Evidence level: C
互信息随通道增加永不下降,这是数学恒等式,不是发现;真正能检验的是”每加一路的增量是否为正”,而这一小步在 400 例以下基本测不准。
正文
四诊合参的现代论证里有一条被反复引用的”信息论保证”:把望、闻、问、切的读数合在一起算,联合互信息 I(证候;望,闻,问,切) 一定不小于任意单通道或双通道的互信息。字面上这句话永远成立,因为互信息对自变量集合是单调的——多给你一个变量,你不会变得更能猜错。所以它压根不需要做实验,也就不能当作”合参有必要”的证据。这是本文要先拆掉的第一层。
第二层才是可检验的:把”逐级增通道”的增量拆开看,每一步的边际增益 Δ_k = I(Y;X_1..X_k) − I(Y;X_1..X_(k-1)) 是不是严格为正、能不能被统计显著地区别于零、以及这个正增量是不是逐级变小。后一句(逐级变小)也有麻烦——互信息作为集合函数是子模的,边际增益递减同样是恒等式,等价于条件互信息不为负。换句话说”越加越不涨”是结构性质,不是经验性质。所以我用一个可解析的生成模型把这两层分开:模型是四通道三态(读到证候 / 读反 / 读不出),加一个受试者层面共同隐因子 G(整体可读性好 vs 差),给定 G 后四通道条件独立——这正是”部分冗余”的最小结构。
先看解析结果。四个场景(通道强 a=0.80、通道弱 a=0.45、强通道+高共源、通道强弱不均)里,逐通道边际增益全部严格为正且单调下降,如通道强场景 +0.3645、+0.1100、+0.0626、+0.0439 比特;子模性检查覆盖 76 组 (A⊂B, c∉B) 组合,违反 0 组。这不是模型配合得好,而是数学上不允许别的结果。于是”合参有增益但递减”在真值层面是恒真命题,任何声称”我们实证了边际递减”的论文,实际实证的是估计量的行为。
顺序这件事更该写进规程。同一组四通道(a 依次 0.85/0.70/0.50/0.30),先加强通道得增益轨迹 +0.3940、+0.0750、+0.0351、+0.0265;先加弱通道变成 +0.0942、+0.1686、+0.1614、+0.1064。两条轨迹完全不同,总量却相同(+0.5305)——终点 I 是同一个数。所以”第几路之后边际趋零”这个拐点不是数据的性质,而是加法顺序的性质。未预注册顺序的拐点结论不可复现。
再看估计层,这里才是真正的坏消息。离散 plug-in 互信息有系统性正偏差,且偏差随通道组合的状态数膨胀:N=50 时单通道偏差 +0.042,四通道偏差 +0.339 比特。于是”加通道涨了”可能纯是偏差涨了。Miller-Madow 校正在大样本下有效,在 N=50 时反而过校正到 −0.018,出现 30.4% 的回合报出负增量——一个真值严格为正的增量被算成负的。N=200 起负增量降到 2.1%,N=500 之后消失。
最刺眼的是零信息通道对照。第四路设为与证候、与隐因子全无关(三态等概率,真 Δ 严格为 0):plug-in 在 N=50 到 2000 全格上 100% 回合报出正增量,数值从 +0.131 衰到 +0.023 仍为正。一个彻底无信息的通道,两千例以内都会被”证明”合参有增益。换成 MM 校正后,N=400 时仍有 75.7% 的 bootstrap 区间排零(声称显著为正),N=800 起才 100% 排零,但点估计仍把真值 0 报成 +0.011——排零的机制是区间被估窄,不是点估计变准。这条对照是本文最硬的产出:它把”零信息通道拒绝率”变成一个可预注册的校准指标。
检测门槛则要按最小增量定,不是按总互信息定。真逐级 Δ 为 +0.1225、+0.0699、+0.0487 时,第二路在 N=100 已有 93.5% 排零率,而最弱的那一步在 N=100 只有 16.0%、N=200 才 76.5%、N=400 起 100%。所以”第几路之后增益趋于零”的定位问题,需要的样本量由最小 Δ 决定;用总量 I 的精度来规划样本量必定高估检出力。
判别层还有一个反直觉的口径问题。小训练集(120 例)朴素贝叶斯下,单通道 AUC 0.9324 反而高于双通道 0.9150,非单调,800 次重复里 97.1% 的回合 Δ2 为负。换成对隐因子求和的真·最优判别器,同一数据变成单调 0.9354→0.9382→0.9443→0.9541。但在”通道强弱不均”场景里,即便最优判别器也非单调(0.9565→0.9435→0.9445→0.9495),而真互信息一路单调涨到 0.5305。结论是明确的:互信息单调并不蕴含判别力单调,AUC 是互信息的非单调函数。所以申报口径若只报 MI,等于只答了半个问题。
外部文献把这条推论顶得更硬。Porter 等人 2023 年对 93 项精神病神经影像分类研究的元分析直接给出:所有模态都能区分患者与对照,但多模态与单模态之间未观察到任何显著差异,功能连接与单模态的效应分布大量重叠(PMID 37563277)。Mu 等人 2026 年的多组学渐进融合框架报告,四个基准数据集里有三个”超过 50% 的病例仅靠单一组学即可准确分类”,其核心机制正是”置信度够了就不再增加模态”(PMID 41319559)。方向上与本文仿真一致:合参的价值不在”必然涨”,而在”知道什么时候已经够了”。中医侧的可测性已有落点——体质多模态分类(PMID 36374441)、四诊辅助 apparatus 评价健康与疗效(PMID 32185979),说明缺的不是设备而是口径。
估计量选择上文献给的路很清楚:Zhang 与 Zheng 的指数衰减偏差估计量,其渐近正态性配合偏差快速衰减才使有限字母表推断可用,作者直陈极大似然(即 plug-in)互信息”极大抬高第一类错误概率”(PMID 25941916);Jiao 等人的极小极大框架给出换算:最优估计量在 n 个样本上约等于 plug-in 用 n ln n 个样本,而 plug-in 达不到该速率(PMID 29375152);连续变量须换几何 k 近邻路线(PMID 29604625)。三条合起来是硬约束:报 Δ 必须用低偏差估计量并同报 plug-in 作对照。
证据与局限
已验证(本文仿真内部):真值层面,四通道边际增益严格为正且递减、子模性 76 组检查零违反;顺序改变使增益轨迹完全不同但总量不变;plug-in 偏差随通道数膨胀(N=50 时 +0.042→+0.339);严格零信息通道在 N≤2000 全格 100% 被估计为正增量;MM 校正在 N=50 过校正并产生 30.4% 假负增量;最弱一步 Δ=+0.0487 需约 400 例才 100% 排零;AUC 在”通道强弱不均”下即使最优判别器也非单调,而 MI 单调。
部分验证:外部文献支持”多模态未必优于最佳单模态”(PMID 37563277 元分析、PMID 41319559 渐进融合),以及”低偏差估计量是必要条件”(PMID 25941916、PMID 29375152、PMID 29604625);但”四诊四通道”的真值结构、真实证候标签的判读者信度、真实通道可读性 a 的取值,均无同批数据支撑——本文的 a=0.45~0.85、pg=0.70~0.95 是设定,不是实测。
待验证:① 零信息通道拒绝率作为预注册校准指标,在真实四诊数据上的可复现性;② 真实标签下”通道强弱不均”是否普遍,若普遍则 AUC 非单调将成常态而非例外;③ 子模性恒真意味着”边际递减”永远不能作为发现,但”递减速率”是否可用于裁定”第几路够用”,需要先确认速率跨病种是否稳定(本文仅做单病种原型)。
该主题尚未进入 L1 对应网络。最近邻为 E-061 段 KB-509「冗余度/协同度(PID)」PMID 33266673,属呼吸-心率两通道偶联、非四诊四通道,且该条已明写口径警告「冗余 = 同一目标有两条通路可预测;协同 = 只有联合才有信息,两者方向可以相反」。
L1 网络状态:EDGE_TABLE / CORRESPONDENCE_NETWORK 检索「四诊 / 互信息 / 融合增益 / 增量 / 子模 / 单调」,命中全在心肺偶联语境(E-061 段 KB-509 冗余度/协同度 PID,PMID 33266673,两通道),本文「四诊四通道融合增益不等式的真值层与估计层分离」无对应边,未获 L1 状态判定,故证据等级封顶 C。
相关阅读
证据等级: C
参考文献: PMID 37563277, PMID 41319559, PMID 25941916, PMID 29375152, PMID 29604625, PMID 36374441, PMID 32185979
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。
The two layers of truth in \”合参必增益\”: the constant half does not count as evidence, while only a small half is testable.
Evidence level: C
Mutual information never decreases as channels increase; this is a mathematical identity, not an empirical discovery. What can actually be tested is whether “the incremental gain per added channel is positive,” yet this single step cannot be reliably measured with fewer than 400 cases.
Body
In modern arguments synthesizing the four diagnostic methods (inspection, listening/smelling, inquiry, palpation), a frequently cited “information-theoretic may support” states that the joint mutual information I(Symptom; Inspection, Listening, Inquiry, Palpation) is always greater than or equal to the mutual information of any single channel or pair of channels. Literally, this statement always holds because mutual information is monotonic with respect to the set of independent variables—adding another variable never makes you more likely to guess incorrectly. Therefore, it requires no experiment and cannot serve as evidence that “synthesis is necessary.” This is the first layer we must dismantle.
The second layer is testable: decompose the incremental gain of “adding channels step-by-step” and examine whether each marginal gain Δ_k = I(Symptom; X_1…X_k) − I(Symptom; X_1…X_(k-1)) is strictly positive, statistically distinguishable from zero, and whether this positive increment decreases step-by-step. The latter point (decreasing step-by-step) also presents difficulties—mutual information as a set function is submodular, so the diminishing marginal gain is an identity equivalent to conditional mutual information being non-negative. In other words, “adding more yields less growth” is a structural property, not an empirical one. Thus, I use an analytically tractable generative model to separate these two layers: the model features four channels and three states (read symptom / read opposite / unreadable), plus a subject-level common latent factor G (overall readability good vs. poor). Given G, the four channels are conditionally independent—this is precisely the minimal structure for “partial redundancy.”
First, the analytical results. Across four scenarios (strong channel a=0.80, weak channel a=0.45, strong channel with high shared source, uneven channel strengths), the marginal gain per channel is strictly positive and monotonically decreasing. For example, in the strong-channel scenario, gains are +0.3645, +0.1100, +0.0626, and +0.0439 bits; the submodularity check covers 76 combinations (A⊂B, c∉B) with zero violations. This is not due to good model fit but because other results are mathematically impossible. Hence, “synthesis yields gain but diminishing returns” is a tautology at the truth level; any paper claiming “we empirically demonstrated diminishing marginal returns” actually tested the behavior of the estimator.
The order of addition matters and should be specified in protocols. For the same four channels (a sequentially 0.85/0.70/0.50/0.30), adding strong channels first yields a gain trajectory of +0.3940, +0.0750, +0.0351, +0.0265; adding weak channels first yields +0.0942, +0.1686, +0.1614, +0.1064. The two trajectories differ completely, yet the total remains identical (+0.5305)—the final I is the same number. Therefore, the inflection point where “marginal gain approaches zero after the nth channel” is not a property of the data but of the addition order. Inflection points without pre-registered order are irreproducible.
Turning to the estimation layer, here lies the real bad news. Discrete plug-in mutual information exhibits systematic positive bias, which inflates with the number of states in the channel combination: at N=50, single-channel bias is +0.042, while four-channel bias reaches +0.339 bits. Thus, “adding channels increases gain” may be purely due to increased bias. Miller-Madow correction works well for large samples but over-corrects at N=50, yielding −0.018 and causing 30.4% of rounds to report negative increments—a strictly positive true increment estimated as negative. At N≥200, negative increments drop to 2.1%, and disappear entirely beyond N=500.
Most striking is the zero-information channel control. Setting the fourth channel to be completely independent of both symptom and latent factor (three states equally probable, true Δ strictly zero): the plug-in estimator reports positive increments in 100% of rounds across N=50 to 2,000, with values decaying from +0.131 to +0.023 yet remaining positive. A completely uninformative channel is “proven” to enhance synthesis within 2,000 cases. With MM correction, at N=400, 75.7% of bootstrap intervals still exclude zero (claiming significant positivity); only at N≥800 do 100% exclude zero, though the point estimate still reports the true value of 0 as +0.011—the mechanism excluding zero is narrower intervals due to estimation, not improved accuracy of the point estimate. This control is our strongest output: it turns “zero-information channel rejection rate” into a pre-registered calibration metric.
Detection thresholds must be set based on the minimum increment, not total mutual information. When true sequential Δ are +0.1225, +0.0699, +0.0487, the second channel achieves 93.5% zero-rejection rate at N=100, whereas the weakest step reaches only 16.0% at N=100, 76.5% at N=200, and 100% starting at N=400. Thus, the question of “after which channel does gain approach zero” depends on sample size determined by the smallest Δ; planning sample size based on total I precision inevitably overestimates detection power.
Discrimination Layer: The Counterintuitive Channel Width Issue
A counterintuitive calibration issue arises at the discrimination layer. Under naive Bayes with a small training set (120 cases), the single-channel AUC of 0.9324 actually exceeds the dual-channel value of 0.9150, violating monotonicity; among 800 repetitions, 97.1% of rounds showed a negative Δ2. Switching to the true optimal discriminator based on summing latent factors renders the same data monotonic: 0.9354→0.9382→0.9443→0.9541. However, in the “uneven channel strength” scenario, even the optimal discriminator remains non-monotonic (0.9565→0.9435→0.9445→0.9495), while true mutual information increases monotonically to 0.5305. The conclusion is clear: monotonicity of mutual information does not imply monotonicity of discriminative power, as AUC is a non-monotonic function of mutual information. Thus, reporting only MI constitutes answering only half the question.
External literature strengthens this inference. Porter et al. (2023) meta-analysis of 93 psychiatric neuroimaging classification studies directly reports: all modalities can distinguish patients from controls, but no significant difference was observed between multimodal and unimodal approaches, with effect distributions of functional connectivity heavily overlapping those of unimodal methods (PMID 37563277). Mu et al. (2026) multi-omics progressive fusion framework reports that three of four benchmark datasets achieve accurate classification using only a single omic, with the core mechanism being “stopping modal addition once confidence suffices” (PMID 41319559). The direction aligns with this simulation: the value of combined reference lies not in “inevitable increase” but in “knowing when it is sufficient.” Measurability on the TCM side has concrete anchors—multimodal constitution classification (PMID 36374441) and four-examination-assisted apparatus evaluation of health and efficacy (PMID 32185979)—indicating the gap is not in equipment but in calibration.
The literature provides clear guidance on estimator selection: Zhang and Zheng’s exponential decay bias estimator, whose asymptotic normality combined with rapid bias decay enables finite-alphabet inference, explicitly states that maximum likelihood (i.e., plug-in) mutual information “dramatically inflates Type I error probability” (PMID 25941916); Jiao et al.’s min-max framework yields the conversion: the optimal estimator on n samples approximates plug-in using n ln n samples, whereas plug-in fails to achieve this rate (PMID 29375152); continuous variables require geometric k-nearest neighbor routing (PMID 29604625). These three constraints combine into hard requirements: reporting Δ must employ low-bias estimators and be contrasted against plug-in.
Evidence and Limitations
Verified (within this paper’s simulation): At the truth level, marginal gains across the four channels are strictly positive and diminishing, with zero violations in 76 submodularity checks; reordering alters the gain trajectory but preserves total quantity; plug-in bias scales with channel count (e.g., +0.042 at N=50 rising to +0.339); strictly zero-information channels are estimated as positive increments in 100% of cases when N≤2000 across all grids; MM correction overcorrects at N=50, generating 30.4% false-negative increments; the weakest step Δ=+0.0487 requires approximately 400 samples to achieve 100% zero-rejection; AUC is non-monotonic even under optimal discriminators when “channel imbalance” exists, whereas mutual information (MI) remains monotonic.
Partially Verified: External literature supports the claim that “multimodal is not necessarily superior to the best unimodal” (PMID 37563277 meta-analysis, PMID 41319559 progressive fusion) and that “low-bias estimators are necessary” (PMID 25941916, PMID 29375152, PMID 29604625); however, the truth structure of “four examinations and four channels,” inter-rater reliability for real syndrome labels, and the parameter values for true channel readability (a) lack supporting data from this batch—the paper’s a=0.45~0.85 and pg=0.70~0.95 are set values, not measured results.
Pending Verification: ① reproducibility of zero-information channel rejection rate as a pre-registered calibration metric on real four-examination data; ② whether “channel imbalance” is common under real labels, implying non-monotonic AUC would be the norm rather than an exception; ③ since submodularity is always true, “diminishing marginal returns” can never constitute a discovery, but whether “deceleration rate” can determine “which route suffices” requires confirmation that rates remain stable across disease types (this paper uses only a single-disease prototype).
This topic has not yet entered the L1 corresponding network. The nearest neighbor is segment E-061 of KB-509, “Redundancy/Collaboration (PID)” PMID 33266673, which pertains to respiratory-heart rate dual-channel coupling rather than the four-examination-four-channel framework; furthermore, this entry explicitly states the scope warning: “redundancy means two pathways can predict the same target; collaboration means information exists only jointly, and their directions may oppose each other.”
L1 Network Status: Retrieving “four examinations / mutual information / fusion gain / increment / submodality / monotonicity” via EDGE_TABLE / CORRESPONDENCE_NETWORK yields hits exclusively within the context of cardiopulmonary coupling (segment E-061 of KB-509, redundancy/collaboration PID, PMID 33266673, two channels). Because the “truth layer and estimation layer separation of the inequality for fusion gain in four examinations and four channels” lacks a corresponding edge, no L1 status determination is obtained; thus, the evidence level is capped at C.
Related Reading
- “He-can-shen” is better than “He-can”: two readings, one always true, one possibly false; the bottleneck is the detection threshold
- Is the ranking of redundancy share versus complementarity share a disease-invariant? Cross-disease stability ruling of the Four-Examination-Four-Channel allocation structure
- Resonance Consciousness Field (Research Pool 1)
Evidence level: C
References: PMID 37563277, PMID 41319559, PMID 25941916, PMID 29375152, PMID 29604625, PMID 36374441, PMID 32185979
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.
Leave a Reply