Blame for reference_core.py

"""Neyaru minimal exact core. Python 3.10+, standard library only.
Run: python3 reference_core.py
The benchmark implementation is separate and reports full overhead.
"""
from dataclasses import dataclass
from math import inf


class Dihedral:
    """Elements r**a s**b encoded a + m*b; group order is 2*m."""
    def __init__(self, m):
        self.m, self.n = m, 2*m
        self.mul = [[((a % m) + (1 if a < m else -1)*(b % m)) % m
                     + m*((a // m) ^ (b // m))
                     for b in range(2*m)] for a in range(2*m)]
        self.inv = [next(b for b in range(self.n)
                         if self.mul[a][b] == 0) for a in range(self.n)]

    def closure(self, generators):
        found, queue = {0}, [0]
        for a in queue:
            for b in generators:
                c = self.mul[a][b]
                if c not in found:
                    found.add(c)
                    queue.append(c)
        return found


def stabilizer(G, values, left):
    """Rare-value filter, exact tests, and certified subgroup closure."""
    buckets = {}
    for x, v in enumerate(values):
        buckets.setdefault(v, []).append(x)
    if len(buckets) == 1:
        return frozenset(range(G.n))
    bucket = min(buckets.values(), key=len)
    anchor = bucket[0]
    candidates = [(G.mul[y][G.inv[anchor]] if left else
                   G.mul[G.inv[anchor]][y]) for y in bucket]
    known, generators = {0}, []
    for a in sorted(candidates):
        if a in known:
            continue
        if all(values[G.mul[a][x] if left else G.mul[x][a]] == values[x]
               for x in range(G.n)):
            generators.append(a)
            known = G.closure(generators)
    return frozenset(known)


@dataclass
class Kernel:
    values: tuple
    L: frozenset
    R: frozenset
    leaf: str = ""
    children: object = None
    choices: object = None

    def witness(self, boundary):
        if self.values[boundary] == inf:
            raise ValueError("infeasible boundary")
        if self.children is None:
            return [(self.leaf, boundary, self.values[boundary])]
        a, b = self.choices[boundary]
        u, v = self.children
        return u.witness(a) + v.witness(b)


def leaf(G, name, values):
    values = tuple(values)
    return Kernel(values, stabilizer(G, values, True),
                  stabilizer(G, values, False), leaf=name)


def compose(G, f, g, quotient=True, learn=True):
    """All role occurrences remain in the witness DAG.

    pairs counts only objective candidates, not total runtime.
    Complete symmetry and orbit costs are charged in the benchmark.
    """
    if quotient:
        H = f.R & g.L
        inner, covered = [], set()
        for a in range(G.n):
            if a not in covered:
                inner.append(a)
                covered.update(G.mul[a][h] for h in H)
        orbits, covered = [], set()
        for z in range(G.n):
            if z in covered:
                continue
            orbit = {}
            for l in sorted(f.L):
                for r in sorted(g.R):
                    x = G.mul[G.mul[l][z]][r]
                    orbit.setdefault(x, (l, r))
            covered.update(orbit)
            orbits.append((z, orbit))
    else:
        inner = list(range(G.n))
        orbits = [(z, {z: (0, 0)}) for z in range(G.n)]
    values, choices, pairs = [inf]*G.n, [None]*G.n, 0
    for z, orbit in orbits:
        best, arg = inf, None
        for a in inner:
            b = G.mul[G.inv[a]][z]
            score = f.values[a] + g.values[b]
            pairs += 1
            if score < best:
                best, arg = score, (a, b)
        for x, (l, r) in orbit.items():
            values[x] = best
            if arg is not None:
                a, b = arg
                # Retrieve child witnesses at NEW boundaries. Never edit
                # a primitive choice merely because aggregate costs agree.
                choices[x] = (G.mul[l][a], G.mul[b][r])
    values = tuple(values)
    L = stabilizer(G, values, True) if learn else f.L
    R = stabilizer(G, values, False) if learn else g.R
    out = Kernel(values, L, R, children=(f, g), choices=choices)
    for x, pair in enumerate(choices):
        if pair is not None:
            a, b = pair
            assert G.mul[a][b] == x
            assert f.values[a] + g.values[b] == values[x]
    return out, pairs


def verify_witness(G, kernel, boundary):
    # Enumerated positions distinguish repeated occurrences of a leaf type.
    product, cost = 0, 0
    for position, (leaf_type, action, action_cost) in enumerate(
            kernel.witness(boundary)):
        assert position >= 0 and leaf_type
        product = G.mul[product][action]
        cost += action_cost
    assert product == boundary and cost == kernel.values[boundary]


def demo():
    G = Dihedral(16)
    c, M = 3, 19

    def seed(j):
        step, size = G.m // (2**j), 2**j
        values = [M]*G.n
        for a in range(size//2 + 1):
            values[(a*step) % G.m] = 0
        values[G.m] = c
        return leaf(G, "f" + str(j), values)

    h, _ = compose(G, seed(1), seed(1))
    print("level  |L|  |R|  quotient_pairs  dense_pairs")
    for j in range(2, 5):
        f = seed(j)
        nxt, reduced_pairs = compose(G, h, f)
        dense, dense_pairs = compose(G, h, f, quotient=False, learn=False)
        assert nxt.values == dense.values and nxt.values != h.values
        expected_K = {a*(G.m//(2**j)) for a in range(2**j)}
        assert set(nxt.L) == expected_K == set(nxt.R)
        for boundary in range(G.n):
            verify_witness(G, nxt, boundary)
        print(j, len(nxt.L), len(nxt.R), reduced_pairs, dense_pairs)
        # Newly learned metadata changes the NEXT composition procedure.
        h = nxt
    assert h.values == tuple([0]*G.m + [c]*G.m)
    print("Exact table equality, new-kernel growth, and witnesses passed.")
    print("Pair counts exclude discovery/orbit overhead; see benchmark.")


if __name__ == "__main__":
    demo()