Foundations of MT-PDCL: Measure-Theoretic Probabilistic Definite Clause Logic

Measure-Theoretic Probabilistic Definite Clause Logic (MT-PDCL) is a generalized foundational framework for probabilistic logic programming that eliminates the finite-domain restriction of standard frameworks. Standard frameworks like PRISM and ProbLog rely on grounding logic programs into discrete propositional representations (e.g., BDDs, SDDs), which restricts exact inference to finite domains and discrete probability distributions. MT-PDCL removes this restriction by explicitly defining stochastic variables over bounded index domains and equipping the interpretation space with standard Borel σ-algebras, allowing logical variables to operate natively over continuous measurable spaces.

Building on Continuous Distribution Semantics, MT-PDCL models probabilistic rules as mutually independent causal events. However, rather than aggregating derivations via finite boolean circuits, declarative entailment is formally defined through exact Lebesgue integration over the continuous measure space. The framework introduces a continuous immediate consequence operator that unifies the integration of continuous prior distributions with the evaluation of exact continuous observations. This approach replaces the combinatorial bottleneck of discrete grounding with exact, algebraic, and structurally differentiable inference, achieving the expressive power of continuous probabilistic models while preserving the pure declarative syntax of definite clause logic.

The paper makes four primary contributions. First, it provides continuous declarative semantics by equipping the interpretation space and explicitly bounded stochastic index domains with standard Borel σ-algebras, removing the finite-domain restriction. Second, it extends classical distribution semantics to infinite domains, defining possible worlds via independent continuous variables and probabilistic causal events, with declarative entailment formally defined as the exact Lebesgue integral of logically valid possible worlds. Third, it generalizes the immediate consequence operator (TK) to operate algebraically over continuous probability densities, using exact integration for continuous marginalization and a continuous Noisy-OR formulation to aggregate disjunctive derivations. Fourth, it formally proves that this continuous operator converges to a well-defined least fixed point and is everywhere differentiable, providing a structurally transparent, non-negative gradient signal for neural-symbolic deep learning.

The syntax of MT-PDCL defines a first-order language L(C, R) where R is partitioned into user-defined logical relations (Ru), built-in base predicates for standard mathematical operators (O), and stochastic predicates (Rs). Terms are either logical variables or constants, with constants including numeric values, categorical symbols, mathematical intervals, and finite sets. Stochastic predicates use a semicolon to separate deterministic indices from generative variables: r(I1,..., Ik; X1,..., Xm). The knowledge base is a tuple K = ⟨Ms, Φ⟩, where Φ is the logic program of definite clauses and Ms is the Stochastic Measure Mapping that assigns probability measures to stochastic predicates based on ground index tuples.

The semantics are built on a universal domain D that is a standard Borel space, constructed as the disjoint union of a continuous space (e.g., Rn) and a countable discrete set. Interpretations map constants to domain elements and predicates to indicator functions, with the requirement that the subset of Dn for which a predicate evaluates to 1 must be measurable. The stochastic state space Ωs is the Cartesian product of domains for all independent random variables across all stochastic predicates and ground indices, with the Kolmogorov Extension Theorem ensuring a valid measurable space with a well-defined joint probability measure. The index domain Ir is established either implicitly through parameter support (e.g., standard deviation must be positive) or explicitly through bounded fields defined by conjunctions of built-in base predicates.

Probabilistic rules are modeled as independent causal events Ej, each succeeding with probability pj. The total universe of uncertainty is the product space Ω = Ωs × ΩE, with a total probability measure μΩ = μs ⊗ μE. A specific possible world ω = (σ, ϵ) resolves all stochastic predicates to constants and all causal events to True/False, leaving a classical deterministic Definite Clause program with a unique Least Model. Ground query probabilities are defined as the Lebesgue integral of the logical indicator function over the space of possible worlds. Non-ground queries return probabilistic answers as tuples ⟨ΩA, PK(Q, ΩA)⟩, where ΩA is a measurable subset of the continuous domain and PK is the exact declarative probability.

The paper provides several illustrative examples. In the continuous spatial detection example, a security system monitors a 10km road with a sensor zone covering [4, 6] and an 80% alarm probability. The exact probability of alarm is computed via Lebesgue integration as 0.16. In the multi-sensor fusion example, exact deterministic observations (temperature 120.5, pressure 4.2) yield a 95% probability of a critical warning and 0% for moderate warning, demonstrating backward compatibility with discrete frameworks when the stochastic measure mapping is empty. In the continuous stochastic optimization example, an agent allocates energy between navigation and communication systems subject to continuous noise, with the optimal allocation derived via calculus as x = 10/3 to navigation and y = 20/3 to communication, yielding a maximum mission success probability of approximately 0.2739. The parameterized distributions example models a component failing at a random time with a drifting thermal signature, computing the alarm probability via marginalization over the failure time prior. The hybrid measure spaces example unifies continuous time with discrete Poisson fault counts, computing a critical failure probability of approximately 0.70. The continuous information cascades example models time-decaying information diffusion through a social network with cyclic paths, with a theorem bounding the approximation error of the Noisy-OR assumption.

The operational semantics introduce the continuous probabilistic consequence operator T: P → P, where P is the space of probability functions mapping ground facts to [0, 1]. For a rule instance evaluated at binding v, the isolated contribution is Pupdate(H Rj, v) = pj · P(B v). The exact operator Texact computes the Lebesgue measure of the geometric union of success regions across all rules and continuous evaluation domains. However, this exact aggregation is generally intractable, so the paper introduces the continuous Noisy-OR operator TK, which assumes Independent Causal Influence (ICI) and computes the joint failure probability as a product integral: Pfail(Rj) = exp(∫ ln(1 - Pupdate(H Rj, v)) dμV(v)). The final probability is 1 minus the product of failure probabilities across all rules.

The paper proves that both Texact and TK are monotonically increasing over the complete lattice P, guaranteeing the existence of a least fixed point via the Knaster-Tarski theorem, and that they are ω-continuous, guaranteeing convergence via Kleene's fixed-point theorem. The paper also analyzes contraction properties using the Banach Fixed-Point Theorem, defining the recursive measure γ as the supremum of expected successful derivations for cyclic predicates. When γ < 1, the operator is a strict contraction with geometric convergence and a computable iteration bound. When γ ≥ 1, the operator is expansive, and the Knaster-Kleene approach is necessary for convergence. The paper proves that Texact is sound and complete with respect to declarative semantics (lfp(Texact)(H) = PK(H)), and that TK is sound and complete under conditional independence. The paper also provides an absolute bound on the approximation gap between exact and operational probabilities using Boole-Fréchet inequalities.

The differentiability of TK is proven, showing that the gradient of the operational probability with respect to neural network weights is non-negative, ensuring stable gradient routing for backpropagation. The paper compares MT-PDCL to existing frameworks including PRISM, ProbLog, DeepProbLog, TensorLog, BLOG, DC-ProbLog, Hybrid ProbLog, and Weighted Model Integration (WMI), highlighting that MT-PDCL uniquely provides exact denotational semantics via measure theory and fixed-point theorems while preserving structural differentiability.

The execution traces section evaluates MT-PDCL on three problems. The infinite-horizon probability of ruin problem (supply chain) computes the probability that inventory never drops below zero, showing that the probability decays as 1/√k and converges to zero by the Sparre Andersen theorem. The probability of safe arrival problem (kinematic agent) computes a backward reachable safe set, expanding geometrically until reaching the full domain [0, 10] with probability 1.0. The continuous signal propagation problem (hierarchical sensor network) computes exact probabilities for a rooted tree, yielding 0.2 for node4 and approximately 0.0416 for node3.

The paper concludes by discussing computability and tractability limitations. Exact analytical solutions are guaranteed only for standard integrable functions and linear arithmetic constraints. The curse of dimensionality applies: every continuous variable adds an integration dimension, and recursive derivations of depth k require k-dimensional integrals with complexity O(mk) for grid-based quadrature. Non-contractive programs (γ ≥ 1) require infinite iterations for exact convergence, making exact probabilities computationally unreachable in finite time. Future work focuses on algorithmic implementation, query-directed inference procedures, Deep-MT-PDCL for neural-symbolic integration, and identifying algebraic systems closed under continuous logical reasoning.

Improvements to AI Systems:

1. Continuous Probabilistic Reasoning over Unbounded Domains

• Replace discrete grounding (BDDs/SDDs) with exact Lebesgue integration, enabling AI to reason natively over continuous variables (e.g., sensor readings, time, spatial coordinates) without discretization error.

• Capability: An AI can answer queries like What is the probability that a moving object enters a zone at any real-valued time? with exact analytical results, not approximations.

2. Differentiable Neural-Symbolic Inference with Non-Negative Gradients

• Use the proven differentiability of the continuous Noisy-OR operator (TK) to backpropagate through probabilistic logic programs, with guaranteed non-negative gradient signals.

• Capability: A neural-symbolic system can learn continuous distribution parameters (e.g., Gaussian means, variances) from data while preserving logical constraints, avoiding vanishing/exploding gradients common in discrete logic.

3. Exact Marginalization over Hybrid Measure Spaces

• Leverage the Kolmogorov Extension Theorem and product measures to combine continuous (e.g., time) and discrete (e.g., Poisson counts) random variables in one framework.

• Capability: An AI can compute joint probabilities like probability of system failure within 5 hours given a discrete number of faults exactly, without Monte Carlo sampling.

4. Fixed-Point Convergence with Contraction Guarantees

• Use the recursive measure γ to detect contractive programs (γ < 1) and apply Banach fixed-point iteration for fast, bounded convergence; fall back to Kleene iteration for expansive cases.

• Capability: An AI can automatically choose the optimal inference algorithm for recursive probabilistic queries (e.g., infinite-horizon inventory ruin probabilities) with a computable error bound.

5. Algebraic Optimization over Continuous Stochastic Programs

• Apply calculus-based optimization directly on the continuous probability density functions (e.g., finding optimal resource allocation via derivatives).

• Capability: An AI can solve constrained stochastic optimization problems (e.g., energy allocation under noise) analytically, yielding closed-form solutions rather than numerical search.

6. Sound Approximation with Boole-Fréchet Error Bounds

• Use the proven gap bound between exact (Texact) and operational (TK) probabilities to quantify approximation error in real time.

• Capability: An AI can trade off computational cost vs. accuracy, reporting probability = 0.42 ± 0.03 with a formal guarantee, crucial for safety-critical decisions.

7. Backward Reachable Set Computation for Continuous Dynamics

• Apply the continuous immediate consequence operator to compute exact probability of reaching safe states in kinematic/dynamical systems.

• Capability: An AI can plan robot trajectories by computing the exact probability of safe arrival over continuous state spaces, expanding the safe set geometrically until convergence.

8. Hierarchical Continuous Signal Propagation

• Use the rooted-tree exact inference to propagate continuous signals (e.g., sensor noise) through multi-level networks with exact probabilities.

• Capability: An AI can perform fault diagnosis in sensor networks where each node's reading is a continuous random variable, yielding exact posterior probabilities for root causes.

9. Infinite-Domain Probabilistic Logic Programming

• Support stochastic predicates with unbounded continuous indices (e.g., "for all real x > 0") without finite grounding.

• Capability: An AI can reason about universal statements like the probability that the system survives for any positive time t as a single Lebesgue integral, not a sum over discrete time steps.

10. Structural Transparency for Explainable AI

• The least fixed point and differentiable operator provide a clear algebraic structure linking inputs (priors, rule probabilities) to outputs (query probabilities).

• Capability: An AI can explain its probabilistic conclusions by showing the exact integral decomposition and the contribution of each rule, enabling auditability in high-stakes domains.