What does The Limits of Guardrail Compression conclude, in numbers, about “The Axis: Cliffs, Concave Curves and the Straight Line under Deterministic Labels”?

Answer: Chapter 2 quantifies the cliff: the cliff in share, 0.331, is 41.7 times the earlier average fall.

Scope of the numbers: Chapter 1 builds an illustrative source of eight kinds of draft and three verdicts and the counting bound: three rules keep 0.960 of the verdict information, two only 0.629. Chapter 4 shows that the critical exchange rate is the reciprocal of the squared maximal correlation, 0.6844 and 1.4610 on the illustrative source while information rises continuously as rules split.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.331, 41.7

epsI(X;Y)share3share2cliffslideratioacc3acc2
0.00.75340.9600.6290.3310.00841.70.8050.678
0.10.58850.9640.6270.3370.00747.30.7580.644
0.20.45520.9670.6260.3410.00751.90.7110.609
0.30.34410.9690.6250.3440.00656.10.6630.575
0.50.17440.9730.6230.3500.00563.90.5690.506
0.80.02860.9760.6200.3560.00574.10.4280.402
Show the full program
# Chapter 2. p22: how big is the cliff? Best hard k-rule groupings of the
# eight draft kinds (all 4140 partitions) as the reviewer's labels are
# blurred by noise eps: p(y|x) -> (1 - eps) p(y|x) + eps / 3.
# cliff = share of I(X;Y) lost going from 3 rules to 2;
# slide = mean share lost per rule removed from 8 rules down to 3.
# Source: the illustrative guardrail source of p11; eps is a scan.
import math

PX = [0.40, 0.18, 0.10, 0.07, 0.10, 0.06, 0.05, 0.04]
PYX = [[0.92, 0.06, 0.02], [0.25, 0.70, 0.05], [0.30, 0.55, 0.15],
       [0.10, 0.60, 0.30], [0.05, 0.15, 0.80], [0.03, 0.12, 0.85],
       [0.02, 0.08, 0.90], [0.01, 0.04, 0.95]]


def mutual(pxy):
    px = [math.fsum(r) for r in pxy]
    py = [math.fsum(c) for c in zip(*pxy)]
    return math.fsum(v * math.log2(v / (px[i] * py[j]))
                     for i, r in enumerate(pxy)
                     for j, v in enumerate(r) if v > 0)


def partitions(items):
    if not items:
        yield []
        return
    first, rest = items[0], items[1:]
    for part in partitions(rest):
        for i in range(len(part)):
            yield part[:i] + [[first] + part[i]] + part[i + 1:]
        yield [[first]] + part


PARTS = list(partitions(list(range(8))))


def curve(pyx):
    best_i, best_a = {}, {}
    for part in PARTS:
        k = len(part)
        pty = [[math.fsum(PX[x] * pyx[x][y] for x in g) for y in range(3)]
               for g in part]
        best_i[k] = max(best_i.get(k, 0.0), mutual(pty))
        best_a[k] = max(best_a.get(k, 0.0),
                        math.fsum(max(r) for r in pty))
    return best_i, best_a


print("  eps  I(X;Y)  share3  share2  cliff  slide  ratio  acc3   acc2")
for eps in [0.0, 0.1, 0.2, 0.3, 0.5, 0.8]:
    pyx = [[(1 - eps) * v + eps / 3 for v in r] for r in PYX]
    bi, ba = curve(pyx)
    full = bi[8]
    s3, s2 = bi[3] / full, bi[2] / full
    cliff = s3 - s2
    slide = (1.0 - s3) / 5
    print("%5.1f  %6.4f   %5.3f   %5.3f  %5.3f  %5.3f  %5.1f  %5.3f  %5.3f"
          % (eps, full, s3, s2, cliff, slide, cliff / slide,
             ba[3], ba[2]))

Single source and citation

Fan, Y. (2026). The Information Bottleneck and Rate-Distortion Limits: How Far a Bilingual Monograph Can Be Compressed into Guardrails before Discriminative Power Falls Away (Version 1.0.0), Chapter 2: The Axis: Cliffs, Concave Curves and the Straight Line under Deterministic Labels. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-250 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026250,
  author    = {Fan, Yafu},
  title     = {The Information Bottleneck and Rate-Distortion Limits: How Far a Bilingual Monograph Can Be Compressed into Guardrails before Discriminative Power Falls Away},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book