---
title: Spatial-Angular Residual Subspace Mamba Block
url: https://www.emergentmind.com/topics/spatial-angular-residual-subspace-mamba-block-sa-rsmb
type: topic
---

# Spatial-Angular Residual Subspace Mamba Block

Searching arXiv for the cited papers to ground the article in current literature.
The Spatial-Angular Residual Subspace Mamba Block (SA-RSMB) is a light-field super-resolution module introduced in the hybrid Mamba-Transformer framework LFMT for modeling shallow spatial-angular correlations in 4D light field (LF) data [2509.04824]. It is designed for Stage I of a dual-stage architecture and addresses a specific limitation of prior Mamba-based LF processing: multi-directional scanning can produce inefficient and redundant feature extraction when applied to complex LF structure. SA-RSMB responds by alternating domain-specific processing in the spatial and angular subspaces, using residual subspace Mamba operators to build spatial-angular representations efficiently before deeper epipolar refinement in Stage II [2509.04824].

## 1. Problem setting and motivation

LF images encode coupled spatial and angular information, yielding 4D data with complex spatial-angular correlations. In the formulation used for LFMT, this structure is treated as high-dimensional input whose non-local dependencies are difficult to capture with conventional CNNs because of their local receptive fields, while Transformer-based approaches incur substantial computational cost when applied directly to large and highly redundant 4D LF inputs [2509.04824].

The motivation for SA-RSMB follows from two premises. First, Mamba-based state space models offer long-range dependency modeling with linear complexity and therefore are attractive for LFSR. Second, existing Mamba-based image strategies often rely on multi-directional scanning, which the LFMT paper characterizes as inefficient and redundant for LF data. The stated design problem is therefore to model non-local dependencies and spatial-angular correlations effectively and efficiently, while avoiding redundant feature extraction and exploiting the intrinsic structure of LF representations [2509.04824].

Within this setting, SA-RSMB is explicitly positioned as a shallow feature extractor rather than a full-depth reconstruction mechanism. Its function is coarse extraction of spatial and angular information early in the network, leaving deeper disparity-aware and epipolar refinement to later modules.

## 2. Position within the LFMT architecture

LFMT adopts a dual-stage non-local spatial-angular modeling strategy. In Stage I, SA-RSMB performs shallow spatial-angular feature extraction. In Stage II, a dual-branch parallel structure composed of the Epipolar Plane Mamba Block (EPMB) and Epipolar Plane Transformer Block (EPTB) performs deep epipolar feature refinement [2509.04824].

This placement is central to the semantics of SA-RSMB. The block is not a generic residual Mamba unit applied arbitrarily to LF tensors; it is a domain-specific front-end mechanism whose output is intended to provide a strong basis for subsequent epipolar-plane modeling. The paper describes three cascaded SA-RSMBs with a residual connection over the stack, so Stage I itself is residual at both intra-block and inter-block levels [2509.04824].

A plausible implication is that the architecture separates coarse correlation discovery from fine structural refinement because state-space modeling alone may not preserve all spatial-angular and disparity information equally well. That interpretation is consistent with the explicit division of labor between Stage I and Stage II, although the precise effectiveness of the partition is established in the paper through ablation rather than by a formal proof.

## 3. Spatial-angular processing sequence

SA-RSMB operates by alternating spatial-domain and angular-domain processing on reshaped views of the same LF feature tensor. The input is

$$
F \in \mathbb{R}^{U \times V \times H \times W \times C},
$$

where \(U, V\) denote angular resolution, \(H, W\) denote spatial resolution, and \(C\) denotes channel count [2509.04824].

The first step is spatial modeling. The input is reshaped into a sub-aperture image stack,

$$
F_{SAI} \in \mathbb{R}^{UV \times H \times W \times C},
$$

and then processed by two cascaded Residual Subspace Mamba Blocks (RSMBs) followed by a convolution and a skip connection:

$$
F_S = \text{Conv}(\text{RSMB}^2(\text{RSMB}^1(F_{SAI}))) + F_{SAI}.
$$

This stage integrates spatial contextual information for each view [2509.04824].

The second step is angular modeling. The spatially enhanced feature is reshaped into macro-pixel image form,

$$
F_{MacPI} \in \mathbb{R}^{HW \times U \times V \times C},
$$

which emphasizes the angular plane. Two additional cascaded RSMBs and a convolution are then applied, again with a skip connection:

$$
F_{SA} = \text{Conv}(\text{RSMB}^2(\text{RSMB}^1(F_{MacPI}))) + F_{MacPI}.
$$

The result is reshaped back and passed to Stage II [2509.04824].

The operational summary given in the source is concise: initial feature \(\rightarrow\) spatial RSMBs \(\rightarrow\) spatial features \(\rightarrow\) angular RSMBs \(\rightarrow\) spatial-angular features. The significance of this sequence is that spatial and angular interactions are not jointly flattened into a single indiscriminate scan; they are modeled in dedicated steps, each aligned with a particular LF subspace.

## 4. Internal operators: RSMB and SSMB

The immediate computational core of SA-RSMB is the Residual Subspace Mamba Block. For input feature \(F \in \mathbb{R}^{B \times H \times W \times C}\), the RSMB is defined as

$$
\begin{aligned}
\overline{F} &= SSMB(\text{LN}(F)) + s_1 \cdot F, \\
\widehat{F} &= CA(\text{LN}(\overline{F})) + s_2 \cdot \overline{F},
\end{aligned}
$$

where SSMB denotes the Subspace Simple Mamba Block, LN denotes Layer Normalization, CA denotes Channel Attention, and \(s_1, s_2\) are learnable residual scaling parameters [2509.04824].

This definition makes two architectural commitments. First, Mamba-based subspace modeling is embedded in an explicitly residualized pathway. Second, channel attention is applied after the Mamba transform, which the paper frames as improved feature aggregation and representation. In effect, RSMB is the residual-and-attention wrapper that turns a subspace Mamba scan into a robust LF feature operator.

The SSMB itself is defined by the Subspace Simple Scanning (Sub-SS) strategy. Its reported characteristics are domain-specific unidirectional scanning, replacement of causal convolutions with regular convolutions, and addition of a parallel symmetric convolution+SiLU branch to compensate for information that may be lost under unidirectional modeling. The resulting outputs are concatenated and projected back to the target channel dimension [2509.04824]. Its computation is given as

$$
\begin{aligned}
F_1 &= SS\text{-}sub( \sigma( \text{Conv}(\text{Linear}(C, C/2)(F_{in})))), \\
F_2 &= \sigma(\text{Conv}(\text{Linear}(C, C/2)(F_{in}))), \\
F_{out} &= \text{Linear}(C, C)(\text{Concat}(F_1, F_2)),
\end{aligned}
$$

with \(\sigma\) denoting SiLU activation [2509.04824].

Taken together, SA-RSMB can be understood as a hierarchy of operators: SA-RSMB alternates subspaces; each subspace stage uses RSMBs; each RSMB is built around an SSMB plus normalization, channel attention, and learnable residual scaling.

## 5. Modeling rationale and claimed advantages

The principal methodological claim for SA-RSMB is decoupled modeling of spatial and angular domains. By explicitly separating these domains, the block is said to avoid the redundancy and inefficient coupling associated with multi-directional scanning, while still progressively building spatial-angular correlations through sequential residual aggregation [2509.04824].

The efficiency argument is tied directly to Sub-SS. Because Sub-SS uses unidirectional, domain-specific scanning rather than bidirectional or four-way scanning, the paper attributes to SA-RSMB reduced parameter count and computation relative to earlier Mamba applications to images. The symmetric compensation branch is introduced precisely because unidirectional modeling may block some information flow; in that sense, the design is not merely sparse but compensatory [2509.04824].

Empirically, the paper reports that removing SA-RSMB in ablation studies causes a significant performance drop in both PSNR and SSIM, and it further states that the overall LFMT framework achieves higher or comparable PSNR/SSIM at lower computational cost than pure Transformer or naive Mamba methods. The same source also attributes faster convergence and more stable learning to the Sub-SS/SA-RSMB design [2509.04824]. Because no numerical deltas are supplied in the provided material, the encyclopedic conclusion is qualitative rather than quantitative: SA-RSMB is presented as both a performance-critical and efficiency-oriented component of LFMT.

A plausible implication is that SA-RSMB is best interpreted as a structure-aware preconditioner for later epipolar reasoning. Its output is not the terminal representation of LF content but an organized spatial-angular feature basis that improves the effectiveness of downstream EPMB/EPTB refinement.

## 6. Nomenclature, related architectures, and interpretive boundaries

The term SA-RSMB is specific to LFMT. A common source of confusion is the tendency to retroactively apply the label to earlier or adjacent Mamba-based subspace modules. The provided literature does not support unrestricted use of that nomenclature.

In particular, the 2024 paper "Mamba-based Light Field Super-Resolution with Efficient Subspace Scanning" introduces MLFSR, which uses bi-directional subspace scanning across spatial, angular, and EPI subspaces, together with the Mamba-based Global Interaction (MGI) module and the local Spatial-Angular Modulator (SAM) [2406.16083]. That source explicitly states that the paper does not use the exact term SA-RSMB, but that the alternation of MGI and SAM can be viewed as functionally realizing the concept. The distinction matters: MLFSR is a relevant precursor in subspace-aware LF Mamba design, but SA-RSMB as a named module belongs to LFMT [2406.16083].

A second contextual comparison arises from "DualMamba: A Lightweight Spectral-Spatial Mamba-Convolution Network for Hyperspectral Image Classification," which operates on spectral-spatial rather than spatial-angular structure. Its lightweight spatial and spectral Mamba blocks, residual design, and adaptive global-local fusion offer an architectural analogy for subspace-specific Mamba processing, but the source frames any relation to SA-RSMB as an extension idea rather than an instantiated LF module [2406.07050]. This suggests a broader design pattern—residualized, subspace-specific state-space modeling with complementary local branches—without collapsing distinct problem domains into a single terminology.

Accordingly, the historically precise view is that SA-RSMB is a Stage I LFMT block for shallow spatial-angular extraction in LFSR, situated within a line of work on efficient subspace scanning but not interchangeable with all prior subspace Mamba constructs.

## 7. Significance in light-field super-resolution

Within LFMT, SA-RSMB is described as the mechanism that makes early spatial-angular exploration both efficient and non-redundant, enabling the framework to integrate the strengths of Mamba and Transformer models across spatial, angular, and epipolar-plane domains [2509.04824]. Its significance is therefore architectural rather than merely local: it operationalizes the first half of LFMT’s dual-stage strategy and defines how shallow LF structure is exposed to later deep refinement.

The broader significance follows from the paper’s framing of the LFSR problem. If multi-directional scanning is genuinely redundant on LF data, then SA-RSMB represents a move toward subspace-aligned scanning policies rather than ever more exhaustive directional traversals. If the combination of unidirectional Sub-SS, residual scaling, channel attention, and alternating spatial-angular reshaping is responsible for the reported PSNR/SSIM gains and complexity reductions, then SA-RSMB also marks a specific answer to a more general question in vision SSM design: how to preserve domain structure while retaining linear-complexity long-range modeling [2509.04824].

In that sense, SA-RSMB is best understood not as an isolated block but as a formalization of domain-decoupled shallow LF modeling. Its contribution lies in the exact way it sequences spatial and angular residual subspace Mamba operations before handing off to epipolar Mamba-Transformer refinement.

Source: https://www.emergentmind.com/topics/spatial-angular-residual-subspace-mamba-block-sa-rsmb