Probabilistically Stable Belief Revision

• Probabilistically stable belief revision is a framework that ties categorical belief to resiliently high credence by enforcing strict probability thresholds under all non-negligible updates.
• It revises beliefs by applying Bayesian conditioning followed by the extraction of a strongest-stable set, ensuring that only robust propositions are retained.
• The framework supports non-monotonic inference, exemplified by Rational Monotonicity, and finds applications in voting theory, comparative probability, and revealed preference.

Searching arXiv for recent and foundational papers on probabilistically stable belief revision and closely related probabilistic belief-change frameworks. Probabilistically stable belief revision is a family of belief-change theories in which categorical belief is tied to resiliently high credence rather than to high credence simpliciter. In a central formulation on finite probability spaces, an event qualifies for acceptance only if every non-negligible, non-contradictory update preserves above-threshold probability, and revision by new evidence proceeds by Bayesian conditioning followed by renewed stability selection. The resulting operators induce a non-monotonic qualitative logic with strong monotonicity properties, especially Rational Monotonicity, while departing from orthodox AGM revision and from simple Lockean thresholding (Mierzewski, 2 Sep 2025).

1. Stability as resiliently high credence

The basic setting is a finite probability space $(\Omega,\mathfrak{A},\mu)$ together with a threshold $t \in [1/2,1)$. Leitgeb’s stability rule defines an event $X \in \mathfrak{A}$ as $(\mu,t)$-stable iff for all $E \in \mathfrak{A}$ such that $X \cap E \neq \emptyset$ and $\mu(E) > 0$, one has $\mu(X \mid E) > t$. The intended reading is that categorical belief is reserved for propositions whose probability remains above threshold under every non-negligible, non-contradictory conditioning. The same framework allows a more general schema with a separate resilience threshold $s \in (0,1]$, but the main formulation fixes the resilience domain implicitly at the positive-probability condition $\mu(E)>0$ (Mierzewski, 2 Sep 2025).

The corresponding strongest-stable-set operator is

$t \in [1/2,1)$0

Its induced categorical belief set is

$t \in [1/2,1)$1

On finite algebras, $t \in [1/2,1)$2 exists and is unique, and the set of stable events is linearly ordered by $t \in [1/2,1)$3. Thus the theory does not identify belief with an arbitrary family of high-probability propositions; it identifies belief with the logically strongest event whose probability is resilient under admissible updates.

This conception differs from threshold acceptance in a decisive way. A proposition may be highly probable at the current state and yet fail stability if some consistent, non-negligible conditioning would depress it below $t \in [1/2,1)$4. Conversely, a stable proposition is not merely probable now; it is robust under the relevant domain of Bayesian revisions.

2. Revision by conditioning and strongest-stable-set selection

Probabilistically stable revision tracks Bayesian conditioning at the quantitative level and then re-applies the stability rule at the qualitative level. For evidence $t \in [1/2,1)$5 with $t \in [1/2,1)$6, define the conditional measure

$t \in [1/2,1)$7

and then define the revision operator

$t \in [1/2,1)$8

Unconditional categorical belief is therefore $t \in [1/2,1)$9, while the revised strongest categorical belief after learning $X \in \mathfrak{A}$0 is $X \in \mathfrak{A}$1. In this sense, the qualitative dynamics are the “shadow” of Bayesian update: conditioning happens first, strongest-stable-set extraction second (Mierzewski, 2 Sep 2025).

This induces a Ramsey-test style consequence relation. Writing $X \in \mathfrak{A}$2 for “after learning $X \in \mathfrak{A}$3, $X \in \mathfrak{A}$4 is defeasibly accepted,” one has

$X \in \mathfrak{A}$5

The relation is explicitly non-monotonic: additional information can alter which stable core is selected, and hence can alter the accepted consequences of a premise.

The same dynamics can be reformulated by selection functions. A selection structure is a triple $X \in \mathfrak{A}$6 with $X \in \mathfrak{A}$7, where $X \in \mathfrak{A}$8 is read as the selected or typical $X \in \mathfrak{A}$9-worlds, here interpreted as the strongest stable set after conditioning on $(\mu,t)$0. The structure is $(\mu,t)$1-representable iff there exists a probability measure $(\mu,t)$2 such that $(\mu,t)$3 when $(\mu,t)$4, and $(\mu,t)$5 when $(\mu,t)$6.

On a finite sample $(\mu,t)$7, each $(\mu,t)$8-plan determines a region in the probability simplex $(\mu,t)$9 via fixed-odds hyperplanes

$E \in \mathfrak{A}$0

for pairs $E \in \mathfrak{A}$1 with $E \in \mathfrak{A}$2. Two measures generate the same $E \in \mathfrak{A}$3-plan iff they lie in the same region of this hyperplane arrangement and have the same support. This geometric characterization makes the stability operator a piecewise-defined object over the simplex rather than a merely syntactic belief-change rule.

3. Representation theorems and comparative-probability structure

At threshold $E \in \mathfrak{A}$4, the strongest-stable-set semantics admits an exact qualitative characterization by axioms on selection functions. The relevant conditions are regularity $E \in \mathfrak{A}$5, reflexivity of selection $E \in \mathfrak{A}$6, strong monotonicity $E \in \mathfrak{A}$7, and the weak fallback scheme $E \in \mathfrak{A}$8, together with a Scott-style cancellation principle. The induced atom-set dominance relation is defined by $E \in \mathfrak{A}$9 iff $X \cap E \neq \emptyset$0, and $X \cap E \neq \emptyset$1 iff $X \cap E \neq \emptyset$2. Extending this to the mixed strict/non-strict relation $X \cap E \neq \emptyset$3, one obtains the measurement-theoretic constraint called (Scott): if balanced sequences $X \cap E \neq \emptyset$4 and $X \cap E \neq \emptyset$5 satisfy $X \cap E \neq \emptyset$6 for all $X \cap E \neq \emptyset$7, then $X \cap E \neq \emptyset$8 for all $X \cap E \neq \emptyset$9 (Mierzewski, 2 Sep 2025).

The main theorem states that a selection structure $\mu(E) > 0$0 satisfies $\mu(E) > 0$1–$\mu(E) > 0$2 and (Scott) iff it is representable as $\mu(E) > 0$3 for some regular $\mu(E) > 0$4 on $\mu(E) > 0$5. For rational thresholds $\mu(E) > 0$6, the same pattern yields the axiom Scott$\mu(E) > 0$7, based on $\mu(E) > 0$8-balanced sequences. The corresponding theorem states that a selection structure satisfies $\mu(E) > 0$9–$\mu(X \mid E) > t$0 and Scott$\mu(X \mid E) > t$1 iff it is representable as $\mu(X \mid E) > t$2 for some regular $\mu(X \mid E) > t$3. Moreover, every representable $\mu(X \mid E) > t$4 is representable at some rational $\mu(X \mid E) > t$5.

The proof has two notable components. First, the axioms are shown to mirror the qualitative structure of the stability test. Second, the representability problem is translated into a strict/non-strict linear system over $\mu(X \mid E) > t$6, and the Motzkin Transposition Theorem is used to show that solvability is equivalent to the relevant cancellation constraint. This is why the framework is tightly connected to comparative probability.

That connection extends beyond revision proper. The paper gives necessary and sufficient conditions for the joint weak representability of a pair of strict and non-strict comparative probability orders and develops an axiomatization for ratio comparisons of the form “event $\mu(X \mid E) > t$7 is at least $\mu(X \mid E) > t$8 times more likely than event $\mu(X \mid E) > t$9.” In the ratio setting, $s \in (0,1]$0 means $s \in (0,1]$1, and the corresponding logic is controlled by a Scott$s \in (0,1]$2 scheme.

4. Logical profile and relation to AGM-style belief change

The logic generated by strongest-stable-set revision is non-monotonic but not weak. It validates Ref, Left Equivalence, Right Weakening, And, and Rational Monotonicity. Rational Monotonicity takes the form: if $s \in (0,1]$3 and $s \in (0,1]$4, then $s \in (0,1]$5. At the same time, Or fails, Hawthorne–Makinson’s weak Or fails, and Cut fails in general (Mierzewski, 2 Sep 2025).

The failure of Or is structurally important. It shows that stable revision does not support unrestricted case reasoning. A very weak Or-like rule survives only under side conditions excluding atypical disjuncts: if for a finite family $s \in (0,1]$6 one has $s \in (0,1]$7 for all $s \in (0,1]$8, then

$s \in (0,1]$9

In the two-premise version, if atyp$\mu(E)>0$0 and atyp$\mu(E)>0$1, where atyp$\mu(E)>0$2, then from $\mu(E)>0$3 and $\mu(E)>0$4 one may infer $\mu(E)>0$5.

The framework also departs from AGM. The $\mu(E)>0$6-induced tracking operator $\mu(E)>0$7 is different from the “stability-induced AGM” operator obtained via spheres of stable sets. Success holds because $\mu(E)>0$8, and Preservation is validated when $\mu(E)>0$9 is consistent with current beliefs, but Inclusion fails. The paper links this divergence to the Kelly–Lin impossibility and treats the Or-rule failure as part of a broader failure of conglomerability-style case reasoning.

A related but distinct probabilistic tradition reaches a similar verdict about AGM-strength dynamics. Goodman and Salow’s question-relative account of belief validates Box− and BoxR in all probability structures, but Diamond−, DiamondR, and Box+ can fail in general; they explicitly characterize their dynamics as much weaker than AGM, though stronger than Lockean threshold belief (Goodman et al., 2023). A plausible implication is that the distance between probabilistic belief and AGM revision is not an idiosyncrasy of stability-based semantics, but a recurrent feature of frameworks that insist on preserving genuinely probabilistic structure.

5. Robustness, threshold phenomena, and neighboring senses of stability

Within strongest-stable-set semantics, stability is synchronic resilience under all non-negligible, consistent conditionings. In adjacent literatures, however, closely related expressions of “probabilistic stability” denote different robustness conditions. In distributed revision of belief commitment in Bayesian networks, stability refers to convergence and consistency of categorical commitments under max-product propagation. In singly connected networks, the algorithm converges to a unique global assignment $t \in [1/2,1)$00 that maximizes $t \in [1/2,1)$01, and robustness can be measured by the log-posterior margin

$t \in [1/2,1)$02

A large $t \in [1/2,1)$03 implies that small changes in local conditionals or evidence likelihoods are unlikely to alter the maximizing assignment. The medical-diagnosis example in that framework exhibits a sharp threshold on an upstream message at $t \in [1/2,1)$04, where two explanations become equally probable; crossing that threshold reversibly flips the commitment, and the lack of hysteresis in trees is attributed to global optimality (Pearl, 2013).

Evidence-theoretic belief base revision uses stability in another sense: robustness of selected subbases or selected intersections under perturbations of the underlying mass function. There the key quantity is again a credibility gap, written as the minimum strict difference between the top credibility and the runner-up over the relevant selection domain. If this gap is positive, then there exists $t \in [1/2,1)$05 such that sufficiently small sup-norm perturbations preserving ordering gaps beyond $t \in [1/2,1)$06 leave the selected family unchanged. Because Dempster’s rule is continuous on the simplex of BBAs away from total conflict, the induced Bel-values vary continuously, and selection is piecewise constant, changing only when a tie is created or broken (Ktari et al., 2020).

A third neighboring use appears in revision of incompletely specified convex probabilistic belief bases by generalized imaging. There, stability refers to the fact that revision is defined even when $t \in [1/2,1)$07, the imaging map is affine on the simplex, convexity is preserved, and revising only the boundary distributions and then taking convex closure yields exactly the same revised set as revising all distributions in $t \in [1/2,1)$08. The framework therefore treats zero-probability evidence without the undefinedness that affects ordinary conditioning (Rens et al., 2016).

These variants are not equivalent, but they are structurally aligned. Each ties belief change to some invariance domain: strongest-stable-set semantics uses invariance under admissible conditioning, Bayesian-network revision uses invariance under small perturbations away from threshold boundaries, evidence-theoretic operators use invariance of an argmax under perturbation of masses, and generalized imaging uses invariance of convex structure under revision.

6. Applications, computation, and broader debates

The strongest-stable-set framework is not confined to abstract revision logic. One application is to simple voting games. For each $t \in [1/2,1)$09, define the blocking game $t \in [1/2,1)$10 with winning coalitions $t \in [1/2,1)$11, where $t \in [1/2,1)$12. Representability of the selection function is then equivalent to simultaneous quota representation of the family $t \in [1/2,1)$13 by a common weight function. Another application is to revealed preference: representable $t \in [1/2,1)$14 yields a cautious, dominance-stable choice function for menus, satisfying $t \in [1/2,1)$15, Aizerman, and (WIIA), while failing $t \in [1/2,1)$16, $t \in [1/2,1)$17, (RIIA), and (WARP) (Mierzewski, 2 Sep 2025).

Algorithmically, checking stability reduces to testing linear inequalities of the form

$t \in [1/2,1)$18

for atoms $t \in [1/2,1)$19 and strict subsets $t \in [1/2,1)$20. The global partition of $t \in [1/2,1)$21 by fixed-odds hyperplanes induces a cell decomposition in which each cell corresponds to a distinct $t \in [1/2,1)$22-plan. Motzkin’s transposition theorem supplies the existence criterion for a representing $t \in [1/2,1)$23, definitional complexity in probability logic is $t \in [1/2,1)$24 with linear constraints for $t \in [1/2,1)$25, and model checking is linear in the number of inequality tests. Thus the formalism is nontrivial logically, but computationally anchored in linear feasibility and separation.

A persistent controversy concerns the role of conditioning in probabilistic revision. One line of work argues that Bayes’ theorem is not a generally applicable revision rule once explicit conditions and implicit conditions are distinguished, and that Jeffrey’s rule is still an updating rule rather than a general revision rule because the new assessment replaces the prior assessment of the target event rather than combining sources symmetrically (Wang, 2013). A related proposal argues that constraining, not conditioning, is the better candidate for probabilistic expansion: constraining is monotone, idempotent, commutative for consistent constraints, and preserves previously accepted probabilistic constraints in a way that conditioning does not (Voorbraak, 2013). Another neighboring approach adds counterfactual probabilities to possibility rankings so that revision by zero-probability sentences is still meaningful: if $t \in [1/2,1)$26 but $t \in [1/2,1)$27, revision proceeds by restricting and renormalizing on the most plausible $t \in [1/2,1)$28-worlds, thereby extending conditionalization through a generalized imaging perspective (Boutilier, 2013).

Taken together, these debates clarify the place of probabilistically stable belief revision. It is not merely a probabilistic gloss on AGM, nor merely threshold acceptance, nor merely Bayesian conditioning under another name. It is a family of qualitative belief-change operators extracted from quantitative credal structure by a resilience criterion, with exact representation theorems, sharply delimited logical behavior, and applications that range from comparative probability to voting theory and revealed preference.