Theory Digest — Apr 11, 2026

1. Motivation: Deep transformers exhibit representation collapse—token embeddings become clustered or low-dimensional as depth increases, degrading expressivity. Understanding when and how this occurs as a function of architectural scaling (depth, width, heads, residual scaling) remains theoretically open, especially when weights vary across layers.

2. Mathematical setup: Consider an attention-only transformer with $L$ layers and $H$ heads. Token embeddings evolve as:

   \[x_i^{\ell+1} = N\left(x_i^\ell + \eta \frac{1}{H} \sum_{h=1}^H B_{\theta_h^\ell}[\mu_{X^\ell}](x_i^\ell)\right)\]

   where $N(x) = x/|x|$ is layer normalization, $\eta > 0$ is residual scale, and

   \[B_\theta[\mu](x) = \frac{1}{Z_A[\mu](x)} \int e^{\beta \langle Ax, y \rangle} V y \, \mu(dy)\]

   with normalizing constant $Z_A\mu = \int e^{\beta \langle Ax, y \rangle} \mu(dy)$.

   Assumptions:

   1. Weights $\theta_h^\ell = (V_h^\ell, A_h^\ell)$ are i.i.d. across layers $\ell \in [L]$ and heads $h \in [H]$ with distribution $\rho_* \in \mathcal{P}(\Theta)$
   2. Token embeddings $x_i^\ell \in S^{d-1}$ live on the unit sphere
   3. Empirical measure $\mu_{X^\ell} = \frac{1}{n} \sum_{k=1}^n \delta_{x_k^\ell}$

3. Toy example: When $d=2$, $n=2$, $H=1$, $\beta=0$, and Gaussian weights, the dynamics reduce to two particles on the circle with random linear combinations at each layer, leading to potential clustering depending on the balance between drift and diffusion.

4. Formal objective: Prove that under joint scalings of depth $L$, residual step $\eta$, and heads $H$, the discrete dynamics converge to a tractable continuous limit, and characterize when representation collapse occurs:

   \[\lim_{L \to \infty, \eta \to 0, H \to \infty} \mathbb{E}[\phi(X^{\eta L})] = \mathbb{E}[\phi(X(T))]\]

   where $X(t)$ solves a limiting SDE.

The method uses homogenization theory to derive a continuous-time limit.

Step 1: Decompose each attention head into mean plus fluctuation:

\[B_{\theta_h^\ell}[\mu](x) = b_{\rho_*}[\mu](x) + \xi_{\theta_h^\ell}[\mu](x)\]

Step 2: After averaging over heads, the update becomes:

\[x_i^{\ell+1} = x_i^\ell + \eta b_{\rho_*}[\mu_{X^\ell}](x_i^\ell) + \frac{\eta}{\sqrt{H}} \tilde{\sigma}[\mu_{X^\ell}](x_i^\ell) \zeta^\ell\]

Step 3: Set macroscopic time $t_L = \eta L$ and scaling parameter:

\[\alpha := \frac{\eta \sigma^2}{H}\]

Step 4: The limiting SDE on $(S^{d-1})^n$ is:

\[dX(t) = b(X(t))dt + \sqrt{\alpha} \int_\Theta G_\sigma(X(t), \theta) W(d\theta, dt)\]

where $W$ is a cylindrical Wiener process and the second term includes Itô corrections from sphere projection.

Step 5: Three scaling regimes emerge:

• Ballistic ($\alpha \eta L \to 0$): deterministic ODE limit
• Diffusive ($\alpha \eta L = \Theta(1)$): stochastic SDE limit
• Super-diffusive ($\alpha \eta L \gg 1$): noise-dominated

Toy example application: For $d=2$, $H=1$, $\eta=1/L$, the method shows whether the two particles cluster depends on the ratio $\sigma^2/(L \cdot 1/L) = \sigma^2$, determining if we’re in ballistic or diffusive regime.

Prior work:

• GLPR24, GLPR25: studied clustering in transformers with fixed weights across layers, showing deterministic convergence to clustered states via interacting particle systems
• FSE+26: extended to random value matrices (but fixed query/key), obtaining stochastic limits
• CNQG25: analyzed single attention block contractions and phase transitions

Delta: This paper studies the first model where all weight matrices (query, key, value) vary independently across layers and heads, reflecting realistic transformer architectures rather than artificial weight-tying assumptions.

Theory-specific assessment:

• The main theorem (homogenization limit) is somewhat predictable given existing stochastic homogenization theory, but the application to transformers with spherical constraints is new
• The proof technique adapts known martingale methods (EK86, SV07) and stochastic modified equations (LTE17) to the transformer setting—this is competent but not groundbreaking
• The bounds are not shown to be tight, and no lower bounds are established
• The Gaussian case analysis (Theorems 2-4) provides explicit characterizations but relies heavily on the special structure

The theoretical contribution is solid but incremental: it extends known homogenization techniques to a new (and more realistic) transformer model. The explicit analysis in the Gaussian case provides useful insights about representation collapse.

Verdict: INCREMENTAL — solid extension of existing homogenization theory to more realistic transformer models, with useful explicit results in special cases.

The proof follows the stochastic modified equations approach of Lie-Trotter-Ethier.

1. Generator matching: Match the discrete one-step generator with the Itô generator:

   \[(\mathcal{L}^\eta \phi)(x) = \frac{\alpha}{2} \int_\Theta \text{Hess}\phi(x)[G_\sigma(x,\theta), G_\sigma(x,\theta)] \rho_*(d\theta)\]

2. Martingale problem formulation: Show the discrete chain solves a martingale problem that converges to the continuous martingale problem for the limiting SDE.

3. Key concentration bound: For the fluctuation term, apply subGaussian concentration:

   \[\mathbb{E}\left\|\frac{1}{\sqrt{H}} \sum_{h=1}^H \xi_{\theta_h^\ell}[\mu](x)\right\|^2 \leq \frac{\sigma^2}{H}\]

4. Weak convergence via tightness: Establish tightness of the discrete chain sequence using:

   \[\mathbb{E}|\phi(X^{\eta}(t+h)) - \phi(X^{\eta}(t))|^2 \leq Ch\]

5. Error bounds: The main approximation bound uses:

   \[\sup_{t \in [0,t_L]} |\mathbb{E}\phi(X(t)) - \mathbb{E}\phi(X^\eta(t))| \leq Ce^{Ct_L}\eta(t_L+1)\max(1,\alpha)\]

6. Gaussian case specialization: When weights are Gaussian, the drift vanishes ($b_{\rho_*} \equiv 0$) and explicit calculations become possible using rotational invariance.

7. Sphere geometry: Handle the spherical constraint via Itô corrections:

   \[P_x^{\perp} = I_d - xx^T\]

   and modified drift terms from the projection $N(x) = x/|x|$.

Purely theoretical. The paper includes some numerical illustrations:

• Figure 2: Compares the simplex dynamics $\gamma(t)$ from Theorem 2 with prior deterministic results from GLPR25, showing similar long-time clustering behavior
• Figure 3: Validates Theorem 3’s logistic equation approximation for the high-temperature regime with varying dimension $d$

Empirical validation would require:

1. Training transformers with the proposed initialization scalings
2. Measuring representation collapse metrics (effective rank, clustering coefficients) during training
3. Comparing against standard initializations on downstream tasks
4. Verifying the predicted phase transitions between ballistic/diffusive regimes

Limitations:

1. Independence assumption: Weights are i.i.d. across layers/heads only at initialization - TECHNICAL (needed for tractability but violated during training)

2. Attention-only architecture: Excludes MLP blocks which are crucial in practice - NATURAL (MLP extension is mentioned in Remark 2.1)

3. Gaussian weight restriction: Sharpest results require exactly Gaussian weights rather than just subGaussian - TECHNICAL (likely removable with more work)

4. Bounded time horizon: Approximation only valid for macroscopic times $t_L = \eta L = O(1)$ - RESTRICTIVE (prevents studying very deep limits)

5. Mean-field assumptions: Theorem 4 requires density bounds that may not hold in practice - RESTRICTIVE (limits applicability of the slow motion results)

6. No training dynamics: Analysis is purely forward-pass, ignoring how weights evolve - RESTRICTIVE (initialization analysis only)

   Open problems:

7. Extend homogenization results to trained transformers where weights develop layer-to-layer correlations
8. Establish matching lower bounds for the approximation error rates

1. Motivation: Contest design problems arise in recommender systems, crowdsourcing, and online platforms where strategic contestants choose effort levels. A designer must allocate prizes to incentivize high-quality outcomes while accounting for strategic behavior.

2. Mathematical setup: There are $n$ strategic content producers. Each producer $i$ chooses quality $q_i \geq 0$ at cost $c(q_i) = q_i^{\beta}$ for $\beta > 0$. The recommender system commits to a rank-based policy $p \in \Delta$ where:

   \[\Delta = \{p \in \mathbb{R}_{\geq 0}^n : p_1 \geq p_2 \geq \cdots \geq p_n \geq 0, \sum_{i=1}^n p_i = 1\}\]

   Under policy $p$, content ranked $i$-th receives recommendation probability $p_i$. Producer $i$’s utility is:

   \[U_i(\mu_i, \mu_{-i}, p) = R_i(\mu_i, \mu_{-i}, p) - \mathbb{E}_{q_i \sim \mu_i}[c(q_i)]\]

   where $R_i$ is expected revenue when recommended. The RS maximizes:

   \[\text{OBJ}(\mu, p) = \alpha W(\mu, p) + (1-\alpha) Q(\mu, p)\]

   where $W(\mu, p)$ is user welfare (expected quality of recommended content) and $Q(\mu, p)$ is platform quality (average content quality).

   Assumptions:

   1. Homogeneous producers with identical cost functions
   2. Complete information setting
   3. Ties broken uniformly at random
   4. Individual rationality: playing $q = 0$ yields non-negative utility

3. Toy example: When $n = 3$, $\beta = 2$, and $\alpha = 0.5$, the RS chooses between HardMax $(p_1, p_2, p_3) = (1, 0, 0)$ and UniformButLast $(1/2, 1/2, 0)$. Under quadratic costs, producers face trade-offs between higher effort (to achieve top rank) versus moderate effort (competing for intermediate ranks).

4. Formal objective: Find policy $p^*$ solving:

   \[\max_{p \in \Delta} \alpha W(\mu, p) + (1-\alpha) Q(\mu, p) \text{ subject to } \mu \text{ is a symmetric MNE}\]

The method involves several key steps:

1. Establish existence/uniqueness of symmetric mixed Nash equilibrium using topological arguments from Reny (1999) and better-reply security conditions.

2. Transform the optimization problem. The key reduction theorem shows the RS’s problem becomes:

   \[\max_p \alpha n \int_0^1 h(x,p)^{1+1/\beta} dx + (1-\alpha) \int_0^1 h(x,p)^{1/\beta} dx\]

   subject to $p_1 \geq p_2 \geq \cdots \geq p_n = 0$, $\sum p_i = 1$, where:

   \[h(x,p) = \sum_{i=1}^n a_i(x) p_i\] \[a_i(x) = \binom{n-1}{i-1} x^{n-i}(1-x)^{i-1}\]

   are Bernstein basis polynomials.

3. For single-minded cases ($\alpha \in {0,1}$), apply Schur-convexity analysis. The function $h(x,p)^r$ is Schur-convex when $r \geq 1$ or $r \leq 0$, Schur-concave otherwise.

4. For convex-minded cases ($\alpha \in (0,1)$), use variation diminishing property. Show the gradient sequence exhibits quasiconvex structure via total positivity of Bernstein polynomial matrices, then apply KKT conditions.

5. Computational algorithm uses branch-and-bound with careful bounds construction, achieving poly$(n, 1/\varepsilon)$ time complexity.

   Toy example application: For $n=3$, $\beta=2$, $\alpha=0.5$, the method reduces to optimizing over $(p_1, p_2, 0)$ with $p_1 + p_2 = 1$, yielding structure $p_1 \geq p_2 = 0$ from the variation diminishing analysis.

Step 1 — Prior work:

• Glazer and Hassin (1988): proved UniformButLast optimal for linear costs and average quality objective
• Fang, Noe, and Strack (2020): studied convex/concave utility functions with corresponding concave/convex costs
• Jagadeesan et al. (2024): analyzed symmetric MNE existence under HardMax policy only

Step 2 — Delta: This paper extends to nonconvex objective functions (convex combinations of user welfare and platform quality, arbitrary posynomials) and provides complete structural characterization $p_1 \geq p_2 = \cdots = p_{n-1} \geq p_n = 0$ across all cases. New proof technique using variation diminishing property and total positivity.

Step 3 — Theory-specific assessment:

• The main structural theorem is somewhat surprising given the nonconvex nature of objectives, but the specific structure follows patterns seen in related majorization theory
• Proof technique genuinely novel: connecting gradient sequence shape to optimal solution structure via Bernstein polynomials and total positivity is creative
• No explicit lower bounds provided; tightness demonstrated only through computational experiments

The reduction from high-dimensional nonconvex optimization to single-dimensional optimization is a significant technical contribution that extends beyond this specific application.

Verdict: SIGNIFICANT — provides first complete analysis of optimal contest design beyond convex objectives with genuinely novel proof techniques, though the specific structural result follows somewhat predictable patterns from majorization theory.

The proof strategy involves several sophisticated techniques:

1. Existence via Better-Reply Security: Use Reny (1999) framework for discontinuous games. Endow strategy space with weak*-topology and L´evy-Prokhorov metric. Show diagonal better-reply security holds despite payoff discontinuities from rank-based rewards.

2. Equilibrium Characterization: Key insight that symmetric MNE has atomless, strictly increasing CDF. Producer’s payoff in support:

   \[U_i(q, \mu_{-i}, p) = \sum_{i=1}^n p_i \binom{n-1}{i-1} F(q)^{n-i}(1-F(q))^{i-1} - q^{\beta}\]

3. Optimization Reduction: Transform using correlation inequalities to show $p_n = 0$ optimal. Express objective purely in terms of Bernstein basis polynomials:

   \[G(p) = \alpha n \int_0^1 h(x,p)^{1+1/\beta} dx + (1-\alpha) \int_0^1 h(x,p)^{1/\beta} dx\]

4. Schur-Convexity Analysis: For single-minded cases, prove $h(x,p)^r$ is Schur-convex when $r \geq 1$, Schur-concave when $0 < r < 1$. Apply majorization theory on probability simplex.

5. Total Positivity and Variation Diminishing: Core technical innovation. Show matrix $(a_i(1-x_l))_{s,l}$ is totally positive by connecting to generalized Vandermonde determinants:

   \[\det = \prod_{l=1}^k x_l^{n-i_k} \prod_{1 \leq l < l' \leq k}(x_{l'} - x_l) \cdot (\text{positive terms})\]

   Use Karlin’s variation diminishing property to prove gradient sequence $(\partial G/\partial p_i)$ is quasiconvex, enabling KKT analysis.

6. Branch-and-Bound Algorithm: Exploit affine structure $h(x,p) = h_0(x) + (p_1 - p_2) \cdot x^{n-1}$ to construct tight upper/lower bounds. Convergence rate depends polynomially on interval shrinkage.

Purely theoretical with computational validation.

Key computational evidence:

• Figure 1: Brute-force verification of optimal policy structure for $n=5$ producers using discretization granularity 0.005 and trapezoidal integration
• Figure 2: Heat maps showing $p_1$ and $p_2$ values across parameter ranges $\alpha \in [0.05,1]$, $\beta \in [0.1,5]$ with 1000 grid points
• Algorithms implemented with value oracle access for integral approximation

Empirical validation would require:

• Real recommender system deployment with strategic content producers
• A/B testing different prize allocation mechanisms
• Measurement of actual user welfare vs platform quality trade-offs
• Validation that producers behave according to symmetric MNE predictions

Limitations:

1. Complete information assumption - RESTRICTIVE (real platforms have incomplete information about producer costs and user preferences)

2. Homogeneous producers with identical cost functions - RESTRICTIVE (significant heterogeneity exists in practice)

3. Single-dimensional quality model - TECHNICAL (multi-dimensional quality spaces would complicate ranking mechanisms but framework could extend)

4. Risk-neutral producers - NATURAL (standard assumption in contest literature)

5. Symmetric equilibrium focus - NATURAL (anonymity and coordination simplicity justify this choice)

6. Fixed participation (no entry/exit costs) - TECHNICAL (individual rationality ensures participation but dynamic entry could be studied)

7. Value oracle assumption for computational algorithm - TECHNICAL (oracle can be approximated by numerical integration)

   Open problems:

8. Characterize optimal mechanisms under incomplete information about producer heterogeneity and extend structural results to asymmetric settings

9. Analyze dynamic versions where producers can adapt strategies over time, potentially with learning about platform policies and competitor behavior

Motivation: Modern LLM evaluation platforms use pairwise human judgments (e.g., ARENA battles) where users compare model outputs. Current leaderboards provide limited uncertainty quantification despite noisy, sparse, and highly non-uniform data collection.

Mathematical setup: Let $T^* \in \mathbb{R}^{d_1 \times d_2}$ be a latent score matrix where $T^*_{j,u}$ represents model $j$’s ability on task category $u$. Each observation is a binary comparison outcome:

\[Y^{(i)} \in \{0,1\}\] \[P(Y^{(i)} = 1 | p_i, q_i, u_i, T^*) = \sigma(\eta_i) = \frac{\exp(\eta_i)}{1 + \exp(\eta_i)}\]

where:

\[\eta_i = T^*_{p_i,u_i} - T^*_{q_i,u_i}\]

The design matrix is $X^{(i)} = (e_{p_i} - e_{q_i})e_{u_i}^T$ so that $\eta_i = \langle T^*, X^{(i)} \rangle$.

Assumptions:

1. $T^*$ has low-rank Tucker structure: $T^* = C^* \times_1 U_1 \times_2 U_2$ with orthonormal factors
2. Zero-sum identification constraint: $\mathbf{1}_{d_1}^T T^* = 0$
3. $\mu$-incoherence condition on factor matrices

4. Bounded entries: $|T^*|_{\infty} \leq B$

   Toy example: When $d_1 = 3$ models, $d_2 = 2$ categories, and $T^* = \begin{pmatrix} 1 & -0.5 \ -0.5 & 1 \ -0.5 & -0.5 \end{pmatrix}$, comparing models 1 vs 2 on category 1 gives $\eta = 1 - (-0.5) = 1.5$, so $P(\text{model 1 wins}) = \sigma(1.5) \approx 0.82$.

   Formal objective: Estimate smooth functionals of the latent tensor:

   \[\psi(T^*) = \langle \Gamma, T^* \rangle \text{ (linear)} \text{ or } \psi(T^*) = \sigma(T^*_{a,t} - T^*_{b,t}) \text{ (nonlinear)}\]

The method is a one-step debiased estimator based on semiparametric efficiency theory.

Step 1: Initial Estimator: Obtain $\hat{T}$ via low-rank tensor estimation with entrywise accuracy:

\[\|\hat{T} - T^*\|_{\infty} \leq C_1 \sqrt{\frac{r\bar{d}\log^c \bar{d}}{n}}\]

Step 2: Information Operator: Define the Fisher information operator $G$ via:

\[\langle GU, V \rangle = \mathbb{E}_*[I(\eta^*)\langle U, X \rangle \langle V, X \rangle]\]

where $I(\eta) = \sigma(\eta)(1-\sigma(\eta))$ is the scalar Fisher information.

Step 3: Efficient Direction: Solve the information equation for the optimal direction:

\[AH^* = P_T \Gamma\]

where $A = P_T G P_T$ and $P_T$ projects onto the identifiable low-rank tangent space. The solution is:

\[H^* = A^{-1} P_T \Gamma\]

Step 4: One-Step Correction: For fold $k$, compute:

\[\hat{\psi}^{(k)} = \psi(\hat{T}^{(k)}) + \frac{K}{N} \sum_{i \in D_k} s_{\eta}(Y_i, \hat{\eta}_i^{(k)}) \langle \hat{H}^{(k)}, X_i \rangle\]

where $s_{\eta}(y,\eta) = \partial_{\eta} \ell(y,\eta)$ is the score function.

Score Whitening Alternative: To avoid non-commutativity issues, use the whitened score:

\[\tilde{s}_{\eta}(y,\eta) = \frac{s_{\eta}(y,\eta)}{I(\eta)}\]

This makes the effective information operator isotropic: $A_0 = \frac{1}{d_*}P_T$, giving the simplified direction:

\[H_{ws}^* = d_* P_T \Gamma\]

Toy Example Application: For our 3×2 matrix example with target $\Gamma = e_1 e_1^T - e_2 e_1^T$ (comparing models 1 vs 2 on category 1), the method projects $\Gamma$ onto the tangent space, solves for the optimal weighting of comparisons based on their Fisher information, and applies the correction to debias the initial estimator.

Step 1 — Prior work:

1. Matrix completion literature (Candès & Recht 2009, Keshavan et al. 2010): established low-rank matrix recovery from uniformly sampled entries with Gaussian noise.
2. Bradley-Terry-Luce models (recent: Fan, Hou & Yu 2024): developed statistical theory for pairwise comparison models but without low-rank structure or tensor formulations.

3. LLM evaluation platforms (ARENA, Chiang et al. 2024): created practical pairwise evaluation systems but with limited statistical theory for uncertainty quantification.

   Step 2 — Delta: This paper adds:

   • First semiparametric efficiency theory for low-rank tensor completion from pairwise comparisons
   • Identification of non-commutativity bottleneck: $G$ and $P_T$ don’t commute under heterogeneous Fisher information
   • Score whitening technique to restore isotropic information geometry
   • Extension from direct entry observations to pairwise contrasts with identification constraints

   Step 3 — Theory-specific assessment:

   • Main theorem: The efficient influence function derivation and efficiency bound $V_{eff} = \langle P_T\Gamma, A^{-1}P_T\Gamma \rangle$ is novel for this setting but follows standard semiparametric methodology
   • Proof technique: The non-commutativity analysis is genuinely new. Standard debiasing assumes $G \propto I$, but here $G$ is anisotropic due to varying Fisher information $I(\eta^*)$ across comparisons. The quantity $C_A$ controlling $|A^{-1}|_{\infty \to \infty}$ is novel.
   • Bounds vs. lower bounds: The paper derives matching upper and lower bounds achieving the semiparametric efficiency bound. The sample complexity $n \gtrsim \bar{d}\log^c \bar{d}$ matches information-theoretic expectations.

   Verdict: SIGNIFICANT — The non-commutativity bottleneck and score whitening solution represent a clear non-obvious advance for tensor completion with structured observations. Most researchers working on low-rank methods or LLM evaluation should read this.

The proof follows a sophisticated five-term decomposition strategy:

1. Main Decomposition: The one-step estimator error decomposes as:

\[\hat{\psi} - \psi(T^*) = (P_n - P_*)\phi^* + R_{emp} + R_{bias}\]

where $\phi^* = s_{\eta}(Y,\eta^*)\langle H^*, X \rangle$ is the efficient influence function.

2. Empirical Process Control: The empirical remainder splits into:

\[R_{emp} = (P_n - P_*)[s_{\eta}(Y,\hat{\eta})\langle \hat{H} - H^*, X \rangle] + (P_n - P_*)[s_{\eta}(Y,\hat{\eta}) - s_{\eta}(Y,\eta^*))\langle H^*, X \rangle]\]

Key inequality for the direction error term:

\[|(P_n - P_*)[s_{\eta}(Y,\hat{\eta})\langle \hat{H} - H^*, X \rangle]| \leq C_A \|\Gamma\|_1 \sqrt{\frac{\bar{d}\log^c \bar{d}}{n}}\]

3. Operator Inversion Analysis: The critical bottleneck emerges from:

\[\hat{H} - H^* = \hat{A}^{-1}\hat{P}_T\Gamma - A^{-1}P_T\Gamma\]

The key technical insight is that while $|A^{-1}|_{op} \lesssim d_*$, the entrywise norm can be much worse:

\[\|A^{-1}\|_{\infty \to \infty} \leq C_A d_* \sqrt{\bar{d}}\]

This comes from $A = P_T G P_T$ where $G$ has heterogeneous Fisher weights $I(\eta^*)$ that don’t commute with the tangent space projection $P_T$.

4. Score Perturbation Control: Uses bounded signal assumption and score regularity:

\[|s_{\eta}(Y,\hat{\eta}) - s_{\eta}(Y,\eta^*)| \leq |\hat{\eta} - \eta^*| \leq \|\hat{T} - T^*\|_{\infty}\]

5. Subspace Perturbation Theory: Leverages sin-theta theorems adapted to pairwise observations:

\[\|\hat{P}_T - P_T\|_{op} \lesssim \frac{\|\hat{T} - T^*\|_F}{\lambda_{min}}\]

6. Score Whitening Innovation: The whitened score $\tilde{s}_{\eta}(y,\eta) = s_{\eta}(y,\eta)/I(\eta)$ satisfies:

\[\mathbb{E}[\partial_{\eta}\tilde{s}_{\eta}(Y,\eta^*)|X] = -1\]

This makes the effective operator isotropic: $A_0 = \frac{1}{d_*}P_T$, eliminating the $C_A$ bottleneck entirely.

Purely theoretical paper with no empirical evaluation.

What would constitute validation:

1. Synthetic experiments on low-rank matrices with varying Fisher information heterogeneity to demonstrate when $C_A$ becomes dimension-dependent vs. when score whitening helps
2. Real ARENA battle data analysis showing improved uncertainty quantification over current leaderboard methods
3. Comparison with naive plug-in estimators and existing BTL ranking methods
4. Computational experiments demonstrating the operator inversion challenges and whitening benefits

Limitations:

1. Bounded entries assumption $|T^*|_{\infty} \leq B$ - TECHNICAL (needed for score regularity but could potentially be relaxed)
2. Low-rank Tucker structure - RESTRICTIVE (rules out some natural LLM performance patterns, though reasonable for capturing major ability factors)
3. Zero-sum identification constraint - NATURAL (standard identifiability requirement for pairwise models)
4. Uniform sampling requirement for score whitening - RESTRICTIVE (real platforms have highly non-uniform data, though IPW extension addresses this)

5. Initial estimator needs entrywise control - TECHNICAL (challenging to achieve in practice but methods exist)

   Open problems:

6. Optimal $C_A$ characterization: Close the gap between the $O(1)$ lower bound and $O(\sqrt{\bar{d}})$ upper bound in Proposition 4.6. When exactly is the non-commutativity bottleneck unavoidable?
7. Beyond Tucker decomposition: Extend to other tensor formats (CP, tensor trains) or completely unstructured but sparse models for LLM evaluation across many contextual dimensions.