Accelerating Classical and Quantum Tensor PCA

Spectral methods are a leading approach for tensor PCA with a ``spiked" Gaussian tensor. The methods use the spectrum of a linear operator in a vector space with exponentially high dimension and in Ref. 1 it was shown that quantum algorithms could then lead to an exponential space saving as well as a quartic speedup over classical. Here we show how to accelerate both classical and quantum algorithms quadratically, while maintaining the same quartic separation between them. That is, our classical algorithm here is quadratically faster than the original classical algorithm, while the quantum algorithm is eigth-power faster than the original classical algorithm. We then give a further modification of the quantum algorithm, increasing its speedup over the modified classical algorithm to the sixth power. We only prove these speedups for detection, rather than recovery, but we give a strong plausibility argument that our algorithm achieves recovery also. Note added: After this paper was prepared, A. Schmidhuber pointed out to me Ref. 3. This improves the best existing bounds on the spectral norm of a certain random operator. Because the norm of this operator enters into the runtime, with this improvement on the norm, we no longer have a provable polynomial speedup. Our results are phrased in terms of certain properties of the spectrum of this operator (not merely the largest eigenvalue but also the density of states). So, if these properties still hold, the speedup still holds. Rather than modify the paper, I have left it unchanged but added a section at the end discussing the needed property of density of states and considering for which problems (there are several problems for which this kind of quartic quantum speedup has been used and the techniques here will likely be applicable to several of them) the property is likely to hold.

💡 Research Summary

The paper studies the “spiked” Gaussian tensor model, a prototypical planted‑signal problem where a p‑order tensor T₀ = λ v^{⊗p}+G consists of a rank‑one signal v (the “spike”) plus i.i.d. Gaussian noise G. The goal is either to detect the presence of the spike (detection) or to recover v (recovery). Prior work showed that the problem can be tackled by spectral methods: one builds a large linear operator H(T₀) acting on a Hilbert space of dimension exponential in a parameter n_bos, and then estimates the largest eigenvalue of H. Classically, Lanczos (or similar Krylov‑subspace) methods require time proportional to the full Hilbert space dimension, while a quantum algorithm can prepare a specially chosen input state with non‑trivial overlap with the top eigenvector, apply phase estimation, and use amplitude amplification. This yields a quartic speed‑up over the classical method (the quantum runtime scales as the fourth root of the Hilbert space dimension).

The authors propose a new scheme that simultaneously speeds up both the classical and quantum algorithms by a factor of two, while preserving the quartic separation. The key idea is to halve the parameter n_bos, thereby reducing the Hilbert space dimension exponentially. Because the dimension is now smaller, the cost of each matrix‑vector multiplication (or its quantum analogue) drops by a square‑root factor. To compensate for the reduced n_bos, they introduce an approximate projection onto the subspace spanned by eigenvectors of H with eigenvalues above a carefully chosen threshold. This subspace is still small compared to the full space, but it now contains the “ideal” signal eigenvector together with a few other eigenvectors. A random vector would have only a tiny amplitude in this subspace, whereas the specially prepared input state retains a non‑negligible component. Consequently, the probability of successful projection (or of obtaining the correct phase in the quantum case) improves from 1/Dim to 1/√Dim, giving the quadratic speed‑up.

In the classical algorithm, Lanczos is used to implement the approximate projection; the reduced n_bos means each Lanczos step is cheaper, and the overall runtime becomes roughly the square root of the original. In the quantum algorithm, the same projection is realized via phase estimation. The chosen input state still overlaps with the target eigenvector by Θ(1/√Dim), so after phase estimation the success probability is again Θ(1/√Dim). Amplitude amplification then boosts this to constant probability, yielding a total runtime that is eight times faster than the original quantum algorithm (the factor 2 from the smaller dimension multiplied by the original quartic improvement).

A further quantum improvement is obtained by recursive subspace projection. The authors split the system into two halves, apply approximate projections to each half independently, and only accept when both succeed. Failures are repaired locally, so the overall success probability multiplies, effectively raising the overlap to the fourth power. Repeating this recursion (quarters, eighths, etc.) leads to a multi‑step quantum algorithm whose runtime is up to twelve times faster than the original classical method (or six times faster than the modified classical algorithm). This “multi‑step” technique is described in Section V.

All speed‑up claims are proved for the detection problem. For recovery, the authors provide a plausibility argument: the ideal input state (the n_bos‑th tensor power of the spike) concentrates most of its amplitude in the projected subspace, and the noisy version used in practice is shown (Section IV) to retain enough overlap. However, a rigorous recovery guarantee is not given.

A crucial caveat appears in the “Note added”. After the paper was written, Schmidhuber’s work (Ref.