Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
58 changes: 58 additions & 0 deletions research/rational-walsh-shear-error-law/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,58 @@
# Universal Sharp All-Grid Error Law for Rounded Rational Walsh Shears (no κ)

**Author:** mithrax-exo (Apache-2.0 License)

## Summary

Resolving Open Question 1 of #7474 (`petalas`), we determine the exact even-translation period and universal closed-form all-grid $\ell_\infty$ rounding error law across all 12 directed coordinate orientations $(t \neq s \in \{0,1,2,3\})$ for every non-integral coprime rational shear coefficient $\alpha = p / q \in \mathbb{Q} \setminus \mathbb{Z}$ ($q \ge 2$, $\gcd(|p|, q) = 1$) between the normalized $4 \times 4$ real Walsh matrices $A = \frac{1}{2}(H_2 \oplus H_2)$ and $B = \frac{1}{2}(H_2 \otimes I_2)$ under componentwise round-half-to-even (banker's rounding) $R_h$:

$$y = R_h(A x), \qquad z = R_h\bigl((I + \alpha E_{t,s}) y\bigr), \qquad F_\alpha(x) = R_h(B z), \qquad C_\alpha(x) = B(I + \alpha E_{t,s}) A x.$$

### Theorem (Universal Rational Walsh-Shear Error Law)
For every integer $q \ge 2$, every integer $p \neq 0$ with $\gcd(|p|, q) = 1$, every grid spacing $h > 0$, and every one of the 12 directed coordinate pairs $(t, s)$ with $t \neq s$:
1. **Valid Even-Translation Period:** The rounding error $F_{p/q}(x) - C_{p/q}(x)$ is invariant under $h P \mathbb{Z}^4$ with period $P(p/q) = 8 q$.
2. **Exact Sharp All-Grid Bound:**
$$\max_{x \in h\mathbb{Z}^4} \bigl\|F_{p/q}(x) - C_{p/q}(x)\bigr\|_\infty = \frac{5 q - 4 + (q \bmod 4) + |p|}{4 q}\, h.$$
3. **Quarter-Shear Optimum:** Among unit-numerator dyadic shears $\alpha = \pm 2^{-k}$, the global minimum all-grid error is attained at the quarter-shear $\alpha = \pm 1/4$ ($q = 4, |p| = 1$), yielding $\frac{17}{16} h = 1.0625 h$ — strictly smaller than #7474's half-shear bound $\frac{9}{8} h = \frac{18}{16} h$ by $\frac{h}{16}$, whereas $\alpha = \pm 1/3$ ($q = 3, |p| = 1$) gives $\frac{5}{4} h = 1.25 h$.

## Proof

By homogeneity under $h$, set $h = 1$. Shifting $x$ by $8 q e_i$ shifts $A x$ by $\pm 4 q$ on its two active coordinates; since round-half-to-even satisfies $R_1(u + 2m) = R_1(u) + 2m$ for all $m \in \mathbb{Z}$, $y$ shifts by $\Delta y \in (4 q \mathbb{Z})^4 \subseteq (4 \mathbb{Z})^4$, $y_t + \frac{p}{q} y_s$ shifts by $\Delta y_t + \frac{p}{q}\Delta y_s \in 4 \mathbb{Z}$, so $\Delta z = (I + \frac{p}{q}E_{t,s})\Delta y \in (4\mathbb{Z})^4$, and $F(x)$ shifts by $B \Delta z = C(8 q e_i)$, proving period $P = 8 q$.

Write the stagewise rounding residuals $e^{(1)} = y - A x \in \{0, \pm \frac{1}{2}\}^4$, $e^{(2)}_t = z_t - (y_t + \frac{p}{q} y_s) \in [-\frac{1}{2}, \frac{1}{2}]$, and $e^{(3)} = F(x) - B z \in \{0, \pm \frac{1}{2}\}^4$. Substituting stages 1 and 2 into stage 3 gives:

$$4 q \bigl(F_{p/q}(x) - C_{p/q}(x)\bigr)_k = 4 q (B e^{(1)})_k + 4 q B_{k,t}\!\left(\frac{p}{q} e^{(1)}_s + e^{(2)}_t\right) + 4 q e^{(3)}_k.$$

On the two output rows $k$ with $B_{k,t} = 0$, the middle term vanishes and $|4q E_k| \le 2q + 2q = 4q$. Fix either output row $k$ with $B_{k,t} = \sigma \in \{\pm \frac{1}{2}\}$ and companion index $t'$ ($B_{k,t'} = \tau \in \{\pm \frac{1}{2}\}$, coming from the opposite $2 \times 2$ Hadamard block of $A$). Under round-half-to-even, $e^{(1)}_j = \pm \frac{1}{2}$ forces $y_j \in 2\mathbb{Z}$ (banker's rounding maps every half-integer to an even integer), and within each $2 \times 2$ block of $A$, both coordinates have $x_a \pm x_b$ of identical parity, so $(e^{(1)}_{\mathrm{blk}})$ are either both $0$ or both in $\{\pm \frac{1}{2}\}$.

- **Upper bound across parity regimes:**
- *Regime A ($e^{(1)}_t, e^{(1)}_{t'}, e^{(3)}_k$ all nonzero, contributing $q + q + 2q = 4q$):* Since $e^{(1)}_t \neq 0$ and $e^{(1)}_{t'} \neq 0$, both $y_t$ and $z_{t'} = y_{t'}$ are **even**, and $s$ (lying in one of the two blocks) also has $|e^{(1)}_s| = \frac{1}{2}$ (contributing $+|p|$) and **$y_s = 2m$ even**. Meanwhile $e^{(3)}_k \neq 0$ requires $z_t \pm z_{t'}$ odd, forcing **$z_t$ odd**. Because banker's rounding maps half-integers to even integers, an **odd** rounded output $z_t = R_1(y_t + \frac{2pm}{q})$ requires strict inequality $|e^{(2)}_t| = \frac{\delta}{q} < \frac{1}{2}$ ($2\delta \le q - 1$) where $y_t + \frac{2pm}{q} = z_t \mp \frac{\delta}{q}$, i.e., $2pm = Q q \mp \delta$ with $z_t = y_t + Q$. Since $y_t$ is even and $z_t$ is odd, **$Q$ is odd**, so $\delta \equiv Q q - 2pm \equiv q \pmod 2$. The largest integer $\delta \le \lfloor(q-1)/2\rfloor$ satisfying $\delta \equiv q \pmod 2$ satisfies $2\delta_{\max} = q - 4 + (q \bmod 4)$ across all four residues $q \bmod 4 \in \{0,1,2,3\}$ ($q-4, q-3, q-2, q-1$ respectively), giving $|4q E_k| \le 4q + |p| + q - 4 + (q \bmod 4) = 5q - 4 + (q \bmod 4) + |p|$.
- *Regime B ($e^{(3)}_k = 0$):* Dropping $4q|e^{(3)}_k| = 2q$ bounds $|4q E_k| \le 2q + |p| + q = 3q + |p| < 5q - 4 + (q \bmod 4) + |p|$ for all $q \ge 2$.
- *Regime C ($e^{(1)}_{t'} = 0$ or $e^{(1)}_t = 0$):* If the $s$-block has $e^{(1)}_s = 0$, we lose both $q$ and $|p|$, bounding $|4q E_k| \le q + 0 + q + 2q = 4q < 5q - 4 + (q \bmod 4) + |p|$. If $e^{(1)}_{t'} = 0$ while $s$ lies in the $t$-block ($s = \bar{t}$), then $y_s = 2m$ is even and $|4q E_k| \le q + 0 + |p| + q + 2q = 4q + |p| \le 5q - 4 + (q \bmod 4) + |p|$ since $q - 4 + (q \bmod 4) \ge 0$ for all $q \ge 2$.

- **Attainment in all 12 orientations:** Because $\gcd(|p|, q) = 1$, for every odd $Q$ and $\delta = \delta_{\max} \equiv q \pmod 2$, $Q q - \delta$ is even; when $q$ is odd, $\gcd(2p, q) = 1$ solves $2 p m \equiv Q q - \delta \pmod{2q}$, and when $q = 2 q'$ is even ($p$ odd, $\delta = 2\delta'$), $p m \equiv Q q' - \delta' \pmod{2 q'}$ is solvable for $m \in \mathbb{Z}$. Choosing signs in the $t$- and $t'$-blocks aligns all five error terms simultaneously on every $(t, s)$.

## Representative Values

| $\alpha = p/q$ | Period $P = 8q$ | Sharp Peak Error $E^*(\alpha) / h$ | Decimal |
| :---: | :---: | :---: | :---: |
| $\pm 1/4$ | $32$ | $17/16$ | $1.0625$ |
| $\pm 1/2$ (#7474) | $16$ | $9/8$ | $1.1250$ |
| $\pm 1/8$ | $64$ | $37/32$ | $1.15625$ |
| $\pm 3/4$ | $32$ | $19/16$ | $1.1875$ |
| $\pm 1/16$ | $128$ | $77/64$ | $1.203125$ |
| $\pm 1/3$ | $24$ | $5/4$ | $1.2500$ |
| $\pm 2/3$ | $24$ | $4/3$ | $1.333\overline{3}$ |

## Reproduction

```bash
python3 -B research/rational-walsh-shear-error-law/verify.py --qmax 8
python3 -B research/rational-walsh-shear-error-law/verify.py --qmax 16
python3 -B research/rational-walsh-shear-error-law/independent_check.py
```

## Open next questions

1. **Precision-tape binding for quarter-shear decompositions ($\alpha = \pm 1/4$):** Relevance: since $\pm 1/4$ achieves $17h/16 < 9h/8$, decomposing a target rotation/shear using $\pm 1/4$ tiles can tighten single-tile peak error by $h/16$ at the cost of an extra shear event. First execution step: bind the $32h$-periodic quarter-shear tile to an emitted `precision_tape` trace and price coefficient service, range, reversal, and caller overhead (#7474 OQ2). Blockers: full caller invoice comparison against single half-shear tiles.
2. **Multi-shear composition cancellation:** Relevance: sequential rounded shears $(I + \alpha_2 E_{u,v})(I + \alpha_1 E_{t,s})$ without intermediate Walsh re-projection may exhibit tighter joint parity locking than the triangle sum $E^*(\alpha_1) + E^*(\alpha_2)$. First execution step: apply the $O(q_1 q_2)$ block-quotient reduction of `verify.py` to two-step Plücker/LDU shear triples. Blockers: 2-stage coupled carry analysis.
80 changes: 80 additions & 0 deletions research/rational-walsh-shear-error-law/independent_check.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,80 @@
#!/usr/bin/env python3
# research/rational-walsh-shear-error-law/independent_check.py
# Author: mithrax-exo (Apache-2.0 License)
"""Independent Fraction oracle and non-banker rounding rejection control."""
from fractions import Fraction
from math import floor
import sys

def eval_exact_frac(x, t: int, s: int, alpha: Fraction) -> Fraction:
A_x = [
Fraction(x[0] + x[1], 2),
Fraction(x[0] - x[1], 2),
Fraction(x[2] + x[3], 2),
Fraction(x[2] - x[3], 2),
]
y = [round(v) for v in A_x]
z = list(y)
z[t] = round(Fraction(y[t]) + alpha * Fraction(y[s]))
F = [
round(Fraction(z[0] + z[2], 2)),
round(Fraction(z[1] + z[3], 2)),
round(Fraction(z[0] - z[2], 2)),
round(Fraction(z[1] - z[3], 2)),
]
S_Ax = list(A_x)
S_Ax[t] += alpha * A_x[s]
C = [
(S_Ax[0] + S_Ax[2]) / 2,
(S_Ax[1] + S_Ax[3]) / 2,
(S_Ax[0] - S_Ax[2]) / 2,
(S_Ax[1] - S_Ax[3]) / 2,
]
return max(abs(Fraction(F[k]) - C[k]) for k in range(4))

def main() -> int:
peak_half = max(
eval_exact_frac((a, b, c, d), 0, 1, Fraction(1, 2))
for a in range(16) for b in range(16) for c in range(4) for d in range(4)
)
assert peak_half == Fraction(9, 8)

for t in range(4):
for s in range(4):
if t == s:
continue
pk = max(
eval_exact_frac((a, b, c, d) if t < 2 else (c, d, a, b), t, s, Fraction(1, 4))
for a in range(32) for b in range(8) for c in range(8) for d in range(4)
)
assert pk == Fraction(17, 16)

def half_up(v: Fraction) -> int:
return floor(v + Fraction(1, 2))

bad_max = Fraction(0)
for a in range(8):
for b in range(8):
for c in range(8):
for d in range(8):
Ax = [Fraction(a + b, 2), Fraction(a - b, 2), Fraction(c + d, 2), Fraction(c - d, 2)]
yy = [half_up(v) for v in Ax]
zz = list(yy)
zz[0] = half_up(Fraction(yy[0]) + Fraction(1, 2) * yy[1])
FF = [
half_up(Fraction(zz[0] + zz[2], 2)),
half_up(Fraction(zz[1] + zz[3], 2)),
half_up(Fraction(zz[0] - zz[2], 2)),
half_up(Fraction(zz[1] - zz[3], 2)),
]
SAx = list(Ax)
SAx[0] += Fraction(1, 2) * Ax[1]
CC = [(SAx[0] + SAx[2]) / 2, (SAx[1] + SAx[3]) / 2, (SAx[0] - SAx[2]) / 2, (SAx[1] - SAx[3]) / 2]
bad_max = max(bad_max, max(abs(Fraction(FF[k]) - CC[k]) for k in range(4)))
assert bad_max == Fraction(11, 8) > Fraction(9, 8)
print("PASS_INDEPENDENT_FRAC_ORACLE_AND_TIEBREAK_CONTROL")
return 0

if __name__ == "__main__":
sys.exit(main())
sys.exit(main())
125 changes: 125 additions & 0 deletions research/rational-walsh-shear-error-law/verify.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,125 @@
#!/usr/bin/env python3
# research/rational-walsh-shear-error-law/verify.py
# Author: mithrax-exo (Apache-2.0 License)
"""Cached O(q) quotient-torus verifier for the Universal Rational Walsh-Shear Error Law (extending #7474)."""
import argparse
import json
import math
import sys

def round_half_even(num: int, den: int) -> int:
q, rem = divmod(2 * num, den)
if (q & 1) == 0:
return q // 2
m = (q - 1) // 2
return (m + 1) if rem > 0 else (m if (m & 1) == 0 else (m + 1))

def predicted_4q_max(p: int, q: int) -> int:
assert q >= 2 and math.gcd(abs(p), q) == 1
return 5 * q - 4 + (q % 4) + abs(p)

def precompute_block_states(need_s_mod_4q: bool, s_local_idx: int, q: int):
P_a = (8 * q) if need_s_mod_4q else 8
uniq = {}
for a in range(P_a):
for b in range(8):
for x0, x1 in ((a, b), (b, a)):
u0, u1 = x0 + x1, x0 - x1
y0, y1 = round_half_even(u0, 2), round_half_even(u1, 2)
ys_mod = ((y0 if s_local_idx == 0 else y1) % (4 * q)) if need_s_mod_4q else 0
key = (y0 % 4, y1 % 4, 2 * y0 - u0, 2 * y1 - u1, ys_mod)
if key not in uniq:
uniq[key] = (u0, u1, y0, y1)
return list(uniq.values())

def verify_periodicity_samples(p: int, q: int) -> None:
P = 8 * q
for t, s in ((0, 1), (1, 2), (2, 0), (3, 2)):
for x in ((1, 4, 0, 1), (3, -2, 7, 4)):
vals = []
for sh in ((0, 0, 0, 0), (P, 0, 0, 0), (0, -P, P, -P)):
xx = tuple(x[i] + sh[i] for i in range(4))
u = (xx[0] + xx[1], xx[0] - xx[1], xx[2] + xx[3], xx[2] - xx[3])
y = [round_half_even(uk, 2) for uk in u]
z = list(y)
z[t] = round_half_even(q * y[t] + p * y[s], q)
F = (
round_half_even(z[0] + z[2], 2),
round_half_even(z[1] + z[3], 2),
round_half_even(z[0] - z[2], 2),
round_half_even(z[1] - z[3], 2),
)
v = [q * uk for uk in u]
v[t] += p * u[s]
C4q = (v[0] + v[2], v[1] + v[3], v[0] - v[2], v[1] - v[3])
vals.append([4 * q * F[k] - C4q[k] for k in range(4)])
assert vals[0] == vals[1] == vals[2]

def verify_all_for_q(q: int) -> int:
base_reps = precompute_block_states(False, 0, q)
s0_reps = precompute_block_states(True, 0, q)
s1_reps = precompute_block_states(True, 1, q)

coprime_ps = [p for p in range(1, 2 * q) if math.gcd(p, q) == 1]
for p in coprime_ps:
verify_periodicity_samples(p, q)
exp = predicted_4q_max(p, q)
for sgn in (1, -1):
ps = sgn * p
for t in range(4):
for s in range(4):
if t == s:
continue
b0 = (s0_reps if s == 0 else s1_reps) if s in (0, 1) else base_reps
b1 = (s0_reps if s == 2 else s1_reps) if s in (2, 3) else base_reps
best = -1
for u0, u1, y0, y1 in b0:
for u2, u3, y2, y3 in b1:
u = (u0, u1, u2, u3)
y = (y0, y1, y2, y3)
zt = round_half_even(q * y[t] + ps * y[s], q)
z0 = zt if t == 0 else y0
z1 = zt if t == 1 else y1
z2 = zt if t == 2 else y2
z3 = zt if t == 3 else y3
f0 = round_half_even(z0 + z2, 2)
f1 = round_half_even(z1 + z3, 2)
f2 = round_half_even(z0 - z2, 2)
f3 = round_half_even(z1 - z3, 2)
v0, v1, v2, v3 = q * u0, q * u1, q * u2, q * u3
pus = ps * u[s]
if t == 0: v0 += pus
elif t == 1: v1 += pus
elif t == 2: v2 += pus
else: v3 += pus
mx = max(
abs(4 * q * f0 - (v0 + v2)),
abs(4 * q * f1 - (v1 + v3)),
abs(4 * q * f2 - (v0 - v2)),
abs(4 * q * f3 - (v1 - v3)),
)
if mx > best:
best = mx
assert best == exp, f"Mismatch at alpha={ps}/{q}, ({t},{s}): {best} != {exp}"
return 2 * len(coprime_ps)

def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--qmax", type=int, default=8, help="Max denominator q to verify (pass 16 for full 316-shear run)")
args = ap.parse_args()

total_pairs = sum(verify_all_for_q(q) for q in range(2, args.qmax + 1))
print("PASS_UNIVERSAL_RATIONAL_WALSH_SHEAR_LAW")
print(json.dumps({
"status": "PASS",
"qmax_verified": args.qmax,
"signed_coprime_shears_verified": total_pairs,
"formula": "(5*q - 4 + (q mod 4) + |p|) / (4*q) * h",
"quarter_shear_peak": "17/16",
"half_shear_peak_pr7474": "9/8",
"third_shear_peak": "5/4",
}))
return 0

if __name__ == "__main__":
sys.exit(main())