Merge pull request #290 from AdaWorldAPI/claude/pillar11-w1-sigker-parity

Pillar-11 W1 + W4: the cross-repo signature parity bridge, and PSD at depth-infinity

Author: AdaWorldAPI

Cargo.toml

@@ -424,6 +424,9 @@ exclude = [
     "crates/burn",
     "crates/wasm-simd-parity",
     "crates/neon-simd-parity",
+    # Cross-repo: its dev-dep is a PATH into a lance-graph sibling checkout.
+    # In-workspace, a missing sibling would fail resolution for EVERY member.
+    "crates/sigker-parity",
     "vendor/chacha20",
 ]
 default-members = [

crates/sigker-parity/.gitignore

@@ -0,0 +1,2 @@
+target/
+Cargo.lock

crates/sigker-parity/Cargo.toml

@@ -0,0 +1,32 @@
+# sigker-parity — the cross-repo gate for the Pillar-11 signature lanes.
+#
+# `ndarray::hpc::pillar::signature::signature_d2_deg3` (hardware: f32, fixed
+# d=2/deg-3, Chen accumulation) and lance-graph `sigker::signature_truncated`
+# (reference: f64, any d/depth) compute the SAME iterated integrals and had
+# zero cross-checks — census finding F-4 of
+# `pillar11-signature-certification-unification-v1`. The workspace's own
+# architecture rule (ndarray = hardware, lance-graph = thinking) blesses the
+# split but demands the parity test it never got. This crate is that test
+# (W1), plus the depth-infinity PSD leg (W4) that needs the same sibling.
+#
+# EXCLUDED from the workspace (see root Cargo.toml `exclude`) because its
+# `sigker` dep is a PATH into a sibling checkout. An unconditional path dep
+# whose target is absent fails manifest resolution for the WHOLE workspace —
+# ndarray CI does not check lance-graph out, so an in-workspace dep here
+# would break every ndarray build on a fresh clone. Excluded, it costs
+# nothing when the sibling is missing and runs on demand:
+#
+#   cargo test --manifest-path crates/sigker-parity/Cargo.toml
+#
+# Same shape as `crates/wasm-simd-parity` and `crates/neon-simd-parity`.
+[package]
+name = "sigker-parity"
+version = "0.0.0"
+edition = "2021"
+publish = false
+
+[dependencies]
+ndarray = { path = "../..", default-features = false, features = ["std", "hpc-extras", "pillar"] }
+
+[dev-dependencies]
+sigker = { path = "../../../lance-graph/crates/sigker" }

crates/sigker-parity/examples/w1_diagnose.rs

@@ -0,0 +1,37 @@
+//! Diagnostic only: is the W1 gap a FORMULA difference or f32 accumulation?
+use ndarray::hpc::pillar::signature::signature_d2_deg3;
+use sigker::signature_truncated;
+
+const NAMES: [&str; 15] = [
+    "s0", "1x", "1y", "2xx", "2xy", "2yx", "2yy", "3xxx", "3xxy", "3xyx", "3xyy", "3yxx", "3yxy", "3yyx", "3yyy",
+];
+
+fn cmp(tag: &str, flat: &[f32], n: usize) {
+    let hw = signature_d2_deg3(flat, n);
+    let pts: Vec<Vec<f64>> = (0..n)
+        .map(|k| vec![flat[2 * k] as f64, flat[2 * k + 1] as f64])
+        .collect();
+    let refr: Vec<f64> = signature_truncated(&pts, 3)
+        .levels
+        .iter()
+        .flat_map(|l| l.iter().copied())
+        .collect();
+    println!("--- {tag} (n={n}) ---");
+    for i in 0..15 {
+        let (h, r) = (hw[i] as f64, refr[i]);
+        let d = (h - r).abs();
+        let rel = if r.abs() > 1e-12 { d / r.abs() } else { d };
+        if rel > 1e-6 {
+            println!("  {:>5}: hw {:+.9}  ref {:+.9}  rel {:.3e}  <== DIFFERS", NAMES[i], h, r, rel);
+        }
+    }
+}
+
+fn main() {
+    // Single segment: closed form, zero accumulation — any gap here is FORMULA.
+    cmp("one segment", &[0.0, 0.0, 1.0, 0.5], 2);
+    // Two segments: Chen composition enters.
+    cmp("two segments", &[0.0, 0.0, 1.0, 0.5, 1.3, -0.2], 3);
+    // Three, exact small values (representable in f32) — still formula-only.
+    cmp("three segments", &[0.0, 0.0, 0.5, 0.25, 0.75, -0.5, 0.25, 0.125], 4);
+}

crates/sigker-parity/examples/w1_sweep.rs

@@ -0,0 +1,60 @@
+//! Pre-registration sweep: how does the hardware-vs-reference error scale,
+//! and under WHICH normalization is it a stable gate?
+use ndarray::hpc::pillar::signature::signature_d2_deg3;
+use sigker::signature_truncated;
+
+struct Rng(u64);
+impl Rng {
+    fn f(&mut self) -> f32 {
+        let mut x = self.0;
+        x ^= x >> 12;
+        x ^= x << 25;
+        x ^= x >> 27;
+        self.0 = x;
+        ((x.wrapping_mul(0x2545_F491_4F6C_DD1D) >> 40) as f32 / (1u32 << 24) as f32) - 0.5
+    }
+}
+
+// level of each of the 15 coefficients
+const LEVEL: [usize; 15] = [0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3];
+
+fn main() {
+    println!("{:>6} {:>12} {:>14} {:>16}", "N", "worst |abs|", "worst /coeff", "worst /levelmax");
+    for &n in &[16usize, 32, 64, 128, 256] {
+        let mut rng = Rng(0x9E37_79B9_7F4A_7C15);
+        let (mut wa, mut wc, mut wl) = (0.0f64, 0.0f64, 0.0f64);
+        for _ in 0..1000 {
+            let (mut x, mut y) = (0.0f32, 0.0f32);
+            let mut flat = Vec::with_capacity(n * 2);
+            let mut pts = Vec::with_capacity(n);
+            for _ in 0..n {
+                flat.push(x);
+                flat.push(y);
+                pts.push(vec![x as f64, y as f64]);
+                x += rng.f();
+                y += rng.f();
+            }
+            let hw = signature_d2_deg3(&flat, n);
+            let refr: Vec<f64> = signature_truncated(&pts, 3)
+                .levels
+                .iter()
+                .flat_map(|l| l.iter().copied())
+                .collect();
+            // characteristic magnitude per level, from the REFERENCE
+            let mut lvmax = [0.0f64; 4];
+            for i in 0..15 {
+                lvmax[LEVEL[i]] = lvmax[LEVEL[i]].max(refr[i].abs());
+            }
+            for i in 0..15 {
+                let d = (hw[i] as f64 - refr[i]).abs();
+                wa = wa.max(d);
+                if refr[i].abs() > 1e-12 {
+                    wc = wc.max(d / refr[i].abs());
+                }
+                let s = lvmax[LEVEL[i]].max(1e-12);
+                wl = wl.max(d / s);
+            }
+        }
+        println!("{n:>6} {wa:>12.3e} {wc:>14.3e} {wl:>16.3e}");
+    }
+}

crates/sigker-parity/examples/w4_concentration_sweep.rs

@@ -0,0 +1,63 @@
+//! Does depth-inf self-kernel concentration shrink like 1/sqrt(N) (a sample-
+//! size effect) or plateau (a genuine heavy tail)? Measure, do not assume.
+use ndarray::hpc::pillar::signature::brownian_path_d2;
+use ndarray::hpc::pillar::SplitMix64;
+use sigker::{signature_kernel_pde, signature_truncated};
+
+const SEED: u64 = 0x5EED_1111_5164_A7AB;
+const N_STEPS: usize = 50;
+
+fn pool(n: usize) -> Vec<Vec<Vec<f64>>> {
+    let mut rng = SplitMix64::new(SEED);
+    (0..n)
+        .map(|_| {
+            let p = brownian_path_d2(&mut rng, N_STEPS);
+            (0..=N_STEPS)
+                .map(|k| vec![p[2 * k] as f64, p[2 * k + 1] as f64])
+                .collect()
+        })
+        .collect()
+}
+
+fn stats(v: &[f64]) -> (f64, f64, f64) {
+    let n = v.len();
+    let h = n / 2;
+    let m1 = v[..h].iter().sum::<f64>() / h as f64;
+    let m2 = v[h..].iter().sum::<f64>() / (n - h) as f64;
+    let mean = v.iter().sum::<f64>() / n as f64;
+    let var = v.iter().map(|x| (x - mean).powi(2)).sum::<f64>() / n as f64;
+    // half-mean gap, and the coefficient of variation that predicts it
+    ((m1 - m2).abs() / mean, var.sqrt() / mean, mean)
+}
+
+fn main() {
+    println!("depth-INFINITY (Goursat PDE)");
+    println!("{:>6} {:>12} {:>10} {:>14} {:>12}", "N", "concentr", "CV", "predicted", "mean K");
+    for &n in &[64usize, 128, 256, 512, 1000] {
+        let p = pool(n);
+        let k: Vec<f64> = p.iter().map(|x| signature_kernel_pde(x, x)).collect();
+        let (c, cv, mean) = stats(&k);
+        // For independent samples the expected half-mean gap ~ CV * sqrt(8/(pi*N))
+        let pred = cv * (8.0 / (core::f64::consts::PI * n as f64)).sqrt();
+        println!("{n:>6} {c:>12.4} {cv:>10.3} {pred:>14.4} {mean:>12.4e}");
+    }
+    println!("\ndepth-3 TRUNCATED (the existing battery's kernel, f64 reference)");
+    println!("{:>6} {:>12} {:>10} {:>14} {:>12}", "N", "concentr", "CV", "predicted", "mean K");
+    for &n in &[64usize, 1000] {
+        let p = pool(n);
+        let k: Vec<f64> = p
+            .iter()
+            .map(|x| {
+                let s = signature_truncated(x, 3);
+                s.levels
+                    .iter()
+                    .flat_map(|l| l.iter())
+                    .map(|v| v * v)
+                    .sum::<f64>()
+            })
+            .collect();
+        let (c, cv, mean) = stats(&k);
+        let pred = cv * (8.0 / (core::f64::consts::PI * n as f64)).sqrt();
+        println!("{n:>6} {c:>12.4} {cv:>10.3} {pred:>14.4} {mean:>12.4e}");
+    }
+}

crates/sigker-parity/src/lib.rs

@@ -0,0 +1,3 @@
+pub fn sibling_is_wired() -> bool {
+    true
+}