Dual-Prior Null-space Learning (DP-NSL)

• The paper introduces DP-NSL, a method that decomposes the reconstruction into a measurement-determined component and a learned null-space part, ensuring exact data consistency.
• DP-NSL employs a pseudo-inverse upsampling operator and Measurement-Consistent Projection to prevent hallucination of structures in unobserved regions.
• Empirical results across CT and MRI datasets demonstrate that DP-NSL outperforms baselines by effectively balancing deterministic observation with adaptive geometric continuity.

Searching arXiv for the primary DP-NSL paper and key related null-space learning references. Dual-Prior Null-space Learning (DP-NSL) denotes a constrained reconstruction framework in which the solution of an ill-posed inverse problem is decomposed into a measurement-determined component and a learned null-space component, with the learning process guided by two complementary priors: a Deterministic Observation Prior that preserves the acquired data exactly, and a Geometric Continuity Prior that regularizes only the unobservable content. In the formulation introduced for arbitrary medical slice super-resolution, DP-NSL reconstructs isotropic volumes from anisotropic clinical scans by combining a deterministic range-space anchor with a null-space estimate passed through a Measurement-Consistent Projection (MCP), so that synthesized details cannot alter the originally observed slices (Song et al., 25 Jun 2026).

1. Definition and conceptual basis

DP-NSL was introduced for arbitrary medical slice super-resolution as a framework for reconstructing dense isotropic volumes from anisotropic acquisitions whose through-plane resolution is much coarser than the in-plane resolution (Song et al., 25 Jun 2026). The low-resolution input is written as

$\mathbf{V}_{LR}\in\mathbb{R}^{S\times H\times W},$

while the target or reconstructed volume is

$$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$

The forward model is

$$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$

with $\mathcal{D}$ denoting the slice downsampling or acquisition operator.

The framework is motivated by two failure modes of unconstrained arbitrary-scale super-resolution: such methods can hallucinate anatomically implausible structures in inter-slice regions, and they can modify the originally observed slices. DP-NSL addresses both by replacing unconstrained residual regression with a constrained recovery process. The term “dual-prior” refers to the use of two complementary sources of structure: one prior specifies what must remain fixed, namely the acquired measurements, and the other specifies how missing anatomy should be interpolated, namely through content-adaptive continuity (Song et al., 25 Jun 2026).

The general null-space logic underlying DP-NSL is older than the term itself. In operational-space control, constrained actions are decomposed into task-space and null-space components,

$$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$

and null-space recovery is posed as learning a projection that preserves null-space motion while annihilating task-space motion (Lin et al., 2016). This suggests that DP-NSL belongs to a broader family of methods in which identifiable learning occurs by separating measurement-constrained structure from underdetermined structure.

2. Null-range-space decomposition and exact measurement consistency

The central mathematical device in DP-NSL is a pseudo-inverse upsampling operator

$$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$

satisfying

$\mathcal{D}\mathcal{U}=\mathbf{I}.$

Under this assumption, any target-resolution volume can be decomposed as

$$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$

The paper denotes the null-space projection by

$$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$

Using a learnable Null-Space Estimator $f_\theta$, the reconstruction is defined as

$$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$0

$$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$1

This is not a free residual model. The network may predict a raw candidate $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$2, but only its null-space-projected component survives in the final output.

The Measurement-Consistent Projection (MCP) is the operator $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$3. Because

$$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$4

the learned null component is invisible to the acquisition model. The final reconstruction therefore satisfies

$$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$5

In the paper’s terminology, this realizes the Deterministic Observation Prior: the acquired slices are reproduced with zero error by construction, not merely encouraged by a loss (Song et al., 25 Jun 2026).

This exact-consistency construction places DP-NSL in direct continuity with inverse-problem null-space networks, where learned corrections are restricted to $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$6 so that data consistency is preserved: $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$7 (Schwab et al., 2018). A plausible implication is that DP-NSL can be read as a domain-specific null-space network in which the forward operator is slice acquisition and the learned correction is further regularized by a second prior specialized to anatomy.

3. The two priors: deterministic observation and geometric continuity

The first prior in DP-NSL is the Deterministic Observation Prior. Intuitively, it states that anything already measured by the scanner must be copied exactly into the reconstruction. Formally, it is realized through the decomposition above and the condition $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$8, which makes MCP a hard projection operator. The paper emphasizes that this is an exact guarantee in implementation for the sampling model used, and verifies it empirically through slice-wise MAE curves that are exactly zero at observed slice indices $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$9 (Song et al., 25 Jun 2026).

The second prior is the Geometric Continuity Prior, implemented through the Mixture-of-Splines (MoS) module. Its motivation is that anatomy exhibits heterogeneous smoothness: homogeneous soft tissue, organ boundaries, vessels, and bone-soft tissue interfaces do not obey a single global continuity class. A single shared implicit decoder is therefore treated as too uniform an assumption. MoS instead uses B-spline experts of different orders to model local continuity adaptively.

Given low-resolution features

$$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$0

and a target coordinate $$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$1, MoS defines

$$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$2

For a spline expert of order $$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$3,

$$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$4

with

$$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$5

Here $$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$6 is the 1D B-spline basis of order $$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$7, $$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$8 are scaling coefficients, $$\mathbf{V}_{LR} = \mathcal{D}(\mathbf{V}_{HR}) + \boldsymbol{\epsilon},$$9 are adaptive knots, $\mathcal{D}$0 are continuous dilations, and

$\mathcal{D}$1

uses a learned sub-voxel offset $\mathcal{D}$2.

The paper uses $\mathcal{D}$3 in the final model. Since a B-spline of order $\mathcal{D}$4 guarantees $\mathcal{D}$5 continuity, higher-order experts impose smoother transitions while lower-order experts better accommodate sharper changes. The experts are mixed dynamically through

$\mathcal{D}$6

and

$\mathcal{D}$7

The paper describes this as a “content-aware level of continuity” (Song et al., 25 Jun 2026).

This dual-prior pattern has analogues in later null-space methods outside medical super-resolution. In imaging inverse problems, GSNR combines a global image prior $\mathcal{D}$8 with a dedicated null-space prior

$\mathcal{D}$9

explicitly separating whole-image regularization from invisible-component regularization (Gualdrón-Hurtado et al., 23 Feb 2026). This suggests that DP-NSL is part of a broader design principle: one prior governs the observable or full-signal manifold, and another governs the invisible or ambiguous component.

4. Architecture and implementation

DP-NSL is implemented as a fully 3D pipeline. A low-resolution anisotropic volume $$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$0 is processed by a 3D EDSR encoder to produce $$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$1. In parallel, the pseudo-inverse upsampler $$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$2 forms the deterministic anchor

$$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$3

The MoS module then queries the encoded features at arbitrary target coordinates to produce a dense high-resolution feature field $$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$4. A decoder converts $$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$5 into a raw null-space estimate $$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$6, MCP projects that estimate into the null space, and the final output is

$$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$7

The decoder is the Local Spatial Consistency Decoder (LSCD). Its role is to mitigate the lack of local coupling in coordinate-wise implicit decoding. LSCD uses an Inception-style split-transform-merge design with channel groups

$$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$8

and produces

$$\mathbf{u}=\mathbf{u}^{ts}+\mathbf{u}^{ns} =\mathbf{A}^\dagger\mathbf{b}+\mathbf{N}\boldsymbol{\pi}, \qquad \mathbf{N}=\mathbf{I}-\mathbf{A}^\dagger\mathbf{A},$$9

where $$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$0 denotes standard 3D convolution and $$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$1 depthwise 3D convolution. According to the paper, LSCD improves over both a pixel-wise decoder and a slice-wise decoder while using fewer FLOPs and parameters than the tested MLP-based option (Song et al., 25 Jun 2026).

Training uses only an $$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$2 reconstruction loss,

$$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$3

No additional measurement-consistency loss is required because consistency is imposed by MCP, not by optimization. Optimization uses Adam with $$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$4, $$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$5, initial learning rate $$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$6, learning-rate halving every 200 epochs, 1000 total epochs, and batch size 8. Training samples are four consecutive low-resolution slices center-cropped to $$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$7 (Song et al., 25 Jun 2026).

A recurring theme across null-space learning is that architectural enforcement can replace soft penalties. In refusal steering for LLMs, AlphaSteer enforces benign-data preservation by parameterizing $$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$8 so that $$\mathcal{U}: \mathbb{R}^{S\times H\times W}\to\mathbb{R}^{\hat S\times H\times W}$$9 holds structurally rather than as a loss (Sheng et al., 8 Jun 2025). A similar structural logic is at work in DP-NSL: the learned branch cannot corrupt measured slices because the projection removes all measurable components.

5. Empirical performance and ablation structure

DP-NSL is evaluated on three CT datasets and one MRI dataset: Colon, Liver, Hepatic Vessels, and IXI. Low-resolution inputs are synthesized by downsampling only along the slice axis. The method is trained on in-scale factors $\mathcal{D}\mathcal{U}=\mathbf{I}.$0 and tested both in-scale and out-of-scale at $\mathcal{D}\mathcal{U}=\mathbf{I}.$1. Baselines include EDSR3D, MetaSR, LTE, HIIF, ArSSR, SAINR, CycleINR, and DC$\mathcal{D}\mathcal{U}=\mathbf{I}.$2SR. Evaluation uses PSNR and SSIM (Song et al., 25 Jun 2026).

On in-scale evaluation, DP-NSL is reported as best on all four datasets and all tested factors. Representative values include:

• Colon: $\mathcal{D}\mathcal{U}=\mathbf{I}.$3 at $\mathcal{D}\mathcal{U}=\mathbf{I}.$4, $\mathcal{D}\mathcal{U}=\mathbf{I}.$5 at $\mathcal{D}\mathcal{U}=\mathbf{I}.$6, $\mathcal{D}\mathcal{U}=\mathbf{I}.$7 at $\mathcal{D}\mathcal{U}=\mathbf{I}.$8
• Liver: $\mathcal{D}\mathcal{U}=\mathbf{I}.$9 at $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$0, $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$1 at $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$2, $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$3 at $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$4
• Hepatic Vessels: $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$5 at $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$6, $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$7 at $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$8, $$\mathbf{V}_{SR} = \mathcal{U}\mathcal{D}(\mathbf{V}_{SR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR} = \mathcal{U}(\mathbf{V}_{LR}) + (\mathbf{I}-\mathcal{U}\mathcal{D})\mathbf{V}_{SR}. \tag{1}$$9 at $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$0
• IXI: $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$1 at $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$2, $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$3 at $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$4, $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$5 at $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$6

The paper highlights gains up to 1.07 dB over the second-best method on Liver at $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$7, 0.71 dB on Colon at $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$8, 0.65 dB on Colon at $$\Pi_{\mathcal N}\triangleq \mathbf{I}-\mathcal{U}\mathcal{D}.$$9, and 0.45 dB on IXI at $f_\theta$0. Out-of-scale generalization remains best on all datasets; for example, on IXI it reaches $f_\theta$1 at $f_\theta$2, $f_\theta$3 at $f_\theta$4, and $f_\theta$5 at $f_\theta$6. On the real paired RPLHR-CT dataset at $f_\theta$7, DP-NSL achieves 34.91 dB PSNR and 0.8925 SSIM (Song et al., 25 Jun 2026).

The component ablation is particularly important for understanding the “dual-prior” claim. On Colon at $f_\theta$8, the plain baseline without MCP, MoS, or LSCD scores 41.82 dB. Adding MCP alone increases this to 42.13, adding MoS alone gives 42.45, adding LSCD alone gives 42.06, combining MCP+MoS yields 42.52, and the full MCP+MoS+LSCD gives 42.54. The paper interprets this as evidence that MCP and MoS are complementary: MoS improves synthesized details, while MCP prevents those details from corrupting measured structure (Song et al., 25 Jun 2026).

The measurement-consistency analysis is unusually direct. Slice-wise MAE drops to exactly zero at observed slice positions $f_\theta$9, matching

$$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$00

The paper contrasts this with unconstrained direct-estimation and residual-estimation baselines, which retain nonzero error on known slices (Song et al., 25 Jun 2026).

6. Broader research context, extensions, and limitations

DP-NSL names and formalizes a design pattern that appears, under different terminology, in several other areas. The table summarizes representative examples already present in the literature.

Paper	Domain	Null-space mechanism
(Lin et al., 2016)	Operational-space control	Preserve $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$01, annihilate $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$02
(Schwab et al., 2018)	Inverse problems	Learn only $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$03 corrections
(Gualdrón-Hurtado et al., 23 Feb 2026)	Imaging inverse problems	Add explicit null-only prior to full-image prior
(Sheng et al., 8 Jun 2025)	LLM activation steering	Force benign activations into steering null space
(Ju et al., 9 Jun 2026)	LLM unlearning	Project LoRA updates into retain-orthogonal subspaces
(Jacome et al., 2 Oct 2025)	Imaging inverse problems	Predict low-dimensional null-space coordinates

These works do not all call themselves DP-NSL, but they repeatedly instantiate the same structural idea: a projector or constrained parameterization protects one subspace, while learning acts only in a complementary one. In that sense, the 2026 DP-NSL paper can be read as a specific medical-imaging realization of a more general null-space-learning paradigm (Song et al., 25 Jun 2026).

Several objective clarifications follow from this context. First, DP-NSL is not simply any method that uses two losses; in the named formulation, the “dual priors” are a Deterministic Observation Prior and a Geometric Continuity Prior, tied to a hard decomposition and an exact projection. Second, MCP’s exact guarantee depends on the acquisition model and on having $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$04. The paper notes that future work is needed for more realistic degradation models involving slice-profile blur, scanner-specific interpolation, partial-volume effects, motion, and noise. Third, the framework is more expensive than lightweight INR baselines, though still cheaper than some heavier attention-based alternatives. Finally, its training remains supervised and assumes access to high-resolution targets (Song et al., 25 Jun 2026).

A broader inference suggested by related work is that DP-NSL admits multiple realizations beyond slice super-resolution. In imaging, GSNR shows that a null-specific prior can be made spectral and graph-based rather than spline-based (Gualdrón-Hurtado et al., 23 Feb 2026). In inverse problems more generally, NPN shows that one can regularize a reconstruction through low-dimensional null-space coordinates $$\mathbf{V}_{HR},\mathbf{V}_{SR}\in\mathbb{R}^{\hat S\times H\times W}.$$05 predicted from measurements (Jacome et al., 2 Oct 2025). In LLMs, AlphaSteer and NSRU show that null-space constraints can protect benign activations or retain-subspace behavior while allowing targeted steering or unlearning (Sheng et al., 8 Jun 2025, Ju et al., 9 Jun 2026). These parallels suggest that DP-NSL is best understood not only as a single medical-imaging architecture, but also as a general strategy for combining an observation-preserving prior with a complementary prior over underdetermined degrees of freedom.