Finite-Type Bayesian Conditioning

Updated 5 July 2026

Finite-Type Bayesian Conditioning is a framework that formalizes Bayesian updating on finite sets using assert and normalization operations with fuzzy predicates.
It internalizes states, predicates, and side-effect-free instruments within a type-theoretic setting, deriving Bayes’ rule from the underlying definitions.
The framework extends to rigorous probabilistic inference, offering insights into posterior prediction under constraints and algebraic extensions of convexity.

Searching arXiv for the cited papers to ground the article. arXiv search: (Adams et al., 2015) arXiv search: (Sargsyan, 18 Jun 2026) Finite-type Bayesian conditioning is the finite set-based realization of Bayesian updating in which a prior state on a finite type is transformed by an “assert-then-normalize” operation induced by a fuzzy predicate. In the type-theoretic formulation of probabilistic reasoning, states, predicates, normalization, and conditioning are internalized as typed constructions, and conditioning is tied to a bijective correspondence between predicates and side-effect-free actions. In this setting, Bayes’ rule is not added externally; it is obtained directly from the definitions of assert and normalization (Adams et al., 2015).

1. Finite types, states, and fuzzy predicates

The basic carrier is a fixed finite set $A$ , referred to as a finite type. A probability state on $A$ is a map

$\omega:A\to[0,1]$

satisfying

$\sum_{a\in A}\omega(a)=1.$

Such states are written $\omega\in D(A)$ , using the finite distribution monad. A fuzzy predicate on $A$ is a map

$p:A\to[0,1].$

A substate $\alpha$ on $A$ is a map $\alpha:A\to[0,1]$ with total weight

$A$ 0

satisfying $A$ 1. Normalization is defined by

$A$ 2

where $A$ 3, so that $A$ 4 is a full state in $A$ 5 (Adams et al., 2015).

Object	Definition	Role
State $A$ 6	$A$ 7, $A$ 8	Prior or posterior distribution
Fuzzy predicate $A$ 9	$\omega:A\to[0,1]$ 0	Soft event or likelihood
Substate $\omega:A\to[0,1]$ 1	$\omega:A\to[0,1]$ 2, $\omega:A\to[0,1]$ 3	Unnormalized updated state
Normalization	$\omega:A\to[0,1]$ 4	Converts substate to state

The logic associated with this setting is quantitative, with fuzzy predicates. This is significant because conditioning is formulated for graded evidence rather than only for sharp events, so likelihood-style updating is handled in the same formal language as Boolean conditioning. A plausible implication is that finite-type Bayesian conditioning serves simultaneously as a semantic account of posterior formation and as a proof-theoretic account of probabilistic inference.

2. Assert maps and the predicate–instrument correspondence

Given a state $\omega:A\to[0,1]$ 5 and a predicate $\omega:A\to[0,1]$ 6, the assert substate is defined pointwise by

$\omega:A\to[0,1]$ 7

Its total weight is

$\omega:A\to[0,1]$ 8

which is $\omega:A\to[0,1]$ 9 whenever conditioning is performed on an event of positive probability (Adams et al., 2015).

In the formal type theory, this assert map arises as a partial map

$\sum_{a\in A}\omega(a)=1.$ 0

with the property that its underlying substate on $\sum_{a\in A}\omega(a)=1.$ 1 is exactly $\sum_{a\in A}\omega(a)=1.$ 2. Equivalently, at the level of partial maps,

$\sum_{a\in A}\omega(a)=1.$ 3

A central structural fact is the bijection between fuzzy predicates on $\sum_{a\in A}\omega(a)=1.$ 4 and side-effect-free instruments $\sum_{a\in A}\omega(a)=1.$ 5, or equivalently asserts $\sum_{a\in A}\omega(a)=1.$ 6. The correspondence is given by

$\sum_{a\in A}\omega(a)=1.$ 7

Conversely, from a suitable map $\sum_{a\in A}\omega(a)=1.$ 8 one recovers the predicate by

$\sum_{a\in A}\omega(a)=1.$ 9

The paper identifies this predicate–instrument correspondence as the key aspect distinguishing the probabilistic type theory from quantum type theory, and it uses that correspondence to derive computation rules for conditioning and related probabilistic calculations (Adams et al., 2015).

The summary given for the finite set-based account states that these constructions are internalized in the COMET type theory via the instrument/assert maps, the partial sum, and the normalization operators, with the bijection between predicates and side-effect-free asserts guaranteeing that “assert-then-normalize” is the correct formal realization of Bayesian conditioning. This suggests that the operational content of conditioning is built into the calculus rather than imposed as an external meta-level rule.

3. Conditioning as normalization of assertion

For a state $\omega\in D(A)$ 0 and a predicate $\omega\in D(A)$ 1 with total weight

$\omega\in D(A)$ 2

the conditioned state is defined by

$\omega\in D(A)$ 3

Pointwise,

$\omega\in D(A)$ 4

In type-theoretic notation,

$\omega\in D(A)$ 5

Bayes’ rule follows immediately from this definition. If $\omega\in D(A)$ 6 is a finite set, $\omega\in D(A)$ 7 is a prior on $\omega\in D(A)$ 8, and $\omega\in D(A)$ 9 is a binary observation event represented by the predicate

$A$ 0

then the asserted substate is

$A$ 1

After normalization,

$A$ 2

Writing $A$ 3, $A$ 4, and $A$ 5, one obtains

$A$ 6

In this formulation, the classical Bayesian posterior is exactly the normalized assert state (Adams et al., 2015).

One common misconception is to treat conditioning in such systems as merely a notation for posterior probabilities already defined elsewhere. In the finite-type treatment, conditioning is instead a derived operation on states and predicates, and Bayes’ rule appears as a theorem-level consequence of that operation.

4. Worked two-variable Boolean example

A fully explicit example is given for two Boolean variables $A$ 7. The prior on $A$ 8 is assumed to be a product prior,

$A$ 9

with

$p:A\to[0,1].$ 0

A test $p:A\to[0,1].$ 1 is observed with likelihood factorization

$p:A\to[0,1].$ 2

where

$p:A\to[0,1].$ 3

The test is treated as a predicate on $p:A\to[0,1].$ 4: $p:A\to[0,1].$ 5

The joint substate after asserting $p:A\to[0,1].$ 6 is

$p:A\to[0,1].$ 7

Its total weight is

$p:A\to[0,1].$ 8

The normalized posterior on $p:A\to[0,1].$ 9 is then

$\alpha$ 0

Marginalizing over $\alpha$ 1 gives

$\alpha$ 2

The expectation over $\alpha$ 3 is

$\alpha$ 4

Hence

$\alpha$ 5

Therefore,

$\alpha$ 6

The example exhibits the full finite-type workflow: prior factorization, likelihood as predicate, assert, normalization, and optional marginalization (Adams et al., 2015).

5. Formalisation and algebraic extensions

A later Cubical Agda formalisation develops finite-type Bayesian conditioning for finitely supported distributions as a higher inductive type $\alpha$ 7 with point-constructors

$\alpha$ 8

together with six path-constructors imposing the usual convex-algebra laws, including $\alpha$ 9-idem, $A$ 0-comm, two boundary laws, $A$ 1-interchange, and $A$ 2-assoc (Sargsyan, 18 Jun 2026).

That work states that the standard convex-algebra interchange axiom, common to probability-monad formalisations since Stone, is provably too weak to support full Bayesian conditioning. The obstruction appears when conditioning is lifted through the $A$ 3-interchange path: after conditioning, the two inner probabilities become

$A$ 4

where the event marginals $A$ 5 are expectations of the likelihood under the corresponding branches. In general $A$ 6, so the standard interchange law cannot be applied.

On a full-support syntactic fragment $A$ 7, recursive conditioning is defined using

$A$ 8

and a recursive $A$ 9 operator that computes expectations $\alpha:A\to[0,1]$ 0 on subtrees and reweights each $\alpha:A\to[0,1]$ 1 node accordingly. To make HIT-lifting possible, the formalisation adds one new path-constructor, $\alpha:A\to[0,1]$ 2, which generalises the ordinary interchange law and specialises back to it when $\alpha:A\to[0,1]$ 3. The paper proves that the standard form is the degenerate case where the two inner weights coincide.

For a finite carrier $\alpha:A\to[0,1]$ 4, the same work defines a conditioning operation by

$\alpha:A\to[0,1]$ 5

and states that this extends uniquely to a total Agda function $\alpha:A\to[0,1]$ 6 satisfying Bayes’ rule at the mass level. It also packages a chain $\alpha:A\to[0,1]$ 7 into an $\alpha:A\to[0,1]$ 8 record and computes $\alpha:A\to[0,1]$ 9 via a conditioning operation $A$ 00 on the joint distribution (Sargsyan, 18 Jun 2026).

Finite-type Bayesian conditioning also appears in more asymptotic or constraint-based forms. In one line of work, an exchangeable sequence on a finite alphabet is conditioned on empirical moment constraints through a “Lanford window”

$A$ 01

under assumptions including finite alphabet, strictly positive baseline $A$ 02-density, convex and closed constraint set, and unique $A$ 03-projection. The resulting theorem states that for every fixed block size $A$ 04,

$A$ 05

where

$A$ 06

and gives the finite-sample bound

$A$ 07

so that for fixed $A$ 08 the leading rate is $A$ 09 (Polson et al., 16 Sep 2025). This suggests a broader finite-type picture in which posterior prediction under constraints is governed by $A$ 10-projection and exponential tilting rather than by elementary event indicators alone.

A related framework studies finite sample spaces $A$ 11 with a reference pmf $A$ 12, a moment map $A$ 13, and a constraint manifold

$A$ 14

There the constrained predictive law for future draws is represented as a mixture over feasible empirical types, and around the information projection

$A$ 15

the predictor admits a discrete–Gaussian mixture representation involving the projected Hessian

$A$ 16

The smallest eigenvalue $A$ 17 measures local curvature, and the total-variation contraction bound depends explicitly on $A$ 18 and $A$ 19 (Polson et al., 23 Oct 2025). The same paper relates this geometry to empirical likelihood, Bayesian exponentially-tilted empirical likelihood, generalized method of moments, and generalized estimating equations.

The scope of finite-type conditioning is clarified by results on brittleness in continuous settings. For Bayesian models on a Polish space $A$ 20, specified only through finitely many marginals or moment constraints, posterior lower and upper bounds after conditioning on finite-precision data can still attain the full deterministic range of a quantity of interest $A$ 21. The same work shows that arbitrarily small Prokhorov or total-variation perturbations of the model may produce arbitrarily large posterior shifts after conditioning (Owhadi et al., 2013). This does not negate the finite set-based theory; rather, it delineates a boundary. A plausible implication is that the algebraic clarity of finite-type Bayesian conditioning does not by itself supply robustness once the setting moves to continuous model classes under finite information.

In that sense, finite-type Bayesian conditioning occupies two roles. First, it is an exact formal account of Bayesian updating on finite carriers via states, predicates, assert maps, and normalization. Second, it functions as a reference model against which later formalisations, asymptotic constraint-based updates, and robustness analyses can be compared.