Back to search
2411.09004

The Geometry of the Deep Linear Network

Govind Menon

correctmedium confidence
Category
Not specified
Journal tier
Strong Field
Processed
Sep 28, 2025, 12:56 AM

Audit review

The paper’s Theorem 10 gives vol(OW) = c_d^{N−1} sqrt(van(Σ^2)/van(Σ^{2/N})) (see the displayed formula and its product form under Theorem 10) . The candidate solution re-derives exactly this expression by pulling back the Frobenius metric to O(d)^{N−1}, diagonalizing a tridiagonal Toeplitz form in each (i,j)-block via the discrete-sine basis, and evaluating the determinant using a Chebyshev/roots-of-unity identity. This is the same structural argument used in the paper’s Section 8 (block tridiagonal metric and Chebyshev diagonalization) , together with the SVD-based parametrization of the balanced manifold M . A consistency check at N=2 yields vol(OW) = c_d ∏_{i<j} √(σ_i+σ_j), which both approaches satisfy. The mention of a denominator van(Σ^{2N}) in the prompt appears to be a typo not present in the provided PDF (which has Σ^{2/N}); with the correct exponent and the square root spanning the entire fraction (as seen from the product expression), the paper and model agree.

Referee report (LaTeX)

\textbf{Recommendation:} minor revisions

\textbf{Journal Tier:} strong field

\textbf{Justification:}

The result is clean and presented within a coherent geometric framework for the DLN. Theorem 10’s entropy/volume formula for group orbits upstairs is derived via a careful pullback-metric computation and an explicit orthonormal basis, leading to a Vandermonde-type expression. The presentation could be further clarified where square-root placement and exponents might be misread in line notation; relying on the product form helps. With minor expository tweaks (explicit parentheses around square roots and reiteration of assumptions on singular values), the exposition will be highly readable.