Dual-Space Multi-Scale Model

• The model defines and couples two complementary spaces (e.g., physical, spectral, or algorithmic) while resolving phenomena at multiple scales.
• It employs strategies such as oversampling, image pyramids, and hierarchical scanning to align dual domains for improved simulation and reconstruction.
• Empirical studies show that proper coupling and scale matching lead to significant gains in accuracy and efficiency across diverse applications.

A dual-space multi-scale model is a modeling pattern in which two distinct but coupled representational spaces, domains, continua, or processes are learned or solved jointly, while the target phenomenon is resolved across multiple scales. In the cited literature, the paired spaces include matrix and fracture continua in subsurface flow, spatial and frequency domains in image deraining, image space and k-space in MRI reconstruction, transform space and classifier space in neuroimaging, low- and high-frequency subspaces in elliptic PDE solvers, and fast and slow reasoning systems in human–AI collaboration (Cheung et al., 2018, Zou et al., 15 Mar 2025, Meng et al., 14 Jan 2025, Baek et al., 3 Mar 2025, Li et al., 2021, Li et al., 18 Jul 2025). The multiscale component is realized through coarse regions and oversampling, image pyramids, hierarchical feature branches, learnable spectral scales, multi-scale scanning paths, or block time partitions (Zhang et al., 2020, Gao et al., 2024, Kumar et al., 2020).

1. Semantic scope and recurring design pattern

In the available literature, “dual-space” does not denote a single standardized object. It denotes a family of constructions in which two complementary spaces are made explicit, and “multi-scale” denotes a second axis of organization that controls locality, hierarchy, or resolution. The duality may be physical, spectral, architectural, or algorithmic.

Paired spaces or processes	Multi-scale mechanism	Representative paper
Matrix and fracture continua	Coarse regions, local spectral modes, oversampling	(Cheung et al., 2018)
Spatial and frequency domains	External image pyramid and internal multi-scale modules	(Zou et al., 15 Mar 2025)
Image space and k-space	Hierarchical HR/LR scanning paths	(Meng et al., 14 Jan 2025)
Transform space and classifier space	Learnable scales in covariance eigenspace	(Baek et al., 3 Mar 2025)
Low- and high-frequency subspaces	Separate DNN submodules	(Li et al., 2021)
Fast and slow reasoning systems	Knowledge, style, and intention hierarchy	(Li et al., 18 Jul 2025)

This range of meanings matters because many architectures that are superficially similar implement different couplings. In subsurface simulation, the duality is in the governing physics; in image restoration, it is often a domain decomposition between spatial and spectral representations; in representation learning, it may be a separation between the space in which the signal is transformed and the space in which the downstream predictor operates (Cheung et al., 2018, Zou et al., 15 Mar 2025, Baek et al., 3 Mar 2025).

2. Mathematical archetypes

A canonical physical formulation appears in the dual continuum model for subsurface flow, where the two spaces are pressures in matrix and fracture continua. The governing system is

$\begin{split} c_1 \frac{\partial p_1}{\partial t} - \nabla \cdot (\kappa_1 \nabla p_1) + \rho \sigma(p_1 - p_2) &= \rho f_1,\ c_2 \frac{\partial p_2}{\partial t} - \nabla \cdot (\kappa_2 \nabla p_2) - \rho \sigma(p_1 - p_2) &= \rho f_2. \end{split}$

The associated multiscale construction in CEM-GMsFEM uses local spectral problems

$$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$

followed by constrained energy minimization on oversampled regions to obtain localized vector-valued basis functions for both continua. The resulting multiscale space is

$$V_{ms} = \text{span}\{\psi_{k,ms}^{(j)} : 1 \leq k \leq L_j, 1 \leq j \leq N\},$$

and the method is analyzed through exponential decay and coarse-mesh-dependent error estimates (Cheung et al., 2018). A related non-local multicontinuum upscaling formulation constructs basis functions by minimizing localized energy subject to mean-value constraints that are one in the targeted continuum and zero in the others, thereby encoding non-local transfer and effective properties in the upscaled space (Zhang et al., 2020).

A second archetype is explicit subspace decomposition. In SD$^2$NN, the solution of a multi-scale PDE is split as

$u(\mathbf{x}) = u_c(\mathbf{x}) + u_f(\mathbf{x}),$

with a low-frequency module for the smooth component and one or more high-frequency MscaleDNN submodules for the oscillatory component. The learned output is

$$y(\mathbf{x}; \boldsymbol{\theta}_1, \boldsymbol{\theta}_2) = y_1(\mathbf{x}; \boldsymbol{\theta}_1) + \alpha\, y_2(\mathbf{x}; \boldsymbol{\theta}_2),$$

and overlap between the two subspaces is penalized by an orthogonality term in the loss. Here the “dual space” is a low-/high-frequency decomposition rather than a pair of physical domains (Li et al., 2021).

A third archetype is a transform-space/classifier-space split. In covariance-based neuroimaging classification, the sample covariance is decomposed as $\Sigma = U \Lambda U^T$, and a covariance-based multi-scale transform is defined by

$C_X(s) = U g(s\Lambda) U^T X.$

For learnable scales $S = [s_1,\ldots,s_J]$, the embeddings are concatenated as

$E = [C_X(s_1), \ldots, C_X(s_J)].$

The dual-space mechanism is the joint optimization of transform-space scales and classifier weights, with gradients propagated through the scale parameters (Baek et al., 3 Mar 2025).

A fourth mathematical reading of duality appears in control. DDIP reformulates mixed-integer MPC with lifted stage coupling and dual variables, then uses forward and backward sweeps to approximate cost-to-go functions with cuts. In this setting, “dual-space” refers to dual information from LP relaxations used to construct Benders cuts, while “multi-scale” refers to block time partitions that group fine time steps into coarser stages (Kumar et al., 2020).

3. Architectural instantiations in learned systems

In computer vision, a common instantiation is a dual-domain network with explicit domain transforms. The DMSR deraining architecture uses multiple Dual-Domain Scale-Aware Modules, each containing a Multi-Scale Progressive Spatial Refinement Module and a Frequency Domain Scale Mixer. It processes image pyramids $$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$0 in a coarse-to-fine fashion, while the frequency branch applies FFT, modulates real and imaginary parts, and returns via IFFT. The stated objective is to exploit joint multi-scale representations from both external and internal domains in parallel while leveraging both spatial and frequency domains (Zou et al., 15 Mar 2025).

In medical segmentation, ISD-MAE uses a dual-branch masked autoencoding scheme. One branch applies intensity masking, the other spatial masking, both feeding a shared encoder–decoder structure with a SegFormer encoder and a symmetric decoder. Reconstruction is supervised with SSIM, while the two branch embeddings are aligned by a contrastive loss,

$$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$1

The duality is therefore between intensity-differentiated and spatial-differentiated views, and the multiscale component is realized through five encoder stages and hierarchical decoder fusion (Ding et al., 2024).

State-space models instantiate the pattern in a different way. MSFMamba separates spatial and spectral feature extraction in remote sensing through MSpa-Mamba and Spe-Mamba, and then bridges the heterogeneous gap between HSI and LiDAR/SAR through Fus-Mamba, which extends Mamba to accommodate dual inputs. Spatial modeling uses multi-scale scanning routes, spectral modeling scans along the channel dimension, and fusion is achieved by cross-modal state-parameter generation in the Fusion SSM (Gao et al., 2024). In MRI reconstruction, a dual-domain hierarchical Mamba processes image space and k-space in parallel. The k-space branch uses circular scanning to preserve spectrum structure, the image branch uses hierarchical scanning in image space, and the outputs are fused after inverse FFT as

$$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$2

The hierarchy is implemented through one high-resolution path and several low-resolution paths in each domain (Meng et al., 14 Jan 2025).

In wireless channel prediction, ChannelKAN uses a Dual-Domain Expansion module to derive frequency-domain and delay-domain CSI representations, a Multi-Scale Frequency Information Enhancement module based on temporal FFT and TopK masking, a hybrid CNN-KAN feature extractor, and a Dual-Domain Fusion module. The duality is frequency versus delay domain; the multi-scale mechanism keeps dominant spectral components at multiple scales to strengthen key features and suppress noise (Jiang et al., 11 May 2026).

A non-perceptual but structurally related example is DPMT for human–AI collaboration. It separates a fast system for macro-action selection from a slow system that performs multi-scale Theory of Mind inference over domain knowledge, cognitive style, and domain intention:

$$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$3

Here the duality is process-level rather than signal-level (Li et al., 18 Jul 2025).

4. Scale management, localization, and coupling mechanisms

The multiscale axis is not uniform across these models. In multiscale finite element methods for dual continua, scale is tied to coarse blocks, oversampled regions, and the spectrum of local operators. The smallest eigenvalues in the auxiliary spectral problem identify dominant high-contrast channels or fracture networks, and oversampling yields localized basis functions whose global-to-local discrepancy decays exponentially with the number of coarse layers. The main error estimate depends on the coarse mesh size $$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$4, the spectral quantity $$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$5, and a coefficient lower bound, while sufficiently many oversampling layers render the error independent of the fine scale $$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$6 and coefficient contrast (Cheung et al., 2018). In the non-local multicontinuum setting, oversampling similarly enables localized bases that still capture non-local transfer (Zhang et al., 2020).

In neural architectures, scale is often split into “external” and “internal” variants. DMSR uses an external image pyramid and internal multi-scale branches inside MPSRM and FDSM; the former progressively integrates features obtained from multiple pooling factors, and the latter mixes spatial kernels of sizes $$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$7, $$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$8, and $$a_Q^{(j)}(\phi_k^{(j)}, v) = \lambda_k^{(j)} s^{(j)}(\phi_k^{(j)}, v),$$9 before frequency-domain modulation (Zou et al., 15 Mar 2025). ISD-MAE uses multi-scale encoder outputs and skip-connected decoder fusion, rather than an explicit image pyramid (Ding et al., 2024). MSFMamba reduces redundancy by applying SSM scans at multiple spatial scales, and the downsampled features are later upsampled and merged (Gao et al., 2024). The dual-domain hierarchical Mamba for MRI uses one full-resolution path and several low-resolution paths in both image and k-space branches to mitigate long-range forgetting while preserving fine detail (Meng et al., 14 Jan 2025). ChannelKAN defines scale through TopK spectral filtering at several levels after temporal rFFT, then aggregates the retained components with learned weights (Jiang et al., 11 May 2026).

Control and reasoning systems realize scale differently. DDIP uses block time partitions so that a stage may contain several fine time steps, allowing forward–backward sweeps at a slower temporal resolution while preserving within-block dynamics (Kumar et al., 2020). DPMT uses a hierarchy of mental characteristics—knowledge, style, and intention—inside the slow system, while the fast system handles low-latency action selection (Li et al., 18 Jul 2025). This suggests that “scale” can refer not only to spatial resolution or spectral bandwidth, but also to temporal abstraction and semantic abstraction.

5. Reported empirical properties across domains

The empirical literature consistently evaluates whether the second space and the multiscale mechanism contribute measurable gains, usually through ablations. The reported findings cover numerical analysis, segmentation, restoration, classification, reconstruction, control, and collaboration.

System	Domain	Reported result
CEM-GMsFEM	Dual continuum subsurface flow	Experiments used contrast up to $$V_{ms} = \text{span}\{\psi_{k,ms}^{(j)} : 1 \leq k \leq L_j, 1 \leq j \leq N\},$$0; with $$V_{ms} = \text{span}\{\psi_{k,ms}^{(j)} : 1 \leq k \leq L_j, 1 \leq j \leq N\},$$1 basis functions and sufficient oversampling, errors are below 1% for $$V_{ms} = \text{span}\{\psi_{k,ms}^{(j)} : 1 \leq k \leq L_j, 1 \leq j \leq N\},$$2 (Cheung et al., 2018)
ISD-MAE	Chest CT segmentation	Dice reaches 90.10% on COVID19 LESION; Table 2 reports 90.10±0.54 on COVID19_LESION and 93.37±1.55 on Lung_nodule_seg (Ding et al., 2024)
DMSR	Single image deraining	State-of-the-art performance across six benchmark datasets; on Rain100H the paper reports +0.7 dB vs. FSNet (Zou et al., 15 Mar 2025)
MSFMamba	Multi-source remote sensing classification	Reported OA values include 77.11% on Berlin, 91.30% on Augsburg, and 92.38% on Houston2018 (Gao et al., 2024)
Dual-domain hierarchical Mamba	MRI reconstruction	Reported complexity examples include 117G FLOPs and 1.87M parameters, versus 342G FLOPs for ReconFormer and 190G FLOPs with 14.3M parameters for MambaIR (Meng et al., 14 Jan 2025)
Covariance-based dual-space model	Alzheimer classification	For CN vs. EMCI on cortical thickness, accuracy is 0.858 versus 0.765 for the best MLP, and target accuracy 0.6 is reached in 50–110 epochs versus 430 epochs (Baek et al., 3 Mar 2025)
ChannelKAN	Massive MIMO-OFDM CSI prediction	Full model ablation reports NMSE 0.0265, SE 6.414, BER 0.00771; removing CNN-KAN raises NMSE to 0.0628 (Jiang et al., 11 May 2026)
DDIP	Multi-scale mixed-integer MPC	On the largest instance, Gurobi needs about 25,000 s while DDIP converges in 662 s with less than 0.14% optimality gap (Kumar et al., 2020)

Two recurring empirical observations are especially stable across the cited works. First, the coupling between the two spaces is usually indispensable: removing frequency mixing in DMSR, removing one masking branch in ISD-MAE, removing Fus-Mamba in MSFMamba, omitting k-space modeling in MRI reconstruction, or discarding the dual-domain branch in ChannelKAN all degrades performance in the respective ablations (Zou et al., 15 Mar 2025, Ding et al., 2024, Gao et al., 2024, Meng et al., 14 Jan 2025, Jiang et al., 11 May 2026). Second, scale allocation must be matched to the problem structure: in dual continuum methods, the number of local spectral basis functions must match the number of high-conductivity channels or fracture networks; in covariance-based neuroimaging, performance improves as more scales are used up to an optimum around $$V_{ms} = \text{span}\{\psi_{k,ms}^{(j)} : 1 \leq k \leq L_j, 1 \leq j \leq N\},$$3; in hierarchical Mamba for MRI, the reported 1+3 HR/LR split gives the best trade-off in the cited ablation (Cheung et al., 2018, Baek et al., 3 Mar 2025, Meng et al., 14 Jan 2025).

6. Misconceptions, limits, and adjacent formulations

A common misconception is that “dual-space” always means spatial domain plus frequency domain. The cited literature directly contradicts that restriction. In subsurface simulation, the two spaces are physical continua; in neuroimaging, they are transform space and classifier space; in SD$$V_{ms} = \text{span}\{\psi_{k,ms}^{(j)} : 1 \leq k \leq L_j, 1 \leq j \leq N\},$$4NN they are low- and high-frequency subspaces; in DDIP they are primal MPC trajectories and dual cut-based value approximations; in DPMT they are fast and slow reasoning systems (Cheung et al., 2018, Baek et al., 3 Mar 2025, Li et al., 2021, Kumar et al., 2020, Li et al., 18 Jul 2025).

A second misconception is that “multi-scale” is synonymous with an image pyramid. Image pyramids are one important realization, but the cited work also uses oversampling layers, coarse mesh partitions, spectral scales, multi-route scanning, block time partitions, and semantic hierarchies. In this sense, scale can refer to resolution, frequency, topology, time, or abstraction level (Zhang et al., 2020, Zou et al., 15 Mar 2025, Gao et al., 2024, Kumar et al., 2020, Li et al., 18 Jul 2025).

The boundary of the concept is also fluid. Closely related work on multi-scale latent factorization in diffusion models decomposes latent features into a low-frequency base signal and a high-frequency residual signal, then generates them in two stages; the reported ImageNet FID values are 2.08 at $$V_{ms} = \text{span}\{\psi_{k,ms}^{(j)} : 1 \leq k \leq L_j, 1 \leq j \leq N\},$$5 and 2.47 at $$V_{ms} = \text{span}\{\psi_{k,ms}^{(j)} : 1 \leq k \leq L_j, 1 \leq j \leq N\},$$6, with a $$V_{ms} = \text{span}\{\psi_{k,ms}^{(j)} : 1 \leq k \leq L_j, 1 \leq j \leq N\},$$7 speed-up over the DiT baseline using the same number of sampling steps (Xu et al., 23 Jan 2025). Scale Space Diffusion formalizes a family of diffusion models with generalized linear degradations, using downsampling within the forward process and a Flexi-UNet that activates only the necessary parts of the network for the current resolution (Mukhopadhyay et al., 9 Mar 2026). These models do not always use “dual-space” as their primary label, but they instantiate adjacent decompositions in which coarse structure and fine detail are separated and then recombined.

Taken together, the literature suggests that the central invariant is not a particular choice of spaces, but a particular strategy: partition the problem into complementary views that are easier to model separately, then re-couple them through explicit constraints, projections, transforms, fusion operators, or hierarchical message passing. In finite element methods this is done by spectral selection and energy minimization; in deep networks by FFT/IFFT, cross-domain fusion, contrastive alignment, or state-space coupling; in control by dual cuts; and in collaboration frameworks by slow-system inference steering fast decisions (Cheung et al., 2018, Zou et al., 15 Mar 2025, Ding et al., 2024, Gao et al., 2024, Kumar et al., 2020, Li et al., 18 Jul 2025).