Image-Space Preconditioning (ISP)

• Image-Space Preconditioning (ISP) is a technique that applies preconditioning directly in the image domain to improve iterative reconstruction methods.
• It preserves key structural features such as boundary corrections and uses FFT-compatible approximations for efficient deblurring and MRI reconstructions.
• By altering the inner product in the image space, ISP reduces condition numbers and accelerates convergence in iterative solvers.

Searching arXiv for the specified papers and closely related ISP/MRI preconditioning work. arxiv_search query: (2002.01429) OR (Milani et al., 7 Jul 2025) OR "image-space preconditioning" MRI preconditioning iterative reconstruction

arxiv_search query: "Image-Space Preconditioning" MRI reconstruction variational formulation

Image-Space Preconditioning (ISP) denotes a class of preconditioning strategies in which the preconditioner acts on the image domain, or on an image-domain coefficient representation, rather than solely on the data domain. In the cited literature, ISP appears in two closely related but distinct technical roles. For frame-based image deblurring, it is realized through a structure-preserving approximation of the inverse blurring operator that retains the boundary-condition correction of the imaging matrix and remains FFT-compatible (2002.01429). For magnetic resonance imaging (MRI), it is formulated as a non-conventional inner product on image space, so that iterative reconstruction algorithms inherit the preconditioner through the geometry of the variational problem itself (Milani et al., 7 Jul 2025). Taken together, these formulations present ISP as a mechanism for improving conditioning, accelerating iterative schemes, and preserving fidelity to the underlying forward model.

1. Definition in inverse problems and finite-dimensional geometry

In frame-based deblurring, the starting point is the continuous convolution model

$$g(x,y)\;=\;\bigl(\kappa*f\bigr)(x,y)\;+\;\eta(x,y),\quad (x,y)\in\Omega,$$

which, after uniform-grid discretization with prescribed boundary conditions, yields a blurring matrix

$$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$

where $\mathcal T_m(\kappa)$ is Block–Toeplitz-with-Toeplitz-Blocks (BTTB) and $\mathcal B_m^{\mathcal X}(\kappa)$ is a small-rank correction determined by $$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$. If $W^*$ is a tight-frame synthesis operator with $W^*W=I$, then

$$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$

and the observation equation becomes $g=A\,c$ with $c=Wf$ (2002.01429).

In MRI, the same general idea is expressed in the language of finite-dimensional inner-product spaces. Let $$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$0 be image space, $$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$1 data space, and $$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$2 the forward operator with $$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$3. The classical regularized least-squares problem is

$$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$4

ISP is introduced by replacing the Euclidean metric on $$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$5 with

$$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$6

while data-space preconditioning (DSP) analogously replaces the Euclidean metric on $$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$7 by $$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$8 (Milani et al., 7 Jul 2025).

This suggests a unifying interpretation: ISP modifies the effective geometry of the inverse problem in image space so that the iterative method follows a better-conditioned metric without altering the underlying reconstruction target.

2. Structure-preserving ISP for frame-based image deblurring

The deblurring construction emphasizes that an effective preconditioner must preserve the same boundary-condition structure as the coefficient matrix. The goal is to build $$K \;=\;\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa),$$9 such that it preserves the boundary-condition structure of $\mathcal T_m(\kappa)$0, is Block-Circulant with Circulant Blocks (BCCB) so that forward and inverse actions can be carried out by 2D FFTs, and approximates $\mathcal T_m(\kappa)$1 or $\mathcal T_m(\kappa)$2 on the “signal-space” subspace (2002.01429).

The construction begins with the standard BCCB approximation

$\mathcal T_m(\kappa)$3

whose eigenvalues are obtained by 2D FFT: $\mathcal T_m(\kappa)$4 Given a regularization parameter $\mathcal T_m(\kappa)$5 and eigenvalues $\mathcal T_m(\kappa)$6 of a chosen regularizing operator $\mathcal T_m(\kappa)$7, one forms the Tikhonov-filtered spectrum

$\mathcal T_m(\kappa)$8

An inverse 2D FFT then yields a generating mask $\mathcal T_m(\kappa)$9, from which the preconditioner is assembled with the same boundary conditions: $\mathcal B_m^{\mathcal X}(\kappa)$0

By construction, $\mathcal B_m^{\mathcal X}(\kappa)$1 shares the structural correction $\mathcal B_m^{\mathcal X}(\kappa)$2 of $\mathcal B_m^{\mathcal X}(\kappa)$3, while remaining circulant in its principal block. The resulting costs for multiplication $\mathcal B_m^{\mathcal X}(\kappa)$4 and for solving $\mathcal B_m^{\mathcal X}(\kappa)$5 are $\mathcal B_m^{\mathcal X}(\kappa)$6. The literature summary further states that a purely circulant preconditioner would ignore the low-rank boundary correction and leave a handful of eigenvalues far from the bulk, slowing convergence, whereas matching the BTTB plus low-rank structure yields tight eigenvalue clustering around $\mathcal B_m^{\mathcal X}(\kappa)$7 for $\mathcal B_m^{\mathcal X}(\kappa)$8 or $\mathcal B_m^{\mathcal X}(\kappa)$9 (2002.01429).

A common misconception is therefore that circulant approximation alone is sufficient. In the formulation under discussion, the small-rank boundary correction is not incidental; it is part of the mechanism by which eigenvalue clustering is obtained.

3. Nonstationary MLBA and wavelet-domain thresholding

In the same deblurring framework, ISP is coupled to a nonstationary preconditioned variant of the modified linearized Bregman algorithm (MLBA). The iteration is carried out in coefficient space. If $$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$0 is the current frame-domain approximation, then

$$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$1

where $$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$2 denotes componentwise soft-thresholding at level $$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$3. The image-space residual is

$$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$4

The iteration chooses $$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$5, sets $$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$6, and defines the discrepancy ratio

$$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$7

While $$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$8, one defines

$$\mathcal X\in\{\text{Periodic},\text{Reflective},\text{Anti-Reflective}\}$$9

and finds $W^*$0 such that

$W^*$1

The preconditioning step is

$W^*$2

followed by

$W^*$3

The cited results state that $W^*$4 decreases monotonically, the iteration stops in finitely many steps $W^*$5, and as $W^*$6 the output converges to the minimum-$W^*$7-norm solution of $W^*$8. The role assigned to soft-thresholding is twofold: it exploits the sparsity of natural images in the wavelet domain, and it stabilizes inversion against noise while promoting sharp edges in the restoration (2002.01429).

The same source characterizes the combination of structure-preserving approximation, FFT-based implementation, and wavelet thresholding as ISP. In that characterization, eigenvalue clustering, fast convergence, low per-step cost, and high restoration quality measured in PSNR/SSIM arise jointly from these components.

4. ISP as an inner product in variational MRI reconstruction

In the MRI formulation, ISP is embedded directly into the variational problem by equipping image space with a nonstandard Hermitian inner product

$W^*$9

and likewise data space with

$W^*W=I$0

The corresponding adjoint of the forward operator is

$W^*W=I$1

and for a real-valued functional $W^*W=I$2, the induced gradient satisfies

$W^*W=I$3

with $W^*W=I$4 the Wirtinger derivative (Milani et al., 7 Jul 2025).

Within this framework, DSP is represented by the choice

$W^*W=I$5

and ISP by

$W^*W=I$6

A common DSP choice is a diagonal density-compensation matrix $W^*W=I$7 in k-space. A classic ISP example from Pruessmann et al. is $W^*W=I$8, where $W^*W=I$9 encodes coil-sensitivity magnitudes.

The regularized least-squares problem is rewritten as

$$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$0

The first-order optimality condition becomes

$$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$1

which reduces, for $$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$2, to

$$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$3

This formulation makes ISP algorithm-independent: it is attached to the metric of $$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$4, not to a particular implementation trick.

5. Algorithmic propagation and convergence consequences

The MRI formulation explicitly states that because ISP lives in the choice of $$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$5, any iterative solver for

$$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$6

automatically inherits the preconditioner $$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$7, without code-level changes to line-searches, momentum, or proximal operators (Milani et al., 7 Jul 2025). In the simplest smooth setting, if

$$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$8

then a preconditioned gradient step is

$$A \;=\; K\,W^* \;=\; \bigl(\mathcal T_m(\kappa)\;+\;\mathcal B_m^{\mathcal X}(\kappa)\bigr)\,W^*,$$9

For conjugate-gradient on the normal equations, the same source states that one replaces inner products by $g=A\,c$0 and uses $g=A\,c$1 rather than $g=A\,c$2.

The principal theoretical claim is spectral clustering: choosing $g=A\,c$3 makes the spectrum of $g=A\,c$4 tightly clustered, drives the condition number toward unity, and accelerates conjugate gradient or gradient descent. More specifically, the local convergence rate of gradient descent is governed by $g=A\,c$5, and a good preconditioner can reduce $g=A\,c$6 from $g=A\,c$7–$g=A\,c$8 down to $g=A\,c$9. The same formulation also attributes stability to the metric interpretation, stating that the preconditioned step remains a genuine gradient step of a convex function in the geometry of $c=Wf$0, thereby preserving monotonicity and convergence guarantees (Milani et al., 7 Jul 2025).

A plausible implication is that the metric-based viewpoint supplies a common language for classical iterative methods and for more recent iterative schemes that mix analytical forward models with learned components, because the preconditioner is introduced at the level of the variational geometry rather than as a solver-specific patch.

6. Applications, relations to DSP, and scope of the term

The two cited uses of ISP arise in different application areas. In image deblurring with space-invariant blur, the emphasis is on BTTB structure, low-rank boundary corrections, BCCB principal blocks, 2D FFT implementation, and wavelet-domain sparsity (2002.01429). In MRI, the emphasis is on the forward model $c=Wf$1, the variational formulation of regularized least squares, and the representation of preconditioning through image-space and data-space metrics (Milani et al., 7 Jul 2025).

The relationship between ISP and DSP is especially explicit in the MRI treatment. DSP acts in data space through $c=Wf$2, often as a density-compensation matrix in k-space, while ISP acts in image space through $c=Wf$3. The cited summary presents them as symmetric choices of non-conventional inner products on $c=Wf$4 and $c=Wf$5, respectively. This symmetry is absent from formulations that treat only data-space weighting, and the reformulation is presented as restoring the natural symmetry of the variational problem.

The practical examples reported for MRI include 2D radial CINE, non-Cartesian CS, and temporal regularization. The reported outcomes are that ISP roughly halved the iteration count to a given residual for Iterative-SENSE without a regularizer, gave approximately $c=Wf$6 speed-up in convergence for $c=Wf$7-spatial CS, produced small but measurable acceleration for $c=Wf$8-temporal CS, and yielded identical images to machine precision when implemented either via the nonstandard inner product or via the equivalent change of variables (Milani et al., 7 Jul 2025).

Across the cited literature, “Image-Space Preconditioning” therefore names a family of techniques rather than a single algorithm. In deblurring, it refers to a structure-preserving preconditioner integrated with nonstationary MLBA and wavelet thresholding. In MRI, it refers to an image-space metric that can be propagated systematically through iterative reconstruction algorithms. This suggests that the common denominator of ISP is not a particular solver architecture, but the relocation of preconditioning into image-domain structure or image-domain geometry.