Rigorous detection thresholds for spiked models

Determine rigorous signal detection thresholds for spiked models as a function of data type; specifically, establish precise detectability conditions for low-rank signals embedded in high-dimensional noise for spiked matrix and spiked tensor settings.

Background

The paper surveys state-of-the-art methods for signal detection in high-dimensional data and highlights limitations of principal component analysis in nearly continuous spectra. Within this context, it points out that theoretical understanding of detection limits in spiked models is incomplete.

The authors emphasize that while random matrix theory provides a mature framework for spiked matrices, extending rigorous detection thresholds to other data modalities—particularly tensors—remains challenging, partly due to connections with spin-glass phenomena and the comparatively less developed state of random tensor theory.

References

Finally, note that, on the mathematical side, some questions about a rigorous signal detection threshold for spiked models remain open, depending on the nature of the data.

— Functional Renormalization for Signal Detection: Dimensional Analysis and Dimensional Phase Transition for Nearly Continuous Spectra Effective Field Theory  (2507.01064 - Finotello et al., 30 Jun 2025) in Section 2: Contribution and Related Work

However, in practical scenarios, it is generally unknown whether some of the spikes fall below this threshold, i.e., whether $t_i<\sqrt{c}$ for some $i$.

— Semi-Blind Channel Estimation for Dynamic NTN Systems via Spiked Random Matrix Theory  (2608.25694 - Zhang et al., 26 Aug 2026) in Section 4, subsection “Scenario with Some $t_i$ Satisfying $0<t_i<\sqrt{c}$”

Neither the variance estimate nor the dimension exponent is required to be optimal; the optimal detection threshold for this kernel family remains to be determined.

— Cubic spectral cancellation and random geometric graph detection: a quadratic-kernel counterexample  (2609.11662 - Luo, 10 Sep 2026) in Section 6, “Spectral cancellation and the scope of the conclusions”

It is not clear if the one sample approach applied here can be extended to deal with the scenario when a big spike presents in the noise part as we previously discussed in \citet{BaoCheongLeeLi2025}. It will be left a a furture research direction.

— Spike Estimation from Heteroscedastic Noise via Random Splitting  (2609.11169 - Bao et al., 10 Sep 2026) in Section 1, Introduction, paragraph discussing the relation to Bao, Cheong, Lee, and Li (2025)

This is a validation on an injected-spike surrogate: it perturbs the population covariance with a known rank-one spike, so it tests the threshold arithmetic rather than whether a real modality's spikes sit near the edge; estimating a real spike's strength relative to $s_c$ is left open (Section~\ref{sec:discussion}).

— Spectrum Is Not Enough: A Random-Matrix Account of Data-Free Preconditioning for Differentially Private Learning  (2610.21312 - Zhang et al., 15 Oct 2026) in Section 4, Section “Pillar II: the failure boundary is a BBP threshold”

This does reveal an avenue of future exploration: given $1\ge p\gg\frac{1}{\log(n)}$ and $1-q\gg1/n2,$ are there any methods that can successfully distinguish or recover when $\theta<1$?

— Distinguishability and Recovery for a Doubly Sparse Spiked Wigner Model  (2610.23033 - Dumitriu et al., 16 Oct 2026) in Section 5, subsection “Distinguishability and Recovery: The Full Picture,” paragraph “PCA”