## 1. Precise statement The claim is correct under the following interpretation. For every **fixed integer** \(r\ge 2\), there are sequences of finite experiments \(E_M,F_M\), indexed by integers \(M\ge2\), on the same theory set \[ \Theta_M,\qquad |\Theta_M|=N_M=M^r, \] such that: 1. **Equivalence on every small menu.** For every nonempty \(A\subseteq\Theta_M\) with \(|A|\le r\), the experiments restricted to parameters \(\theta\in A\) are Blackwell-equivalent. Thus there are stochastic simulation kernels in both directions that work for every \(\theta\in A\). - The kernels may depend on \(A\) and \(M\), but not on the true \(\theta\). - Equivalently, the experiments have identical optimal Bayes risks for every prior supported on \(A\) and every finite decision problem on that restricted parameter set. 2. **Different global identification performance.** Under the uniform prior on all of \(\Theta_M\), let \[ \operatorname{Acc}(E_M)=\sup_\delta \Pr\{\delta(\text{report})=\theta\}. \] Then \[ \operatorname{Acc}(E_M)\longrightarrow1, \qquad \operatorname{Acc}(F_M)\longrightarrow0. \] Important qualifications: - The theory set is finite at each stage, but its size grows without bound. - “At most \(r\) rival theories” means that the truth is known to belong to a specified menu of at most \(r\) theories. It does **not** mean a decision problem with at most \(r\) available actions on the full theory set. - The experiments are not globally Blackwell-equivalent. ## 2. Checking the argument and independently deriving the result ### Initial audit Every substantive mathematical step in the supporting text checks out: - The optimal identification probability is indeed \(\mathbb E[1/K]\), under the stated uniform prior. - The signed weights annihilate polynomials of degree at most \(r-1\). - Their positive and negative parts give probability laws with matching moments. - The weights at \(1\) and \(M\) dominate asymptotically and have opposite signs. - Restricted-menu intersection probabilities depend only on moments through degree \(s-1\). - The rest of the report is simulable once the intersection is known. The last two points need a pointwise argument for each possible true theory—not merely equality after averaging over the menu. Here is that argument, followed by the construction. ### A. Start with the experiment’s likelihood Let \(p_k=\Pr(K=k)\). If the reported subset is \(S\subseteq\Theta\), then \[ P_p(S\mid\theta) = \mathbf 1_{\{\theta\in S\}} \frac{p_{|S|}}{\binom{N-1}{|S|-1}}. \] Under the uniform prior, this likelihood is the same for every \(\theta\in S\), and zero otherwise. Hence the posterior is uniform on \(S\). Consequently, conditional on \(K=k\), optimal identification succeeds with probability \(1/k\), giving \[ \operatorname{Acc}(p)=\mathbb E_p[1/K]. \] ### B. Derive exactly what a small menu can observe Fix a menu \(A\) of size \(s\), and suppose \(\theta\in A\). Let \[ T=S\cap A. \] For a particular nonempty \(B\subseteq A\), write \(b=|B|\). If \(\theta\notin B\), then \(P(T=B\mid\theta)=0\). If \(\theta\in B\), then \[ \begin{aligned} P(T=B\mid\theta,K=k) &=\frac{\binom{N-s}{k-b}}{\binom{N-1}{k-1}}\\ &=\frac{(k-1)_{\underline{b-1}}\, (N-k)_{\underline{s-b}}} {(N-1)_{\underline{s-1}}}, \end{aligned} \] where \(x_{\underline d}=x(x-1)\cdots(x-d+1)\). The numerator is a polynomial in \(k\) of degree at most \[ (b-1)+(s-b)=s-1. \] Therefore, if two laws \(p,q\) on \(K\) have matching moments through degree \(r-1\), then for every \(s\le r\), \[ P_p(T=B\mid\theta)=P_q(T=B\mid\theta) \] for every \(\theta\in A\) and every \(B\subseteq A\). This establishes equality of the **restricted channels**, not just their prior-averaged output distributions. ### C. Show that the intersection loses no information on that menu For any report \(S\) satisfying \(S\cap A=B\), and any \(\theta\in B\), \[ P_p(S\mid\theta)=\frac{p_{|S|}}{\binom{N-1}{|S|-1}}, \] which does not depend on which member of \(B\) is true. Thus the conditional distribution of \(S\) given \(T=B\) is independent of the true theory within \(B\). It can be used as a simulation kernel. Explicitly, to simulate the \(q\)-experiment from the \(p\)-experiment on menu \(A\): 1. Compute \(B=S\cap A\). 2. Generate a new report according to the \(q\)-experiment’s conditional distribution given intersection \(B\). Because the intersection channels agree, this reproduces the \(q\)-experiment for every \(\theta\in A\). Reversing the roles gives the other simulation. So matching moments through degree \(r-1\) is sufficient for the required local Blackwell equivalence. ### D. Construct moment-matched laws with extreme identification probabilities Take \[ x_j=M^j,\quad j=0,\ldots,r,\qquad N=M^r, \] and \[ w_j=\frac1{\prod_{\ell\ne j}(x_j-x_\ell)}. \] By Lagrange interpolation, for every polynomial \(f\) of degree at most \(r-1\), \[ \sum_{j=0}^r w_j f(x_j)=0. \] Indeed, the sum is the coefficient of \(t^r\) in the Lagrange interpolation formula for \(f(t)\), and that coefficient must vanish. Define \(a_j=(-1)^r w_j\). Since \[ \operatorname{sign}(w_j)=(-1)^{r-j}, \] we have \[ \operatorname{sign}(a_j)=(-1)^j. \] Because \(\sum_j a_j=0\), their positive and negative totals are equal: \[ C=\sum_j a_j^+=\sum_j(-a_j)^+>0. \] Define \[ p_G(x_j)=\frac{a_j^+}{C}, \qquad p_B(x_j)=\frac{(-a_j)^+}{C}. \] Then, for \(m=0,\ldots,r-1\), \[ \mathbb E_G[K^m]=\mathbb E_B[K^m]. \] These laws therefore yield locally Blackwell-equivalent experiments. Now examine their asymptotics. For fixed \(r\), \[ |w_j| =(1+o(1))M^{-r(r+1)/2-j(j-1)/2}. \] In particular, \[ |w_0|\sim |w_1|, \qquad |w_j|=o(|w_0|)\quad(j\ge2). \] Thus \[ p_G(K=1)\longrightarrow1, \qquad p_B(K=M)\longrightarrow1. \] It follows that \[ \operatorname{Acc}(G)\ge p_G(K=1)\longrightarrow1. \] Moreover, every support point of \(p_B\) is at least \(M\), so \[ \operatorname{Acc}(B)=\mathbb E_B[1/K]\le\frac1M\longrightarrow0. \] This completes the proof. ## 3. Attempts to break it ### A concrete example: \(r=2\), \(M=2\), \(N=4\) The nodes are \(1,2,4\), with weights \[ w_0=\frac13,\qquad w_1=-\frac12,\qquad w_2=\frac16. \] Normalizing gives: | Experiment | Distribution of \(K\) | \(\mathbb E[K]\) | Identification accuracy | |---|---|---:|---:| | \(G\) | \(K=1\) with probability \(2/3\); \(K=4\) with probability \(1/3\) | \(2\) | \(3/4\) | | \(B\) | \(K=2\) always | \(2\) | \(1/2\) | For any two-theory menu: - With probability \(2/3\), the intersection is the singleton containing the truth. - With probability \(1/3\), the intersection is the whole two-theory menu. That statement holds for both experiments. Their two-theory restrictions are therefore equivalent, although their global identification probabilities differ. For general \(M\) with \(r=2\), the construction simplifies to \[ G:\quad \Pr(K=1)=\frac{M}{M+1},\quad \Pr(K=M^2)=\frac1{M+1}, \] and \[ B:\quad K=M. \] Their accuracies are \[ \operatorname{Acc}(G)=1-\frac1M+\frac1{M^2}, \qquad \operatorname{Acc}(B)=\frac1M. \] ### Important boundaries and possible misreadings **1. A fixed finite theory set cannot support the stated limit.** Under a uniform prior on \(N\) theories, any experiment permits accuracy at least \(1/N\), simply by guessing. The construction necessarily uses growing \(N\). **2. Equivalence does not extend to all small-action decision problems.** The four-theory example already distinguishes the experiments using only two actions. Partition the theories into two groups of two, and ask which group contains the truth. - Experiment \(G\) achieves accuracy \[ \frac23+\frac13\cdot\frac12=\frac56. \] - Experiment \(B\) achieves accuracy \[ \frac13+\frac23\cdot\frac12=\frac23. \] So “few rival theories” cannot be replaced by “few possible decisions.” **3. The global accuracy conclusion is prior-sensitive.** With a point-mass prior, both experiments identify the truth perfectly before any observation. The stated global comparison uses the uniform prior. The local Blackwell equivalence, by contrast, holds for **every** prior supported on the chosen small menu. **4. Menu-dependent simulation is essential.** There is no single global pair of simulation kernels. Such kernels would imply global Blackwell equivalence and hence equal identification accuracy. **5. The integer and nonvacuity conditions are satisfied.** Taking integer \(M\ge2\) makes every \(M^j\) an admissible subset size. Also \(M^r>r\), so the small-menu condition never includes the entire theory set. I find no counterexample to the precisely stated theorem and no unproved mathematical step remaining. ## 4. Novelty The construction uses established mathematics: - **Blackwell’s comparison-of-experiments theorem** supplies the connection between simulation kernels and agreement across decision problems. - The displayed weights are the standard **divided-difference/Lagrange interpolation weights**: an \(r\)-th divided difference annihilates polynomials of degree below \(r\). - Splitting a signed measure that annihilates low-degree polynomials into positive and negative parts is a standard **truncated-moment matching construction**. The mechanism is therefore a variant of familiar moment-matching and limited-order indistinguishability ideas. I do **not** know a reliable reference for this exact formulation—equivalence on every \(r\)-theory menu together with identification probabilities tending to opposite extremes. Standard ingredients alone do not establish that the exact theorem has previously appeared. Accordingly, I would classify it as a **special construction or variant of known results**, while leaving the publication priority of this precise formulation unresolved. ## 5. Fairness of the one-sentence claim The sentence is a fair summary **under the supporting text’s menu-restriction interpretation**. The proof establishes exact local equivalence, not merely approximate agreement. The qualifications that should accompany it are: - it concerns a sequence of experiments on growing finite theory sets; - the menu is specified and known when comparing restricted experiments; - “small” counts possible true theories, not actions; - the global identification comparison uses the uniform prior; - it is an existence result for specially constructed experiments, not a general characterization of theory choice. These qualifications limit its interpretation, but they do not undermine the mathematical result. {"verdict":"correct with caveats","novelty":"special case or variant of known results","fair_summary":"yes","confidence":4,"one_line":"The construction rigorously gives exact Blackwell equivalence on every fixed-size theory menu alongside opposite global identification limits, provided the theory set grows and equivalence is understood menu by menu."}