关于「标准库 DPLL:单元传播、分支与「易、难、易」」,《规则叠加的悬崖》给出的定量结论是什么?

结论:DPLL 用赋值化简、单元传播与分支回溯判定规则册有没有解,单元传播推出的步骤本身就是一份可读的死锁说明。p21 显示,一本只有八个开关的规则册,第二周加一条规则后约束比只是 1.375,单元传播七步就证明了死锁,因为单开关硬规定把其余规则钉死了。p22 复现了「易、难、易」:死路数的中位数从 0.0 升到 26.5 再回落到 9.0。p23 显示单元传播把工作量压低两个数量级。可是,悬崖究竟在哪里,能不能不做实验就算出一个界?下一章用解数的期望、2-SAT 的精确阈值与子句长度给出答案。

数据口径:书中规则册与数据均为示意;全部计算由书中程序生成;本书属非个人化经营方法,不构成法律、合规或经营决策意见。 第四章给一本 24 个开关的示意规则册做体检:裸约束比 0.958,载荷 0.475,死锁时约束比中位数只有 1.292,最小冲突核横跨三个规则族。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.0, 26.5, 9.0

n = 50, 60 formulas per ratio; medians and 90th percentiles
alphaP(sat)guessesdead endsdead ends p90max
2.01.00024.00.00.00
3.01.00018.00.01.04
3.61.00016.01.09.948
4.00.83318.55.538.556
4.30.60022.516.046.964
4.60.30027.026.544.663
5.00.01725.026.038.944
6.00.00016.017.025.030
8.00.0009.510.514.019
10.00.0008.09.011.915
查看完整验算程序
# Chapter 2. p22: easy-hard-easy - search effort peaks at the cliff
# Model: random 3-SAT on n = 50 switches, 60 formulas per ratio
# (illustrative), DPLL with unit propagation and a weighted branching
# rule. Effort = number of guesses (branch points) the solver opens.
# Under-constrained sets are solved almost without guessing, heavily
# over-constrained sets are refuted quickly, and the median effort
# peaks near the 50 per cent point (Mitchell et al. 1992).

import random
import statistics
from collections import Counter

N, TRIALS = 50, 60
ALPHAS = [2.0, 3.0, 3.6, 4.0, 4.3, 4.6, 5.0, 6.0, 8.0, 10.0]


def assign(cls, lit):
    out = []
    for c in cls:
        if lit in c:
            continue
        if -lit in c:
            c = [x for x in c if x != -lit]
            if not c:
                return None
        out.append(c)
    return out


def dpll(cls, st):
    while True:
        unit = next((c[0] for c in cls if len(c) == 1), None)
        if unit is None:
            break
        cls = assign(cls, unit)
        if cls is None:
            st[1] += 1
            return False
    if not cls:
        return True
    st[0] += 1
    w = Counter()
    for c in cls:
        for x in c:
            w[x] += 1 << max(0, 6 - len(c))
    lit = max(w, key=lambda x: (w[x] + w[-x], w[x], -abs(x), x))
    for b in (lit, -lit):
        nxt = assign(cls, b)
        if nxt is None:
            st[1] += 1
        elif dpll(nxt, st):
            return True
    return False


rng = random.Random(20261115)
print("n = %d, %d formulas per ratio; medians and 90th percentiles"
      % (N, TRIALS))
print("alpha  P(sat)  guesses  dead ends  dead ends p90  max")
for a in ALPHAS:
    m = round(a * N)
    efforts, dead, ok = [], [], 0
    for _ in range(TRIALS):
        f = []
        for _ in range(m):
            vs = rng.sample(range(1, N + 1), 3)
            f.append([v if rng.random() < 0.5 else -v for v in vs])
        st = [0, 0]
        if dpll(f, st):
            ok += 1
        efforts.append(st[0])
        dead.append(st[1])
    q = statistics.quantiles(dead, n=10)
    print("%5.1f  %6.3f  %7.1f  %9.1f  %13.1f  %4d"
          % (a, ok / TRIALS, statistics.median(efforts),
             statistics.median(dead), q[-1], max(dead)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《布尔可满足性相变阈值、促销合规与门禁叠加的约束比临界与全盘死锁预警母典》第 2 章「标准库 DPLL:单元传播、分支与「易、难、易」」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-252 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026252,
  author    = {Fan, Yafu},
  title     = {Boolean Satisfiability Phase Transitions: the Critical Constraint Ratio at Which Stacked Promotion, Compliance and Gate Rules Deadlock, and How to See It Coming},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论