Two-Timescale Alternating MAP Framework

• The paper introduces a two-timescale alternating MAP framework that decouples slow sparse signal estimation from fast grid refinement to enhance super-resolution in compressive sensing and XL-MIMO.
• It employs a tanh-based variational Bayesian inference for sparse signal estimation, effectively isolating support detection from continuous parameter refinement.
• The fast grid update uses BFGS optimization with Armijo line search to fine-tune active parameters, achieving super-resolution with reduced computational complexity.

Searching arXiv for the cited papers to ground the article in the relevant literature. Two-timescale alternating maximum a posteriori (MAP) denotes an alternating optimization pattern in which different subsets of variables are updated on distinct temporal or algorithmic scales. In the literature considered here, the phrase is used explicitly for robust super-resolution compressive sensing (SR-CS), where a slow-timescale sparse signal estimation module is coupled to a fast-timescale grid refinement module (Zhou et al., 9 Aug 2025). Closely related alternating or two-timescale structures also appear in joint visibility region (VR) detection and channel estimation for XL-MIMO, where fast inner updates of latent variables are combined with slower outer updates of a dynamic polar-domain grid (Xu et al., 2024), and in a proximal bundle method for MAP inference, where fast inner block-coordinate Frank-Wolfe iterations improve a proximal subproblem while slower outer iterations refresh the proximal center (Swoboda et al., 2018).

1. Definition and alternating structure

In robust SR-CS, the observation model is

$$\mathbf{y}=\mathbf{A}(\boldsymbol{\theta})\mathbf{x}+\mathbf{w},$$

where $\mathbf{y}\in\mathbb{C}^{M}$ is the measurement vector, $\mathbf{x}\in\mathbb{C}^{N}$ is sparse, $$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$ is a sensing or dictionary matrix determined by a grid parameter vector $\boldsymbol{\theta}$, and $$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$ is Gaussian noise. The proposed two-timescale alternating MAP framework separates the problem into a slow-timescale sparse signal estimation module and a fast-timescale super-resolution grid update module (Zhou et al., 9 Aug 2025).

At the slow timescale, the sparse signal estimation module infers the posterior of the sparse coefficients, support variables, precision variables, and noise precision on a dense grid. This stage identifies the active support set $\hat{\mathcal S}$. At the fast timescale, the grid update module refines only the active grid parameters $\boldsymbol{\theta}_{\hat{\mathcal S}}$ and the corresponding gains $\mathbf{x}_{\hat{\mathcal S}}$. The outer iteration is explicitly described as: run sparse signal estimation with the current grid, extract $\hat{\mathbf{x}}$, $\mathbf{y}\in\mathbb{C}^{M}$0, and $\mathbf{y}\in\mathbb{C}^{M}$1, and then run super-resolution grid update on the active set for a few inner iterations (Zhou et al., 9 Aug 2025).

This organization is motivated by the off-grid and densely packed regimes. A dense grid improves representational flexibility but also makes dictionary columns highly correlated. The two-timescale design therefore uses a slower probabilistic stage to identify which atoms are relevant and a faster low-dimensional stage to move those atoms off-grid toward the true parameter values. A plausible implication is that the framework treats support discovery and continuous parameter refinement as coupled but nonidentical optimization tasks.

2. Bayesian and MAP formulation in super-resolution compressive sensing

The probabilistic model is defined by

$\mathbf{y}\in\mathbb{C}^{M}$2

with Gaussian likelihood

$\mathbf{y}\in\mathbb{C}^{M}$3

and Gamma prior

$\mathbf{y}\in\mathbb{C}^{M}$4

The sparse prior is a three-layer Bernoulli-Gamma-Tanh model,

$\mathbf{y}\in\mathbb{C}^{M}$5

with

$\mathbf{y}\in\mathbb{C}^{M}$6

$\mathbf{y}\in\mathbb{C}^{M}$7

and the coefficient prior

$\mathbf{y}\in\mathbb{C}^{M}$8

The MAP target is essentially

$\mathbf{y}\in\mathbb{C}^{M}$9

with no prior assumed on $\mathbf{x}\in\mathbb{C}^{N}$0, so the grid update is effectively ML/MAP without prior regularization (Zhou et al., 9 Aug 2025).

The role of the tanh prior is central. It is described as a smooth approximation to an $\mathbf{x}\in\mathbb{C}^{N}$1-like penalty. Because $\mathbf{x}\in\mathbb{C}^{N}$2 rises quickly near zero and saturates for large arguments, it strongly penalizes small coefficients while barely affecting large ones. The paper contrasts this with Gaussian or Laplace priors, which correspond roughly to $\mathbf{x}\in\mathbb{C}^{N}$3 or $\mathbf{x}\in\mathbb{C}^{N}$4 penalties and are less aggressive in suppressing spurious neighboring activations under dense, highly correlated dictionaries (Zhou et al., 9 Aug 2025).

The same paper frames this formulation against standard grid-based compressed sensing. If the grid is too coarse, basis mismatch, energy leakage, and loss of resolution for closely spaced components appear. If the grid is made very dense, ordinary OMP or Gaussian-prior sparse Bayesian learning can become unreliable. The alternating MAP construction is therefore designed specifically for off-grid and closely spaced components (Zhou et al., 9 Aug 2025).

3. Slow-timescale sparse signal estimation via tanh-VBI

The sparse signal estimation module uses mean-field variational Bayesian inference with factorized approximation,

$\mathbf{x}\in\mathbb{C}^{N}$5

and minimizes the Kullback–Leibler divergence. For a single factor,

$\mathbf{x}\in\mathbb{C}^{N}$6

Because the tanh term makes the posterior of $\mathbf{x}\in\mathbb{C}^{N}$7 non-Gaussian, the method uses successive linear approximation around the previous posterior mean $\mathbf{x}\in\mathbb{C}^{N}$8,

$\mathbf{x}\in\mathbb{C}^{N}$9

with

$$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$0

This yields

$$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$1

where

$$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$2

Hence

$$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$3

with

$$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$4

The paper emphasizes that this adaptive covariance differs from ordinary variational Bayesian inference because it changes with the current posterior mean through $$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$5 (Zhou et al., 9 Aug 2025).

The remaining posteriors have closed form. The precision variables follow

$$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$6

the support posterior is Bernoulli,

$$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$7

and the noise precision update is

$$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$8

A diagonal approximation is suggested for the trace term in the $$\mathbf{A}(\boldsymbol{\theta})\in\mathbb{C}^{M\times N}$$9 update for efficiency. The final MAP estimates are taken from these posteriors, especially $\boldsymbol{\theta}$0 and $\boldsymbol{\theta}$1, and the support $\boldsymbol{\theta}$2 is extracted from $\boldsymbol{\theta}$3 (Zhou et al., 9 Aug 2025).

This slow-timescale stage is the support-selection engine of the framework. The paper states that tanh-VBI yields much sparser estimates and much better support detection than the original variational Bayesian inference based on Gaussian-like sparsity. This is directly tied to the subsequent effectiveness of low-dimensional grid refinement (Zhou et al., 9 Aug 2025).

4. Fast-timescale grid refinement and super-resolution effect

Once the active support set $\boldsymbol{\theta}$4 has been identified, the super-resolution grid update refines only the active variables. Given $\boldsymbol{\theta}$5 and $\boldsymbol{\theta}$6, the gains are updated by minimizing

$\boldsymbol{\theta}$7

which yields the closed-form LMMSE solution

$\boldsymbol{\theta}$8

Given $\boldsymbol{\theta}$9, the grid update solves

$$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$0

which is nonconvex. The paper uses BFGS, a quasi-Newton method that approximates the inverse Hessian, together with Armijo backtracking for the step size (Zhou et al., 9 Aug 2025).

The update is written as

$$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$1

while BFGS maintains an approximation $$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$2 and updates it through the standard secant formula. The descent direction is selected adaptively:

$$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$3

Then

$$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$4

with $$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$5 chosen by Armijo backtracking. The paper also notes a variable projection viewpoint: for each candidate $$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$6, the corresponding $$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$7 is recomputed, so the line search is done on the reduced objective (Zhou et al., 9 Aug 2025).

The super-resolution effect is attributed to the fact that only the active support set is refined, so the update is low-dimensional. Frequent BFGS refinement of these few active parameters allows the algorithm to slide the active atoms off-grid toward the true parameter values, achieving super-resolution with much lower cost than full-dictionary optimization. This also clarifies a common misunderstanding: the framework does not refine the entire dense dictionary at each step; it refines the active grid components selected by the slow-timescale inference stage (Zhou et al., 9 Aug 2025).

5. Related alternating MAP and two-timescale formulations

The term “alternating MAP” also appears in XL-MIMO channel estimation with spatial non-stationarity and near-field propagation. There, the joint posterior is built from a likelihood, a hierarchical sparse prior for polar-domain channel coefficients, a Gamma prior on the noise precision, and a 2D Markov prior for VR matrices. Because the full MAP is intractable, the method alternates among three modules: estimate $$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$8 and $$\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\kappa^{-1}\mathbf{I})$$9 given $\hat{\mathcal S}$0 and $\hat{\mathcal S}$1, estimate $\hat{\mathcal S}$2 given $\hat{\mathcal S}$3, $\hat{\mathcal S}$4, and $\hat{\mathcal S}$5, and refine $\hat{\mathcal S}$6 given $\hat{\mathcal S}$7, $\hat{\mathcal S}$8, and $\hat{\mathcal S}$9. The paper explicitly describes this as an alternating or two-timescale structure with fast inner updates of latent variables and slower outer updates of the grid (Xu et al., 2024).

The three XL-MIMO modules are low-complexity inverse-free variational Bayesian inference for channel estimation, structured expectation propagation for VR detection under a 2D Markov prior, and gradient-ascent grid refinement with Armijo step sizes. The outer loop repeats initialization, channel estimation, VR detection, grid update, and convergence check. This is described as nested alternating optimization: inner iterations refine probabilistic latent variables, while outer iterations alternate between channel, VR, and grid updates (Xu et al., 2024).

A related but not identical two-timescale structure appears in MAP inference for discrete structured energies. In that setting, MAP inference is converted into maximizing a concave, nonsmooth Lagrangian dual lower bound, and the solver FWMAP combines a proximal bundle framework with a multi-plane block-coordinate Frank-Wolfe inner solver. The paper states that this makes the method conceptually close to alternating optimization or a two-timescale scheme: fast inner iterations improve a proximal subproblem, while slower outer iterations refresh the proximal center. Its main alternation is therefore between optimizing a proximalized dual and relocating the trust region, rather than between Bayesian latent-variable blocks and continuous grid parameters (Swoboda et al., 2018).

Paper	Application	Alternating structure
"Robust Super-Resolution Compressive Sensing: A Two-timescale Alternating MAP Approach" (Zhou et al., 9 Aug 2025)	SR-CS	slow-timescale sparse signal estimation; fast-timescale grid refinement
"Joint Visibility Region Detection and Channel Estimation for XL-MIMO Systems via Alternating MAP" (Xu et al., 2024)	XL-MIMO	channel estimation, VR detection, and grid update with slower outer grid updates
"MAP inference via Block-Coordinate Frank-Wolfe Algorithm" (Swoboda et al., 2018)	discrete MAP inference	fast inner BCFW steps; slower proximal-center updates

Taken together, these works suggest that “two-timescale alternating MAP” is less a single fixed algorithm than a structural template: one layer stabilizes or sparsifies the latent representation, while another layer updates a lower-dimensional or slower-varying set of global parameters.

6. Complexity, empirical behavior, and interpretive boundaries

For robust SR-CS, the paper does not give a full global proof, but characterizes the method as an alternating MAP/VBI scheme with a monotonic-style variational update in the sparse signal estimation step and quasi-Newton refinement with Armijo line search in the grid update step. The outer loop is reported to converge quickly within only a few iterations. If $\boldsymbol{\theta}_{\hat{\mathcal S}}$0 is the number of outer iterations, $\boldsymbol{\theta}_{\hat{\mathcal S}}$1 the number of inner tanh-VBI iterations, $\boldsymbol{\theta}_{\hat{\mathcal S}}$2 the number of inner grid-update iterations, $\boldsymbol{\theta}_{\hat{\mathcal S}}$3 the dictionary size, and $\boldsymbol{\theta}_{\hat{\mathcal S}}$4, then the total complexity is

$\boldsymbol{\theta}_{\hat{\mathcal S}}$5

with a stated possibility of reducing the sparse signal estimation cost further by inverse-free VBI techniques (Zhou et al., 9 Aug 2025).

The empirical findings in that paper are presented for channel extrapolation in TDD massive MIMO-OFDM. The proposed method is compared with MUSIC-CS, QNOMP, and DMRA, and is reported to achieve lower NMSE for channel extrapolation, lower normalized RMSE for angle and delay estimation, faster convergence than MUSIC-CS, and comparable runtime to QNOMP and DMRA but much better robustness. The paper attributes this to the combination of repeated support identification under a dense grid, stronger sparsity promotion from the tanh prior, and BFGS-based grid optimization using approximate second-order information (Zhou et al., 9 Aug 2025).

For XL-MIMO, the proposed alternating MAP method combines near-field polar-domain modeling, a 2D Markov prior for clustered VRs, inverse-free variational inference, structured EP, and gradient-based off-grid correction. Its per-outer-iteration complexity is

$\boldsymbol{\theta}_{\hat{\mathcal S}}$6

and the paper reports convergence within about 20 outer iterations. It also states that the dynamic grid improves both channel NMSE and VR error rate over fixed-grid and on-grid methods, and that the 2D Markov prior gives an additional gain over an i.i.d. prior, especially in low SNR, many-path settings, and scenarios with pronounced 2D clustered VRs (Xu et al., 2024).

A further interpretive boundary concerns convergence guarantees. In the proximal bundle formulation for discrete MAP inference, the paper notes that exact inner solves would make the whole method equivalent to a convergent proximal point scheme, while the practical algorithm uses inexact inner optimization, refreshes the proximal center every 10 iterations, and does not rigorously enforce shrinking inner error. This does not negate the two-timescale interpretation, but it distinguishes practical alternating schedules from fully conservative schemes (Swoboda et al., 2018).

In this literature, then, two-timescale alternating MAP is best understood as a disciplined decomposition principle for hard MAP or MAP-like estimation problems with off-grid structure, latent support variables, or large structured dual spaces. Its recurring pattern is to reserve the slower layer for globally organizing variables such as support, VRs, or proximal centers, and the faster layer for comparatively cheap local refinement steps that exploit the reduced active set or block structure.