Finite Rate Innovation Principles

• Finite Rate Innovation (FRI) principles are a signal processing framework that recovers continuous, non-bandlimited signals using a finite set of parameters per interval.
• FRI employs specialized sampling kernels and annihilating-filter techniques to enable reconstruction at sub-Nyquist rates, even in noisy conditions.
• The framework supports various applications from imaging to radar by enabling efficient recovery methods that bypass conventional bandwidth limitations.

Finite Rate of Innovation (FRI) principles constitute a unifying signal processing framework that enables the exact recovery of classes of continuous, generally non-bandlimited signals that are specified by a finite set of parameters per interval—termed the "innovation rate." Fundamental to FRI is the insight that the minimal sampling rate for perfect reconstruction is dictated not by the signal's bandwidth but by the number of degrees of freedom per unit time, frequently enabling sub-Nyquist acquisition. The theory leverages specialized sampling kernels, annihilating-filter or Prony-type estimators, and, more recently, model-based and data-driven learning paradigms, admitting generalized signal models, sampling architectures, and robust recovery in the presence of noise and system nonidealities.

1. Signal Models and Rate of Innovation

FRI signals are characterized by possessing a finite number of degrees of freedom (innovations) per observation interval, irrespective of their bandwidth. Canonical forms include streams of $L$ weighted, time-shifted pulses: $x(t) = \sum_{\ell=1}^L a_\ell\,g(t-t_\ell),$ where $g$ is a known pulse shape and $\{a_\ell, t_\ell\}$ are unknown amplitudes and time shifts, yielding $2L$ real degrees of freedom in $[0, T]$ and hence an innovation rate $\rho = 2L/T$ (Ben-Haim et al., 2010). This paradigm extends to unions of shift-invariant subspaces, piecewise polynomials or harmonics, periodic pulse trains, and higher-dimensional settings. In the multidimensional case, e.g., MRI or imaging, the innovation rate generalizes to degrees of freedom per unit area or volume (Ongie et al., 2015, Shastri et al., 2019). On circulant graphs, $K$-sparse vertex signals inherit a graph-FRI structure: any $K$-sparse vector can be exactly recovered from $2K$ graph-Fourier samples (Kotzagiannidis et al., 2016).

2. Sampling Kernels and Acquisition Architectures

FRI recovery hinges on the design of analog (or digital) sampling kernels that ensure a linear measurement operator maps the analog signal to a structured set of inner products from which innovations can be extracted.

• Kernel Requirements: The kernel $h(t)$ must satisfy a generalized Strang–Fix (moment-reproducing or exponential-reproducing) property, so that the measurement sequence admits a sum-of-exponentials structure (Tur et al., 2010, Shastri et al., 2019, Nitsure et al., 28 Sep 2025).
• Compact Support and Alias Cancellation: In both 1-D and 2-D, alias cancellation conditions are imposed in the Fourier domain to guarantee that each measurement contributes only to a specific set of spectral components (Tur et al., 2010, Shastri et al., 2019). The Paley–Wiener theorem is used to derive compactly supported kernels that satisfy these properties (Shastri et al., 2019). Sum-of-sincs (SoS), exponential-reproducing (e-spline), and sum-of-modulated-spline (SMS) kernels are prominent practical choices.
• Hardware-Implementable and Learnable Kernels: “Learnable Kernels for FRI” demonstrates that jointly optimizing the kernel and recovery encoder in an end-to-end fashion yields analog-implementable filters (e.g., two-pole Sallen-Key), with superior reconstruction accuracy and noise robustness (Nitsure et al., 28 Sep 2025).

Panels of acquisition architectures encompass uniform sampling, event-driven sampling (e.g., time encoding machines—TEMs), modulo ADCs, low-rate spectrographic observations, and graph-spectral projections (Kamath et al., 2021, Mulleti et al., 2022, Kotzagiannidis et al., 2016).

3. Reconstruction Algorithms: Spectral Estimation, Annihilating Filter, and Generalizations

The FRI recovery problem is universally cast as the estimation of innovations from a finite set of generalized measurements that exhibit an underlying sum-of-exponentials (or, more broadly, shift-invariant) structure.

3.1 Classical Pipeline

• Spectral/Prony Reduction: The linear measurements are mapped (via moment generation or kernel inversion) to a finite sequence $m_n$ satisfying $m_n = \sum_{\ell=1}^K b_\ell u_\ell^n$. A length-$(K+1)$ annihilating filter $h$ (with $h_0 = 1$) is found via SVD nullspace extraction on a Hankel/Toeplitz data matrix such that $\sum_{j=0}^K h_j m_{n+j}=0$ (Legros et al., 2022, Tur et al., 2010).
• Parameter Recovery: The $u_\ell$ are retrieved from the roots of the annihilating polynomial $H(z)=\sum_{j=0}^K h_j z^{-j}$, and innovations (delays, nodes) follow from $\arg(u_\ell)$; amplitudes are recovered from a Vandermonde system (Tur et al., 2010, Ongie et al., 2015).
• Higher-dimensional and Graph Settings: In 2-D, edge-curve recovery employs a 2D version of annihilating filters via block Toeplitz matrices and robust SVD-based aggregation (Ongie et al., 2015, Shastri et al., 2019). The graph-FRI case maps K-sparse vertex signals to sum-of-exponentials in the GFT, enabling Prony-based support reconstruction (Kotzagiannidis et al., 2016).
• Noise Handling: Cadzow denoising (alternating Toeplitz/low-rank projection) is standard for enhancing robustness. Total-least-squares and structured matrix denoisers further improve estimation in perturbative settings (Legros et al., 2022, Simeoni et al., 2020).

3.2 Generalizations

• Arbitrary Linear Measurements (genFRI): The generalised FRI (genFRI) framework replaces explicit moment operators with arbitrary linear measurements, enforcing low-rank structure via rank constraints or regularization on the Toeplitz (or Hankel) matrices of the measurements. Efficient solvers such as proximal gradient descent with plug-and-play Cadzow denoising achieve practical convergence guarantees and scalable recovery (Simeoni et al., 2020, Liu et al., 2 Feb 2026).
• Fully Nonlinear or Nonparametric Acquisition: Least-squares (LS) optimization under a stability hypothesis unifies and generalizes prior FRI approaches, extending recovery guarantees to nonlinear sampling (e.g., memoryless nonlinearities, event-driven ADCs), arbitrary signal manifolds, and composite operators (Michaeli et al., 2011, Florescu, 2023). Provided the number of measurements matches the innovation rate and the parametric map is injective and stable, recovery is provably unique.
• Time Encoding and Event-Driven Sampling: TEM-based FRI frameworks recover innovation parameters from the timing of events (e.g., integrate-and-fire, zero-crossing, area modulation), often requiring only mild smoothness and monotonicity of the kernel and precise area rules for spike intervals (Kamath et al., 2021, Naaman et al., 2021, Florescu, 2023). Recent theoretical advances show that even for unknown or general kernels, innovation parameters can be extracted from local sequences of integrals without explicit kernel parametrization or global matrix methods.

4. Noise, Performance Bounds, Conditioning, and Instability

In noise-free scenarios, FRI techniques enable exact signal reconstruction at the innovation rate; in practical settings, several limitations arise:

• Cramér–Rao Bound (CRB) and Ziv–Zakai Bound (ZZB): The fundamental mean-square error (MSE) achievable for innovation recovery scales linearly with the innovation rate and noise variance:

$\mathsf{MSE} \ge K\,\sigma_c^2,$

implying that in continuous-time noise, the per-innovation error is bounded by $\sigma_c^2$ (Ben-Haim et al., 2010, Liu et al., 2 Feb 2026).

• Sampled System Bounds: When measurements are finite-dimensional samples, the Fisher information matrix for the parameters depends on both sampling and signal model geometry, and MSE is minimized when the sampling subspace captures all innovation directions. Oversampling (beyond the innovation rate) can dramatically decrease error in infinite union-of-subspaces models (Ben-Haim et al., 2010).
• Breakdown Regimes and Subspace-Swaps: Subspace-based algorithms (e.g., Prony, matrix pencil) exhibit phase transitions in noise, known as subspace swaps: below a breakdown PSNR (dependent on pulse spacing and kernel choice) the algorithm catastrophically fails, with error variance exploding (Leung et al., 2022, Leung et al., 2019). Learning-based methods (e.g., deep unfolding, encoder-decoders) can delay this breakdown regime by 15–30 dB (Leung et al., 2022, Leung et al., 2019).
• Numerical Conditioning: Classical annihilating-filter approaches hinge on the conditioning of Vandermonde matrices, which degrades if pulses are closely spaced. Compactly supported SoS kernels and stable numerical design of sampling/reconstruction filters are critical for large-scale or high-innovation-rate problems (Tur et al., 2010).

5. Learning-Based and Jointly Optimized FRI Frameworks

Recent advances leverage data-driven learning to enhance classical FRI robustness and practical deployability:

• Deep Neural Techniques: Deep unfolding of denoising procedures and encoder-decoder architectures learn innovation extraction directly from data, improving accuracy near the breakdown SNR or in presence of unknown kernels. Unfolded Cadzow or Projected Wirtinger Gradient Descent (PWGD) networks can match classical performance at high SNR and extend operating regions substantially (Leung et al., 2022, Leung et al., 2019).
• Joint Sampling and Reconstruction Optimization: Methods such as "Learning to Sample" jointly select Fourier sampling locations and optimize neural recovery models (e.g., LISTA, sparse unrolled ISTA), achieving improved mean-squared error and even breaking theoretical $2K$ sample limits in noiseless cases (Mulleti et al., 2021).
• Learnable and Hardware-Realizable Kernels: End-to-end optimization that includes both kernel parameterization (e.g., as sum-of-exponentials realizable via Sallen–Key filters) and compact convolutional neural encoders achieves hardware validation with low model complexity, reduced sampling rates, and robust recovery under noise (Nitsure et al., 28 Sep 2025).

6. Extensions: Event-Driven, Modulo, Graph, and Multidimensional FRI

The FRI principles generalize beyond uniform scalar sampling:

• Modulo Sampling: FRI signals can be sampled after modulo folding, which mitigates dynamic range limitations and allows for reliable identification from ambiguous folded samples using unwrapping and ambiguous sample annihilating-filter techniques with optimal sample complexity (Mulleti et al., 2022).
• Graph-FRI: Signal sparsity in the vertex domain of circulant graphs is recovered using minimal graph-Fourier samples, leveraging the DFT-diagonalized graph Laplacian. Sampling and reconstruction mirror classical FRI but in the spectral (dual) domain of the underlying topology (Kotzagiannidis et al., 2016).
• 2-D and Higher Dimensions: Super-resolution MRI and spectral imaging methods apply 2-D FRI principles, modeling edge curves or spectral ridges as the zero level set of bandlimited polynomials, and reconstructing from low-frequency k-space or time-frequency representations using 2-D annihilating masks and robust SVD techniques (Ongie et al., 2015, Legros et al., 2022).
• Time Encoding Machines: IF-TEM and C-TEM enable FRI sampling with asynchronous, event-driven measurements, offering energy efficiency and robustness to clock uncertainties. Advanced generalized models enable kernel-agnostic innovation recovery via local integral ratios (Naaman et al., 2021, Florescu, 2023).

7. Applications, Impact, and Open Directions

The impact of FRI spans radar, ultrasound, biological imaging, cognitive radio, and ultralow-power ADC design. Application demonstrations include:

• Ultrasound Imaging: Two orders of magnitude sampling-rate reductions with sub-millimeter localization accuracy are achieved by FRI processing of echo time series (Tur et al., 2010).
• Super-Resolution MRI: Subpixel edge mask estimation from low-pass data enables accurate recovery of anatomical boundaries beyond the nominal bandwidth (Ongie et al., 2015).
• Full-Space DOA Estimation: FRI models and proximal-gradient algorithms facilitate gridless multiuser angle estimation with STAR-RIS architectures in mmWave MIMO systems, quantified by Ziv–Zakai bounds (Liu et al., 2 Feb 2026).
• Resource-Constrained Edge Devices: Joint hardware-algorithm pipeline deployment (e.g., Sallen–Key filter + CNN encoder) validates FRI in real signals with edge-optimized architectures (Nitsure et al., 28 Sep 2025).

FRI remains an active field, with ongoing developments in deep model-based methods, kernel learning, optimal sampling under noise or system imperfections, and generalizations to nonlinear, manifold-structured, or hybrid acquisition modalities. Open questions include optimal sensor/co-design under practical constraints, handling nonidealities in analog hardware, further characterization of breakdown phenomena, and scalable algorithms for massive-scale innovation models.