Adaptive Rotation Scheme (ARS)

• ARS is a class of adaptive algorithms that optimally adjusts rotation parameters in matrices or control systems for improved performance across diverse domains.
• It integrates techniques like Hadamard rotations, Clifford algebra, and gradient-based optimization to robustly handle nonstationarity, outliers, and hardware constraints.
• ARS enhances system efficiency, reducing quantization errors in deep models and maximizing signal-to-noise ratios in wireless communications and adaptive optics.

An Adaptive Rotation Scheme (ARS) is a class of algorithmic and architectural strategies that enable dynamic, context-dependent adjustment of rotation parameters in complex systems. ARS solutions are used across a diverse range of domains, including deep model quantization, signal processing, wireless communications, and control of physical devices such as antennas and adaptive optics. By adaptively selecting or optimizing rotation operators—either on data representations, network parameters, physical devices, or control matrices—ARS increases performance, robustness, or efficiency in the face of domain shifts, nonstationarity, outliers, or hardware constraints.

1. Mathematical Foundations and General Formulations

ARS is characterized by the presence of learnable or optimizable rotation transformations that can be dynamically selected or constructed in response to observed system states or input statistics. Mathematically, a typical ARS operation can be abstracted as:

• For a data or parameter matrix $X$, applying a rotation $R$ to obtain $X'$: $X' = X R$
• For a control or actuation matrix $C$, applying basis transformations: $C_{\text{rot}} = R_{m} C_{0} R_{s}^{-1}$, where $R_{m}$ and $R_{s}$ are rotation operators acting in different domains (e.g., actuator and sensor spaces (Arcidiacono et al., 2010))

The construction, selection, or parameterization of $R$ depends on application context:

• In post-training quantization of neural models, $R$ can be a normalized Hadamard matrix or a carefully constructed outlier-aware orthogonal rotation (Yang et al., 5 Aug 2025).
• In geometric adaptive filtering, $R$ is a rotor (multivector) in Clifford algebra, estimated via stochastic gradient descent to best align correspondences (Lopes et al., 2016).
• In MIMO wireless, $R$ may represent antenna orientation matrices, panel-level or element-level rotations, or IRS surface normals (Feng et al., 16 Oct 2025, Zheng et al., 8 Jan 2026, Peng et al., 16 Dec 2025).
• In domain generalization, ARS optimizes over the Lie group SO(3) to identify the hardest orientations for each sample during training (Liu et al., 4 Feb 2025).

2. Algorithmic Instantiations

ARS across fields typically comprises stages for (a) estimation (of signal, error, or loss landscapes), and (b) optimization or selection of the rotation, followed by application of the resulting transformation:

• Post-Training Quantization (PTQ): ARS in PTQ chooses per-layer activation rotations (Hadamard for mild outliers, or outlier-aware for salient high channels) by monitoring the fluctuation metric $J$; the rotation is then jointly applied to the activations and weights, followed by quantization (Yang et al., 5 Aug 2025).
• Geometric Algebra-Based Filtering: An online least-mean-squares update in the Clifford algebra is performed to iteratively adapt the rotation parameter $r$ minimizing $J(r)$, ensuring $r\tilde{r} = 1$ at each iteration (Lopes et al., 2016).
• Wireless Beamforming & Channel Estimation: ARS alternates between channel estimation/fitting (e.g., via MUSIC and LS) and maximizing array gain or SNR by updating orientation parameters $\{\theta_n, \alpha_m, \beta_n\}$ using gradient ascent, feasible-direction, particle swarm optimization, or genetic algorithms (Xiong et al., 25 Jun 2025, Zheng et al., 8 Jan 2026, Feng et al., 16 Oct 2025, Peng et al., 16 Dec 2025).
• Intricate Orientation Learning: In rotation-robust 3D learning, ARS alternately mines worst-case SO(3) rotations for each training instance (via projected gradient ascent on loss) and updates the representation via contrastive objectives and knowledge distillation (Liu et al., 4 Feb 2025).
• Adaptive Optics Control: For multi-conjugate/adaptive optics, ARS numerically rotates the DM–WFS control matrix in real time according to measured pupil rotation, updating when the deviation threshold is exceeded (Arcidiacono et al., 2010).

3. Representative Applications

ARS is deployed in a variety of scenarios:

Domain	ARS Mechanism	Primary Objective
Neural Quantization	Layer-wise adaptive activation rotations	Reduce quantization error under low-bit constraints
Point Cloud Learning	SO(3) mining of maximally confusing orientations	Generalize to unseen rotations and domains
Wireless Antennas	Joint optimization of antenna/panel angles	Maximize sum rate, SNR, array/interference performance
IRS/Metasurfaces	Cooperative IRS orientation & beamforming	Signal/noise ratio, secrecy, interference management
Adaptive Optics	Real-time rotation of control/interaction matrix	Maintain low wavefront error despite field rotation

Deep Model Quantization

ARS, as in LRQ-DiT (Yang et al., 5 Aug 2025), selects between Hadamard and outlier-aware rotations for each Transformer layer based on a root-mean-square metric $J$. This rotation, applied prior to quantization, disperses outlier effects and enables tighter quantization intervals. Experimentally, combining ARS with log-based quantization (TLQ) achieves significant FID improvements, e.g., PixArt-α W3A6 FID=57.76 (TLQ+ARS) versus 65.61 (TLQ only).

Geometric Adaptive Filtering

The GA-LMS ARS in (Lopes et al., 2016) incrementally estimates the rotor mapping source to target via a stochastic descent on the Clifford-algebraic least-squares cost. Its per-iteration complexity (54 RM, 39 RA) is lower than SVD-based registration for large correspondence sets, and it allows efficient streaming or online registration.

Rotatable Antenna Arrays

ARS is used to coordinate antenna orientation in response to channel conditions:

• (Feng et al., 16 Oct 2025) links rotation vectors (ARV) and complex weights (AWV) in a joint optimization; alternating SCA/PSO is used for multi-beam/multi-user cases.
• (Zheng et al., 8 Jan 2026) introduces cross-linked rotation (row/column coupling) and solves the sum-rate maximization using MMSE/feasible-direction AO and genetic algorithms, yielding 128% sum-rate gains over fixed arrays.

Rotatable IRS and Secure Metasurfaces

• (Peng et al., 16 Dec 2025) optimizes the orientation of IRS surfaces to maximize SNR under cascaded LoS and Rician channels, using closed-form expressions in 2D and PSO/AO in 3D.
• (Li et al., 17 Nov 2025) exploits ARS for array pose in dual-BS secure communications, yielding secrecy rates up to 22% higher than non-rotatable baselines; solutions use GRQ for discrete online rotation and MADDPG for offline policy learning.

Adaptive Optics

In MCAO systems such as LINC-NIRVANA (Arcidiacono et al., 2010), ARS rotates the control matrix in slope-mode space as the pupil image rotates. Real-time implementation enables closed-loop operation with wavefront error bounded below 20 nm RMS by updating every few seconds.

4. Theoretical Insights and Complexity

ARS often yields closed-form updates in single-user or LoS cases, and efficient global optimization can be assured via convex relaxations or eigen-decomposition (e.g., Rayleigh quotient maximization in security/power problems (Li et al., 17 Nov 2025)). In overparameterized joint-optimization settings (e.g., multi-user MIMO), alternating minimax, SCA, or PSO achieve empirically rapid convergence. Complexity per iteration is governed by the structure of the rotation (Hadamard: $O(C \log C)$; dense: $O(C^2)$; panel/element rotations: $O(M+N)$ updates).

5. Empirical Performance

Consistent quantitative improvements are observed across domains:

• In DiT quantization, ARS+TLQ reduces FID from >150 (Hadamard only) to ≈53 (with outlier-aware ARS) (Yang et al., 5 Aug 2025).
• In rotatable arrays, ARS yields up to 128% sum-rate gain over fixed orientation, and element-level cross-linked ARS is within ≲14% of the fully flexible (per-antenna) optimum (Zheng et al., 8 Jan 2026).
• In multi-IRS relaying, ARS enables SNR scaling from $O(N^2)$ to $O(N^4)$ in double rotatable deployments (Peng et al., 16 Dec 2025).
• In rotation-robust 3D learning, ARS mining reduces MMD and t-SNE cluster overlap, yielding higher accuracy and domain generalization (Liu et al., 4 Feb 2025).
• In AO, ARS limits Strehl loss to ≲2% for pupil rotation ≤0.3°, maintaining $\sigma_{\rm WFE} \lesssim 20$ nm (Arcidiacono et al., 2010).

6. Limitations and Future Prospects

While ARS universally enhances adaptability, key challenges remain:

• Memory and Computation: Dense outlier-aware rotations ($O(C^2)$ storage) may be prohibitive in very large models or arrays (Yang et al., 5 Aug 2025).
• Hardware Constraints: Mechanical actuation latency and discrete rotation quantization can limit real-time adaptation in practical arrays; continuous or hybrid actuators are an active area of research (Li et al., 17 Nov 2025, Zheng et al., 8 Jan 2026).
• Hyperparameter Sensitivity: Thresholds (e.g., in outlier detection for DiT ARS, batch-sizes for contrastive learning ARS) and step-sizes require empirical tuning.
• Scalability: For massive IRSs, fully cooperative ARS may encounter prohibitive communication or computation requirements; distributed or hierarchical ARSs are a plausible avenue.

Extensions under exploration include integration with learned scale/shift per channel, continuous-rotation control via gradient-based methods, distributed ARS across multi-cell systems, and robustification to non-ideal or adversarial environments.

7. Comparative Summary of ARS Methods

Reference	Domain	ARS Rotation Method	Core Objective	Key Algorithmic Trait
(Yang et al., 5 Aug 2025)	DiT Quantization	Hadamard, Outlier-Aware	Reduce quant error	Dynamic per-layer selection
(Lopes et al., 2016)	Geometric Filtering	Clifford Rotor	Min LS error in rotation	Instantaneous GA-LMS
(Feng et al., 16 Oct 2025)	Rotatable Antenna Arrays	Element/panel rotation	Max beam/interf. perf	Joint ARV/AWV opt., AO+PSO
(Peng et al., 16 Dec 2025)	Cooperative IRSs	IRS 3D normal rotation	Max SNR in cascaded link	PSO, closed-form 2D, AO
(Li et al., 17 Nov 2025)	Secure Metasurfaces	Array pose, T-RIS	Max secrecy rate	Rayleigh Quotient, MADDPG
(Liu et al., 4 Feb 2025)	3D Domain Generalization	SO(3) per-sample mining	Rotation-robust learning	Alternating inner maximization
(Xiong et al., 25 Jun 2025)	Channel Estimation	RA per-element orientation	Max channel gain, AoA	MUSIC, LS, per-element gradient
(Arcidiacono et al., 2010)	Adaptive Optics	WFS/DM control matrix	Min residual WFE	Double-basis control rotation
(Zheng et al., 8 Jan 2026)	Cross-linked RA MIMO	Coupled row/col rotation	Max sum-rate	MMSE + feasible-dir, GA for discretes

The ARS paradigm thus provides a unified conceptual and algorithmic framework for leveraging rotation as an adaptive, context-sensitive degree of freedom in both computational and physical systems. Its implementations span mathematical optimizations, learning-based control, signal processing, data mining on Lie groups, and real-time low-level hardware actuation. Each application exploits tailored rotation subspaces and adaptive protocols to blend performance, robustness, and practical constraints.