Answers are grounded in this paper and cite their sources.
The inverse reconstruction of material properties from optical extinction efficiency (Qext) is challenged by the apparent high dimensionality of Mie scattering spectra. We show that Qext exhibits highly compressible but representation-dependent structure across a six-polymer dielectric library. The discrete cosine transform (DCT) consistently provides more efficient representations (up to 6× higher) than the fast Fourier transform (FFT), particularly at high-fidelity energy-retention thresholds. Spectral entropy and modal analysis identify a pronounced fixed-basis complexity maximum near r ∼ 0.1 μm, where substantially more modes are required and the spectra become least compressible. This feature remains comparatively stable under 10% additive Gaussian noise. To distinguish transform-domain complexity from the dimensionality of the underlying spectral ensemble, we perform complementary singular-value decomposition (SVD) analyses. The complete library requires only 11, 16, and 38 globally optimized modes (relative to the 620-point computational grid) to retain 90%, 95%, and 99% of the singular-value energy, respectively, while local SVD ranks remain small near r ∼ 0.1 μm. Principal-angle and projection-distance analyses reveal an order-of-magnitude increase in local subspace reorientation across this crossover, whereas a second large-radius reorganization is accompanied by substantial growth in local SVD rank, indicating that fixed-basis complexity can arise from both geometric reorientation and increased local dimensionality. Consistent with this interpretation, a single globally fixed SVD basis loses reconstruction fidelity most rapidly near r ≈ 0.13 μm for the 95% and 99% energy representations. Representational sparsity alone, however, does not determine the minimum measurement budget. By mapping the DCT representation onto a sparse measurement architecture and performing a direct reconstruction using the fixed M = 165 subspace and P = 204 uniformly sampled wavelengths, we demonstrate a 42% reduction in spectral measurement count relative to a 350-point uniformly sampled reference grid. These results separate fixed-basis representation cost, adaptive linear dimensionality, and spectral-manifold geometry, and connect them to reduced-order sensing design.
No takes yet. Share an insight, caveat, or question.
Proity Nayeeb Akbar (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: