「合参越好」这句话有两种读法:一种恒真、一种可以为假,而真正卡住的是检测门槛

Written by

in

Language

Evidence level: C

I(X₁, X₂; Z) ≥ max{I(X₁; Z), I(X₂; Z)} 这个不等式是数学恒等式,不需要数据检验;真正需要检验的是「多一条通道到底多给多少比特」,而这个量在有限样本下几乎完全由「组合后有多少个格子」决定,不由机制本身决定。

正文

四诊合参的信息论表述,网上和文献里都爱写成一条不等式:多通道联合对证候的互信息,不会低于任何单通道。很多人把它当成一条「应该被实验证实」的规律,去做逐级加通道、看增益是否递增。这一步在方法上就跨错了门槛:那条不等式是链式法则的直接推论 I(X,Y;Z) = I(X;Z) + I(Y;Z|X),而条件互信息非负,所以并集互信息在数学上必然不减。它不可能被数据推翻,因此它也不可能被数据证实——测出来「成立」只是复述了一遍定义。把它写进预注册当假设,等于用数据给一条恒等式盖章。

真正有内容、而且可以为假的是另一个量:把三条通道各自单独携带的信息加起来,再减掉联合携带的信息,剩下的部分叫共信息(co-information)。三种融合结构的符号完全不同。若三条通道信息大体重复(比如脉象与舌象都由同一个潜在体质状态驱动),联合互信息约等于单通道互信息,共信息为正。若三条通道彼此条件独立且各自携带不同侧面,联合等于各自之和,共信息近零。若三条通道只有在联合看时才携带信息(经典例子:Z 是两路独立信号的异或,任一路单独与 Z 完全无关),那么单通道互信息为零而联合互信息满值,共信息为负——这就是协同。用合成数据摆一次,协同结构下两条通道各自与目标的互信息都是 0.000 bit,联合起来是 1.000 bit,共信息 −1.000 bit。所以在协同结构里,「单通道对」的信息不是「略少」,而是「一点没有」。这两种极端的差距,不是用「单调递增是否成立」能问出来的。

还有第二条硬约束:增益的递减是先天上界,不是待验证的经验事实。设四条通道各自独立地以同一准确率对目标做判别,通道之间混淆互不重叠,那么每多一条通道,联合互信息必然增加、但增加量必然越来越小。用合成数据核一遍,单通道准确率 0.40 时四条通道信息量依次 0.078 → 0.153 → 0.223 → 0.292 bit,增量 0.078 → 0.075 → 0.070 → 0.069;准确率 0.70 时增量 0.640 → 0.403 → 0.284 → 0.197。所以「严格单调递增」这个提法本身要改写:并集互信息单调不减(恒真),增益单调递减(也有上界保证),唯一可能为真的是「第四条通道还能不能测出增量」这个问题——它已经不在机制层面,而在估计层面。

估计层面才是真正的墙。直方图法算互信息,就是把经验频率代进公式再求对数,这个估计量有系统性上偏,偏差量为 (kx−1)(kz−1)/(2n ln2) bit(kx 是通道组合后的格数,kz 是证候数,n 是样本量)。注意这个偏差只与格数和样本量有关,与数据的真实分布完全无关——也就是说,只要加通道,格数就涨,偏差就涨,估计器会额外送你一段分数。这正是「合参总是更好」在实验里几乎从不失败的真实原因。

拿具体档位算一次。脉 3 档、舌 4 档、问 3 档、闻 3 档,四类证候,四条通道全叠起来的累积格数是 3 → 12 → 36 → 108。从 1 档到 2 档,格数 3→12,偏差系数是 (12−3)(4−1)/(2 ln2) = 19.5,即偏差增量为 19.5/n bit;从 2 档到 3 档系数 51.9;从 3 档到 4 档系数 155.8。而在一个「通道各有侧重、彼此不冗余」的真值模型里,三档真增量分别是 0.198、0.019、0.0018 bit。两相对照,偏差等于真增量的临界样本量是:第 1→2 档 99 例、第 2→3 档 2777 例、第 3→4 档 86562 例。要让偏差压到真增量的一成以内,分别是 985 例、27774 例、865617 例。也就是说,前两档在几百到几千例的常见队列里还想谈,第四档在现实样本量下原理上不可判——不是「研究没做好」,而是数据量不够付估计器的过路费。

实测验证也一致。在真增量为零的设定里(第三条是纯噪声通道),原地估计在 n = 100、200、500、1000、2000 五档下,把「加噪声通道」判成有增益的比例全是 100%。加上 Miller–Madow 型上偏修正后,n = 100 降到 0%,但 n = 500 反而升到 70.7%、n = 1000 升到 83.7%——修正项本身在格数大幅增长时会被过度校正,把真增量削成负数,

于是这条补救路线也塌了:修正后不是「偏差变小」,而是「偏差换了个方向」,从系统性高估翻成系统性低估。而一旦真增量被削成负数,任何双侧判据都会把它读成「与零显著不同」,判定结果与未修正时同样错,只是错法相反。

为什么修正项救不了:它瞄准的是错的量。 Miller–Madow 修的是 $(\hat{k}-1)/(2n\ln 2)$ 这一阶项,其中 $\hat{k}$ 是样本里实际观测到的非空格数,不是理论格数。这两者在稀疏样本下差得很远:理论 108 格、n = 500 时实际非空可能只有二三十格。于是修正项既依赖样本量、又依赖填充率,而填充率本身是数据的函数——修正量不再是数据分布的无关项,「只与格数和样本量有关」这个前提在真实数据上根本不成立。Paninski 系统比较过高阶修正与核方法,结论是 Miller–Madow 只在极小格数区间优于朴素式,格数一涨就交棒给别的估计器,而这正是四诊合参所处的区间(DOI 10.1162/089976603321780272)。

真正的出路有两条,都不在「修正」上。

第一条是换估计器。k 近邻类估计(Kozachenko–Leonenko 熵估计与 Kraskov 互信息估计)不做直方图分箱,因此没有格数膨胀这一项;Paninski 的结论是这类方法在「意外地小的样本区间」里仍然有效,且大范围优于 Miller–Madow 式(同上 DOI)。代价是它估的是连续量,要求把 3 档、4 档这类有序离散量先做合理编码,编码本身又引入选择自由度——这换掉了一个偏差,换来了一个自由度,必须在预注册里锁死。

第二条是换问法。别问「第四条通道多给多少 bit」,改问「第四条通道带来的增量,在本样本量下能不能被检出」。这是个检测门槛问题,答案可以直接算:给定 $n$ 与格数,估出该配置下的最小可检增量,再与真值模型的预期增量比。原文那张临界样本量表(99 / 2777 / 86562 例)就是这张门槛表的雏形——它的正确用法不是「判断研究做得好不好」,而是设计期就该拿来定样本量。Berisha 等对数字医学的维度诅咒做过系统论述:多模态高维数据是双刃剑,样本量与模型复杂度必须同时规划,否则性能估计本身不可信(PMID 34711924)。

判据四格。 一:某档位的实际样本量低于该档临界样本量的 3 倍 ⇒ 该档增益不可判,只能报「本样本量下无法区分」,不得报「无增益」——这两句话在统计上完全不同。二:改用 k-NN 估计器后,增量方向与直方图法一致、且效应量差距在一个数量级内 ⇒ 原结论稳健,可用于跨研究比较。三:两种估计器给出方向相反的增量 ⇒ 该档增益是估计器伪迹,此前的「合参更好」证据须按估计器重新分类。四:门槛表算出来某档在任何现实样本量下都不可判(如第 3→4 档需 8.6 万例)⇒ 该档应从「待验证」降级为「原理上不可测」,对应的研究提案一律撤除,不得继续消耗样本。

一句话收口:合参的增益不是测不出来,是在常见样本量下必然测不准;把这一点写进设计,比事后争论「修正项够不够好」有用得多。

相关阅读

证据与局限

已验证:链式法则保证并集互信息单调不减,条件互信息非负,此不等式恒真、不可被数据检验(信息论基本结论);直方图法估互信息存在与格数、样本量成正比的上偏,偏差量 $(k_x-1)(k_z-1)/(2n\ln 2)$(Paninski 2003 系统论述,DOI 10.1162/089976603321780272);Miller–Madow 一阶修正仅在极小格数区间优于朴素式,格数增大后失准(同上 DOI);k 近邻类估计器不做分箱、在意外小的样本区间仍有效(同上 DOI);多模态高维数据的样本量—复杂度必须联合规划,否则性能估计不可信(PMID 34711924)。

部分验证:四诊多模态融合的信息论路线已有实际尝试(PMID 21954670 用多信息融合做智能辨证;近年国内综述提出多粒度学习 + 多源融合框架),但均未报告「各通道增量 vs 样本量」的门槛分析,也未说明所用互信息估计器;互信息估计器的选型对下游推断结果影响显著,这一结论在基因调控网络推断中被系统验证(PMID 22242113 比较经验估计、Miller–Madow 等多种估计器,结论是估计器选择直接改变推断出的网络),但尚未在四诊数据上重复。

待验证:本文的三档临界样本量(99 / 2777 / 86562 例)由合成真值模型加解析偏差公式算出,未在真实四诊数据上校验;「改用 k-NN 估计器后方向是否保持」纯属预期,没有任何实测;Miller–Madow 修正后 n = 500 时 70.7% 的反向判读率出自本文合成实验,其机制(双侧判据 + 负数真增量)是推断而非直接测量,须复核;第 3→4 档「原理上不可测」的结论依赖「格数按 3→12→36→108 增长」这一编码假设,若改用连续编码或降维,该结论可能改写。该主题尚未进入 L1 对应网络,本文证据等级封顶 C。

参考文献: DOI 10.1162/089976603321780272, PMID 34711924, PMID 21954670, PMID 22242113


免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

合参越好:恒真与可假之间的检测门槛

Evidence level: C

I(X₁, X₂; Z) ≥ max{I(X₁; Z), I(X₂; Z)} is a mathematical identity, requiring no data validation; what actually needs testing is “how many bits are added by one extra channel,” and this quantity, under finite samples, is almost entirely determined by “how many cells result from combination,” not by the mechanism itself.

Body

The information-theoretic formulation of the Four-Examination Synthesis is frequently expressed online and in literature as a single inequality: the mutual information of the syndrome across multiple channels jointly is never lower than that of any single channel. Many treat this as a law “to be experimentally verified,” proceeding by adding channels sequentially to observe whether the gain increases. This step crosses the methodological threshold incorrectly: the inequality is a direct consequence of the chain rule, I(X,Y;Z) = I(X;Z) + I(Y;Z|X), and because conditional mutual information is non-negative, the joint mutual information is mathematically may support not to decrease. It cannot be refuted by data, and therefore it also cannot be confirmed by data—measuring that it “holds” merely restates the definition. Embedding it in a pre-registered hypothesis amounts to stamping an identity with data.

The quantity that carries substantive content and can be falsified is another: the sum of the information carried individually by each of the three channels, minus the information jointly carried. The remainder is called co-information. The symbols for the three fusion structures differ entirely. If the information across the three channels largely repeats (for example, pulse and tongue signs both driven by the same underlying constitutional state), the joint mutual information is approximately equal to the single-channel mutual information, and co-information is positive. If the three channels are conditionally independent of each other and each carries a different aspect, the joint equals the sum of the individuals, and co-information is near zero. If the three channels carry information only when viewed jointly (a classic example: Z is the XOR of two independent signals, so each individually is completely unrelated to Z), then the single-channel mutual information is zero while the joint mutual information reaches its maximum, and co-information is negative—this is synergy. Using synthetic data to illustrate once: under the synergistic structure, the mutual information of each pair of channels with the target is 0.000 bit, while jointly they yield 1.000 bit, and co-information is −1.000 bit. Thus, in the synergistic structure, the information of a “single-channel pair” is not “slightly less,” but “none at all.” The gap between these two extremes cannot be asked by whether “monotonic increase holds.”

There is a second hard constraint: the diminishing gain is an intrinsic upper bound, not an empirical fact to be tested. Suppose four channels independently discriminate the target with the same accuracy, and the confusion among them does not overlap. Then adding each additional channel necessarily increases the joint mutual information, but the amount of increase necessarily becomes smaller. Using synthetic data to verify: when single-channel accuracy is 0.40, the information quantities of four channels are successively 0.078 → 0.153 → 0.223 → 0.292 bit, with increments 0.078 → 0.075 → 0.070 → 0.069; when accuracy is 0.70, the increments are 0.640 → 0.403 → 0.284 → 0.197. Therefore the phrase “strictly monotonically increasing” itself must be rewritten: joint mutual information is monotonically non-decreasing (always true), the gain is monotonically decreasing (also may support by an upper bound), and the only question that could possibly be true is “can the fourth channel still measure an increment?”—this has already moved from the mechanism level to the estimation level.

The estimation level is the true wall. The histogram method computes mutual information by plugging empirical frequencies into the formula and taking logarithms; this estimator has a systematic upward bias, with bias equal to (kx−1)(kz−1)/(2n ln2) bit (where kx is the number of bins after channel combination, kz is the number of syndromes, and n is the sample size). Note that this bias depends only on the number of bins and the sample size, completely independent of the true data distribution—meaning that simply adding channels increases the number of bins, which increases the bias, and the estimator effectively grants extra points. This is precisely why “synthesis is always better” almost never fails in experiments.

Let us compute with specific tiers. Pulse has 3 bins, tongue 4, questioning 3, and smelling 3; with four syndromes, the cumulative number of bins when all four channels are stacked is 3 → 12 → 36 → 108. From tier 1 to tier 2, bins go from 3→12, and the bias coefficient is (12−3)(4−1)/(2 ln2) = 19.5, so the bias increment is 19.5/n bit; from tier 2 to tier 3 the coefficient is 51.9; from tier 3 to tier 4 it is 155.8. In a true model where “each channel has its own focus and they do not overlap,” the real increments across three tiers are 0.198, 0.019, and 0.0018 bit respectively. Comparing the two, the critical sample size at which bias equals the true increment is: tier 1→2 requires 99 cases, tier 2→3 requires 2777 cases, and tier 3→4 requires 86562 cases. To suppress the bias to within one-tenth of the true increment, the required sample sizes are 985, 27774, and 865617 respectively. Thus, for the first two tiers, discussion is still possible within common cohorts of a few hundred to a few thousand cases; but for the fourth tier, under realistic sample sizes, it is theoretically undecidable—not because “the study was not done well,” but because the data volume is insufficient to pay the estimator’s toll.

Experimental verification also agrees. Under the assumption of zero true increment (the third channel is pure noise), the raw estimator at n = 100, 200, 500, 1000, and 2000 always judges “adding the noise channel” as having gain, with a proportion of 100%. After applying the Miller–Madow upward-bias correction, n = 100 drops to 0%, but n = 500 rises to 70.7% and n = 1000 rises to 83.7%—the correction term itself, when the number of bins grows substantially, over-corrects and turns the true increment into a negative value.


Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *