphisms to form a positive Markov category if and only if:

1. Each identity morphism idX has an initial dilation ι : X → X ⊗ E whose marginal is non-creative
2. Specific deterministic conditions are satisfied

Applications and Examples

Semiring-Valued Kernels

The paper studies in detail the Kleisli category of the distribution monad DR over a semiring R:

• When R is an integral domain with no zero sums, the category is positive
• When R satisfies specific cancellation conditions, the category is causal
• Special semirings are constructed to demonstrate the separation of positivity and causality

Quasi-Borel Spaces and Privacy

In quasi-Borel spaces:

• Random singleton sets {X} and the empty set ∅ are equal in distribution
• This embodies information hiding and privacy protection characteristics
• Relates to fresh name generation in theoretical computer science

Quantum Information

In the category of finite-dimensional Hilbert spaces and quantum channels:

• Stinespring dilations are examples of initial dilations
• The no-cloning theorem corresponds to identity morphisms having only trivial dilations

Related Work

Markov Category Theory

The paper builds on the foundational Markov category theory established by Fritz and others, which has been successfully applied to:

• Sufficient statistic theory
• d-separation criteria for Bayesian networks
• Ergodic system theory
• Probabilistic programming theory

Dilation Theory

The concept of dilations has important applications in quantum information:

• Stinespring dilation theory
• Quantum self-testing theory
• Quantum purification axioms

Quasi-Borel Spaces

Quasi-Borel spaces were introduced by Heunen and colleagues as a cartesian-closed extension of the standard Borel space category, with important applications in higher-order probabilistic programming.

Conclusions and Discussion

Main Conclusions

1. Hierarchical relationships: Establishes a complete hierarchy of information flow axioms: existence of conditional probabilities → causality → positivity
2. Structure vs. property: Proves that positive Markov category structure is a property of symmetric monad categories rather than additional structure
3. Privacy interpretation: Interprets the failure of positivity in certain probabilistic models as a manifestation of privacy and information hiding

Limitations

1. Construction complexity: Some counterexample constructions are complex, requiring deep algebraic knowledge
2. Computational complexity: The paper focuses primarily on theoretical properties with limited discussion of computational complexity
3. Application scope: While providing a theoretical framework, verification methods in concrete applications require further development

Future Directions

1. Representable category characterization of causality: Seek characterizations of causality similar to the strongly affine monad characterization of positivity
2. Additional counterexamples: Construct more natural examples demonstrating axiom separation
3. Computational aspects: Develop algorithms and computational methods for verifying these axioms
4. Application extension: Apply the theory to concrete problems in machine learning and statistical inference

In-Depth Evaluation

Strengths

1. Theoretical depth: The paper resolves important open problems in Markov category theory and establishes a complete theoretical framework
2. Methodological innovation: Unifies treatment of different probabilistic models through dilation theory, providing new technical tools
3. Cross-disciplinary connections: Organically combines probability theory, category theory, quantum information, and theoretical computer science
4. Concrete applications: Demonstrates theoretical practicality through concrete examples such as quasi-Borel spaces

Weaknesses

1. Technical threshold: Requires substantial background in category theory and algebraic geometry, limiting the audience
2. Practical utility: While theoretically complete, the application value in practical probabilistic computation remains to be verified
3. Example complexity: Some key example constructions are overly technical, affecting intuitive understanding

Impact

1. Theoretical contribution: Establishes a solid foundation for Markov category theory, advancing categorical probability theory
2. Methodological value: The systematic development of dilation theory provides new tools for related fields
3. Interdisciplinary impact: Connects research in probability theory, quantum information, and theoretical computer science

Applicable Scenarios

1. Theoretical probability: Provides new perspectives for axiomatic research in probability theory
2. Quantum information: Dilation theory has direct applications in quantum channel and quantum measurement theory
3. Probabilistic programming: Provides theoretical foundations for the semantics of higher-order probabilistic programming languages
4. Statistical inference: Provides categorical-theoretic frameworks for Bayesian inference and causal inference

References

The paper cites 49 important references spanning category theory, probability theory, quantum information, and theoretical computer science, particularly:

• Fritz's foundational work on Markov category theory
• Related literature on quasi-Borel space theory
• Applications of dilation theory in quantum information
• Research on probabilistic programming and fresh name generation