--- title: Statistical Similarity Trap Insights url: https://www.emergentmind.com/topics/statistical-similarity-trap type: topic --- # Statistical Similarity Trap Insights Searching arXiv for the cited papers to ground the article in current records. “Statistical similarity trap” does not denote a single standardized construct across the literature. Rather, the phrase and closely related variants have been used for several failure modes in which observed similarity, closeness after preprocessing, or smooth aggregate fit is mistaken for validity, robustness, or substantive association. In benchmark evaluation, Mania et al. argue that high prediction agreement among models changes the effective severity of test-set reuse, because the usual worst-case multiplicity analysis is too loose when model errors are strongly correlated [1905.12580]. In robust comparison of distributions, Álvarez-Esteban et al. show that trimming beyond the true contamination level can make empirical samples appear closer than expected, producing an overfitting effect [1205.1950]. In binary similarity analysis, Chung et al. show that raw Jaccard/Tanimoto overlap can be large under independence when marginal probabilities are extreme [1903.11372]. Related formulations arise when omitted structure reverses aggregate associations, when RLHF-trained LLMs collapse to low-entropy “safe” modes, and when bulk image metrics reward blurry weather forecasts that miss rare extremes [2507.03628] [2512.10121] [2509.09195]. This suggests a unifying pattern: naive similarity measures are often confounded by latent dependence, marginal prevalence, smoothing, or imbalance. ## 1. Terminological scope and unifying pattern The literature uses the term in several technically distinct ways. In each case, the “trap” is not similarity itself, but the inference that high similarity or apparent closeness is automatically informative. | Paper | Domain | Core trap | |---|---|---| | [1905.12580] | Test-set reuse | Worst-case multiplicity ignores correlated model errors | | [1205.1950] | Trimmed sample comparison | Over-trimming makes samples closer than expected | | [1903.11372] | Binary co-occurrence | Raw Jaccard is inflated by marginals | | [2507.03628] | Regression and aggregation | Hidden variables create spurious or reversed associations | | [2512.10121] | Long-form LLM generation | Low-entropy smoothing suppresses expert burstiness | | [2509.09195] | Extreme-weather forecasting | Bulk metrics reward blurry predictions with zero rare-event skill | A common misconception is to treat similarity as intrinsically evidential. The cited works instead show that similarity must be interpreted relative to a null model, contamination model, causal structure, or task-specific objective. Depending on the setting, similarity can either mitigate an apparent problem, as in correlated benchmark reuse, or conceal a genuine one, as in rare-event forecasting and oversmoothed text generation. ## 2. Benchmark reuse, correlated errors, and similarity-aware generalization In the test-set reuse setting, let $D$ be the data distribution on examples $z=(x,y)$, let $S=\{z_1,\dots,z_n\}$ be an i.i.d. test set of size $n$, let $f_1,\dots,f_k$ be classifiers with $0$–$1$ losses $q_i(z)=1\{f_i(x)\neq y\}$, let $E_S[q_i]=(1/n)\sum_j q_i(z_j)$ be empirical error, and let $E_D[q_i]=E_{z\sim D}q_i(z)$ be population error. The quantity of interest is $$ P\!\left[\max_{i\le k}\left|E_S[q_i]-E_D[q_i]\right|\ge \epsilon\right]. $$ The vanilla union bound yields $$ P\!\left[\max_i |E_S[q_i]-E_D[q_i]|\ge \epsilon\right] \le \sum_{i=1}^k P\!\left[|E_S[q_i]-E_D[q_i]|\ge \epsilon\right], $$ which, with Bernoulli loss, gives the familiar non-asymptotic rate $$ \max_i |E_S[q_i]-E_D[q_i]| \le \sqrt{\frac{\tfrac12\log(2k/\delta)}{n}} $$ with probability at least $1-\delta$. Equivalently, if all models have similar error, one can only safely reuse the test set $O(\exp(n\epsilon^2))$ times before losing statistical validity. Mania et al. observe, however, that benchmark models are often highly correlated in their mistakes. They define pairwise agreement $$ \eta_{ij}:=P_{z\sim D}[q_i(z)=q_j(z)], $$ assume for simplicity a uniform lower bound $\eta_{ij}\ge \eta$, and introduce an “$\eta$-similarity cover” $M\subseteq\{q_1,\dots,q_k\}$ with minimum size $N_\eta$. Their main non-adaptive theorem gives $$ P\!\left[\max_{i\le k}|E_S[q_i]-E_D[q_i]|\ge \epsilon\right] \le 2N_\eta e^{-n\epsilon^2/2} +2k\exp\!\left[-(n\epsilon/4)\log\!\left(1+\frac{\epsilon}{4(1-\eta)}\right)\right]. $$ A corresponding confidence bound is $$ \max_{i\le k}|E_S[q_i]-E_D[q_i]| \le \max\!\left\{ \sqrt{\frac{2\log(4N_\eta/\delta)}{n}}, \sqrt{\frac{32(1-\eta)\log(4k/\delta)}{n}} \right\}. $$ When $\eta\approx 1-O(\epsilon)$, the dependence on $k$ can be exponentially better than the vanilla bound. The empirical motivation is ImageNet ILSVRC. Recht et al. and Mania et al. measure pairwise similarities among $66$ published ImageNet models, including AlexNet, VGG, ResNet, Inception, DenseNet, and SqueezeNet. If two error indicators were independent, agreement would be $$ \text{baseline}_{ij}=\mu_i\mu_j+(1-\mu_i)(1-\mu_j), $$ and because most $\mu_i$ lie around $24\%$ error, the baseline is about $0.62$. Empirically, the $2145$ model-pairs exhibit mean agreement approximately $0.85$. Figure A.1 further reports that $27.8\%$ of validation images are correctly classified by all $66$ models, $55.9\%$ are correct on at least $60$ models, and $4.7\%$ are misclassified by all $66$. For $n=50{,}000$, $\epsilon=1\%$, and $\delta=5\%$, the vanilla union bound gives $k\approx 2.6\times 10^5$ models testable, while the similarity-aware bound with pairwise $\eta\approx 0.85$ and minimal covering $N_\eta\approx 66$ gives $k\approx 10^6$, approximately a $4\times$ gain; under a further “naive-Bayes” assumption, one even obtains $k$ up to approximately $10^8$ [1905.12580]. The paper simultaneously states important limitations: deliberate design of disagreeing models, dataset drift or domain shift, clever adaptive strategies based on test feedback, and higher-order dependencies beyond pairwise overlap can all invalidate the optimistic interpretation of similarity. ## 3. Trimmed similarity of probability measures and overfitting by over-trimming Álvarez-Esteban et al. define two probabilities $P$ and $Q$ to be $\alpha$-similar if there exists a common core $P_0$ and contaminations $P_1',P_2'$ such that $$ P=(1-\epsilon_1)P_0+\epsilon_1P_1',\qquad Q=(1-\epsilon_2)P_0+\epsilon_2P_2', $$ with $0\le \epsilon_1,\epsilon_2\le \alpha$. The associated trimming operator is $$ \mathcal{R}_\alpha(P):=\{\,Q\in\mathcal{P}:Q\ll P \text{ and } dQ/dP\le 1/(1-\alpha)\ \text{$P$-a.s.}\,\}. $$ This set is convex and weakly compact. The paper also gives an equivalent parametrization by monotone maps $h\in\mathcal{C}_\alpha$ with $h(0)=0$, $h(1)=1$, and $h'(t)\le 1/(1-\alpha)$. Similarity admits a minimal-distance characterization. For a metric $d(\cdot,\cdot)$ on $\mathcal{P}$ for which $\mathcal{R}_\alpha(P)$ is compact, similarity at level $\alpha$ is equivalent to $$ \inf\{\,d(\tilde P,\tilde Q):\tilde P\in\mathcal{R}_\alpha(P),\ \tilde Q\in\mathcal{R}_\alpha(Q)\,\}=0. $$ With the $L_2$-Wasserstein distance $\mathcal{W}_2$, define $$ D_\alpha(P,Q):=\min_{R_1\in\mathcal{R}_\alpha(P),\,R_2\in\mathcal{R}_\alpha(Q)}\mathcal{W}_2(R_1,R_2). $$ Proposition 2.4 shows that $D_\alpha(P,Q)=0$ if and only if $P$ and $Q$ are $\alpha$-similar. The “trap” appears when trimming is set too aggressively. If $P_n,Q_n$ are empirical laws of two samples and $D_\alpha(P,Q)=0$, then $D_\alpha(P_n,Q_n)\to D_\alpha(P,Q)$ almost surely. But Theorem 2.6 shows that if $d_{TV}(P,Q)<\alpha$ and one trims at level $\alpha_n\ge d_{TV}(P,Q)+r_n/\sqrt{n}$ with $r_n\to\infty$, then $$ \sqrt{n}\,\mathcal{W}_2(P_{n,\alpha_n},Q_{n,\alpha_n})\to 0 $$ in probability. Under true homogeneity $P=Q$ with $\alpha_n=0$, by contrast, $\mathcal{W}_2(P_n,Q_n)=O_P(n^{-1/2})$ but not $o_P(n^{-1/2})$. Over-trimming therefore produces empirical closeness that is tighter than the benchmark case of exact equality. That accelerated collapse is the statistical similarity trap in this formulation [1205.1950]. The proposed remedy is bootstrap-based assessment. One computes optimally trimmed empirical laws, forms a pooled trimmed law $R_n=\tfrac12 P_{n,\alpha}+\tfrac12 Q_{n,\alpha}$, resamples from $R_n$, and compares the bootstrap statistic $$ S^*=\sqrt{n'/2}\,\mathcal{W}_2(P_{n'}^*,Q_{n'}^*) $$ to the observed $$ S_0=\sqrt{n/2}\,\mathcal{W}_2(P_{n,\alpha},Q_{n,\alpha}). $$ Theorem 3.1 states consistency: if $d_{TV}(P,Q)<\alpha$, then $p^*\to 1$ in probability; if $d_{TV}(P,Q)>\alpha$, then $p^*\to 0$. ## 4. Raw overlap coefficients and the need for null-centered similarity For binary vectors $y_i,y_j\in\{0,1\}^m$, Chung et al. define the Jaccard/Tanimoto similarity $$ T(y_i,y_j)= \frac{\sum_{k=1}^m y_{i,k}y_{j,k}} {\sum_{k=1}^m [y_{i,k}+y_{j,k}-y_{i,k}y_{j,k}]}. $$ The trap is to rank pairs by large observed $T(y_i,y_j)$ and treat those pairs as significantly co-occurring. Under independence, large raw overlap can arise from marginal abundances alone. Under the null model $$ y_{i,k}\stackrel{i.i.d.}{\sim}\mathrm{Bernoulli}(p_i),\qquad y_{j,k}\stackrel{i.i.d.}{\sim}\mathrm{Bernoulli}(p_j), $$ independently, Proposition 1 gives $$ \mathbb{E}[T(y_i,y_j)] = \frac{p_i p_j}{p_i+p_j-p_i p_j}. $$ The centered coefficient is therefore $$ T^c(y_i,y_j)=T(y_i,y_j)-\frac{p_i p_j}{p_i+p_j-p_i p_j}, $$ so that under independence $\mathbb{E}[T^c]=0$. Large positive $T^c$ indicates more co-occurrence than expected from marginals; large negative $T^c$ indicates less. The paper derives an exact null distribution from the multinomial law of the four cell counts and also gives an asymptotic approximation. With $$ q_1=p_i p_j,\qquad q_2=p_i+p_j-2p_i p_j,\qquad \mu=\frac{q_1}{q_1+q_2}, $$ Proposition 2 states $$ \sqrt{m}\,(T(y_i,y_j)-\mu)\xrightarrow{d}\mathcal{N}(0,\sigma^2), \qquad \sigma^2=\frac{q_1 q_2(1-q_2)}{(q_1+q_2)^3}. $$ Because direct enumeration becomes prohibitively slow once $m\gtrsim 500$, Chung et al. develop two faster procedures: a bootstrap with total cost $O(Bm)$, typically $O(m^2)$ when $B=5m$, and the Measure Concentration Algorithm (MCA), which restricts summation to a high-probability multinomial “ball.” With $\varepsilon=10^{-5}$, MCA gives a rigorous two-sided bound on the true p-value. Simulations show that exact, bootstrap, and MCA p-values are well calibrated, whereas asymptotic p-values can be anti-conservative or conservative when $m$ is moderate. In the Vanuatu birds data, raw Jaccard $T$ correlates strongly ($r=0.43$) with the product of marginal frequencies, and centering with significance testing recovers $374$ significant pairs at q-value $<0.10$ [1903.11372]. ## 5. Hidden structure, omitted variables, and Simpson reversals A related use of the trap concerns hidden structure in regression and contingency tables. Consider the true model $$ y_i=\beta x_i+\gamma z_i+\epsilon_i,\qquad E[\epsilon_i\mid x_i,z_i]=0, $$ but suppose only $(x_i,y_i)$ are observed and one fits $$ y_i=\tilde\beta x_i+u_i,\qquad u_i\equiv \gamma z_i+\epsilon_i. $$ Then the OLS estimator satisfies $$ E[\hat\beta]=\beta+\gamma\frac{\mathrm{Cov}(x,z)}{\mathrm{Var}(x)}. $$ The bias term is therefore $$ \mathrm{Bias}(\hat\beta)=E[\hat\beta]-\beta= \gamma\frac{\mathrm{Cov}(x,z)}{\mathrm{Var}(x)}. $$ Unless $\gamma=0$ or $\mathrm{Cov}(x,z)=0$, the omitted variable induces bias [2507.03628]. Simpson’s paradox is the contingency-table analogue. Let $Z$ be a stratifying variable with levels $j=1,\dots,J$ and $$ E[Y\mid X=x,Z=j]=\alpha_j+\beta_j x. $$ The paradox arises when all conditional slopes have one sign, for example $\beta_j>0$ for every $j$, but the marginal slope $$ \beta_{\mathrm{marginal}}\equiv \frac{dE[Y\mid X=x]}{dx} =\frac{\mathrm{Cov}(X,Y)}{\mathrm{Var}(X)} $$ has the opposite sign. For binary $X,Y$, this is equivalent to $$ P(Y=1\mid X=1,Z=j)>P(Y=1\mid X=0,Z=j)\quad \text{for each } j, $$ yet $$ P(Y=1\mid X=1)