Universality of exponential decay of the $d$-th SIR eigenvalue

§ Problem Statement

Setup

Let $d\in\mathbb N$ and $p\ge d$ . Consider the multiple-index regression model

Y=f(PX)+\varepsilon,

where $X\in\mathbb R^p$ is Gaussian with law $N(0,I_p)$ , $P\in\mathbb R^{d\times p}$ has orthonormal rows ( $PP^\top=I_d$ ), $f:\mathbb R^d\to\mathbb R$ is measurable, and $\varepsilon\in\mathbb R$ is independent of $X$ with $\mathbb E[\varepsilon]=0$ and $\mathbb E[\varepsilon^2]<\infty$ . Define the SIR population matrix

M:=\operatorname{Cov}\!\big(\mathbb E[X\mid Y]\big)\in\mathbb R^{p\times p},

and let $\lambda_1(M)\ge\cdots\ge\lambda_p(M)\ge0$ be its eigenvalues. The quantity of interest is the $d$ -th eigenvalue $\lambda_d(M)$ .

This setup follows Huang et al. (2023).

Unsolved Problem

Problem 2. Characterize the largest class $\mathcal F$ of link functions (or laws of random link functions) such that, for every $d$ (and uniformly over $p\ge d$ , admissible $P$ , and admissible noise laws), there exist constants $C,\beta>0$ independent of $d$ for which

\lambda_d\!\big(\operatorname{Cov}(\mathbb E[X\mid Y])\big)\le C e^{-\beta d}.

For deterministic $f$ , this is a pointwise bound on $M$ ; for random $f$ , require the bound with high probability over the draw of $f$ (equivalently, specify a probability level tending to $1$ as $d\to\infty$ ). In particular, determine whether this exponential decay of $\lambda_d(M)$ holds beyond Gaussian-process-based random-link settings.

§ Discussion

Loading discussion…

§ Significance & Implications

In Huang, Tian, and Lin (arXiv:2305.04340v2), this universality question is explicitly posed as open in the discussion of gSNR decay. Resolving whether exponential decay is universal, rather than tied to specific random-function assumptions, is central to understanding when SIR degrades as structural dimension grows.

§ Known Partial Results

Huang et al. (2023): In the canonical arXiv v2 source, the universality question is open and no general beyond-model theorem is claimed there. Later manuscript versions report Gaussian-process-specific exponential-decay progress, but this does not resolve universality beyond GP-type assumptions.

§ References

[1]

On the Structural Dimension of Sliced Inverse Regression

Dongming Huang, Songtao Tian, Qian Lin (2023)

arXiv preprint

📍 Section 5 (Discussions), paragraph beginning “Our findings raise several open questions,” second open question on exponential decay of gSNR (arXiv v2).

Canonical source version for this problem statement.

Link ↗arXiv ↗

§ Tags

sir-eigenvalues gaussian-process high-dimensional-statistics dimension-reduction universality