Asymmetric Reduction and Restoration (AsymRnR)

• AsymRnR is a framework that employs targeted, non-uniform reduction of undesired features and explicit restoration of critical structure in complex domains.
• In neural architectures like AS-Mamba, methodologies such as SSM-based directional suppression and frequency-decoupled restoration significantly enhance artifact reduction and image quality.
• AsymRnR principles extend to transformer acceleration and nonlinear dynamics, where adaptive token pruning and bifurcation-induced stabilization yield improved performance and stability.

Asymmetric Reduction and Restoration (AsymRnR) is a unifying principle and suite of methodologies for the targeted suppression (“reduction”) of undesired features and the precise recovery (“restoration”) of desired structure or function in complex systems, utilizing mathematically and physically asymmetric mechanisms. Originating in disparate contexts—computational imaging, generative machine learning, and nonlinear wave dynamics—all AsymRnR methodologies exploit non-uniform properties of the underlying problem (e.g., spatial, frequency, or feature redundancy asymmetries) to enable performance or stability that is provably unattainable with symmetric, uniform, or “one-size-fits-all” approaches. Crucial innovations include content-adaptive reduction schedules in neural architectures, physical-directional decoupling in signal processing, and non-symmetric bifurcation-induced stabilization in nonlinear dynamics.

1. Core Principles of Asymmetric Reduction and Restoration

AsymRnR frameworks are distinguished by three interrelated features:

• Targeted Asymmetry: The reduction strategy is intentionally non-uniform; it acts more strongly or exclusively on those components (spatial/frequency directions, feature types, or modes) identified as redundant, unstable, or undesirable by means of intrinsic measurements or modeling.
• Explicit Restoration: Lost information or destabilized structure is not left as collateral; instead, restoration is addressed with complementary mechanisms (iterative refinement, cache-based token reinflation, or nonlinear stabilization), often leveraging the same asymmetries that motivated reduction.
• Mathematical Decoupling and Design: AsymRnR methodologies often decouple problem dimensions—frequency bands, query/key/value token streams, or waveguide directions—allowing specialization of reduction and restoration to the physical or informational properties that underlie observed artifacts or inefficiencies (Ning et al., 6 Feb 2026, Sun et al., 2024, Zezyulin, 2024).

2. AsymRnR in Neural Architectures for Artifact Reduction

In the context of medical imaging, AsymRnR is operationalized in the AS-Mamba architecture for metal artifact reduction in computed tomography (CT). The framework incorporates asymmetric reduction and restoration at multiple architectural layers:

• Directional Streak Suppression via SSM-based Mamba Blocks: Exploiting the nearly linear propagation of streak artifacts, AS-Mamba utilizes State Space Model (SSM) recurrences to promote memory along the trajectory of artifacts. Each MambaBlock enforces an asymmetric update:

$$F_m = F_{in} + \alpha \cdot \text{SSM}_s(\mathrm{LN}(F_{in})), \ F_{out} = F_m + \beta \cdot \text{FFN}(\mathrm{LN}(F_m))$$

with feature flattening along candidate streak paths and unequal treatment of feature directions (Ning et al., 6 Feb 2026).

• Frequency-Decoupled Restoration: The architecture splits the input into high-frequency (streaks) and low-frequency (shading) bands via discrete Haar wavelets. High-frequency suppression employs the above SSM-driven branch, while low-frequency restoration applies a learned correction to the amplitude of the Fourier spectrum:

$$\hat A(u,v) = A(u,v) + [\text{FFT}(f_e * \text{IFFT}(A))](u,v)$$

only affecting the spectrum magnitude and preserving global shading (Ning et al., 6 Feb 2026).

• Iterative MANet and Contrastive Perceptual Losses: Following synthesis via inverse wavelet transform, iterative proximal updates and a self-guided VGG-19-based contrastive loss ($L_S$) enforce both global artifact suppression and fine-detail matching, with epoch-dependent weight shifting the focus over training.

The efficacy of AS-Mamba is empirically validated, outperforming previous methods such as ACDNet in PSNR, SSIM, and clinical scores (e.g., PSNR of 45.04 dB vs 44.19 dB; clinical score of 4.00/5 vs 3.85 for ACDNet) (Ning et al., 6 Feb 2026).

3. AsymRnR for Transformer Acceleration in Generative Video

In large-scale video Diffusion Transformers (DiTs), AsymRnR enables efficient model acceleration through token pruning and restoration strategies tailored to attention redundancy:

• Feature-wise Redundant Token Reduction: At each self-attention layer, feature-wise reduction operators $\mathcal{R}_Q, \mathcal{R}_{KV}$ remove tokens deemed redundant by a block- and time-dependent similarity metric:

$$\sigma_{b,t}^A = \frac{2}{n(n-1)}\sum_{1\le i<j\le n} -\|a_i - a_j\|_2$$

Rather than uniform pruning, an adaptive schedule is constructed:

$$r_{b,t}^A = \begin{cases} r_{\max} \frac{\sigma_{b,t}^A - \sigma_{\min}}{ \sigma_{\max} - \sigma_{\min}}, & \sigma_{\min} \leq \sigma_{b,t}^A \leq \sigma_{\max} \ 0, & \sigma_{b,t}^A < \sigma_{\min} \ r_{\max}, & \sigma_{b,t}^A > \sigma_{\max} \end{cases}$$

Assigning much higher pruning to almost-linearly dependent (redundant) values than to less-redundant queries preserves quality (Sun et al., 2024).

• Matching Cache-based Restoration: To minimize the computational overhead of bipartite soft-matching, which merges source-destination token pairs, AsymRnR employs a matching cache $\mathcal{S}^{(t)}$ updated only every $s$ steps, exploiting empirical temporal redundancy.
• Controlled Quality-Performance Trade-off: Perturbation analysis shows pruning in high-redundancy directions introduces only $O(\epsilon)$ errors in attention; experiments confirm near-zero drop in VBench quality even at 30% token reduction. On CogVideoX 2B/5B, up to 1.17× speedups are obtained with VBench quality loss $<0.002$ (Sun et al., 2024).

4. AsymRnR in Nonlinear Wave Dynamics

Mathematical AsymRnR is manifested in stability restoration of nonlinear modes in complex optical potentials:

• Reduction by Asymmetry in Double-Well Potentials: In non-Hermitian, parity-symmetric but not parity-time-symmetric potentials (Wadati form $$\hat A(u,v) = A(u,v) + [\text{FFT}(f_e * \text{IFFT}(A))](u,v)$$0), the linear regime only supports unstable eigenmodes (complex-valued spectrum).
• Restoration via Nonlinear Asymmetry: Above a threshold amplitude, bifurcation analysis yields stable, asymmetric stationary modes $$\hat A(u,v) = A(u,v) + [\text{FFT}(f_e * \text{IFFT}(A))](u,v)$$1, realized via saddle–node (fold) bifurcations in the nonlinear Schrödinger equation:

$$\hat A(u,v) = A(u,v) + [\text{FFT}(f_e * \text{IFFT}(A))](u,v)$$2

The two-mode reduction translates this into a non-Hermitian dimer system with exact thresholds for existence and stability of asymmetric modes. Spectral stability is demonstrated by moving eigenvalues onto the imaginary axis for sufficient nonlinear amplitude, restoring stability (Zezyulin, 2024).

• Existence of Periodically Breathing Solutions: The two-mode projection admits closed periodic orbits in the Stokes space, corresponding to periodically oscillating intensity in the original NLSE dynamics, further illustrating a dynamical restoration regime only accessible via asymmetric reduction.

5. Empirical and Theoretical Evaluation

Robust evidence across domains substantiates the necessity of explicit asymmetric design:

• Medical Imaging: Ablation studies in AS-Mamba highlight that symmetric (unified) architectures (SDI-Net, UWP-Net) are strictly dominated by frequency-decoupled, directionally asymmetric strategies, with additive gains in PSNR and SSIM.
• Transformer Acceleration: Experiments in video DiT models show that reducing only value tokens outweighs prior state-of-the-art (ToMe); indiscriminate (query) reduction severely degrades generation.
• Non-Hermitian Potentials: Saddle–node bifurcation and restriction to asymmetric nonlinear states delineate clear parameter thresholds. Restoration bands are accurately predicted by two-mode theory ($$\hat A(u,v) = A(u,v) + [\text{FFT}(f_e * \text{IFFT}(A))](u,v)$$3), and numerical simulations confirm agreement.

A summary of the AsymRnR approaches across domains is organized below.

Domain/Model	Core Reduction	Restoration Mechanism	Quality/Performance Outcome
AS-Mamba (Ning et al., 6 Feb 2026)	SSM/Mamba along artifact direction	Fourier amplitude correction, iterative refinement	PSNR ↑ 45.04 dB, SSIM ↑ 0.9918
DiT (Sun et al., 2024)	Token pruning via redundancy measure	Token restoration via matching cache	1.13× speedup, <0.002 VBench loss
Nonlinear Optics (Zezyulin, 2024)	Linear-mode exclusion via asymmetry	Nonlinear bifurcation-induced stabilization	Real-stable modes only for $$\hat A(u,v) = A(u,v) + [\text{FFT}(f_e * \text{IFFT}(A))](u,v)$$4

6. Mathematical and Physical Foundations

A commonality of AsymRnR systems is their reliance on structural or statistical non-uniformities:

• In AS-Mamba, the linear structure of metal-induced artifacts is congruent with the SSM recurrence formalism, which is parametrized asymmetrically by data-driven selection in state transition and input matrices at each spatial position.
• In video DiTs, the feature-wise discrepancy in query, key, and value redundancy is exploited through layer/timestep-dependent schedules, enforcing adaptive, content-aware reduction.
• In optical waveguides, the exclusion of symmetric stationary states arises due to the parity symmetry (but not PT-symmetry) of the complex potential; only through a nonlinear shift can asymmetric, stable states exist.

This suggests that AsymRnR effectiveness is fundamentally linked to the presence of non-uniform or directionally biased data, dynamics, or artifacts, and that the optimal design of reduction and restoration mechanisms must be informed by analytic or empirical modeling of these asymmetries.

7. Implications and Extensions

The AsymRnR paradigm demonstrates that domain-specific, physically and informationally grounded asymmetries enable capabilities beyond symmetric approaches. In imaging, this yields state-of-the-art artifact suppression and structure preservation; in machine learning, hardware-efficient acceleration at constant or near-constant quality; in wave physics, the existence of dynamically stable states otherwise forbidden.

A plausible implication is that further generalization to other domains—signal processing, communications, or inverse problems—could yield methodologies that systematically exploit model and data asymmetries to achieve both computational and qualitative advantages, contingent on rigorous characterization of reduction-restoration interplay. Ongoing research may focus on learning or discovering optimal asymmetry structures for new modalities, and on theoretical bounds that relate the degree of informational asymmetry to achievable restoration fidelity or acceleration.