Skip to content

Experiment plan #31

@isPANN

Description

@isPANN
  1. Technical Claims (Critical)

1.1 "Provably minimizing" Claim ✅ Fixed

  • Issue: Abstract claims "provably minimizing γ" but uses greedy heuristic for NP-hard problem
  • Status: Removed "provably", changed to "low-γ"

1.2 Measure Choice Lacks Theoretical Justification

  • Issue: Δρ(c) = "number of newly fixed variables" not justified as good measure for SAT
  • Issue: Table 3 shows NumUnfixedTensors (worse Knuth variance) outperforms NumUnfixedVars, undercutting "principled γ
    minimization" narrative
  • Suggestion: Honestly acknowledge γ is a heuristic guide, not a theoretical guarantee
  • Suggestion: Show correlation between γ and actual subtree sizes on small instances

1.3 Completeness/Soundness Argument Incomplete

  • Issue: How is completeness ensured when branches overlap or local configs are globally inconsistent?
  • Issue: Do you ensure disjointness or rely on propagation to eliminate redundancy?
  • Suggestion: Add proof sketch or more explicit discussion

  1. HOB/CnC Integration

2.1 Cutoff Condition is Ad Hoc

  • Issue: |σ_dec| × |σ_all| > θ has no rationale—why use product vs sum?
  • Issue: No experiments varying θ
  • Issue: No analysis of how OB-specific features (γ plateaus, region sizes) interact with cutoff
  • Suggestion: Add θ sensitivity experiment, or acknowledge borrowed heuristic from CnC

2.2 Cube Difficulty Shift Not Analyzed

  • Issue: Table 6 shows difficulty shifts from cubing to conquering (Avg. Dec. per cube rises)
  • Issue: No systematic analysis of this trade-off or optimization for end-to-end time
  • Suggestion: Analyze trade-off or acknowledge as limitation

  1. Experimental Evaluation

3.1 Narrow Benchmark Scope

  • Issue: Only factorization circuits (single domain) and tiny random 3-CNF (150 vars)
  • Missing: SAT competition benchmarks, graph-coloring, verification, cryptographic circuits
  • Suggestion: Add more benchmark types OR explicitly scope the contribution

3.2 Weak/Missing Baselines

  • Issue: DPLL is a straw man in 2025; NaiveBranch is custom weak baseline
  • Missing: AlphaMapleSAT, other advanced cubers, SAT+CAS pipelines, strong lookahead solvers
  • Suggestion: Add AlphaMapleSAT comparison or explain why not included

3.3 Wall-clock Performance ✅ Partially Addressed

  • Issue: OB often slower than CDCL (Table 5: 32-bit 0.61s vs 0.28s; 40-bit 11.37s vs 1.97s)
  • Status: Reframed to focus on conflicts; added amortization example (191350169: 8.5× speedup)
  • Remaining: Could add more analysis of when overhead is/isn't amortized

3.4 Single-Machine Evaluation

  • Issue: Only MacBook Air, no parallel CnC execution
  • Suggestion: Acknowledge as proof-of-concept OR add parallel experiments

  1. Algorithmic Details Insufficient for Reproduction

4.1 Region Selection Unclear

  • Issue: "Region" definition (graph neighborhood, max depth, max_tensors) buried in details
  • Issue: Algorithm 1 OBSolve has caption but no pseudocode
  • Suggestion: Add complete pseudocode; clarify region construction rules, tie-breaking

4.2 Tensor Construction Vague

  • Issue: How CNF/Tseitin gates map to tensors not clearly specified
  • Issue: Contraction order selection not explained
  • Suggestion: Add detailed specification

4.3 Greedy Merge Details Missing

  • Issue: Stopping condition, tie-breaking, branch weight reuse not spelled out
  • Suggestion: Add details in main text (Appendix Example 3 is too toy)

  1. Scalability Analysis Insufficient

5.1 Runtime Breakdown Missing

  • Issue: No quantitative analysis of time spent in contraction vs greedy merge vs propagation
  • Suggestion: Add profiling data for 40-bit and 44-bit instances

5.2 Region Size Behavior at Scale

  • Issue: No evidence that typical region sizes remain small on larger/denser instances
  • Issue: Real formulas can have large local treewidth
  • Suggestion: Analyze or acknowledge limitation

5.3 Memory Usage

  • Issue: No discussion of memory overhead from caching branch tables per region
  • Suggestion: Add memory analysis

  1. Missing Related Work

6.1 Citations/Discussions to Add
Work: AlphaMapleSAT (2401.13770)
Relevance: MCTS-based cubing; head-to-head comparison needed

Work: ImitSAT (2509.25411)
Relevance: ML-guided CDCL branching; potentially complementary

Work: Exact SAT algorithms (2101.08637, 2507.14504)
Relevance: Branching via nonstandard measures

Work: Shati et al. 2021
Relevance: SAT-based optimal decision trees

Work: Smirnov et al. 2021
Relevance: PB optimization via implicit hitting sets (related to set-cover view)

Work: Torralba 2018
Relevance: Dominance techniques for search pruning


  1. Presentation Issues

7.1 Table Formatting

  • Issue: Some tables have extraction artifacts (Table 1 headers/bit-length alignment)
  • Suggestion: Review and fix formatting

7.2 Figure 2(b)

  • Issue: "Pareto-optimal frontier" not rigorously defined; only one instance shown
  • Suggestion: Add aggregate plot across all instances of same bit-length

7.3 Solver Configuration

  • Issue: Not stated whether single-threaded, whether inprocessing disabled for fair comparison
  • Suggestion: Add solver configuration table for reproducibility

7.4 Eq. (3) Clarity

  • Issue: |σ_dec| × |σ_all| notation is opaque
  • Suggestion: Add sentence explaining why product (vs sum or just |σ_dec|)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions