Technical Summary: Representation-Covariant Scale Exponents under Fisher-Induced Whitening

1. Problem Statement

The paper addresses a fundamental identifiability issue in high-dimensional statistical modeling: how parameter estimates change when the coordinate system of the data is transformed. Specifically, it investigates the KA (Kerimov–Alekberli) spectral model, which posits that spectral projections of the Fisher Information score vector follow a Generalized Gaussian Distribution (GGD). In this model, the scale and shape parameters of the GGD are assumed to scale as power laws of the corresponding Fisher eigenvalues, governed by exponents $\nu$ (scale) and $\alpha$ (shape).

The central problem arises when researchers apply Fisher-induced whitening (normalizing spectral projections by the square root of the eigenvalue, $\sqrt{\lambda_i}$), a common preprocessing step. The paper asks: When spectral projections are observed in this whitened coordinate system, which inferred exponent values are intrinsic to the data-generating process, and which are merely representation-dependent artifacts?

2. Methodology

The authors employ a rigorous combination of algebraic derivation, likelihood theory, and extensive Monte Carlo simulation to characterize the behavior of the exponents $\nu$ and $\alpha$ under whitening.

• Model Definition: The raw KA model defines spectral projections $Z_{it}$ conditionally on eigenvalues $\lambda_{it}$ as $Z_{it} \sim \text{GGD}(0, c\lambda_{it}^\nu, a\lambda_{it}^\alpha)$. This implies a log-linear relationship between the log-magnitude of projections and the log-eigenvalue ratios.
• Whitening Transformation: The whitened coordinate is defined as $Z^w_{it} = Z_{it}\lambda_{it}^{1/2}$. The authors derive that under this transformation, the scale exponent shifts to $\nu_w = \nu + 1/2$, while the shape exponent remains $\alpha_w = \alpha$.
• Estimation Approaches:
  • Ordinary Least Squares (OLS): Applied to the log-linear regression of projection magnitudes.
  • Profile Likelihood: The scale and shape parameters are profiled out to focus on the joint estimation of $(\nu, \alpha)$.
  • Adaptive Model Selection: An AIC-based selector is proposed to choose between a shape-free OLS model ($M_0$) and a full profile GGD model ($M_1$).
• Validation Framework:
  • Algebraic Invariants: Verification of exact identities (e.g., the shift of $\nu$) across 36,000 simulated datasets (360 parameter cells $\times$ 100 replications).
  • Synthetic Demonstration: A pipeline validation using data calibrated to BIS (Bank for International Settlements) credit-to-GDP indicators.
  • Prospective Protocol: Definition of the "LLM-0" audit criteria for future application to Large Language Model (LLM) data.

3. Key Contributions

A. Exact Finite-Sample Coordinate Identity (Proposition 1)

The paper establishes a distribution-free, algebraic identity for the OLS estimator of the scale exponent $\nu$. For any finite sample, any distribution, and any shape parameter, the OLS estimate in the whitened system ($\hat{\nu}^w_{OLS}$) is exactly related to the raw estimate ($\hat{\nu}^{raw}_{OLS}$) by:
$$\hat{\nu}^w_{OLS} = \hat{\nu}^{raw}_{OLS} + \frac{1}{2}$$
This result holds regardless of the underlying data distribution, provided the regressor is non-degenerate.

B. Profile Likelihood Covariance Theorem (Theorem 1)

The authors extend the identity to the Maximum Likelihood Estimation (MLE) framework. They prove that the profile log-likelihood transforms such that the profile MLE pairs satisfy:
$$\hat{\nu}^P_w = \hat{\nu}^P_{raw} + \frac{1}{2} \quad \text{and} \quad \hat{\alpha}^P_w = \hat{\alpha}^P_{raw}$$
This confirms that $\nu$ is representation-covariant (shifting by a universal constant of $+1/2$ under whitening), while $\alpha$ is representation-invariant.

C. Sign-Identifiability Boundary (Proposition 2)

The paper clarifies a critical distinction between sign recovery and unbiased estimation. It demonstrates that perfect sign recovery ($P(\text{sign}(\hat{\nu}) = \text{sign}(\nu)) = 1$) does not imply unbiased estimation ($E[\hat{\nu}] = \nu$).

• The critical boundary for sign flipping is identified as the precision-amplitude null at $\nu = -1/2$, not the Gaussian null at $\nu = 0$.
• At $\nu = -1/2$, positive shape contamination (from $\alpha > 0$) can bias the OLS estimator toward zero, potentially flipping the sign of the estimate even if the true parameter is negative.

D. Adaptive Model Selection

An adaptive AIC selector is introduced to choose between a shape-free model ($M_0$) and the profile GGD model ($M_1$).

• The selector is designed to be liberal (Type-I error rate $\approx 11.1\%$ at $\alpha=0$) to avoid the high efficiency loss (4–8$\times$ increase in RMSE) associated with using the shape-free OLS estimator when shape variation ($\alpha > 0$) is present.
• The selector achieves perfect power (1.00) for $\alpha \geq 0.20$.

4. Results

• Algebraic Verification: Monte Carlo benchmarks over 36,000 runs confirmed the Proposition 1 identity with a maximum deviation of $\leq 5.6 \times 10^{-9}$ (machine precision).
• Estimation Efficiency: The profile estimator reduced the Root Mean Square Error (RMSE) by a factor of 4 to 8 compared to OLS when shape variation was present. The efficiency gain was most pronounced for shorter time series ($T=100$) and diminished but persisted as $T$ increased.
• Bias Analysis: OLS estimators exhibited significant positive bias (shape contamination) as $\alpha$ increased. In contrast, the profile estimator remained centered near zero, with bias bounded only by the grid quantization step.
• Sign Recovery: The profile estimator achieved perfect sign recovery ($P=1.00$) across all test cells where $\nu \neq 0$, whereas OLS ranged from 0.950 to 0.997. However, the paper emphasizes that perfect sign recovery did not equate to zero RMSE due to grid quantization limits.
• Synthetic Demonstration: A pipeline test using BIS-calibrated synthetic data passed all "Gate" checks (Gate 1 deviation $\leq 5.96 \times 10^{-12}$), confirming the executability of the estimation pipeline on policy-relevant data structures.

5. Significance and Claims

The paper explicitly limits its scope to the theoretical and methodological characterization of the KA model's parameters under transformation.

• Theoretical Contribution: It provides a complete representation-covariance theory, establishing that $\nu$ is not an intrinsic structural parameter but a representation-dependent quantity that requires a $+1/2$ correction when comparing systems with different normalizations.
• Practical Implication: Researchers comparing scale exponents across systems (e.g., different financial datasets or different neural network architectures) must apply the $1/2$ correction to avoid comparing artifacts.
• Methodological Validation: The work validates the "frozen" estimation pipeline (Gate 1 $\to$ Check A/B $\to$ Gate 2) on synthetic data.
• Modest Claims on Empirical Findings:
  • The paper does not claim to have estimated $\nu$ or $\alpha$ from actual financial panel data or Large Language Models.
  • The BIS demonstration is explicitly labeled a "pipeline validation exercise," not an empirical finding.
  • The "LLM-0" protocol is defined strictly as a prospective specification for future work. The paper states that actual estimation of exponents from language model data is reserved for a subsequent publication, contingent on passing the LLM-0 audit.
  • The paper does not claim to validate the GGD model for any real-world data source; it only validates the algebraic identities and the behavior of the estimators under the assumed model.

In conclusion, the paper establishes the theoretical infrastructure necessary for future empirical work, clarifying how to correctly interpret spectral power-law exponents in the presence of Fisher-induced whitening.