RatCG Method Overview

• RatCG method encompasses three distinct algorithms that leverage rational structures for sequence analysis, regularization, and adaptive guidance.
• It integrates Rational Krylov-based Tikhonov regularization to achieve optimal convergence rates in solving linear inverse problems.
• R-CGR ensures lossless, reversible sequence encoding, while RAAG adapts guidance to stabilize and accelerate sampling in generative models.

The term “RatCG method” denotes several distinct, technically unrelated algorithms introduced under this acronym in diverse research domains. Three principal instantiations have appeared in the literature: (1) the Rational Chaos Game Representation (RatCG) for sequence analysis and lossless geometric encoding, (2) the RATIO-Aware Adaptive Guidance (RatCG/RAAG) for fast, robust sampling in flow-based generative models, and (3) the Rational Krylov-based CG (RatCG) for the regularization of linear inverse problems. Each method is characterized by fundamentally different motivations, derivations, and application domains, but all share an explicit emphasis on rational structure or adaptive scaling within their respective frameworks.

1. Rational Krylov-Based RatCG for Inverse Problems

The RatCG method for solving ill-posed linear inverse problems is a hybrid regularization algorithm that enhances classical Krylov subspace techniques with rational function-based Tikhonov regularization. Let $A:X\to Y$ be a bounded linear operator between Hilbert spaces, and let $y^\delta$ satisfy $\|y^\delta-y\|\leq\delta$ for the (unknown) $y=Ax^\dagger$. The method defines a sequence of regularization parameters $\alpha_1,\ldots,\alpha_m$ ($\alpha_i>0,\; \alpha_i\neq \alpha_j$), and for each, the corresponding Tikhonov solution

$x_{\alpha_i} = (A^*A+\alpha_i I)^{-1}A^*y^\delta.$

The $n$-th RatCG iterate $x_n$ is the least-squares minimizer over the mixed Krylov–Tikhonov space $KR^n$, which combines Tikhonov solutions and Krylov subspace elements. For even $n=2k$,

$$KR^n = \mathrm{span}\{ x_{\alpha_1},\,A^*y^\delta,\,(A^*A)A^*y^\delta,\,x_{\alpha_2},\ldots,\,x_{\alpha_k} \}.$$

For odd $n=2k+1$, the span is similarly augmented by the next Krylov vector. This construction embeds both $\{(A^*A+\alpha_iI)^{-1}A^*y^\delta\}$ (rational Krylov vectors) and classical $(A^*A)^kA^*y^\delta$ directions.

The short-recurrence RatCG algorithm alternates between

• Tikhonov steps in the direction $(A^*A+\alpha_kI)^{-1}A^*y^\delta$, and
• classical conjugate gradients for the normal equations (CGNE) steps in the $(A^*A)v$ direction,

with orthogonalization performed in the $A^*A$-induced inner product. At each iteration, a small least squares problem is solved in the current basis.

The method employs the discrepancy principle as a stopping rule: find the smallest $n_*$ such that $\|A x_{n_*} - y^\delta\| < \tau\delta$ but $\|A x_{n_*-1} - y^\delta\| \ge \tau\delta$ for a prescribed $\tau>1$. Under a standard source condition $x^\dagger = (A^*A)^\mu\omega$ for $\mu>0$, and provided $\alpha_i\geq c_0>0$, RatCG achieves the optimal regularization rate

$$\|x_{n_*} - x^\dagger\| = O\left( \delta^{\mu/(\mu+1/2)} \right).$$

This rate matches the best attainable under the given source smoothness. The theoretical analysis involves spectral polynomial factorizations, error splitting, and balancing regularization parameters. No numerical experiments are reported in the primary theoretical references, but previous empirical work (Kindermann & Zellinger 2024, Chen & Pereverzyev 2015) demonstrates substantially faster convergence and lower reconstruction error than classical CGNE on standard test problems (Kindermann, 15 Jan 2026).

2. Rational Chaos Game Representation (RatCG) for Sequence Encoding

In bioinformatics and computational biology, RatCG refers to Rational Chaos Game Representation—a geometric sequence mapping technique centered on lossless, information-preserving representations. The method addresses the failure of standard Chaos Game Representation (CGR) to permit complete sequence recovery from geometric mappings. RatCG, implemented in Explicit Path CGR ("R-CGR"), encodes biological sequences as explicit rational paths, ensuring perfect reversibility.

Each alphabet symbol $s_i$ is assigned a rational "corner" $C_i=(\alpha_i,\beta_i)\in\mathbb{Q}^2$, where $\alpha_i,\beta_i$ are rational approximations to $\cos(2\pi i/k),\sin(2\pi i/k)$ with a common denominator $q=2^{\lceil \log_2(4k) \rceil}$. A sequence is mapped iteratively: starting at $p_0=(0,0)$, each point is updated as the rational midpoint between the previous point and the corner associated with the current symbol: $$x_j = x_{j-1}q + c_x q_{j-1},\quad y_j = y_{j-1}q + c_y q_{j-1},\quad q_j = 2q_{j-1}q,$$ with each $(x_j, y_j, q_j)$ immediately reduced by their greatest common divisor. To control denominator growth, optional continued-fraction approximations are applied when denominators exceed a threshold, with all approximations fully recorded to guarantee reversibility.

The decoding algorithm trivially reconstructs the original sequence by reading the stored symbol trace. The method's lossless property is formalized: perfect recovery is guaranteed because every encoded step retains both geometric and symbolic information, and the decoder simply concatenates the stored symbols in order.

A toy example for DNA (alphabet $\Sigma=\{A,T,G,C\}$, $k=4$, $q=16$) demonstrates the construction of explicit rational midpoints for a given short string, e.g. $S=$“AT”, with all intermediate rational numerators and denominators shown.

Empirical results on a synthetic 7-class DNA+protein dataset demonstrate competitive or superior classification performance of RatCG (R-CGR) compared to minimizer-based CGR (Spike2CGR), ProtT5, SeqVec, and ESM2 embeddings, while uniquely retaining full lossless recoverability. For example, VGG16 achieves 79.07% accuracy using R-CGR images, surpassing both traditional CGR images and language-model embeddings. This property enables deep learning on geometric representations without information loss, opening avenues for interpretable and hybrid modeling (Ali, 22 Sep 2025).

3. RATIO-Aware Adaptive Guidance (RatCG/RAAG) in Generative Modeling

RatCG/RAAG ("RATIO-Aware Adaptive Guidance") in diffusion and flow-based generative models is an adaptive scheduling method that mitigates instability in fast (low-step) classifier-free guidance (CFG) sampling. CFG, the standard tool for controllable synthesis in high-dimensional generative models, is subject to instabilities due to a pronounced "RATIO spike" in the earliest reverse steps.

The RATIO metric at each step $t$ is defined as

$$\mathrm{RATIO}_t = \frac{\|\delta_t\|_2^2}{\|v_u(x_t)\|_2^2},$$

where $v_u(x_t)$ is the unconditional velocity, $v_c(x_t,c)$ is the conditional velocity given prompt $c$, and $\delta_t = v_c(x_t,c)-v_u(x_t)$ is the velocity gap. At the initialization step ($t=1$), the denominator is of order $d$ (dimensionality), but the numerator is dominated by the mean shift $\|\mu_c-\mu_u\|^2 = O(d)$, leading typically to $\mathrm{RATIO}_1 = O(1)$. This produces a strong guidance signal and exposes the system to error amplification.

Theoretical analysis demonstrates that under fixed CFG scale $w>1$, any $w \cdot p > 1$ (where $p$ is the instantaneous RATIO) yields exponential amplification of small perturbations—manifesting as semantic collapse and loss of sample quality.

To address this, RatCG/RAAG introduces an adaptive guidance schedule: $$w_t = 1 + (w_{\max}-1)\exp(-\alpha\,\mathrm{RATIO}_t),$$ with $w_{\max}$ the maximum guidance (e.g., $7.0$ in SD3.5), and $\alpha$ a decay-rate parameter determined by grid or greedy N-step search. When $\mathrm{RATIO}_t \approx 0$ (weak conditional signal), full guidance is leveraged; when $\mathrm{RATIO}_t$ spikes, $w_t \rightarrow 1$ and guidance is suppressed.

Operationally, at each reverse step, the model computes $v_u$, $v_c$, $\delta$, and RATIO, evaluates $w_t$ according to the schedule, then updates using $v_{\mathrm{cfg}} = v_u + w_t\delta$. The method requires no retraining, negligible extra computation, is compatible with existing flow frameworks, and involves no additional network inference.

Empirical results across multiple leading flow-based models (Stable Diffusion 3.5, Lumina-Next, WAN2.1) show up to $3\times$ speedup for text-to-image synthesis (e.g., $10$-step RAAG matches or surpasses $30$-step constant-CFG in ImageReward and CLIPScore), robustness to hyperparameter choices, and superior stability compared with alternative guidance schedules (e.g., CFG-Zero*, Guidance Scheduler, B-CFG) in low-step regimes. The exponential decay schedule consistently outperforms linear, quadratic, piecewise, and sigmoid baselines (Zhu et al., 5 Aug 2025).

4. Theoretical Properties and Guarantees

The regularization RatCG method for inverse problems, when paired with the discrepancy principle and sufficiently large regularization parameters, is proven to achieve optimal-order convergence rates under standard Hölder-type source conditions. The rational Krylov augmentation is central to attaining these rates and faster practical convergence, as it embeds richer spectral filtering via Tikhonov solutions without reducing stability. The method is stable under parameter variation, does not rely on data-driven re-selection of regularization strengths, and allows for efficient inner iteration even when direct inversion is infeasible.

The R-CGR method in sequence analysis possesses provable correctness with respect to lossless recovery: at every encoding step, sufficient geometric and symbolic information is stored to guarantee perfect reconstruction. Optional continued-fraction approximations for denominator reduction, while lossy with respect to real-valued geometry, are lossless for sequence encoding as every approximation step is stored in metadata. Thus, the representation realizes a bijection between sequences and geometric path traces.

The RAAG guidance schedule is derived from a theoretical analysis of dynamical instability caused by high initial RATIO values. The schedule's exponential suppression of guidance in high-RATIO regimes is both theoretically motivated (eliminating unstable exponential error modes) and empirically validated in fast sampling settings.

5. Practical Implementation and Parameter Selection

For inverse problems, the practical implementation of RatCG alternates Tikhonov and CG steps within a mixed rational Krylov space, using a geometric or otherwise non-degenerating sequence of $\alpha_k$ values (e.g., $\alpha_k = \alpha_1\rho^{k-1}$, $\rho>1$). The method does not require online tuning of $\alpha_k$; as long as the lower bound $\alpha_k \geq c_0>0$ is maintained, regularization is effective.

In Explicit Path CGR, the denominator $q = 2^{\lceil \log_2(4k) \rceil}$ is chosen to ensure rational corners for all symbols. Denominator growth is managed either by algorithmic reduction (gcd at each step) or by continued-fraction compression up to a threshold $P$.

For RAAG, the core hyperparameters are $w_{\max}$ and $\alpha$, set once by grid or greedy search. The schedule and update rule are fully explicit and add negligible computational cost compared to the model evaluation.

A summary table of RatCG instantiations across domains:

Domain	Core Principle	Key Application/Guarantee
Linear inverse problems	Rational Krylov space	Optimal-order regularization with discrepancy stop
Sequence encoding (R-CGR)	Rational CGR paths	Lossless/invertible geometric representation
Generative modeling (RAAG)	RATIO-adaptive guidance	Stable, fast classifier-free sampling

6. Empirical Findings and Comparative Performance

For linear inverse problems, direct empirical studies are not included in the principal theoretical work (Kindermann, 15 Jan 2026), but earlier studies indicate that RatCG achieves faster convergence and lower error on standard test suites compared to classical CGNE or aggregation methods.

In sequence classification, R-CGR enables both visual interpretability and competitive performance: VGG16 on R-CGR images obtains 79.07% accuracy, ResNet50+logistic regression 76.50%, compared to lower or similar results for Spike2CGR (75.75%), ProtT5 (73.48%), SeqVec (78.25%), and ESM2 (74.51%). The ability to overlay learned filters on the exact geometric path enhances interpretability.

For flow-based generation, RAAG provides up to $3\times$ acceleration in image generation and $2\times$ in video generation, matching or exceeding semantic fidelity compared to much longer standard CFG runs. The method exhibits broad robustness to sequence length, guidance strength, and ODE stepper, with superior ablation performance relative to alternative adaptive guidance strategies.

7. Significance and Interpretative Remarks

The shared “RatCG” acronym belies the fundamental diversity of the methods it denotes. The unifying theme is the explicit exploitation of rational structures—whether in basis construction, sequence mapping, or adaptive control. Each RatCG instantiation addresses limitations of prevailing techniques (Krylov convergence and regularization instability, geometric non-invertibility, guidance-induced sampling instability) through theoretically grounded, practical algorithms with demonstrable performance benefits. The continued mathematical development and cross-pollination of rational Krylov ideas, adaptive scheduling, and sequence encoding strategies remain active avenues of research in their respective domains (Kindermann, 15 Jan 2026, Ali, 22 Sep 2025, Zhu et al., 5 Aug 2025).