updates?

codes/experiment_v0.2/config.py

@@ -1,24 +1,27 @@
 """Edit globals here; no CLI parser is used."""
+
 from datetime import datetime
 from pathlib import Path
 
-SEED = 7
+SEED = 114514
 KAPPA = 1e-3
-NUM_SAMPLES = 10**4                  # requested default
+NUM_SAMPLES = 10**6  # requested default
 LIPSCHITZ_PAIRS = 12_000
 LIPSCHITZ_RESERVOIR = 4_096
-MAJORANA_STAR_STATES = 16            # only for visualization
-MAX_STAR_DEGREE = 63                 # avoid unstable huge root-finding plots
+MAJORANA_STAR_STATES = 16  # only for visualization
+MAX_STAR_DEGREE = 63  # avoid unstable huge root-finding plots
 
-BACKEND = "auto"                     # auto | jax | numpy
-JAX_PLATFORM = ""                    # "", "cpu", "gpu"; set before importing JAX
-RESULTS_DIR = Path("./results") / f"exp-{datetime.now():%Y%m%d-%H%M%S}"
+BACKEND = "auto"  # auto | jax | numpy
+JAX_PLATFORM = "gpu"  # "", "cpu", "gpu"; set before importing JAX
+RESULTS_DIR = (
+    Path.joinpath(Path.cwd(), Path("./results")) / f"exp-{datetime.now():%Y%m%d-%H%M%S}"
+)
 
 # Chosen so the three families have comparable intrinsic dimensions:
 # sphere S^(m-1), CP^(d_A d_B - 1), and Sym^N(C^2) ~ CP^N.
-SPHERE_DIMS = [16, 64, 256, 1024]
-CP_DIMS = [(4, 4), (8, 8), (16, 16), (32, 32)]
-MAJORANA_N = [15, 63, 255, 1023]
+SPHERE_DIMS = [1<<i for i in range(4, 12)]
+CP_DIMS = [(1<<i, 1<<i) for i in range(4, 12)]
+MAJORANA_N = [(1<<i)-1 for i in range(4, 12)]
 
 # Batch sizes are the main speed knob; reduce CP batches first if memory is tight.
 BATCH = {"sphere": 32_768, "cp": 256, "majorana": 65_536}

codes/experiment_v0.2/main.py

@@ -8,12 +8,17 @@ from pathlib import Path
 
 import numpy as np
 
+import sys
+
+# Add the parent directory to sys.path
+sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
 import config
 
 if config.JAX_PLATFORM:
     os.environ["JAX_PLATFORM_NAME"] = config.JAX_PLATFORM
 
-from sampling_pipeline import (  # noqa: E402
+from sampling_pipeline import (
     plot_cross_space_comparison,
     plot_family_summary,
     plot_histogram,
@@ -22,7 +27,11 @@ from sampling_pipeline import (  # noqa: E402
     simulate_space,
     write_summary_csv,
 )
-from spaces import ComplexProjectiveSpace, MajoranaSymmetricSpace, UnitSphereSpace  # noqa: E402
+from spaces import (
+    ComplexProjectiveSpace,
+    MajoranaSymmetricSpace,
+    UnitSphereSpace,
+)
 
 
 def main() -> None:
@@ -54,9 +63,21 @@ def main() -> None:
         plot_tail(result, space, outdir)
 
         if space.family == "majorana" and space.N <= config.MAX_STAR_DEGREE:
-            star_seed = int(seeds[len(spaces) + i].generate_state(1, dtype=np.uint32)[0])
-            from pipeline import _sample_stream  # local import to avoid exporting internals
-            states, _ = _sample_stream(space, config.MAJORANA_STAR_STATES, min(config.MAJORANA_STAR_STATES, config.BATCH["majorana"]), star_seed, config.BACKEND, keep_states=True)
+            star_seed = int(
+                seeds[len(spaces) + i].generate_state(1, dtype=np.uint32)[0]
+            )
+            from sampling_pipeline import (
+                _sample_stream,
+            )  # local import to avoid exporting internals
+
+            states, _ = _sample_stream(
+                space,
+                config.MAJORANA_STAR_STATES,
+                min(config.MAJORANA_STAR_STATES, config.BATCH["majorana"]),
+                star_seed,
+                config.BACKEND,
+                keep_states=True,
+            )
             plot_majorana_stars(space, states, outdir)
 
     results.sort(key=lambda r: (r.family, r.intrinsic_dim))
@@ -76,8 +97,14 @@ def main() -> None:
 
     print("family     dim    mean(bits)   part_diam(bits)   norm_proxy_q99")
     for r in results:
-        q = f"{r.normalized_proxy_q99:.6g}" if r.normalized_proxy_q99 == r.normalized_proxy_q99 else "nan"
-        print(f"{r.family:8s} {r.intrinsic_dim:5d}  {r.mean:11.6f}  {r.partial_diameter:16.6f}  {q:>14s}")
+        q = (
+            f"{r.normalized_proxy_q99:.6g}"
+            if r.normalized_proxy_q99 == r.normalized_proxy_q99
+            else "nan"
+        )
+        print(
+            f"{r.family:8s} {r.intrinsic_dim:5d}  {r.mean:11.6f}  {r.partial_diameter:16.6f}  {q:>14s}"
+        )
     print(f"\nWrote results to: {outdir.resolve()}")
 
 

codes/experiment_v0.2/sampling_pipline.py

@@ -1,324 +0,0 @@
-from __future__ import annotations
-
-import csv
-import math
-from dataclasses import dataclass, field
-from pathlib import Path
-from typing import Sequence
-
-import matplotlib.pyplot as plt
-import numpy as np
-from tqdm.auto import tqdm
-
-from spaces import HAS_JAX, MetricMeasureSpace, jax, random
-
-
-@dataclass
-class SystemResult:
-    """Compact record of one simulated metric-measure system."""
-    family: str
-    label: str
-    slug: str
-    intrinsic_dim: int
-    num_samples: int
-    kappa: float
-    mass: float
-    observable_max: float
-    values: np.ndarray
-    partial_diameter: float
-    interval_left: float
-    interval_right: float
-    mean: float
-    median: float
-    std: float
-    empirical_lipschitz_max: float
-    empirical_lipschitz_q99: float
-    normalized_proxy_max: float
-    normalized_proxy_q99: float
-    theory: dict[str, float] = field(default_factory=dict)
-
-
-def partial_diameter(samples: np.ndarray, mass: float) -> tuple[float, float, float]:
-    """Shortest interval carrying the requested empirical mass."""
-    x = np.sort(np.asarray(samples, float))
-    n = len(x)
-    if n == 0 or not (0.0 < mass <= 1.0):
-        raise ValueError("Need nonempty samples and mass in (0,1].")
-    if n == 1:
-        return 0.0, float(x[0]), float(x[0])
-    m = max(1, int(math.ceil(mass * n)))
-    if m <= 1:
-        return 0.0, float(x[0]), float(x[0])
-    w = x[m - 1 :] - x[: n - m + 1]
-    i = int(np.argmin(w))
-    return float(w[i]), float(x[i]), float(x[i + m - 1])
-
-
-def empirical_lipschitz(
-    space: MetricMeasureSpace,
-    states: np.ndarray,
-    values: np.ndarray,
-    rng: np.random.Generator,
-    num_pairs: int,
-) -> tuple[float, float]:
-    """Estimate max and q99 slope over random state pairs."""
-    n = len(states)
-    if n < 2 or num_pairs <= 0:
-        return float("nan"), float("nan")
-    i = rng.integers(0, n, size=num_pairs)
-    j = rng.integers(0, n - 1, size=num_pairs)
-    j += (j >= i)
-    d = space.metric_pairs(states[i], states[j])
-    good = d > 1e-12
-    if not np.any(good):
-        return float("nan"), float("nan")
-    r = np.abs(values[i] - values[j])[good] / d[good]
-    return float(np.max(r)), float(np.quantile(r, 0.99))
-
-
-def _sample_stream(
-    space: MetricMeasureSpace,
-    n: int,
-    batch: int,
-    seed: int,
-    backend: str,
-    keep_states: bool,
-) -> tuple[np.ndarray | None, np.ndarray]:
-    """Sample values, optionally keeping state vectors for Lipschitz estimation."""
-    vals = np.empty(n, dtype=np.float32)
-    states = np.empty((n, space.state_dim), dtype=np.float32 if space.family == "sphere" else np.complex64) if keep_states else None
-    use_jax = backend != "numpy" and HAS_JAX
-    desc = f"{space.slug}: {n:,} samples"
-    if use_jax:
-        key = random.PRNGKey(seed)
-        for s in tqdm(range(0, n, batch), desc=desc, unit="batch"):
-            b = min(batch, n - s)
-            key, sub = random.split(key)
-            x, y = space.sample_jax(sub, b)
-            vals[s : s + b] = np.asarray(jax.device_get(y), dtype=np.float32)
-            if keep_states:
-                states[s : s + b] = np.asarray(jax.device_get(x), dtype=states.dtype)
-    else:
-        rng = np.random.default_rng(seed)
-        for s in tqdm(range(0, n, batch), desc=desc, unit="batch"):
-            b = min(batch, n - s)
-            x, y = space.sample_np(rng, b)
-            vals[s : s + b] = y
-            if keep_states:
-                states[s : s + b] = x.astype(states.dtype)
-    return states, vals
-
-
-def simulate_space(
-    space: MetricMeasureSpace,
-    *,
-    num_samples: int,
-    batch: int,
-    kappa: float,
-    seed: int,
-    backend: str,
-    lipschitz_pairs: int,
-    lipschitz_reservoir: int,
-) -> SystemResult:
-    """Main Monte Carlo pass plus a smaller Lipschitz pass."""
-    vals = _sample_stream(space, num_samples, batch, seed, backend, keep_states=False)[1]
-    mass = 1.0 - kappa
-    width, left, right = partial_diameter(vals, mass)
-
-    r_states, r_vals = _sample_stream(space, min(lipschitz_reservoir, num_samples), min(batch, lipschitz_reservoir), seed + 1, backend, keep_states=True)
-    lip_rng = np.random.default_rng(seed + 2)
-    lip_max, lip_q99 = empirical_lipschitz(space, r_states, r_vals, lip_rng, lipschitz_pairs)
-    nmax = width / lip_max if lip_max == lip_max and lip_max > 0 else float("nan")
-    nq99 = width / lip_q99 if lip_q99 == lip_q99 and lip_q99 > 0 else float("nan")
-
-    return SystemResult(
-        family=space.family,
-        label=space.label,
-        slug=space.slug,
-        intrinsic_dim=space.intrinsic_dim,
-        num_samples=num_samples,
-        kappa=kappa,
-        mass=mass,
-        observable_max=space.observable_max,
-        values=vals,
-        partial_diameter=width,
-        interval_left=left,
-        interval_right=right,
-        mean=float(np.mean(vals)),
-        median=float(np.median(vals)),
-        std=float(np.std(vals, ddof=1)) if len(vals) > 1 else 0.0,
-        empirical_lipschitz_max=lip_max,
-        empirical_lipschitz_q99=lip_q99,
-        normalized_proxy_max=nmax,
-        normalized_proxy_q99=nq99,
-        theory=space.theory(kappa),
-    )
-
-
-def write_summary_csv(results: Sequence[SystemResult], out_path: Path) -> None:
-    """Write one flat CSV with optional theory fields."""
-    extras = sorted({k for r in results for k in r.theory})
-    fields = [
-        "family", "label", "intrinsic_dim", "num_samples", "kappa", "mass",
-        "observable_max_bits", "partial_diameter_bits", "interval_left_bits", "interval_right_bits",
-        "mean_bits", "median_bits", "std_bits", "empirical_lipschitz_max", "empirical_lipschitz_q99",
-        "normalized_proxy_max", "normalized_proxy_q99",
-    ] + extras
-    with out_path.open("w", newline="") as fh:
-        w = csv.DictWriter(fh, fieldnames=fields)
-        w.writeheader()
-        for r in results:
-            row = {
-                "family": r.family,
-                "label": r.label,
-                "intrinsic_dim": r.intrinsic_dim,
-                "num_samples": r.num_samples,
-                "kappa": r.kappa,
-                "mass": r.mass,
-                "observable_max_bits": r.observable_max,
-                "partial_diameter_bits": r.partial_diameter,
-                "interval_left_bits": r.interval_left,
-                "interval_right_bits": r.interval_right,
-                "mean_bits": r.mean,
-                "median_bits": r.median,
-                "std_bits": r.std,
-                "empirical_lipschitz_max": r.empirical_lipschitz_max,
-                "empirical_lipschitz_q99": r.empirical_lipschitz_q99,
-                "normalized_proxy_max": r.normalized_proxy_max,
-                "normalized_proxy_q99": r.normalized_proxy_q99,
-            }
-            row.update(r.theory)
-            w.writerow(row)
-
-
-def plot_histogram(r: SystemResult, outdir: Path) -> None:
-    """Per-system histogram with interval and theory overlays when available."""
-    v = r.values
-    vmin, vmax = float(np.min(v)), float(np.max(v))
-    vr = max(vmax - vmin, 1e-9)
-    plt.figure(figsize=(8.5, 5.5))
-    plt.hist(v, bins=48, density=True, alpha=0.75)
-    plt.axvspan(r.interval_left, r.interval_right, alpha=0.18, label=f"shortest {(r.mass):.0%} interval")
-    plt.axvline(r.observable_max, linestyle="--", linewidth=2, label="observable upper bound")
-    plt.axvline(r.mean, linestyle="-.", linewidth=2, label="empirical mean")
-    if "page_average_bits" in r.theory:
-        plt.axvline(r.theory["page_average_bits"], linestyle=":", linewidth=2, label="Page average")
-    if "hayden_cutoff_bits" in r.theory:
-        plt.axvline(r.theory["hayden_cutoff_bits"], linewidth=2, label="Hayden cutoff")
-    plt.xlim(vmin - 0.1 * vr, vmax + 0.25 * vr)
-    plt.xlabel("Entropy observable (bits)")
-    plt.ylabel("Empirical density")
-    plt.title(r.label)
-    plt.legend(frameon=False)
-    plt.tight_layout()
-    plt.savefig(outdir / f"hist_{r.slug}.png", dpi=180)
-    plt.close()
-
-
-def plot_tail(r: SystemResult, space: MetricMeasureSpace, outdir: Path) -> None:
-    """Upper-tail plot for the entropy deficit from its natural ceiling."""
-    deficits = r.observable_max - np.sort(r.values)
-    n = len(deficits)
-    ccdf = np.maximum(1.0 - (np.arange(1, n + 1) / n), 1.0 / n)
-    x = np.linspace(0.0, max(float(np.max(deficits)), 1e-6), 256)
-    plt.figure(figsize=(8.5, 5.5))
-    plt.semilogy(deficits, ccdf, marker="o", linestyle="none", markersize=3, alpha=0.45, label="empirical tail")
-    bound = space.tail_bound(x)
-    if bound is not None:
-        plt.semilogy(x, bound, linewidth=2, label="theory bound")
-    plt.xlabel("Entropy deficit (bits)")
-    plt.ylabel("Tail probability")
-    plt.title(f"Tail plot: {r.label}")
-    plt.legend(frameon=False)
-    plt.tight_layout()
-    plt.savefig(outdir / f"tail_{r.slug}.png", dpi=180)
-    plt.close()
-
-
-def plot_family_summary(results: Sequence[SystemResult], family: str, outdir: Path) -> None:
-    """Original-style summary plots, one family at a time."""
-    rs = sorted([r for r in results if r.family == family], key=lambda z: z.intrinsic_dim)
-    if not rs:
-        return
-    x = np.array([r.intrinsic_dim for r in rs], float)
-    pd = np.array([r.partial_diameter for r in rs], float)
-    sd = np.array([r.std for r in rs], float)
-    md = np.array([r.observable_max - r.mean for r in rs], float)
-
-    plt.figure(figsize=(8.5, 5.5))
-    plt.plot(x, pd, marker="o", linewidth=2, label=r"shortest $(1-\kappa)$ interval")
-    plt.plot(x, sd, marker="s", linewidth=2, label="empirical std")
-    plt.plot(x, md, marker="^", linewidth=2, label="mean deficit")
-    plt.xlabel("Intrinsic dimension")
-    plt.ylabel("Bits")
-    plt.title(f"Concentration summary: {family}")
-    plt.legend(frameon=False)
-    plt.tight_layout()
-    plt.savefig(outdir / f"summary_{family}.png", dpi=180)
-    plt.close()
-
-    good = [r for r in rs if r.normalized_proxy_q99 == r.normalized_proxy_q99]
-    if good:
-        x = np.array([r.intrinsic_dim for r in good], float)
-        y1 = np.array([r.normalized_proxy_max for r in good], float)
-        y2 = np.array([r.normalized_proxy_q99 for r in good], float)
-        plt.figure(figsize=(8.5, 5.5))
-        plt.plot(x, y1, marker="o", linewidth=2, label="width / Lipschitz max")
-        plt.plot(x, y2, marker="s", linewidth=2, label="width / Lipschitz q99")
-        plt.xlabel("Intrinsic dimension")
-        plt.ylabel("Normalized proxy")
-        plt.title(f"Lipschitz-normalized proxy: {family}")
-        plt.legend(frameon=False)
-        plt.tight_layout()
-        plt.savefig(outdir / f"normalized_{family}.png", dpi=180)
-        plt.close()
-
-
-def plot_cross_space_comparison(results: Sequence[SystemResult], outdir: Path) -> None:
-    """Direct comparison of the three spaces on one figure."""
-    marks = {"sphere": "o", "cp": "s", "majorana": "^"}
-
-    plt.figure(figsize=(8.8, 5.6))
-    for fam in ("sphere", "cp", "majorana"):
-        rs = sorted([r for r in results if r.family == fam], key=lambda z: z.intrinsic_dim)
-        if rs:
-            plt.plot([r.intrinsic_dim for r in rs], [r.partial_diameter for r in rs], marker=marks[fam], linewidth=2, label=fam)
-    plt.xlabel("Intrinsic dimension")
-    plt.ylabel("Partial diameter in bits")
-    plt.title("Entropy-based observable-diameter proxy: raw width comparison")
-    plt.legend(frameon=False)
-    plt.tight_layout()
-    plt.savefig(outdir / "compare_partial_diameter.png", dpi=180)
-    plt.close()
-
-    plt.figure(figsize=(8.8, 5.6))
-    for fam in ("sphere", "cp", "majorana"):
-        rs = sorted([r for r in results if r.family == fam and r.normalized_proxy_q99 == r.normalized_proxy_q99], key=lambda z: z.intrinsic_dim)
-        if rs:
-            plt.plot([r.intrinsic_dim for r in rs], [r.normalized_proxy_q99 for r in rs], marker=marks[fam], linewidth=2, label=fam)
-    plt.xlabel("Intrinsic dimension")
-    plt.ylabel("Normalized proxy")
-    plt.title("Entropy-based observable-diameter proxy: normalized comparison")
-    plt.legend(frameon=False)
-    plt.tight_layout()
-    plt.savefig(outdir / "compare_normalized_proxy.png", dpi=180)
-    plt.close()
-
-
-def plot_majorana_stars(space: MetricMeasureSpace, states: np.ndarray, outdir: Path) -> None:
-    """Scatter Majorana stars in longitude/latitude coordinates."""
-    if not hasattr(space, "majorana_stars") or len(states) == 0:
-        return
-    pts = np.vstack([space.majorana_stars(s) for s in states])
-    x, y, z = pts[:, 0], pts[:, 1], np.clip(pts[:, 2], -1.0, 1.0)
-    lon, lat = np.arctan2(y, x), np.arcsin(z)
-    plt.figure(figsize=(8.8, 4.6))
-    plt.scatter(lon, lat, s=10, alpha=0.35)
-    plt.xlim(-math.pi, math.pi)
-    plt.ylim(-math.pi / 2, math.pi / 2)
-    plt.xlabel("longitude")
-    plt.ylabel("latitude")
-    plt.title(f"Majorana stars: {space.label}")
-    plt.tight_layout()
-    plt.savefig(outdir / f"majorana_stars_{space.slug}.png", dpi=180)
-    plt.close()

codes/experiment_v0.2/spaces.py

@@ -63,7 +63,9 @@ class MetricMeasureSpace:
     def observable_max(self) -> float:
         raise NotImplementedError
 
-    def sample_np(self, rng: np.random.Generator, batch: int) -> tuple[np.ndarray, np.ndarray]:
+    def sample_np(
+        self, rng: np.random.Generator, batch: int
+    ) -> tuple[np.ndarray, np.ndarray]:
         raise NotImplementedError
 
     def sample_jax(self, key: Any, batch: int) -> tuple[Any, Any]:
@@ -106,7 +108,9 @@ class UnitSphereSpace(MetricMeasureSpace):
     def observable_max(self) -> float:
         return math.log2(self.dim)
 
-    def sample_np(self, rng: np.random.Generator, batch: int) -> tuple[np.ndarray, np.ndarray]:
+    def sample_np(
+        self, rng: np.random.Generator, batch: int
+    ) -> tuple[np.ndarray, np.ndarray]:
         x = rng.normal(size=(batch, self.dim)).astype(np.float32)
         x /= np.linalg.norm(x, axis=1, keepdims=True)
         return x, entropy_bits_from_probs(x * x, np).astype(np.float32)
@@ -154,8 +158,29 @@ class ComplexProjectiveSpace(MetricMeasureSpace):
     def observable_max(self) -> float:
         return math.log2(self.d_a)
 
-    def sample_np(self, rng: np.random.Generator, batch: int) -> tuple[np.ndarray, np.ndarray]:
-        g = (rng.normal(size=(batch, self.d_a, self.d_b)) + 1j * rng.normal(size=(batch, self.d_a, self.d_b)))
+    def sample_np(
+        self, rng: np.random.Generator, batch: int
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """
+        Sample haars-random pure states on C^(d_A d_B), observable = entanglement entropy.
+
+        Parameters
+        ----------
+        rng : np.random.Generator
+            Random number generator.
+        batch : int
+            Number of samples to generate.
+
+        Returns
+        -------
+        x : np.ndarray
+            Shape (batch, d_a * d_b), complex64.
+        y : np.ndarray
+            Shape (batch,), float32.
+        """
+        g = rng.normal(size=(batch, self.d_a, self.d_b)) + 1j * rng.normal(
+            size=(batch, self.d_a, self.d_b)
+        )
         g = (g / math.sqrt(2.0)).astype(np.complex64)
         g /= np.sqrt(np.sum(np.abs(g) ** 2, axis=(1, 2), keepdims=True))
         rho = g @ np.swapaxes(np.conj(g), 1, 2)
@@ -164,8 +189,10 @@ class ComplexProjectiveSpace(MetricMeasureSpace):
 
     def sample_jax(self, key: Any, batch: int) -> tuple[Any, Any]:
         k1, k2 = random.split(key)
-        g = (random.normal(k1, (batch, self.d_a, self.d_b), dtype=jnp.float32)
-             + 1j * random.normal(k2, (batch, self.d_a, self.d_b), dtype=jnp.float32)) / math.sqrt(2.0)
+        g = (
+            random.normal(k1, (batch, self.d_a, self.d_b), dtype=jnp.float32)
+            + 1j * random.normal(k2, (batch, self.d_a, self.d_b), dtype=jnp.float32)
+        ) / math.sqrt(2.0)
         g = g / jnp.sqrt(jnp.sum(jnp.abs(g) ** 2, axis=(1, 2), keepdims=True))
         rho = g @ jnp.swapaxes(jnp.conj(g), -1, -2)
         lam = jnp.clip(jnp.linalg.eigvalsh(rho).real, 1e-30, 1.0)
@@ -177,7 +204,9 @@ class ComplexProjectiveSpace(MetricMeasureSpace):
     def theory(self, kappa: float) -> dict[str, float]:
         d = self.d_a * self.d_b
         beta = self.d_a / (math.log(2.0) * self.d_b)
-        alpha = (math.log2(self.d_a) / math.sqrt(HAYDEN_C * (d - 1.0))) * math.sqrt(math.log(1.0 / kappa))
+        alpha = (math.log2(self.d_a) / math.sqrt(HAYDEN_C * (d - 1.0))) * math.sqrt(
+            math.log(1.0 / kappa)
+        )
         tail = sum(1.0 / k for k in range(self.d_b + 1, d + 1))
         page = (tail - (self.d_a - 1.0) / (2.0 * self.d_b)) / math.log(2.0)
         return {
@@ -193,7 +222,12 @@ class ComplexProjectiveSpace(MetricMeasureSpace):
     def tail_bound(self, deficits: np.ndarray) -> np.ndarray:
         beta = self.d_a / (math.log(2.0) * self.d_b)
         shifted = np.maximum(np.asarray(deficits, float) - beta, 0.0)
-        expo = -(self.d_a * self.d_b - 1.0) * HAYDEN_C * shifted**2 / (math.log2(self.d_a) ** 2)
+        expo = (
+            -(self.d_a * self.d_b - 1.0)
+            * HAYDEN_C
+            * shifted**2
+            / (math.log2(self.d_a) ** 2)
+        )
         out = np.exp(expo)
         out[deficits <= beta] = 1.0
         return np.clip(out, 0.0, 1.0)
@@ -230,7 +264,13 @@ class MajoranaSymmetricSpace(MetricMeasureSpace):
         k = np.arange(self.N + 1, dtype=np.float32)
         p = np.abs(c) ** 2
         rho11 = (p * k).sum(axis=1) / self.N
-        coef = np.sqrt((np.arange(self.N, dtype=np.float32) + 1.0) * (self.N - np.arange(self.N, dtype=np.float32))) / self.N
+        coef = (
+            np.sqrt(
+                (np.arange(self.N, dtype=np.float32) + 1.0)
+                * (self.N - np.arange(self.N, dtype=np.float32))
+            )
+            / self.N
+        )
         off = (np.conj(c[:, :-1]) * c[:, 1:] * coef).sum(axis=1)
         rho = np.zeros((len(c), 2, 2), dtype=np.complex64)
         rho[:, 0, 0] = 1.0 - rho11
@@ -253,8 +293,12 @@ class MajoranaSymmetricSpace(MetricMeasureSpace):
         rho = rho.at[:, 1, 0].set(jnp.conj(off))
         return rho
 
-    def sample_np(self, rng: np.random.Generator, batch: int) -> tuple[np.ndarray, np.ndarray]:
-        c = (rng.normal(size=(batch, self.N + 1)) + 1j * rng.normal(size=(batch, self.N + 1)))
+    def sample_np(
+        self, rng: np.random.Generator, batch: int
+    ) -> tuple[np.ndarray, np.ndarray]:
+        c = rng.normal(size=(batch, self.N + 1)) + 1j * rng.normal(
+            size=(batch, self.N + 1)
+        )
         c = (c / math.sqrt(2.0)).astype(np.complex64)
         c /= np.linalg.norm(c, axis=1, keepdims=True)
         lam = np.clip(np.linalg.eigvalsh(self._rho1_np(c)).real, 1e-30, 1.0)
@@ -262,8 +306,10 @@ class MajoranaSymmetricSpace(MetricMeasureSpace):
 
     def sample_jax(self, key: Any, batch: int) -> tuple[Any, Any]:
         k1, k2 = random.split(key)
-        c = (random.normal(k1, (batch, self.N + 1), dtype=jnp.float32)
-             + 1j * random.normal(k2, (batch, self.N + 1), dtype=jnp.float32)) / math.sqrt(2.0)
+        c = (
+            random.normal(k1, (batch, self.N + 1), dtype=jnp.float32)
+            + 1j * random.normal(k2, (batch, self.N + 1), dtype=jnp.float32)
+        ) / math.sqrt(2.0)
         c = c / jnp.linalg.norm(c, axis=1, keepdims=True)
         lam = jnp.clip(jnp.linalg.eigvalsh(self._rho1_jax(c)).real, 1e-30, 1.0)
         return c, entropy_bits_from_probs(lam, jnp)
@@ -273,11 +319,19 @@ class MajoranaSymmetricSpace(MetricMeasureSpace):
 
     def majorana_stars(self, coeffs: np.ndarray) -> np.ndarray:
         """Map one symmetric state to its Majorana stars on S^2."""
-        a = np.array([((-1) ** k) * math.sqrt(math.comb(self.N, k)) * coeffs[k] for k in range(self.N + 1)], np.complex128)
+        a = np.array(
+            [
+                ((-1) ** k) * math.sqrt(math.comb(self.N, k)) * coeffs[k]
+                for k in range(self.N + 1)
+            ],
+            np.complex128,
+        )
         poly = np.trim_zeros(a[::-1], trim="f")
         roots = np.roots(poly) if len(poly) > 1 else np.empty(0, dtype=np.complex128)
         r2 = np.abs(roots) ** 2
-        pts = np.c_[2 * roots.real / (1 + r2), 2 * roots.imag / (1 + r2), (r2 - 1) / (1 + r2)]
+        pts = np.c_[
+            2 * roots.real / (1 + r2), 2 * roots.imag / (1 + r2), (r2 - 1) / (1 + r2)
+        ]
         missing = self.N - len(pts)
         if missing > 0:
             pts = np.vstack([pts, np.tile(np.array([[0.0, 0.0, 1.0]]), (missing, 1))])

latex/chapters/chap2.tex

@@ -175,7 +175,31 @@ In this section, we will try to use the results from previous sections to estima
 
 From the previous discussion, we see that the only remaining for finding observable diameter of $\C P^n$ is to find the lipchitz function that is isometric with consistent push-forward measure.
 
-To find such metric, we need some additional results from previous sections.
+To find such metric, we need some additional results.
+
+\begin{defn}
+  \label{defn:riemannian-metric}
+
+  Let $M$ be a smooth manifold. A \textit{\textbf{Riemannian metric}} on $M$ is a smooth covariant tensor field $g\in \mathcal{T}^2(M)$ such that for each $p\in M$, $g_p$ is an inner product on $T_pM$.
+
+  $g_p(v,v)\geq 0$ for each $p\in M$ and each $v\in T_pM$. equality holds if and only if $v=0$.
+
+\end{defn}
+
+% TODO: There is a hidden chapter on group action on manifolds, can you find that?
+
+\begin{theorem}
+  \label{theorem:riemannian-submersion}
+
+  Let $(\tilde{M},\tilde{g})$ be a Riemannian manifold, let $\pi:\tilde{M}\to M$ be a surjective smooth submersion, and let $G$ be a group acting on $\tilde{M}$. If the \textbf{action} is
+  \begin{enumerate}
+    \item isometric: the map $x\mapsto \varphi\cdot x$ is an isometry for each $\varphi\in G$.
+    \item vertical: every element $\varphi\in G$ takes each fiber to itself, that is $\pi(\varphi\cdot p)=\pi(p)$ for all $p\in \tilde{M}$.
+    \item transitive on fibers: for each $p,q\in \tilde{M}$ such that $\pi(p)=\pi(q)$, there exists $\varphi\in G$ such that $\varphi\cdot p = q$.
+  \end{enumerate}
+  Then there is a unique Riemannian metric on $M$ such that $\pi$ is a Riemannian submersion.
+
+\end{theorem}
 
 A natural measure for $\C P^n$ is the normalized volume measure on $\C P^n$ induced from the Fubini-Study metric. \cite{lee_introduction_2018} Example 2.30
 
@@ -233,8 +257,175 @@ Using the projection map and Hopf's fibration, we can estimate the observable di
 
 \section{Use entropy function as estimator of observable diameter for complex projective spaces}
 
-In this section, we wish to use observable diameter to estimate the statics of thermal dynamics of some classical systems.
+In this section we describe a Monte Carlo pipeline for comparing concentration phenomena across three metric-measure spaces using real-valued entropy observables. The goal is not to compute the exact observable diameter
+$$
+\operatorname{ObsDiam}_{\mathbb{R}}(X;-\kappa)
+=
+\sup_{f \in \operatorname{Lip}_1(X,\mathbb{R})}
+\operatorname{diam}(f_*\mu_X;1-\kappa),
+$$
+but to estimate it by choosing a specific observable $f:X\to \mathbb{R}$ and then measuring the partial diameter of its push-forward distribution. Thus all numerical quantities below should be interpreted as \emph{entropy-based observable-diameter proxies}, not exact observable diameters in Gromov's sense \cite{MGomolovs,shioya2014metricmeasuregeometry}.
 
+The screen is $\mathbb{R}$ equipped with the Euclidean metric, and for a fixed $\kappa \in (0,1)$ we set
+$$
+\alpha = 1-\kappa.
+$$
+Given sampled values $y_1,\dots,y_N \in \mathbb{R}$ of the observable, the code sorts them and computes the shortest interval $[a,b]$ containing at least $\lceil \alpha N \rceil$ samples. Its width
+$$
+b-a
+$$
+is the empirical partial diameter of the push-forward measure on $\mathbb{R}$.
+
+To compare this width with the true observable diameter, the code also estimates an empirical Lipschitz constant of the chosen observable. If $x_i,x_j \in X$ are sampled states and $f(x_i),f(x_j)$ are the corresponding observable values, then the sampled slopes are
+$$
+\frac{|f(x_i)-f(x_j)|}{d_X(x_i,x_j)},
+$$
+where $d_X$ is the metric of the ambient space. The code records both the maximum sampled slope and the $0.99$-quantile of these slopes. Dividing the empirical partial diameter by these sampled Lipschitz constants gives two normalized proxies:
+$$
+\frac{\operatorname{diam}(f_*\mu_X;1-\kappa)}{L_{\max}}
+\qquad \text{and} \qquad
+\frac{\operatorname{diam}(f_*\mu_X;1-\kappa)}{L_{0.99}}.
+$$
+If the chosen observable were exactly $1$-Lipschitz, these normalized quantities would coincide with the raw width. In practice they should be viewed only as heuristic lower-scale corrections.
+
+\subsection{Random sampling using standard uniform measure on the unit sphere}
+
+The first family of spaces is the real unit sphere
+$$
+S^{m-1}
+=
+\left\{
+x=(x_1,\dots,x_m)\in \mathbb{R}^m : \|x\|_2=1
+\right\},
+$$
+equipped with the geodesic distance
+$$
+d_{S}(x,y)=\arccos \langle x,y\rangle
+$$
+and the normalized Riemannian volume measure. This is the standard metric-measure structure used in concentration of measure on spheres \cite{lee_introduction_2018,romanvershyni,shioya2014metricmeasuregeometry}.
+
+Sampling is performed by drawing a standard Gaussian vector $g\in \mathbb{R}^m$ and normalizing:
+$$
+x=\frac{g}{\|g\|_2}.
+$$
+This produces the uniform distribution on $S^{m-1}$.
+
+The observable is a Shannon entropy built from the squared coordinates:
+$$
+f_{\mathrm{sphere}}(x)
+=
+-\sum_{i=1}^m x_i^2 \log_2(x_i^2).
+$$
+Since $(x_1^2,\dots,x_m^2)$ is a probability vector, $f_{\mathrm{sphere}}$ takes values in $[0,\log_2 m]$, and the code records $\log_2 m$ as the natural upper bound of the observable.
+
+For each chosen dimension $m$, the experiment generates $N$ independent samples $x^{(1)},\dots,x^{(N)}$, computes the values
+$$
+f_{\mathrm{sphere}}(x^{(1)}),\dots,f_{\mathrm{sphere}}(x^{(N)}),
+$$
+and then evaluates the shortest interval containing mass at least $1-\kappa$. This gives an empirical observable-diameter proxy for the sphere family. The code also computes the empirical mean, median, standard deviation, and the normalized proxies obtained from sampled Lipschitz ratios.
+
+\subsection{Visualized the concentration of measure phenomenon on complex projective space}
+
+The second family is complex projective space
+$$
+\mathbb{C}P^{d_A d_B-1},
+$$
+viewed as the space of pure states in $\mathbb{C}^{d_A}\otimes \mathbb{C}^{d_B}$ modulo global phase. Geometrically, this space is equipped with the Fubini--Study metric and its associated normalized volume measure \cite{lee_introduction_2018,Bengtsson_Zyczkowski_2017}. Numerically, a projective point is represented by a unit vector
+$$
+\psi \in \mathbb{C}^{d_A d_B},
+\qquad
+\|\psi\|=1,
+$$
+and distances are computed by the Fubini--Study formula
+$$
+d_{FS}([\psi],[\phi])
+=
+\arccos |\langle \psi,\phi\rangle|.
+$$
+
+Sampling is implemented by drawing a complex Gaussian matrix
+$$
+G \in \mathbb{C}^{d_A \times d_B},
+$$
+with independent standard complex normal entries, and then normalizing it so that
+$$
+\psi = \frac{\operatorname{vec}(G)}{\|\operatorname{vec}(G)\|}.
+$$
+This is equivalent to Haar sampling on the unit sphere in $\mathbb{C}^{d_A d_B}$ and hence induces the standard unitarily invariant measure on $\mathbb{C}P^{d_A d_B-1}$ \cite{Bengtsson_Zyczkowski_2017,Nielsen_Chuang_2010}.
+
+The real-valued observable is the bipartite entanglement entropy. Writing
+$$
+\rho_A = \operatorname{Tr}_B |\psi\rangle\langle \psi|,
+$$
+the code defines
+$$
+f_{\mathrm{CP}}([\psi])
+=
+S(\rho_A)
+=
+-\operatorname{Tr}(\rho_A \log_2 \rho_A).
+$$
+Equivalently, if $\lambda_1,\dots,\lambda_{d_A}$ are the eigenvalues of $\rho_A$, then
+$$
+f_{\mathrm{CP}}([\psi])
+=
+-\sum_{i=1}^{d_A}\lambda_i \log_2 \lambda_i.
+$$
+This observable takes values in $[0,\log_2 d_A]$.
+
+For each dimension pair $(d_A,d_B)$, the experiment samples $N$ independent Haar-random pure states, computes the entropy values, and then forms the empirical push-forward distribution on $\mathbb{R}$. The shortest interval containing mass at least $1-\kappa$ is reported as the entropy-based observable-diameter proxy. In addition, the code plots histograms, upper-tail deficit plots for
+$$
+\log_2 d_A - S(\rho_A),
+$$
+and family-wise comparisons of partial diameter, standard deviation, and mean deficit. When available, these plots are overlaid with the Page average entropy and with Hayden-style concentration scales, which serve as theoretical guides rather than direct outputs of the simulation \cite{Hayden,Hayden_2006,Pages_conjecture_simple_proof}.
+
+\subsection{Random sampling using Majorana Stellar representation}
+
+The third family is the symmetric subspace
+$$
+\operatorname{Sym}^N(\mathbb{C}^2),
+$$
+which is naturally identified with $\mathbb{C}P^N$ after projectivization. In this model, a pure symmetric $N$-qubit state is written in the Dicke basis as
+$$
+|\psi\rangle
+=
+\sum_{k=0}^{N} c_k |D^N_k\rangle,
+\qquad
+\sum_{k=0}^{N}|c_k|^2 = 1.
+$$
+The projective metric is again the Fubini--Study metric
+$$
+d_{FS}([\psi],[\phi])=\arccos |\langle \psi,\phi\rangle|.
+$$
+
+Sampling is performed by drawing a standard complex Gaussian vector
+$$
+(c_0,\dots,c_N)\in \mathbb{C}^{N+1}
+$$
+and normalizing it. This gives the unitarily invariant measure on the projective symmetric state space.
+
+The observable used by the code is the one-particle entropy of the symmetric state. From the coefficient vector $(c_0,\dots,c_N)$ one constructs the one-qubit reduced density matrix $\rho_1$, and then defines
+$$
+f_{\mathrm{Maj}}([\psi])
+=
+S(\rho_1)
+=
+-\operatorname{Tr}(\rho_1 \log_2 \rho_1).
+$$
+Since $\rho_1$ is a qubit state, this observable takes values in $[0,1]$.
+
+To visualize the same states in Majorana form, the code also associates to a sampled symmetric state its Majorana polynomial and computes its roots. After stereographic projection, these roots define $N$ points on $S^2$, called the Majorana stars \cite{Bengtsson_Zyczkowski_2017}. The resulting star plots are included only as geometric visualizations; they are not used to define the metric or the observable. The metric-measure structure used in the actual simulation remains the Fubini--Study metric and the unitarily invariant measure on the projective symmetric state space.
+
+Thus, for each $N$, the simulation produces:
+\begin{enumerate}
+  \item a sample of symmetric states,
+  \item the corresponding one-body entropy values,
+  \item the shortest interval containing mass at least $1-\kappa$ in the push-forward distribution on $\mathbb{R}$,
+  \item empirical Lipschitz-normalized versions of this width,
+  \item and a separate Majorana-star visualization of representative samples.
+\end{enumerate}
+
+Taken together, these three families allow us to compare how entropy-based concentration behaves on a real sphere, on a general complex projective space carrying bipartite entanglement entropy, and on the symmetric subspace described by Majorana stellar data.
 
 
 \ifSubfilesClassLoaded{