Resolution and Robustness Bounds for Reconstructive Spectrometers

TL;DR

This work establishes a physically grounded framework for reconstructive spectrometers by deriving a Fisher-information bound on noise-induced reconstruction error, $\sigma_\epsilon^2 \mathrm{Tr}(\mathbf{G}^{+})$ with $\mathbf{G}=\mathbf{A}^T\mathbf{A}$, and linking it to physically meaningful parameters such as the spectral correlation length $\Gamma_{\mathrm{corr}}$, mean transmittance $T_0$, and channel counts $M,N$. It introduces analytic scaling functions $J(a)$ and $\tilde{J}(a)$ for the correlated, Toeplitz-like correlation structure with $a=\Gamma_{\mathrm{corr}}/\Delta\omega$, and derives explicit expressions for the effective spectral resolution $\Delta\omega_{\min}$, including under- and over-determined regimes and a high-SNR condition for super-resolution $\Delta\omega_{\min}<\Gamma_{\mathrm{corr}}$. The theory is validated through random-matrix simulations and full-wave on-chip FDTD results, revealing trade-offs between $\Gamma_{\mathrm{corr}}$ and $T_0$ and predicting an optimal device size in diffusive regimes. An inverse-design study shows potential to push beyond the Lorentzian-speckle framework by engineering non-Lorentzian correlations, highlighting avenues to further enhance robustness and resolution in practical reconstructive spectrometers.

Abstract

Reconstructive spectrometers are a promising emerging class of devices that combine complex light scattering with inference to enable compact, high-resolution spectrometry. Thus far, the physical determinants of these devices' performance remain under-explored. We show that under a broad range of conditions, the noise-induced error for spectral reconstruction is governed by the Fisher information. We then use random matrix theory to derive a closed-form relation linking the variance bound to a set of key physical parameters: the spectral correlation length, the mean transmittance, and the number of frequency and measurement channels. The analysis reveals certain fundamental trade-offs between these physical parameters, and establishes the conditions for a spectrometer to achieve ``super-resolution'' below the limit set by the spectral correlation length. Our theory is confirmed using numerical validations with a random matrix model as well as full-wave simulations. These results establish a physically-grounded framework for designing and analyzing performant and noise-robust reconstructive spectrometers.

Figures (6)

• Figure 1: (a) Schematic of a reconstructive spectrometer with $M$ output ports. (b) Contents of a typical transmission matrix $\mathbf{A}$, of size $M = 50$ and $N = 100$, generated from the RMT model fyodorov2016randomSM. Each of the $N$ columns contains the output intensities at a given frequency. (c) Simulated spectral reconstruction under zero noise (upper panel) and noise level $\sigma_\epsilon = 0.002$ (lower panel). The ground truth (black line) is reconstructed via the transmission matrix from (b) using a pseudoinverse (orange stars), or a neural network (NN) trained on spectrum-intensity pairs (blue circles). (d) Noise-induced mean squared error (MSE) versus $\operatorname{Tr}[\mathbf{G}^{+}]$, for an ensemble of $\mathbf{A}$ matrices with $\sigma_\epsilon = 0.002$. The samples are generated by varying the RMT model's coupling strength $\gamma \in [0.002, 0.01]$, keeping all other parameters constant SM.
• Figure 2: (a) Structure of the matrix $\mathbf{C} = \mathbf{a}^T\mathbf{a}$ for a typical scatterer, showing its near-Toeplitz form. The $\mathbf{A}$ matrix is generated by random matrix theory (RMT) using a Mahaux--Weidenmüller model fyodorov2016random. (b) Calculated correlation function between columns of the $\mathbf{A}$ matrix from (a), showing a Lorentzian decay with frequency spacing. The FWHM is denoted by $\Gamma_{\mathrm{corr}}$. (c) Variation of $\operatorname{Tr}[\mathbf{G}^{+}]$ with $\Gamma_{\textrm{corr}}$, using an RMT ensemble in which the coupling parameter $\gamma$ varies from 0.005 to 0.1, with $N_{\mathrm{modes}} = 500$ internal cavity modes. The results calculated directly from the $\mathbf{A}$ matrices (stars) agree with the analytic results based on $\Gamma_{\textrm{corr}}$ and the mean transmittance $T_0$ (solid curves), for both the under-determined and over-determined regimes. (d) Effective resolution $\Delta \omega_{\textrm{min}}$ and spectral correlation length $\Gamma_{\textrm{corr}}$ versus coupling strength $\gamma$. Each data point is averaged over 200 samples, using $M=50$, $N=100$, and $N_{\mathrm{modes}} = 500$. The effective resolution $\Delta \omega_{\textrm{min}}$ is calculated from Eq. \ref{['eq:Emin_underdet']} using $\delta^2_{\mathrm{th}} = 4\times10^{-4}$. Over a range of $\gamma$, the device exhibits super-resolution ($\Delta \omega_{\textrm{min}} < \Gamma_{\textrm{corr}}$).
• Figure 3: (a) Schematic of the on-chip reconstructive spectrometer with a $L\times L$ scattering region. (b) Spectral correlation function from FDTD simulations for $L=30\,\mu\textrm{m}$ (blue circles), fitted to a Lorentzian (red dashes) to extract $\Gamma_{\rm corr}$. (c) Scaling of $\Gamma_{\rm corr}$ (upper) and mean transmittance $T_0$ (lower) with cavity size $L$. Data points (averaged over 10 samples) agree well with the diffusive scaling laws $\Gamma_{\rm corr} \propto L^{-2}$ and $T_0 \propto L^{-1}$ (dashed lines). (d) Reconstruction error metric $\operatorname{Tr}[\mathbf{G}^{+}]$ versus $L$. The results obtained directly from the simulated $\mathbf{A}$ matrices (blue stars) are close to the theoretical predictions based on Eq. \ref{['eq:inv-trace-mean']} (blue line). The minimum of the latter matches the optimal cavity size $L_{\mathrm{opt}}$ derived from the fits in (c) (vertical dashes). (e) Plot of the effective resolution $\Delta\omega_{\textrm{min}}$ (orange circles) and $\Gamma_{\rm corr}$ (red circles) versus $L$, for detection threshold $\delta_{\mathrm{th}}^2 = 2 \times 10^{-4}$ and noise level $\sigma_{\epsilon} = 10^{-2}$. (f) Effective resolution $\Delta\omega_{\textrm{min}}$ versus detection threshold $\delta_{\mathrm{th}}^2$, using two choices of cavity size: $L_{\mathrm{opt}} = 17.9\,\mu\textrm{m}$ (blue circles), which minimizes $\operatorname{Tr}[\mathbf{G}^{+}]$ as shown in (d); and $L=25\,\mu\textrm{m}$, which minimizes $\Delta\omega_{\textrm{min}}$ at $\delta_{\mathrm{th}}^2 = 2 \times 10^{-4}$, as shown in (e). Here we again use $\sigma_{\epsilon} = 10^{-2}$.
• Figure S1: (a)--(c) Noise-induced MSE versus $\operatorname{Tr}[\mathbf{G}^{+}]$, obtained from RMT simulations with pseudoinverse and NN reconstruction schemes. Each data point is an ensemble average over fixed RMT parameters; the different subplots correspond to different $M \in \{20, 100, 150\}$, with fixed $N = 100$, and in each subplot we vary $\gamma \in [0.005, 0.01]$. The number of cavity modes is $P = 500$. The dashes show the Cramér--Rao bound $\sigma_\epsilon^{2}\,\mathrm{Tr}(\mathbf{G}^{+})$, which quantitatively fits the pseudoinverse results. (d) Bias term for different $M$, with $N = 100, \gamma = 0.005$ and all other parameters the same as in (a)--(c). The pseudoinverse-based reconstruction experiences significantly increasing bias as $M$ is decreased below $N$; for NN reconstruction, the bias increases less sharply.
• Figure S2: Variation of the scaling function $J(a)$ with the normalized correlation length $a = \Gamma_{\text{corr}}/\Delta \omega$. The solid blue line represents the exact analytical form for the over-determined case ($M \ge N$), which converges to the asymptotic exponential behavior $\sim e^{\pi a}$ (orange dashed line) for large correlations. The green solid line depicts the modified function $\tilde{J}(a)$ for an under-determined system ($M/N = 0.5$), exhibiting suppressed error growth due to spectral smoothing. All functions converge to $\pi$ at $a=0$, representing the baseline error for uncorrelated channels.