omega-tensor / README.md
phanerozoic's picture
Document init_state_tensor.parquet and add its config
e23008a verified
metadata
configs:
  - config_name: default
    data_files: omega_tensor.parquet
  - config_name: init_state
    data_files: init_state_tensor.parquet
dataset_info:
  - config_name: default
    features:
      - name: op1
        dtype: int64
      - name: op2
        dtype: int64
      - name: op3
        dtype: int64
      - name: op4
        dtype: int64
      - name: op1_name
        dtype: string
      - name: op2_name
        dtype: string
      - name: op3_name
        dtype: string
      - name: op4_name
        dtype: string
      - name: halting_count
        dtype: int64
      - name: omega
        dtype: float64
      - name: has_swp
        dtype: bool
      - name: has_mov
        dtype: bool
      - name: has_and
        dtype: bool
      - name: has_shr
        dtype: bool
      - name: involution_count
        dtype: int64
      - name: coupling
        dtype: string
      - name: avg_cycle
        dtype: float64
      - name: max_cycle
        dtype: int64
      - name: entropy_bits
        dtype: float64
      - name: reachable_outputs
        dtype: int64
    splits:
      - name: train
        num_examples: 38416
  - config_name: init_state
    features:
      - name: op1
        dtype: int8
      - name: op2
        dtype: int8
      - name: op3
        dtype: int8
      - name: op4
        dtype: int8
    splits:
      - name: train
        num_examples: 38416
license: mit
task_categories:
  - other
tags:
  - halting-problem
  - computability
  - algorithmic-information-theory
  - exhaustive-enumeration
  - FPGA
size_categories:
  - 1M<n<10M

Omega Tensor: Exact Halting Fractions for 38,416 Instruction Sets

Exhaustive enumeration of all 1,679,616 programs of length 8 for every possible instruction set on a two-register, 8-bit, bounded machine with JNZ-to-zero control flow. Each row is one ISA. Each halting count is exact.

The repository contains two tables:

  • omega_tensor.parquet (38,416 rows) — canonical single-initial-state table used in the paper, with the halting fraction at A=0, B=0 and cross-observable columns (cycle length, output entropy, reachability, structural flags).
  • init_state_tensor.parquet (38,416 rows, 260 columns) — extends the above to every initial A value in 0..255 at B=0. Halting counts only.

Contents — omega_tensor.parquet

38,416 rows. Each row represents one ordered ISA defined by four opcode assignments (opcodes 1-4) drawn from 14 primitive operations. Opcode 0 is always INC. Opcode 5 is always JNZ.

Column Type Description
op1-op4 int Operation IDs for opcodes 1-4
op1_name-op4_name str Human-readable operation names
halting_count int Exact count of halting programs (out of 1,679,616)
omega float halting_count / 1,679,616
has_swp, has_mov, has_and, has_shr bool Structural flags
involution_count int Number of involution operations in the ISA
coupling str Register coupling type: bidirectional, one-way, or none
avg_cycle float Mean cycle length among non-halting programs
max_cycle int Maximum cycle length among non-halting programs
entropy_bits float Shannon entropy of the output distribution (bits)
reachable_outputs int Number of distinct outputs produced by halting programs

Statistics — omega_tensor.parquet

  • Range: [0.290710, 0.816916]
  • Mean: 0.612418
  • Distinct values (6 decimal places): 1,148
  • Total programs executed: 64,527,984,256 (38,416 x 1,679,616)
  • Avg cycle length range: [23.4, 216.1] (mean 116.2)
  • Max cycle length: 255 for all ISAs
  • Output entropy range: [0.80, 5.44] bits (mean 2.50)
  • Reachable outputs: [9, 211] / 256 (mean 53.7)
  • Outputs reachable by ALL ISAs: 9

Extremes

  • Minimum (0.2907): INC INC INC INC INC JNZ. Five copies of INC. 71% of programs loop forever.
  • Maximum (0.8169): INC AND AND AND AND JNZ. Four copies of AND. Only 18% loop.

Method

Each ISA was evaluated by exhaustive enumeration of all 6^8 = 1,679,616 programs of length 8. Each program was executed from initial state (A=0, B=0, PC=0) for up to 256 steps. Programs that reached PC >= 8 were counted as halting. The computation used a Numba-JIT-compiled Python interpreter at 0.26 seconds per ISA (approximately 2.8 hours total for all 38,416 ISAs).

Every halting count is exact. No sampling, no approximation, no estimation.

Verification

Known ISA values verified against independent FPGA silicon measurements and CPU cross-checks:

  • ISA-A (DEC SWP ADD XOR): 978,929 halting (dataset) = 978,929 (FPGA) = 978,929 (CPU)
  • ISA-B (NEG MOV SUB AND): 1,235,107 = 1,235,107 = 1,235,107
  • ISA-G (SHR MOV SUB AND): 1,248,490 = 1,248,490 = 1,248,490

Contents — init_state_tensor.parquet

38,416 rows, one per ISA (same ISA space as omega_tensor.parquet). 260 columns: four opcode slots plus 256 halting counts, one per initial A value.

Column Type Description
op1-op4 int8 Operation IDs for opcodes 1-4
h_000 through h_255 int32 Halting counts at initial state (A=N, B=0) for N=0..255

Join on (op1, op2, op3, op4) to recover names, cycle length, entropy, and structural columns from omega_tensor.parquet. The canonical value omega from that table equals h_000 / 1,679,616.

Statistics — init_state_tensor.parquet

  • Total cells: 38,416 x 256 = 9,834,496 exact halting counts
  • Total program evaluations: ~16.5 billion
  • Omega(init=0) mean: 0.612418 (matches omega_tensor.parquet)
  • Omega(init=128) mean: 0.366431, range [0.236044, 0.681003]
  • Ratio Omega(0)/Omega(128): mean 1.7429, median 1.6430, range [0.8336, 3.1466]
  • 99.92% of ISAs have Omega(0) > Omega(128); 0.08% have the inverse (INC-dominant ISAs)
  • 7.91% of ISAs have a perfectly reflection-symmetric bowl (Omega(a) = Omega(256-a))
  • No ISA has its Omega minimum at init_a = 128 exactly; 3.23% have argmin in [126, 130]
  • Bowl depth Omega(0) - Omega(min) ranges [0.058, 0.513], mean 0.260

Method — init_state_tensor.parquet

Same interpreter and enumeration protocol as the canonical tensor, extended over all 256 values of the initial accumulator A (B held at 0). 9,834,496 (ISA, init_a) cells, each the exact halting count over the full 1,679,616-program space. Computed with the same Numba-JIT interpreter under an 18-worker multiprocessing pool, 110.3 wall hours, zero errors. Every count is exact.

Cross-validation: for every ISA, the h_000 column equals the canonical halting_count from omega_tensor.parquet.

Citation

Based on: The Finite Halting Fraction Across Machine Models