Recursive Flow (RC-Flow): Methods & Applications

• RC-Flow is a family of recursive, flow-based iterative methods that improve stability and efficiency in solving inverse problems and generative tasks.
• It employs recursive operator composition and real-data augmentation to ensure convergence and mitigate model collapse across various applications.
• RC-Flow has demonstrated state-of-the-art performance in image generation, variational inequality solving, and MIMO channel estimation with provable fixed-point guarantees.

Recursive Flow (RC-Flow) denotes a family of recursive, flow-based iterative procedures for solving inverse problems, learning generative models, or enforcing constraints in dynamical systems. RC-Flow frameworks are characterized by recursive self-training or recursive operator composition, often employing learned or analytically constructed flow fields to map between distributions or to update estimands. Empirically and theoretically, RC-Flow methods span applications from deep generative modeling to monotone variational inequalities and under-determined signal reconstruction, yielding provable stability and improved computational efficiency through the design of closed-loop fixed-point systems.

1. Core Frameworks and Mathematical Formulation

RC-Flow emerges in several problem domains, sharing key algorithmic and mathematical structures:

• Generative Modeling (Rectified Flow RC-Flow): Let $p_0(z)=\mathcal{N}(0,I)$ denote a base distribution (e.g., noise) and $p_1(x)$ the target data distribution in $\mathbb{R}^d$. One trains a time-dependent vector field $v_\theta(t,x)$ so that the ODE

$$\frac{d\phi_x(t)}{dt} = v_\theta(t, \phi_x(t)), \quad \phi_x(0)=z$$

transports $z\sim p_0$ towards $\phi_x(1)\sim p_1$. RC-Flow, also termed “Reflow,” recursively re-trains $v_\theta$ on synthetic samples generated by previous flow fields, iteratively refining the learned transport map (Zhu et al., 2024).

• Monotone Variational Inequalities (Recursive Safe Monotone Flow): Given $\mathrm{VI}(F,C)$ with $F: \mathbb{R}^n \to \mathbb{R}^n$ monotone and feasible set $C = \{x \mid g(x) \leq 0, h(x) = 0\}$, RC-Flow constructs a fast-slow dynamical system:

$$\dot{x} = \bar{F}(x,u,v),\quad \epsilon \dot{u} = \max\{ -\beta u, \nabla g(x) \bar{F}(x,u,v) + \alpha g(x) \},\quad \epsilon \dot{v} = \nabla h(x) \bar{F}(x, u, v) + \alpha h(x),$$

where $$\bar{F}(x,u,v) = -F(x) - \nabla g(x)^T u - \nabla h(x)^T v$$, and $(u,v)$ are fast Lagrange-multiplier states. The system contracts towards primal-dual KKT points (Allibhoy et al., 2023).

• MIMO Channel Estimation: RC-Flow composes a pre-trained conditional flow-matching denoiser, a data-fidelity proximal projection, and an anchored trajectory rectification step in a fixed-point iteration:
  • Denoising: $\mathbf{V}=\mathrm{Model}_\theta(\mathbf{H},t)$, update $\widetilde{\mathbf{H}} = \mathbf{H} - t\mathbf{V}$.
  • Proximal projection: minimizes $$||\mathbf{Y} - \mathbf{H}\mathbf{P}||_F^2 / (2\sigma^2) + ||\mathbf{H} - \widetilde{\mathbf{H}}||_F^2/(2w)$$.
  • Rectification: interpolates between anchor and projected estimates.
  • The process is encoded as a composite operator $\mathcal{T}$ with provable fixed-point existence and contraction properties (Jiang et al., 22 Jan 2026).

The recursive processes allow RC-Flow to adapt, refine, or enforce problem structure over multiple iterations or time-scales.

2. Algorithmic Details and Implementation

RC-Flow implementations are highly structured, with problem-dependent operator design:

• Rectified Flow Models (Reflow):
  • At iteration $k$, synthetic pairs $(z_i^{(k)}, x_i^{(k)})$ are generated by forward flow.
  • Each new $v_{\theta_{k+1}}$ is trained to regress these synthetic couplings using a conditional flow-matching loss:

  $$L_{RF}(\theta) = \mathbb{E}_{t\sim U[0,1], (z,x)\sim p^{(k)}} \| v_\theta(t, t z + (1-t) x) - (x-z) \|_2^2.$$ - Advanced variants mix real and synthetic data, regenerate coupling data periodically, and incorporate controlled stochasticity to prevent model collapse (Zhu et al., 2024).

• Recursive Safe Monotone Flow:

  • Employs singular perturbation: a slow state (primal variable) and a fast state (Lagrange multipliers).
  • No online QP is required, as the fast ODEs asymptotically solve the required QP controller.
  • Careful choices for the separation parameter $\epsilon$, gain parameters $\alpha, \beta$, and system conditioning are essential for contracting dynamics and invariance to feasible sets (Allibhoy et al., 2023).
• MIMO Channel Estimation:
  • Alternates model-based denoising and proximal steps, with outer “serial restarts” to update anchor points for trajectory rectification.
  • Step sizes and anchor interpolation weights are adaptively scheduled according to problem SNR and desired speed/accuracy trade-off.
  • Efficient closed-form projections exploit spectral decompositions for computational efficiency (Jiang et al., 22 Jan 2026).

3. Theoretical Properties: Stability, Convergence, and Model Collapse

RC-Flow methods exhibit rigorous theoretical properties:

• Fixed-Point Existence and Contraction (MIMO RC-Flow):
  • Under boundedness and spectral contraction of the constituent operators, the composite RC-Flow operator $\mathcal{T}$ admits a fixed point by Brouwer’s theorem.
  • The total Jacobian norm of the recursive map satisfies $\|\mathbf{J}_\mathcal{T}\|_* < 1$, implying linear (geometric) convergence to the unique estimator in the basin of contraction (Jiang et al., 22 Jan 2026).
• Singular Perturbation and KKT Stability (Monotone VI RC-Flow):
  • The separation of time scales ensures fast convergence of multipliers to quasi-steady-state, embedding the precise KKT conditions in the slow flow dynamics.
  • Local exponential stability of the KKT point is proven under strong monotonicity and polyhedral constraints, with global attractivity when flow gain and constraint regularity conditions are satisfied (Allibhoy et al., 2023).
• Model Collapse and Its Avoidance (Generative RC-Flow):
  • Purified synthetic training in recursive reflow induces exponential decay in the learned map's rank and spectral norm, culminating in model collapse: $v_{\theta_j}(t,x)\rightarrow 0$ and degenerate, constant outputs.
  • Integrating real data at each iteration ensures that the data covariance spectrum remains bounded from below, provably preventing collapse. Formal lower bounds are established by Weyl’s inequality on the spectrum of the mixed covariance (Zhu et al., 2024).

4. Practical Applications and Empirical Performance

RC-Flow frameworks are validated in generative modeling, numerical optimization, and signal processing:

• Image Generation (Rectified Flow RC-Flow):
  • In standard benchmarks (e.g., CIFAR-10), baseline rectified flow collapses after repeated self-training, with FID $\gg 100$ in 1–2-step sampling regimes.
  • Real-data augmented reflow (RCA, OCAR) maintains FID scores as low as 5.6, reduces the number of NFEs required for sharp, colorful sample generation, and preserves principal data components over iterations (Zhu et al., 2024).
• Variational Inequality Solvers:
  • RC-Flow produces “anytime” primal-dual trajectories that continuously satisfy or nearly satisfy safety constraints, converging exponentially to KKT points. Practical invariance means state trajectories do not exit expanded feasible sets even with nonzero $\epsilon$ (Allibhoy et al., 2023).
• MIMO Channel Estimation:
  • RC-Flow achieves state-of-the-art NMSE in low-SNR regimes, outperforming both score-based generative baselines and traditional methods (e.g., LMMSE, L-DAMP). For a $16\times 64$ system at SNR 0 dB, RC-Flow delivers $-12.7$ dB NMSE in 1 ms/sample (A100 GPU), surpassing score-based generators by two orders of magnitude in latency and up to 2.7 dB in NMSE (Jiang et al., 22 Jan 2026).

Method	Params (M)	Latency (ms)	NMSE@0dB
LMMSE	-	0.1	–8
L-DAMP	1.2	0.2	–9
Score-based	4.0	1000	–10
RC-Flow	3.9	1.0	–12.7

5. Collapse Mitigation via Real-Data Augmentation

A recurring phenomenon in recursive self-training is model collapse—the exponential degradation of the learned mapping due to shrinking spectral support of synthetic training data. Empirical and theoretical analysis confirms:

• For linear denoising autoencoders and rectified flows, the spectral norm of the learned transformation vanishes exponentially without real-data intervention.
• Injecting a nonzero fraction of real samples at each iteration sustains the eigenvalue spectrum, with lower bounds ensuring non-collapsing mappings. The severity of collapse is modulated by the fraction and effective covariance of real data incorporated.
• Specific algorithms such as RCA Reflow, OCAR, and OCAR-S, operationalize this principle using reverse trajectory sampling and stochastically-injected real-data pairs, achieving stable training and high downstream task performance (Zhu et al., 2024).

6. Relations Across Methodological Domains

While RC-Flow was introduced in distinct domains, core principles unify the procedures:

• Recursive operator composition—alternating flow priors, projections, or controller updates—underpins both generative and optimization-focused RC-Flow frameworks.
• Fixed-point and contraction theory govern the stability and convergence rates of both classes.
• Real-data augmentation strategies for preventing collapse in generative RC-Flow have conceptual parallels in the enforcement of practical safety in constraint satisfaction for optimization RC-Flow.

The term “Recursive Flow” therefore encompasses algorithm families that leverage recursive applications of flow-based operators—either learned or analytically defined—to achieve superior sample, estimator, or trajectory quality with formal convergence guarantees and robust empirical performance (Zhu et al., 2024, Allibhoy et al., 2023, Jiang et al., 22 Jan 2026).