Super-Resolution Real-Space Inversion

• Super-resolution real-space inversion is a technique that reconstructs high-resolution objects directly in real space from degraded measurements using an explicit forward model.
• It employs various inversion strategies, from exact spectral estimation to convex optimization, to overcome limitations posed by blur, undersampling, or noise.
• Applications include point-source recovery, sky mapping, MRI, and ultrafast scattering, demonstrating significant resolution improvements over conventional methods.

A super-resolution real-space inversion algorithm, in the sense used across several inverse-problem literatures, denotes a reconstruction method that estimates a high-resolution object directly in real space from degraded measurements by enforcing an explicit forward model and a regularizing prior. The reconstructed variable may be a set of off-grid point locations, a sky map, a complex-valued MR image, a pair-distance distribution, or a positive source density, but the common feature is that the unknown is represented in real space and the super-resolved estimate is obtained by inversion rather than by interpolation alone. This usage is exemplified by off-the-grid Fourier inversion for point sources (Huang et al., 2015), regularized map-making for scanning telescopes (Orieux et al., 2011), measurement-aware pair-density recovery in ultrafast scattering (Natan, 2021), and real-space point-source deconvolution by superposition of virtual emitters (Martínez et al., 2018).

1. Scope and conceptual boundaries

The term does not denote a single algorithm. It instead covers a family of methods whose forward operators differ by modality and whose inversion strategies range from exact harmonic retrieval to quadratic regularized optimization, alternating projections, convex sparse recovery, and nonlinear source fitting.

Representative paper	Reconstructed variable	Inversion character
"Super-Resolution Off the Grid" (Huang et al., 2015)	point-source locations and weights	random low-frequency Fourier measurements; Jennrich/SVD tensor decomposition
"An Algorithm for Exact Super-resolution and Phase Retrieval" (Chen et al., 2013)	continuous spike train	autocorrelation recovery, then deterministic sorting/disentangling
"Super-resolution in map-making based on a physical instrument model and regularized inversion" (Orieux et al., 2011)	high-resolution sky map	quadratic regularized inversion with conjugate gradient
"Super-Resolution with Structured Motion" (Litterio et al., 21 May 2025)	high-resolution image $J$	convex inversion with $\ell_1$ or TV under motion occupancy and box integration
"Real-Space Inversion and Super-Resolution of Ultrafast Scattering" (Natan, 2021)	pair-density difference	NSK dictionary with $\ell_1$- or $\ell_2$-regularized deconvolution
"Superresolution method for data deconvolution by superposition of point sources" (Martínez et al., 2018)	positive target function $R(x)$	equal-intensity virtual point-source fit in real space

A useful boundary of the concept appears in papers that are only partially real-space. The 4D Flow MRI method formulates the degradation as $y_j = SHx_j + n_j$ in the spatial domain, but evaluates the closed-form Tikhonov solution analytically in the Fourier domain, so it is best described as a real-space inverse problem with a Fourier-domain solver (Turenne et al., 25 Sep 2025). The nonlinear SIM method expresses the reconstructed image as a spatial-domain weighted recombination of raw images, yet relies on Fourier-domain parameter estimation and compensation, so it is explicitly hybrid rather than purely spatial (Zhang et al., 2023). The one-step diffusion method IDaS-SR is relevant to inverse-problem super-resolution, but its inversion is carried out primarily in latent diffusion trajectory space rather than in pixel space (Weng et al., 27 Apr 2026). The raw/RGB fusion model for real-scene SR is physically motivated and real-space in its imaging interpretation, but its reconstruction is a learned two-branch CNN rather than an explicit optimization-based inverse solver (Xu et al., 2021).

2. Canonical forward models

Despite their diversity, these methods share a forward-model-first structure. A generic statement appears in the scattering-statistics framework, which poses the inverse problem as

$y = \Gamma x + b,$

with feasible set

$\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$

Here the unknown $x$ is a real-space signal or image, while $\Gamma$ specifies the measurement physics (Dokmanić et al., 2016).

Several representative super-resolution models instantiate this template in different ways. In off-grid point-source recovery, the signal is

$\ell_1$0

and the bandlimited Fourier measurements are

$\ell_1$1

with off-grid locations $\ell_1$2 and minimum separation

$\ell_1$3

The inversion target is the set of real-space coordinates $\ell_1$4, not a gridded spectrum (Huang et al., 2015).

In magnitude-only sparse phase retrieval, the continuous spike train

$\ell_1$5

is observed through low-pass Fourier intensities

$\ell_1$6

The real-space inversion problem is then transferred to the autocorrelation spike measure, whose support consists of pairwise differences $\ell_1$7, after which the original locations and amplitudes are disentangled (Chen et al., 2013).

Spatial imaging papers use explicit degradation operators. In scanning astronomy, the map-making model is

$\ell_1$8

or, after factorization,

$\ell_1$9

where $\ell_1$0 is a high-resolution sky map, $\ell_1$1 is a convolution operator, and $\ell_1$2 is a sparse pointing matrix (Orieux et al., 2011). In 4D Flow MRI, the low-resolution complex image obeys

$\ell_1$3

with blur $\ell_1$4, decimation $\ell_1$5, and additive white Gaussian noise (Turenne et al., 25 Sep 2025). In structured-motion SR, each low-resolution frame satisfies

$\ell_1$6

where $\ell_1$7 is a motion occupancy map, $\ell_1$8 is the box kernel induced by pixel integration, and $\ell_1$9 is the unknown high-resolution image (Litterio et al., 21 May 2025).

Ultrafast scattering introduces a two-stage model. First, the measured difference scattering $\ell_2$0 is inverted to a distorted real-space signal $\ell_2$1. Second, this signal is represented as

$\ell_2$2

where $\ell_2$3 is a dictionary of Natural Scattering Kernels and $\ell_2$4 is the real-space coefficient vector (Natan, 2021).

3. Principal algorithmic paradigms

One major paradigm is exact or stable spectral estimation followed by direct read-out in real space. In "Super-Resolution Off the Grid," random low-frequency measurements are embedded into a rank-$\ell_2$5 tensor,

$\ell_2$6

and a symmetric Jennrich/SVD procedure is applied: truncated SVD on one slice, whitening, eigendecomposition of $\ell_2$7, and recovery of the exponential feature matrix $\ell_2$8. The coordinate rows correspond to

$\ell_2$9

so the source locations are read off by

$R(x)$0

The method is explicitly off-grid and avoids the exponentially large multidimensional Hankel grid (Huang et al., 2015).

A related but distinct exact paradigm appears in the magnitude-only setting. There the data

$R(x)$1

are treated as ordinary low-pass Fourier samples of the autocorrelation measure. A Hankel matrix is formed from $R(x)$2, the matrix pencil eigenvalues recover the unlabeled set

$R(x)$3

and a second deterministic stage sorts pairwise products $R(x)$4, reconstructs the distance matrix, and solves the resulting 1D distance-geometry problem. This yields exact recovery in the noiseless case up to the unavoidable global phase, translation, and reflection ambiguities (Chen et al., 2013).

A second paradigm is regularized real-space inversion of blur, sampling, and motion. In the SPIRE/Herschel map-making method, the estimate is defined by

$R(x)$5

with $R(x)$6 derived from first-derivative energy and solved by conjugate gradient on the normal equations

$R(x)$7

The structured-motion method uses the same operator-first viewpoint, but replaces quadratic smoothness by either $R(x)$8 or TV: $R(x)$9 or

$y_j = SHx_j + n_j$0

In both cases, super-resolution is treated as deconvolution plus inversion of coded sampling geometry [(Orieux et al., 2011); (Litterio et al., 21 May 2025)].

A third paradigm replaces explicit sparsity in the image domain by priors in a transformed or dictionary domain. The scattering-statistics method alternates between enforcing measurement consistency and enforcing multiscale statistical consistency in scattering space through

$y_j = SHx_j + n_j$1

while the super-resolution projection itself is

$y_j = SHx_j + n_j$2

Ultrafast scattering instead constructs an NSK dictionary and solves

$y_j = SHx_j + n_j$3

with $y_j = SHx_j + n_j$4 or $y_j = SHx_j + n_j$5. The source-superposition method SUPPOSe represents the object as

$y_j = SHx_j + n_j$6

and minimizes the real-space residual

$y_j = SHx_j + n_j$7

over the source coordinates using a Genetic Algorithm (Dokmanić et al., 2016, Natan, 2021, Martínez et al., 2018).

4. Real-space parameterizations and priors

The real-space unknown can be continuous, discrete, sparse, smooth, or statistically constrained. The choice of parameterization is therefore not incidental; it determines what kind of super-resolution is feasible.

Off-grid point-source methods parameterize the unknown as Dirac masses and rely on geometric separation. In (Huang et al., 2015), the crucial assumption is a positive minimum Euclidean separation $y_j = SHx_j + n_j$8, which governs the required Fourier cutoff but not the number of measurements or runtime. In (Chen et al., 2013), exact recovery requires distinct amplitude magnitudes $y_j = SHx_j + n_j$9, noncollision of differences $y = \Gamma x + b,$0, and noiseless measurements.

Quadratic inverse solvers encode smoothness. The SPIRE/Herschel method penalizes

$y = \Gamma x + b,$1

which is appropriate for extended, relatively smooth emission (Orieux et al., 2011). The 4D Flow MRI method uses an $y = \Gamma x + b,$2-$y = \Gamma x + b,$3 prior centered at an interpolated estimate,

$y = \Gamma x + b,$4

so its prior is proximity to an upsampled initial guess rather than sparsity, total variation, or fluid-dynamics regularization (Turenne et al., 25 Sep 2025).

Convex structured-motion SR chooses between sparsity in the image itself and sparsity in its gradient,

$y = \Gamma x + b,$5

depending on whether the scene is sparse or piecewise smooth (Litterio et al., 21 May 2025). The scattering-statistics method uses invariant multiscale statistics exposed by the scattering transform rather than an explicit norm penalty on $y = \Gamma x + b,$6 (Dokmanić et al., 2016). Ultrafast scattering uses sparsity in pair-distance space, motivated by the expectation that transient structure changes can be represented by a relatively small number of features in $y = \Gamma x + b,$7 (Natan, 2021). SUPPOSe imposes positivity and equal-intensity virtual sources, so intensity is represented by local density of source positions rather than by free amplitudes (Martínez et al., 2018).

A different but related prior design appears in real-scene SR from raw images. There the paper argues that raw data are preferable because raw is linear in scene radiance, blur and sensor noise are more naturally modeled in raw space, and demosaicing and super-resolution are coupled sampling-resolution problems. The learned system separates recovery of a high-resolution linear image $y = \Gamma x + b,$8 from learned color correction guided by a low-resolution processed RGB image $y = \Gamma x + b,$9, which is an inversion-oriented factorization of the camera pipeline rather than a purely end-to-end RGB SR model (Xu et al., 2021).

5. Resolution mechanisms and guarantees

The phrase “super-resolution” does not have a single quantitative meaning across these papers. In each case it is tied to a different limit: cutoff frequency, diffraction-limited inversion, box-filter noninvertibility, focal-plane sampling, or PSF width.

For off-grid Fourier inversion, (Huang et al., 2015) defines super-resolution as exact or stable real-space recovery from Fourier samples only up to frequency scale $\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$0 up to logarithmic factors. Its summary table states the cutoff as

$\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$1

and the algorithm uses a number of measurements and runtime polynomial in $\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$2 and $\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$3, with no dependence on $\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$4. In noise, the stated location error is permutation-invariant max Euclidean error, linear in $\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$5 and polynomial in $\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$6 and $\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$7 (Huang et al., 2015).

For magnitude-only phase retrieval, (Chen et al., 2013) states exact noiseless recovery of an $\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$8-sparse signal from

$\mathcal A \bydef \{u : \|\Gamma u - y\| \le \epsilon\}.$9

low-pass magnitude measurements. The sampling count matches the worst-case number $x$0 of distinct autocorrelation spikes.

In scanning astronomy, the quantitative claim is bandwidth recovery rather than exact support recovery. The proposed inversion restores spatial frequencies over a bandwidth about four times that possible with coaddition, and in the PMW simulations the recovered power spectral density follows the truth up to roughly $x$1, whereas the naive per-integration Shannon limit is about $x$2 (Orieux et al., 2011).

The structured-motion formulation emphasizes invertibility by coding. Convolution with a box is generally non-invertible, but the paper states that sparse priors and known motion can still permit perfect reconstructions of sparse signals using convex optimization. It demonstrates factors as large as $x$3 in the interlaced grid case and near-perfect recovery of a $x$4 sparse target from a single blurred $x$5 image in simulation when the motion is suitably pseudo-random (Litterio et al., 21 May 2025).

In ultrafast scattering, the nominal diffraction-limited resolution is

$x$6

but the practical super-resolution limit is determined by SNR and minimum separation. In the 3-atom noisy simulation with $x$7, the nominal limit is $x$8, whereas the inferred robust minimum separation is

$x$9

and the average recovery error for a well-separated distance remains $\Gamma$0 across the stated SNR range. For the CHD experiment with $\Gamma$1, the paper reports resolution below $\Gamma$2 against a nominal diffraction limit of about $\Gamma$3 (Natan, 2021).

SUPPOSe reports super-resolution in terms of localization uncertainty of virtual-source clouds. The paper derives

$\Gamma$4

which yields an optimal number of virtual sources

$\Gamma$5

Experimentally it reports $\Gamma$6 resolution for the microscope and a fivefold improvement in the spectral resolution for the spectrometer (Martínez et al., 2018).

6. Boundary cases, misconceptions, and methodological tensions

A recurrent misconception is that any super-resolution method with a spatial-domain output is a real-space inversion algorithm. The cited literature draws a sharper distinction. A classical real-space inversion algorithm estimates the unknown directly through a forward operator in image or object coordinates. By that criterion, the map-making method, motion-coded convex inversion, ultrafast scattering NSK deconvolution, and SUPPOSe are direct instances [(Orieux et al., 2011); (Litterio et al., 21 May 2025); (Natan, 2021); (Martínez et al., 2018)]. By contrast, IDaS-SR explicitly states that it is not a pure pixel-space inverse solver; it is a latent-space inversion onto a diffusion manifold, with

$\Gamma$7

and anchored inversion latent

$\Gamma$8

Its relevance is therefore to inverse-problem SR rather than to literal real-space inversion (Weng et al., 27 Apr 2026).

A second misconception is that “real-space” excludes frequency-domain computation. Several papers are explicit that this is not so. The 4D Flow MRI method solves a spatial-domain inverse problem but implements the solution using the BCCB diagonalization $\Gamma$9 and the Fast Super-Resolution framework of Zhao et al. and its 3D extension by Tuador et al. (Turenne et al., 25 Sep 2025). JSFR-NL-SIM reconstructs the super-resolved image through the spatial-domain recombination

$\ell_1$00

yet depends on Fourier-domain parameter estimation, attenuation filtering, and optimization functions $\ell_1$01, so it is best understood as a hybrid real-space/frequency-space inversion framework (Zhang et al., 2023).

A third tension concerns exactness versus plausibility. Off-grid tensor methods and magnitude-only autocorrelation inversion state exact noiseless guarantees under separation and genericity assumptions [(Huang et al., 2015); (Chen et al., 2013)]. The scattering-statistics method states that it may worsen MSE even while recovering more realistic multiscale structure, and explicitly notes that convergence proofs are future work (Dokmanić et al., 2016). SUPPOSe gives an upper bound on localization uncertainty but does not guarantee the global minimum because the optimization is GA-based (Martínez et al., 2018). Real-scene raw/RGB fusion improves generalization by modeling the camera pipeline in raw space, but it remains a learned inverse operator rather than an explicit solver with theorem-level recovery guarantees (Xu et al., 2021).

Across these formulations, the stable core of the topic is therefore methodological rather than taxonomic. A super-resolution real-space inversion algorithm reconstructs a physically meaningful real-space variable under an explicit measurement model, exploits structure such as separation, sparsity, smoothness, motion coding, multiscale statistics, or positivity, and derives super-resolution from that structure. Whether the computation is exact, convex, alternating, closed-form, hybrid, or learned depends on the modality, but the decisive distinction is that the forward model governs the reconstruction, not merely the output representation.