Blame for vector_realization.py

"""Explicit D4 realization on the unit square; no affine-box search.

The fundamental boundary sector is q(t)=(1/2,t), 0<t<1/2.
For every bin of t choose one sign.  A D4 image is selected precisely
when its determinant times the bin sign is positive.  All corner and
edge-midpoint points are added.  Reflection in any normal axis with
angle 0, pi/4, pi/2, or 3pi/4 reverses determinant and preserves the bin,
so every line in those four directional audits is hit exactly.

Each open fundamental bin has eight images, exactly four selected;
the barrier length is therefore exactly 2.  The code evaluates finite
formulas in float64.  Sampling-based line statistics are diagnostics,
not opacity proofs or certified continuum error bounds.
"""
from __future__ import annotations

import math
from typing import Mapping

import numpy as np

AUDIT_NORMALS = np.arange(4, dtype=float) * (math.pi / 4)
FIXED_POINTS = np.array([
    [-.5, -.5], [-.5, 0.], [-.5, .5], [0., -.5],
    [0., .5], [.5, -.5], [.5, 0.], [.5, .5],
])

# h=x*y*(x*x-y*y), r2=x*x+y*y.  The last three positive costs
# distinguish the two determinant classes.  Lower-degree moments alone
# are exactly balanced by D4 symmetry and would be insufficient tests.
POLYNOMIAL_TESTS = {
    "mass": {(0, 0): 1.},
    "x": {(1, 0): 1.},
    "y": {(0, 1): 1.},
    "x_squared": {(2, 0): 1.},
    "xy": {(1, 1): 1.},
    "1+32h": {(0, 0): 1., (3, 1): 32., (1, 3): -32.},
    "1+64h_r2": {(0, 0): 1., (5, 1): 64., (1, 5): -64.},
    "1+128h_r4": {(0, 0): 1., (7, 1): 128., (5, 3): 128.,
                    (3, 5): -128., (1, 7): -128.},
}
NONTRIVIAL_TESTS = ("1+32h", "1+64h_r2", "1+128h_r4")


def _validate_bits(bits):
    b = np.asarray(bits, dtype=np.int8)
    if b.ndim != 1 or not len(b) or not np.all((b == -1) | (b == 1)):
        raise ValueError("bits must be a nonempty one-dimensional +/-1 array")
    return b


def group_sectors(bin_count):
    """Return all eight D4 images, shape (8,M,2,2), and determinants."""
    if int(bin_count) != bin_count or bin_count < 1:
        raise ValueError("bin_count must be a positive integer")
    m = int(bin_count)
    cuts = np.linspace(0., .5, m + 1)
    sectors, determinants = [], []
    for vertical in (True, False):
        for sx in (-1, 1):
            for sy in (-1, 1):
                segments = np.empty((m, 2, 2), dtype=float)
                if vertical:
                    segments[:, :, 0] = sx * .5
                    segments[:, 0, 1] = sy * cuts[:-1]
                    segments[:, 1, 1] = sy * cuts[1:]
                    determinant = sx * sy
                else:
                    segments[:, 0, 0] = sx * cuts[:-1]
                    segments[:, 1, 0] = sx * cuts[1:]
                    segments[:, :, 1] = sy * .5
                    determinant = -sx * sy
                sectors.append(segments)
                determinants.append(determinant)
    return np.asarray(sectors), np.asarray(determinants, dtype=np.int8)


def square_barrier(bits, include_fixed=True):
    """Construct all selected closed segments; fixed points are degenerate pieces."""
    b = _validate_bits(bits)
    sectors, determinant = group_sectors(len(b))
    selected = sectors[determinant[:, None] * b[None, :] > 0]
    if include_fixed:
        selected = np.concatenate((selected, np.repeat(FIXED_POINTS[:, None, :], 2, axis=1)))
    return selected


def segment_lengths(segments):
    s = np.asarray(segments, dtype=float)
    return np.linalg.norm(s[:, 1] - s[:, 0], axis=1)


def integrate_monomial(segments, x_power, y_power):
    """Exact axis-aligned monomial integrals, evaluated in float64, per segment."""
    s = np.asarray(segments, dtype=float)
    if s.shape[-2:] != (2, 2):
        raise ValueError("segments must end in dimensions (2,2)")
    i, j = int(x_power), int(y_power)
    if i < 0 or j < 0:
        raise ValueError("monomial powers must be nonnegative")
    vertical = np.abs(s[..., 1, 0] - s[..., 0, 0]) < 1e-14
    horizontal = np.abs(s[..., 1, 1] - s[..., 0, 1]) < 1e-14
    if not np.all(vertical | horizontal):
        raise ValueError("this exact integrator accepts only axis-aligned pieces")
    xlo = np.minimum(s[..., 0, 0], s[..., 1, 0])
    xhi = np.maximum(s[..., 0, 0], s[..., 1, 0])
    ylo = np.minimum(s[..., 0, 1], s[..., 1, 1])
    yhi = np.maximum(s[..., 0, 1], s[..., 1, 1])
    v = s[..., 0, 0] ** i * (yhi ** (j + 1) - ylo ** (j + 1)) / (j + 1)
    h = s[..., 0, 1] ** j * (xhi ** (i + 1) - xlo ** (i + 1)) / (i + 1)
    return np.where(vertical, v, h)


def integrate_polynomial(segments, polynomial: Mapping):
    s = np.asarray(segments, dtype=float)
    out = np.zeros(s.shape[:-2], dtype=float)
    for (i, j), coefficient in polynomial.items():
        out += coefficient * integrate_monomial(s, i, j)
    return out


def cost_model(bin_count, polynomial):
    """Cost = half_full_cost + dot(bits, coefficients), exactly as finite formulas."""
    sectors, determinant = group_sectors(bin_count)
    integrals = integrate_polynomial(sectors, polynomial)
    full_cost = float(integrals.sum())
    coefficients = .5 * (determinant[:, None] * integrals).sum(axis=0)
    return {
        "full_cost": full_cost,
        "half_full_cost": full_cost / 2,
        "coefficients": coefficients,
        "independent_bit_variance": float(coefficients @ coefficients),
    }


def choose_balanced_bits(bin_count, seed, candidates=32):
    """Reproducible finite search; this is not an optimizer or a monotonicity guarantee."""
    if candidates < 1:
        raise ValueError("candidates must be positive")
    rng = np.random.default_rng(seed)
    possibilities = 2 * rng.integers(0, 2, size=(candidates, bin_count), dtype=np.int8) - 1
    coefficient = np.stack([cost_model(bin_count, POLYNOMIAL_TESTS[n])["coefficients"]
                            for n in NONTRIVIAL_TESTS])
    deviations = possibilities @ coefficient.T
    score = np.max(np.abs(deviations), axis=1)
    best = int(np.argmin(score))
    return possibilities[best], {
        "seed": int(seed), "candidates": int(candidates),
        "selected_candidate_index": best, "selected_max_cost_error": float(score[best]),
        "selection_rule": "minimum maximum absolute error over the three nontrivial polynomial costs",
    }


def projection_union(segments, normal, tolerance=1e-13):
    """Merge actual segment projections; return components and missed square length."""
    s = np.asarray(segments, dtype=float)
    n = np.asarray(normal, dtype=float)
    n = n / np.linalg.norm(n)
    values = s @ n
    intervals = np.sort(values, axis=1)
    intervals = intervals[np.argsort(intervals[:, 0], kind="stable")]
    merged = []
    for lo, hi in intervals:
        if merged and lo <= merged[-1][1] + tolerance:
            merged[-1][1] = max(merged[-1][1], float(hi))
        else:
            merged.append([float(lo), float(hi)])
    half_width = .5 * float(np.abs(n).sum())
    cursor = -half_width
    gaps = []
    for lo, hi in merged:
        lo, hi = max(lo, -half_width), min(hi, half_width)
        if hi < -half_width or lo > half_width:
            continue
        if lo > cursor + tolerance:
            gaps.append([cursor, lo])
        cursor = max(cursor, hi)
    if cursor < half_width - tolerance:
        gaps.append([cursor, half_width])
    return {"components": merged, "target": [-half_width, half_width], "gaps": gaps,
            "uncovered_length": float(sum(b-a for a,b in gaps))}


def square_line_sample(power=19, seed=20261029):
    """Fixed scrambled Sobol sample, rejected to the square line cylinder.

    Sampling diagnostics use the accepted low-discrepancy points.  They
    are not independent Monte Carlo samples, and no IID confidence band
    is assigned to them.
    """
    from scipy.stats import qmc
    points = qmc.Sobol(d=2, scramble=True, seed=seed).random_base2(power)
    theta = math.pi * points[:, 0]
    offset = math.sqrt(2.) * (points[:, 1] - .5)
    normals = np.column_stack((np.cos(theta), np.sin(theta)))
    half_width = .5 * np.abs(normals).sum(axis=1)
    accepted = np.abs(offset) < half_width
    theta, offset, normals = theta[accepted], offset[accepted], normals[accepted]
    direction = np.column_stack((-normals[:, 1], normals[:, 0]))
    origin = offset[:, None] * normals
    lower = (-.5 - origin) / direction
    upper = (.5 - origin) / direction
    lo = np.minimum(lower, upper).max(axis=1)
    hi = np.maximum(lower, upper).min(axis=1)
    endpoints = np.stack((origin + lo[:, None]*direction,
                          origin + hi[:, None]*direction), axis=1)
    rectangle_area = math.pi * math.sqrt(2.)
    return {
        "theta": theta, "offset": offset, "endpoints": endpoints,
        "input_count": int(2**power), "accepted_count": int(accepted.sum()),
        "seed": int(seed), "sobol_power": int(power),
        "estimated_cylinder_area": float(rectangle_area * accepted.mean()),
        "exact_cylinder_area": 4.,
    }


def endpoint_labels(endpoints, bin_count, fixed_tolerance=2e-13):
    """D4 bin and orientation character, independent of the segment construction."""
    e = np.asarray(endpoints, dtype=float)
    absolute = np.abs(e)
    t = absolute.min(axis=-1)
    cell = np.clip(np.floor(2 * bin_count * t).astype(np.int64), 0, bin_count - 1)
    vertical = absolute[..., 0] >= absolute[..., 1]
    sx = np.where(e[..., 0] >= 0, 1, -1)
    sy = np.where(e[..., 1] >= 0, 1, -1)
    determinant = sx * sy * np.where(vertical, 1, -1)
    fixed = (t < fixed_tolerance) | (np.abs(absolute[..., 0]-absolute[..., 1]) < fixed_tolerance)
    return cell, determinant.astype(np.int8), fixed


def selected_endpoints(endpoints, bits):
    b = _validate_bits(bits)
    cell, determinant, fixed = endpoint_labels(endpoints, len(b))
    return (determinant * b[cell] > 0) | fixed


def line_hit_counts(endpoints, bits):
    return selected_endpoints(endpoints, bits).sum(axis=-1)


def line_weights(sample):
    theta, offset = sample["theta"], sample["offset"]
    return {
        "constant": np.ones(len(theta)),
        "1+0.5cos2theta": 1 + .5 * np.cos(2*theta),
        "1+0.5sin2theta": 1 + .5 * np.sin(2*theta),
        "1+2offset_squared": 1 + 2 * offset**2,
    }


def line_diagnostics(sample, bits):
    counts = line_hit_counts(sample["endpoints"], bits)
    weights = line_weights(sample)
    weighted = {name: [float(np.sum(w * (counts == k)) / np.sum(w)) for k in range(3)]
                for name, w in weights.items()}
    cell, _, _ = endpoint_labels(sample["endpoints"], len(bits))
    target = np.array([.25, .5, .25])
    return {
        "hit_fractions": weighted["constant"], "weighted_hit_fractions": weighted,
        "largest_weighted_hit_error": float(max(np.max(np.abs(np.array(v)-target))
                                                  for v in weighted.values())),
        "same_bin_endpoint_fraction": float(np.mean(cell[:, 0] == cell[:, 1])),
        "sampled_counts": [int(np.count_nonzero(counts == k)) for k in range(3)],
    }


def point_membership_from_segments(points, segments, tolerance=2e-12):
    """Independent geometric membership for a modest batch of boundary points."""
    points = np.asarray(points, dtype=float).reshape(-1, 2)
    s = np.asarray(segments, dtype=float)
    out = np.zeros(len(points), dtype=bool)
    for a, b in s:
        lower, upper = np.minimum(a, b), np.maximum(a, b)
        out |= np.all((points >= lower-tolerance) & (points <= upper+tolerance), axis=1)
    return out