Blame for research/code/test_coded_channels.py

#!/usr/bin/env python3
"""Exact tests of optimal positive-mask coordinate tomography.

Python 3.10+, standard library only. All ranks, sensors, reconstruction and
noise certificates use rational arithmetic. The mask-partition routine uses
the shortest exchange augmentation for representable matroid partition and
checks every returned partition explicitly. No novel matroid algorithm is
claimed. The manuscript attributes the partition criterion to Edmonds.
"""
from collections import deque
from fractions import Fraction as F
from itertools import product
from math import prod
import json
import random


def rank(vectors):
    """Rank of vectors, each represented as a sequence of rational numbers."""
    basis = {}
    for vector in vectors:
        row = list(map(F, vector))
        for pivot, prior in sorted(basis.items()):
            if row[pivot]:
                scalar = row[pivot]
                row = [a - scalar * b for a, b in zip(row, prior)]
        pivot = next((i for i, value in enumerate(row) if value), None)
        if pivot is not None:
            scalar = row[pivot]
            basis[pivot] = [value / scalar for value in row]
    return len(basis)


def independent_rows(rows):
    chosen = []
    previous = 0
    for i, row in enumerate(rows):
        current = rank([rows[j] for j in chosen] + [row])
        if current > previous:
            chosen.append(i)
            previous = current
    return chosen


def solve_linear(rows, rhs):
    """A consistent linear system, with all nonpivot variables set to zero."""
    assert rows and len(rows) == len(rhs)
    n = len(rows[0])
    matrix = [list(map(F, row)) + [F(value)] for row, value in zip(rows, rhs)]
    pivots = []
    lead = 0
    for col in range(n):
        selected = next((i for i in range(lead, len(matrix)) if matrix[i][col]), None)
        if selected is None:
            continue
        matrix[lead], matrix[selected] = matrix[selected], matrix[lead]
        scalar = matrix[lead][col]
        matrix[lead] = [value / scalar for value in matrix[lead]]
        for i in range(len(matrix)):
            if i != lead and matrix[i][col]:
                scalar = matrix[i][col]
                matrix[i] = [a - scalar * b for a, b in zip(matrix[i], matrix[lead])]
        pivots.append(col)
        lead += 1
        if lead == len(matrix):
            break
    for row in matrix:
        assert any(row[:n]) or row[n] == 0, 'inconsistent linear system'
    answer = [F(0)] * n
    for i, col in enumerate(pivots):
        answer[col] = matrix[i][-1]
    assert [sum(a * b for a, b in zip(row, answer)) for row in rows] == list(rhs)
    return answer


def inverse(matrix):
    n = len(matrix)
    assert all(len(row) == n for row in matrix)
    columns = [solve_linear(matrix, [F(i == j) for i in range(n)]) for j in range(n)]
    return [[columns[j][i] for j in range(n)] for i in range(n)]


def matvec(matrix, vector):
    return [sum((F(a) * b for a, b in zip(row, vector)), F(0)) for row in matrix]


def coordinate_operator(shape):
    """Every row sums a coordinate fiber; columns are labeled grid cells."""
    points = list(product(*(range(q) for q in shape)))
    rows, labels = [], []
    for axis in range(len(shape)):
        other_shape = [q for i, q in enumerate(shape) if i != axis]
        for kept in product(*(range(q) for q in other_shape)):
            rows.append([int(tuple(x for i, x in enumerate(point) if i != axis) == kept)
                         for point in points])
            labels.append((axis, kept))
    return points, rows, labels


def independent_partition(matrix, channels):
    """Shortest exchange augmentation; independent sets cover the columns."""
    count = len(matrix[0])
    columns = [tuple(row[j] for row in matrix) for j in range(count)]
    groups = [set() for _ in range(channels)]
    tests = 0

    def independent(indices):
        nonlocal tests
        tests += 1
        return rank([columns[j] for j in sorted(indices)]) == len(indices)

    for element in range(count):
        queue = deque([element])
        previous = {element: None}
        terminal = None
        while queue and terminal is None:
            current = queue.popleft()
            for color, group in enumerate(groups):
                if current in group:
                    continue
                if independent(group | {current}):
                    terminal = (current, color)
                    break
                for displaced in sorted(group):
                    if displaced not in previous and independent((group - {displaced}) | {current}):
                        previous[displaced] = (current, color)
                        queue.append(displaced)
        assert terminal is not None, 'requested channel count does not admit a partition'
        revised = [group.copy() for group in groups]
        current, color = terminal
        revised[color].add(current)
        while previous[current] is not None:
            predecessor, color = previous[current]
            revised[color].remove(current)
            revised[color].add(predecessor)
            current = predecessor
        assert set().union(*revised) == set(range(element + 1))
        assert sum(map(len, revised)) == element + 1
        assert all(independent(group) for group in revised)
        groups = revised

    # Enlarge the first independent set to a basis by moving elements from
    # the others. Removing an element cannot spoil their independence.
    full_rank = rank(columns)
    while len(groups[0]) < full_rank:
        move = next((element for group in groups[1:] for element in sorted(group)
                     if independent(groups[0] | {element})), None)
        assert move is not None
        source = next(j for j in range(1, channels) if move in groups[j])
        groups[source].remove(move)
        groups[0].add(move)
    assert all(independent(group) for group in groups)
    return [sorted(group) for group in groups], tests


def sensor_plan(matrix, groups, low=F(1, 4), high=F(1)):
    """Exactly N scalar rows, using one baseline and positive two-level masks."""
    assert 0 < low < high
    n = len(matrix[0])
    baseline_rows = independent_rows(matrix)
    selected = [baseline_rows]
    masks = [[F(1)] * n]
    operator = [[F(x) for x in matrix[i]] for i in baseline_rows]
    for group in groups[1:]:
        restricted = [[row[j] for j in group] for row in matrix]
        choices = independent_rows(restricted)
        assert len(choices) == len(group)
        selected.append(choices)
        mask = [high if j in group else low for j in range(n)]
        masks.append(mask)
        operator.extend([[F(a) * w for a, w in zip(matrix[i], mask)] for i in choices])
    assert len(operator) == n
    assert rank(operator) == n
    return {'selected': selected, 'masks': masks, 'operator': operator,
            'low': low, 'high': high}


def decode(matrix, groups, plan, observations):
    """Reconstruct from only the N selected scalar measurements."""
    n = len(matrix[0])
    rb = plan['selected'][0]
    r = len(rb)
    base_particular = solve_linear([matrix[i] for i in rb], observations[:r])
    full_baseline = matvec(matrix, base_particular)
    answer = [F(0)] * n
    pointer = r
    for group, choices in zip(groups[1:], plan['selected'][1:]):
        values = observations[pointer:pointer + len(choices)]
        pointer += len(choices)
        isolated = [(value - plan['low'] * full_baseline[i]) /
                    (plan['high'] - plan['low']) for value, i in zip(values, choices)]
        local = solve_linear([[matrix[i][j] for j in group] for i in choices], isolated)
        for j, value in zip(group, local):
            answer[j] = value
    assert pointer == n
    residual = [full_baseline[i] - sum(F(matrix[i][j]) * answer[j]
                for group in groups[1:] for j in group) for i in rb]
    first = solve_linear([[matrix[i][j] for j in groups[0]] for i in rb], residual)
    for j, value in zip(groups[0], first):
        answer[j] = value
    return answer


def norm_inf(matrix):
    return max(sum(abs(x) for x in row) for row in matrix)


def case(shape):
    points, matrix, row_labels = coordinate_operator(shape)
    n = prod(shape)
    blind = prod(q - 1 for q in shape)
    r = n - blind
    assert rank(matrix) == r
    channels = (n + r - 1) // r
    groups, oracle_tests = independent_partition(matrix, channels)
    assert len(groups[0]) == r
    assert sum(map(len, groups)) == n
    plan = sensor_plan(matrix, groups)
    rng = random.Random(1000 + n)
    recovered = 0
    for _ in range(8):
        z = [F(rng.randint(1, 80), rng.randint(1, 17)) for _ in points]
        observations = matvec(plan['operator'], z)
        assert decode(matrix, groups, plan, observations) == z
        recovered += 1

    # A certified linear-noise bound, with a worst-case perturbation attaining
    # it: pick the sign of the largest absolute-row-sum inverse row.
    inv = inverse(plan['operator'])
    gain = norm_inf(inv)
    z = [F(i + 2, 7) for i in range(n)]
    epsilon = min(z) / (10 * gain)
    worst_row = max(range(n), key=lambda j: sum(abs(x) for x in inv[j]))
    error = [epsilon * (1 if x > 0 else -1 if x < 0 else 0) for x in inv[worst_row]]
    observations = matvec(plan['operator'], z)
    noisy = [a + b for a, b in zip(observations, error)]
    estimate = decode(matrix, groups, plan, noisy)
    actual_error = max(abs(a - b) for a, b in zip(estimate, z))
    assert actual_error == gain * epsilon
    assert min(estimate) > 0

    # Fewer channels cannot be injective, independently of how masks are
    # chosen. Sample several positive masks to check the rank obstruction.
    lower_rank_tests = []
    if channels > 1:
        assert (channels - 1) * r < n
        for seed in range(3):
            rg = random.Random(seed)
            stacked = []
            for j in range(channels - 1):
                weights = [F(1) if j == 0 else F(rg.randint(1, 9)) for _ in points]
                stacked.extend([[F(a) * w for a, w in zip(row, weights)] for row in matrix])
            observed_rank = rank(stacked)
            assert observed_rank <= (channels - 1) * r < n
            lower_rank_tests.append(observed_rank)

    all_temperature_checks = 0
    if len(set(shape)) == 1 and len(shape) >= 2:
        q = shape[0]
        z0 = [F(2 if sum(point) % q == 0 else 5) for point in points]
        z1 = [F(2 if sum(point) % q == 1 else 5) for point in points]
        assert z0 != z1
        for row in matrix:
            values0 = sorted(z0[j] for j, a in enumerate(row) if a)
            values1 = sorted(z1[j] for j, a in enumerate(row) if a)
            assert values0 == values1
            for power in range(1, q + 3):
                assert sum(x ** power for x in values0) == sum(x ** power for x in values1)
                all_temperature_checks += 1
        for z in (z0, z1):
            assert decode(matrix, groups, plan, matvec(plan['operator'], z)) == z

    return {'shape': list(shape), 'cells': n, 'unmasked_rank': r,
            'blind_dimension': blind, 'optimal_channels': channels,
            'partition_sizes': list(map(len, groups)),
            'partition': groups, 'partition_independence_oracle_calls': oracle_tests,
            'scalar_measurements': len(plan['operator']),
            'selected_fiber_rows': plan['selected'],
            'positive_mask_levels': [str(plan['low']), str(plan['high'])],
            'selected_operator_rank': rank(plan['operator']),
            'exact_rational_recoveries': recovered,
            'inverse_inf_norm': str(gain),
            'condition_number_inf': str(norm_inf(plan['operator']) * gain),
            'sharp_noise_certificate': True,
            'fewer_channel_sample_ranks': lower_rank_tests,
            'unmodulated_multiset_power_equalities': all_temperature_checks,
            'coded_recovery_of_both_ambiguous_fields': True if all_temperature_checks else None}


def main():
    cases = [case(shape) for shape in ((3, 3), (4, 4), (3, 3, 3), (2, 3, 4))]
    cases.extend(case((2,) * dim) for dim in range(1, 6))
    return {'status': 'PASS', 'arithmetic': 'exact integers and fractions',
            'cases': cases,
            'scope': ['Mask channels interrogate the same fixed positive field.',
                      'Masks are calibrated, externally supplied additional observations.',
                      'Optimality is within the stated linear exponential-domain sensor model.',
                      'The matroid partition criterion is established mathematics.',
                      'No practical image or neural-network superiority is tested.']}


if __name__ == '__main__':
    print(json.dumps(main(), indent=2))