Quantum-Logic Tsetlin Machines: Interpretable Quantum Machine Learning with Commuting Projector Clauses

AI summary

Overview

Research area: Quantum machine learning, specifically interpretable quantum-logic learning built on the Tsetlin Machine framework.

Technical level: Intermediate. The paper assumes familiarity with density operators, projectors, Pauli observables, and the basic Tsetlin Machine clause/automaton structure, but the core idea is explained cleanly enough for a reader with a standard quantum-information background.

Scope: The paper defines and tests a hybrid model — the Quantum-Logic Tsetlin Machine (QL-TM) — in which classical Tsetlin automata learn clauses whose literals are quantum projectors inside commuting measurement contexts, evaluated on Bell states, phase-flip syndromes, randomized stabilizer tasks, mixed literal pools, context-budget ablations, and finite-shot noisy simulations.

What This Paper Is About

Tsetlin Machines classify by learning human-readable clauses over Boolean literals, but no clear definition exists for what such a machine should learn when its input is a quantum state rather than a bitstring. The author sidesteps the difficulty of quantizing the automata themselves (whose reward/penalty updates are irreversible) and instead quantizes the logic: literals become projectors, clauses become products of commuting projectors, and a clause's activation becomes a Born probability Tr(ρ C_j) in [0,1] rather than a binary truth value. The goal is a mathematically controlled, interpretable bridge between Tsetlin clause learning and quantum logic — explicitly not a claim of quantum advantage.

Key Contributions

1. Projector clauses with Born-probability activations. Defines projector literals (P_i, its complement I − P_i, or excluded), an explicit commuting-context restriction, clause products C_j = ∏ L_ji, and clause activation c_j(ρ) = Tr(ρ C_j), with multiclass scores S_y(ρ) = Σ_j w_yj c_j(ρ) using weights in {−1, +1}.
2. A classical-reduction theorem. Proves (Theorem 1) that when all base projectors are diagonal in the computational basis, every QL-TM clause activation on a basis state equals the corresponding Boolean conjunction, so the model reduces exactly to an ordinary Boolean Tsetlin Machine.
3. Pauli-stabilizer clause semantics. Shows (Proposition 1) that a clause of the form ∏ (I + s_i g_i)/2 over commuting Pauli operators projects onto the joint eigenspace with eigenvalues s_i, so learned clauses can be read as stabilizer/syndrome propositions.
4. A controlled experimental programme. Reports contextual expressivity, randomized 16-class stabilizer generalization, recovery of true generators from mixed pools with distractors, a context-budget degradation ladder matching a predicted 2^(b−k) trend, and finite-shot/noise robustness, using two clause learners plus a nearest-prototype baseline and a ridge one-vs-rest classifier.

Main Findings

• Diagonal contexts lose phase information; QL contexts do not. On Bell and phase-flip tasks in the exact/noiseless setting, QL projector contexts reach 1.000 ± 0.000 mean accuracy for all four models (ProjectorClauseMiner, TinyProjectorTsetlin, RidgeOneVsRest, NearestPrototype). Diagonal Z-only contexts collapse: Bell accuracy is 0.500 ± 0.055 for ProjectorClauseMiner, TinyProjectorTsetlin, and NearestPrototype, and 0.449 ± 0.032 for RidgeOneVsRest; phase-flip accuracy is 0.239 ± 0.028 for three of the models and 0.201 ± 0.020 for RidgeOneVsRest.

• Learned clauses match known physics. For Bell states the learners recover P_ZZ+ ∧ P_XX+, P_ZZ+ ∧ P_XX−, P_ZZ− ∧ P_XX+, and P_ZZ− ∧ P_XX− for |Φ+⟩, |Φ−⟩, |Ψ+⟩, |Ψ−⟩ respectively. For phase-flip syndromes they recover P_{X1X2}+ ∧ P_{X2X3}+ (None), P_{X1X2}− ∧ P_{X2X3}+ (Z1), P_{X1X2}− ∧ P_{X2X3}− (Z2), and P_{X1X2}+ ∧ P_{X2X3}− (Z3). Diagonal clauses either merge Bell phase pairs or degenerate to TRUE for all phase-flip classes.

• Randomized 16-class stabilizer tasks are solved only by the true context. With k = 4 generators (16 classes, chance 0.0625), the true QL context reaches 1.000 ± 0.000 for both n = 5 and n = 6, using 4.00 literals per clause. Diagonal Z contexts reach 0.150 ± 0.052 (n = 5) and 0.130 ± 0.064 (n = 6); wrong non-diagonal QL contexts stay at 0.062 ± 0.006 for both, with 0.00 literals per clause.

• Mixed pools preserve accuracy and recover all true generators. With a pool containing true literals, wrong-context literals, diagonal Z literals, and 32 random Pauli-observable distractors, the mixed pool reaches 1.000 ± 0.000 accuracy with true-generator coverage 1.00 for both n = 5 and n = 6, using 7.00 and 7.83 selected literals respectively (versus 4.00 with true literals only). Removing the true literals drops accuracy to 0.419 ± 0.427 (n = 5) and 0.351 ± 0.331 (n = 6) with coverage 0.00.

• Post-hoc pruning compresses clauses without loss. Removing redundant non-true literals while keeping the true generators compresses clauses back to the k = 4 stabilizer form with unchanged exact/noiseless accuracy.

• Context-budget degradation follows the predicted ladder. With only b of four true generators available, expected accuracy 2^(b−4) is matched closely for both n = 5 and n = 6: b = 0 gives 0.049 ± 0.011 (expected 0.0625), b = 1 gives 0.113 ± 0.013 (expected 0.1250), b = 2 gives 0.240 ± 0.023 (expected 0.2500), b = 3 gives 0.495 ± 0.027 (expected 0.5000), and b = 4 gives 1.000 ± 0.000 (expected 1.0000).

• Noise tolerance is dominated by shot count, not depolarizing strength. In the harshest plotted slice (depolarizing p = 0.6, coherent rotation scale 0.2, readout noise 0.02), TinyProjectorTsetlin reaches 0.973 for Bell states and 0.944 for phase-flip syndromes at only 16 shots; by 64 shots both tasks recover 1.000. Exact depolarizing noise mainly shrinks confidence without reordering the stabilizer syndrome, so degradation appears when finite-shot and readout uncertainty make estimated projector probabilities ambiguous.

• Explicit non-claims. The paper states it does not claim quantum advantage, predictive superiority over variational quantum classifiers or quantum kernels, or state-of-the-art end-to-end QML performance.

Methodology in Plain English

The author keeps the learner classical and changes only what the rules are made of. Instead of asking "is bit x_i equal to 1?", each literal asks a quantum yes/no question of the form "is the state inside this subspace?", represented by a projector. Because arbitrary quantum questions do not combine like ordinary AND (projectors may fail to commute), each clause is confined to a set of projectors that do commute — a "commuting context" — so multiplying them yields a genuine projector with well-defined conjunction semantics. Activating a clause then means computing the probability that the state passes all of that clause's questions, Tr(ρ C_j), which is a number between 0 and 1 used as a soft vote.

Learning uses a deliberately minimal Tsetlin-style rule. Each literal has a finite-state automaton with 2N states; the literal is included when the automaton state exceeds N. Target-class clauses get Type-I-like feedback (reward high-probability literals toward the include region); firing non-target clauses get Type-II-like feedback (search for a low-probability literal that suppresses the clause and reward including it); opposite-sign literals of the same observable are resolved after each update by keeping the stronger inclusion state. Training uses a product-of-marginals surrogate, ∏ Tr(ρ L_i), while prediction uses the true joint activation Tr(ρ C_j); the author notes these differ in general, but coincide in ordering for the deterministic stabilizer eigenstate tasks where relevant probabilities lie in {0, 1}.

Experiments compare ProjectorClauseMiner (a transparent margin-based clause learner), TinyProjectorTsetlin (the automata prototype), a nearest-prototype baseline, and a dependency-free ridge one-vs-rest classifier, using a random stratified 70/30 split. Fixed settings include N = 16, threshold τ = 0.55, 8 epochs for context/noise runs and 10 epochs for random/mixed stabilizer runs, 80 samples per class with 10 seeds for Bell/phase-flip feature baselines, and 6 random tasks × 3 seeds with 50 (random stabilizer) or 60 (mixed pool, context budget) samples per class. Noise sweeps cover depolarizing p ∈ {0, 0.05, 0.1, 0.2, 0.4, 0.6}, shots ∈ {exact, 16, 32, 64, 128, 256}, coherent rotation scale ∈ {0, 0.1, 0.2, 0.3}, and readout noise ∈ {0, 0.02, 0.05}. Implementation uses NumPy, pandas, and matplotlib with no dependency on TMU or pyTsetlinMachine. The author notes hyperparameters were fixed within experiment families and not tuned against test sets, and that no formal paired significance testing is reported.

Why This Matters

The paper offers a precise, provable way to lift a symbolic, inspectable learning model into quantum state space, which is unusual: most QML models learn continuous circuit parameters or implicit kernel similarities rather than explicit logical propositions. The diagonal-reduction theorem shows the classical Tsetlin Machine is a special case rather than something discarded, and the stabilizer correspondence gives learned clauses a direct physical reading as syndrome patterns. The context-budget ladder is the strongest piece of evidence, because the hierarchy is fixed by available stabilizer information before any learning occurs — which argues that performance tracks meaningful quantum propositions rather than raw feature dimensionality.

Real-world applications implied by the framework:

• Quantum error correction and syndrome decoding — learned clauses read directly as stabilizer syndrome propositions over commuting Pauli projectors.
• Interpretable diagnostics for quantum hardware — rules stated as projector conjunctions are human-auditable, unlike dense parameter vectors.
• State certification and verification — clauses act as explicit joint-eigenspace tests on prepared states such as Bell and GHZ-type states.
• Hybrid quantum-classical pipelines where measurement budget is the constraint — each clause sits in one commuting context, so its literals are estimable from compatible measurement counts, suggesting a measurement-cost objective.

Industry relevance: the work targets settings where interpretability, discrete rule learning, and hardware-near implementation matter — the original motivations cited for Tsetlin Machines — and where quantum teams need auditable classifiers rather than black-box circuits. The author is affiliated with QuantumAI Lab, Fractal Analytics, Mumbai, and code is released publicly.

Future Directions

• Direct joint-context learning. The product-of-marginals quantity is only a training surrogate; the paper states exact joint-projector feedback and joint-context estimation for finite-shot, noisy, or general mixed states remain unvalidated beyond the deterministic stabilizer setting.
• Native sparsity and context selection. Mixed-pool clauses retained redundant distractors (7.00 and 7.83 literals versus 4.00 true), so explicit sparsification and multi-clause learning with context selection are flagged as needed for a full-scale implementation.
• Hyperparameter and significance studies. A systematic sensitivity analysis over N, τ, clause multiplicity, and feedback schedules is stated as future work, as is formal paired significance testing.
• Hardware-grounded and broader QML evaluation. The study is entirely simulated and does not model calibration drift or device-specific measurement constraints; comparisons on less structured state-classification tasks against variational quantum classifiers and quantum kernels are left open.

Target Audience

Researchers working on interpretable or logic-based machine learning who want to extend symbolic models into quantum settings; quantum-information scientists interested in stabilizer formalism, syndrome semantics, and error-correction-adjacent learning; and QML practitioners who need classifiers whose learned rules can be inspected as explicit quantum propositions. It is also useful for readers evaluating claims about "quantum advantage," since the paper is explicit about what its controlled benchmarks do and do not establish.

Authors’ abstract

Tsetlin Machines (TMs) learn interpretable Boolean clauses using finite-state automata. We introduce the Quantum-Logic Tsetlin Machine (QL-TM), which replaces Boolean literals with quantum propositions represented by projectors while retaining classical include/exclude automata. Clauses are restricted to commuting measurement contexts and activate through the Born probability of their joint projector. We prove an exact reduction to ordinary Boolean TM clauses in diagonal computational-basis contexts and connect Pauli-projector clauses to stabilizer and syndrome semantics. Controlled experiments on Bell states, phase-flip syndromes, randomized 16-class stabilizer tasks, mixed literal pools, context-budget ablations, and finite-shot noise show that correct non-diagonal contexts recover physically meaningful clauses, while diagonal or wrong contexts lose the relevant phase/syndrome information. The context-budget results closely follow the predicted separability ladder 2^(b-k) as true stabilizer generators are removed. The contribution is a controlled bridge between Tsetlin clause learning and quantum logic, not a claim of quantum advantage.