VariSAC: Optimization & V2X Connectivity Methods

• The paper introduces VariSAC as a unified variance-reduction technique that maintains simultaneous data and coordinate sketches, yielding linear convergence in composite empirical risk minimization.
• It shows how VariSAC reduces to established methods like SAGA and SEGA by tuning sampling probabilities, providing actionable insights for efficient optimization.
• In wireless networking, VariSAC represents a GNN-augmented Soft Actor-Critic framework that optimizes continuous V2X connectivity in RIS-aided, ISAC-enabled environments.

VariSAC denotes more than one unrelated construct in the arXiv literature. In one usage, it is a variance-reduced stochastic approximation method for composite empirical-risk minimization that performs variance reduction simultaneously over data indices and coordinate indices (Hanzely et al., 2019). In another, it is a graph neural network-augmented Soft Actor-Critic framework for assured, time-continuous connectivity in RIS-aided, ISAC-enabled vehicle-to-everything systems (Tang et al., 8 Sep 2025). A separate but similarly spelled method, VSAC, is a RANSAC-type robust estimator for homography and fundamental-matrix estimation and should not be conflated with either VariSAC usage (Ivashechkin et al., 2021).

1. Nomenclature and scope

The term is not monosemous. The supplied literature uses “VariSAC” in two technically distinct settings, and it also contains the closely related but different acronym “VSAC.”

Name	Domain	Core characterization
VariSAC	Optimization	Variance-reduced stochastic approximation over data and coordinates
VariSAC	Wireless networking	V2X assured connectivity in RIS-aided ISAC via GNN-augmented reinforcement learning
VSAC	Geometric vision	Efficient and accurate estimator for $H$ and $F$

This terminological overlap is a common source of confusion. A plausible implication is that any technical discussion of VariSAC must specify the problem class immediately: convex composite optimization in the sense of regularized empirical risk minimization, or RIS-assisted V2X control in the sense of continuous-connectivity optimization. The 2021 VSAC paper belongs to neither category; it addresses two-view geometry, introduces independent inliers, adaptive SPRT, and Gaussian-elimination-based minimal solvers, and uses a different acronym without the “ri” component (Ivashechkin et al., 2021).

2. VariSAC in composite empirical-risk minimization

In the optimization usage, VariSAC is presented as a special case of a general variance-reduction framework for regularized empirical risk minimization (Hanzely et al., 2019). The problem is

$$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$

The assumptions are explicit. Each $f_j:\mathbb{R}^d\to\mathbb{R}$ is convex and $L_j$-smooth, where $L_j\in\mathbb{R}^{d\times d}$ is positive semidefinite. The regularizer $\psi$ is proper, closed, convex, and admits an easy proximal operator. The average objective $f(x)=\tfrac1n\sum_j f_j(x)$ is $\sigma$-quasi-strongly-convex. The method uses the Jacobian matrix of partial gradients,

$$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$

so that $F$0, with $F$1 the all-ones vector.

The defining structural feature is the simultaneous maintenance of two sketches at every iteration: a “data” sketch $F$2 satisfying $F$3, and a “coordinate” sketch $F$4 satisfying $F$5. In the concrete sampling construction, a random subset $F$6 with $F$7 induces

$F$8

while a random coordinate subset $F$9 with $$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$0 induces

$$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$1

This gives unbiased projectors onto the chosen indices. The algorithm initializes $$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$2 and $$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$3, samples $$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$4 and $$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$5, evaluates $$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$6 and $$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$7, updates the Jacobian estimate by

$$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$8

forms the gradient estimator

$$\min_{x\in\mathbb{R}^d} F(x) \quad\text{where}\quad F(x)=\frac1n\sum_{j=1}^n f_j(x)+\psi(x).$$9

and then applies the proximal step

$f_j:\mathbb{R}^d\to\mathbb{R}$0

The paper states that $f_j:\mathbb{R}^d\to\mathbb{R}$1, so the estimator is unbiased. It also states that $f_j:\mathbb{R}^d\to\mathbb{R}$2 as $f_j:\mathbb{R}^d\to\mathbb{R}$3 because $f_j:\mathbb{R}^d\to\mathbb{R}$4 and $f_j:\mathbb{R}^d\to\mathbb{R}$5. This places VariSAC within the class of memory-based variance-reduced proximal methods, but with an explicit coupling of data subsampling and coordinate subsampling.

3. Variance reduction, convergence theory, and reductions to known methods

The derivation emphasizes sketch-and-project structure (Hanzely et al., 2019). The Jacobian update is written as

$f_j:\mathbb{R}^d\to\mathbb{R}$6

The gradient-estimation error is decomposed into an “old memory bias” term and a correction term, and the mean-square error is bounded by

$f_j:\mathbb{R}^d\to\mathbb{R}$7

The main theorem gives linear convergence under the stated assumptions, together with additional structural conditions on a positive-definite weight operator $f_j:\mathbb{R}^d\to\mathbb{R}$8 commuting with both sketches and on the initial Jacobian guess $f_j:\mathbb{R}^d\to\mathbb{R}$9. With the Lyapunov function

$L_j$0

the theorem concludes that

$L_j$1

that is, $L_j$2 decays linearly at rate $L_j$3.

A defining theoretical feature is that several previously separate methods arise as special cases. If coordinate sketching is turned off, $L_j$4, VariSAC reduces to a mini-batch SAGA with arbitrary data sampling, with complexity

$L_j$5

where $L_j$6. If data sketching is turned off, $L_j$7, VariSAC reduces to SEGA with arbitrary coordinate sampling, with complexity

$L_j$8

where $L_j$9. In the fully combined case, the stated complexity is

$L_j\in\mathbb{R}^{d\times d}$0

The practical guidance is likewise explicit. Data-sampling probabilities may be chosen as $L_j\in\mathbb{R}^{d\times d}$1 or $L_j\in\mathbb{R}^{d\times d}$2-diagonal to minimize the ESO constant, while coordinate-sampling probabilities may be chosen as $L_j\in\mathbb{R}^{d\times d}$3. The paper further states that with these choices VariSAC “automatically interpolates between SAGA (big $L_j\in\mathbb{R}^{d\times d}$4) and SEGA (big $L_j\in\mathbb{R}^{d\times d}$5) and in all regimes enjoys the best-known linear rates.” This suggests that the method’s main contribution is not merely a new update rule, but a unifying abstraction in which data-side and coordinate-side variance reduction become instances of the same sketch-based mechanism.

4. VariSAC in RIS-aided ISAC-enabled V2X systems

In the networking usage, VariSAC is a framework for assured, time-continuous connectivity in RIS-aided, ISAC-enabled V2X systems (Tang et al., 8 Sep 2025). The stated motivation is that existing methods typically optimize either instantaneous physical-layer reliability or per-packet success in isolation, and do not simultaneously unify the continuous multiplexed semantics of V2I temporal reliability and V2V probabilistic latency-bounded delivery, nor capture the strong spatio-temporal dependencies induced by dynamic topologies, RIS reconfigurations, and heterogeneous node capabilities in urban vehicular environments.

The system model contains one multi-antenna base station at $L_j\in\mathbb{R}^{d\times d}$6, one RIS with $L_j\in\mathbb{R}^{d\times d}$7 passive reflecting elements at $L_j\in\mathbb{R}^{d\times d}$8, and $L_j\in\mathbb{R}^{d\times d}$9 single-antenna vehicles, among which $\psi$0 are designated sensing targets. The spectrum is divided into $\psi$1 V2I subchannels and $\psi$2 sensing subchannels, while $\psi$3 V2V links reuse V2I subchannels through scheduling variables $\psi$4. The RIS phase-shift matrix is

$\psi$5

with $\psi$6 and amplitudes $\psi$7.

The communication model uses composite V2I and V2V channel gains. For V2I,

$\psi$8

and for V2V link $\psi$9,

$f(x)=\tfrac1n\sum_j f_j(x)$0

The corresponding SINRs are

$f(x)=\tfrac1n\sum_j f_j(x)$1

with rates $f(x)=\tfrac1n\sum_j f_j(x)$2 and $f(x)=\tfrac1n\sum_j f_j(x)$3. Sensing is modeled through $f(x)=\tfrac1n\sum_j f_j(x)$4, subject to $f(x)=\tfrac1n\sum_j f_j(x)$5.

The central modeling contribution is the Continuous Connectivity Ratio (CCR). V2I continuity is defined through sliding-window indicators $f(x)=\tfrac1n\sum_j f_j(x)$6 and $f(x)=\tfrac1n\sum_j f_j(x)$7, while V2V connectivity is expressed as the delivery probability

$f(x)=\tfrac1n\sum_j f_j(x)$8

The unified CCR objective is

$f(x)=\tfrac1n\sum_j f_j(x)$9

The associated optimization problem maximizes this quantity over channel assignment $\sigma$0, power $\sigma$1, and RIS configuration $\sigma$2, subject to rate, SNR, channel-assignment, power, and phase-shift constraints. This formulation makes continuous reliability, rather than instantaneous throughput alone, the primary optimization target.

5. GNNRA architecture, SAC control, and optimization variables

The control architecture combines a graph representation of the vehicular environment with entropy-regularized reinforcement learning (Tang et al., 8 Sep 2025). Nodes are $\sigma$3vehicles, BS, RIS$\sigma$4, and edges are $\sigma$5V2I links, V2V links, sensing links$\sigma$6. Heterogeneous state vectors $\sigma$7 are linearly projected to obtain $\sigma$8, after which a two-layer GCN with a residual adapter produces a flattened embedding $\sigma$9.

The first graph-convolution layer is

$$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$0

with normalized adjacency matrix $$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$1 and $$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$2. A bottleneck residual adapter is inserted to mitigate oversmoothing, and the second GCN layer is then applied to $$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$3. The final representation is flattened and fed to a Soft Actor-Critic agent.

The resource-allocation problem is cast as an MDP $$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$4. The state at slot $$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$5 is

$$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$6

Each vehicle state includes spatial coordinates, V2V channel and interference terms, remaining load, a sensing indicator, and a sensing channel if the indicator is active. The base-station state is $$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$7, and the RIS state is $$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$8. The action is

$$J(x)=\bigl[\nabla f_1(x)\;\;\nabla f_2(x)\;\dots\;\nabla f_n(x)\bigr]\in\mathbb{R}^{d\times n},$$9

so the agent jointly selects channel allocation, V2V power, and RIS configuration.

The reward function is aligned with CCR:

$F$00

The SAC objective is the standard maximum-entropy objective

$F$01

with explicit update rules for the temperature $F$02, the critics, the actor, and the target critics. The paper’s algorithm pseudocode initializes GNNRA parameters, SAC networks, target critics, temperature $F$03, and replay buffer $F$04; encodes each raw state with GNNRA; samples actions from the actor; stores transitions; updates the critics by minimizing $F$05; updates the actor via the stochastic policy gradient; updates $F$06; and performs target-network soft updates.

The stated design trade-offs are also integral to the method. GNNRA is described as critical for capturing heterogeneous, dynamic spatial relations among vehicles, BS, and RIS; the entropy term of Soft Actor-Critic is described as promoting robust exploration in the high-dimensional action space; the sliding-window length $F$07 trades off strictness of continuous V2I service against resource-allocation complexity; increasing V2I power improves continuous sensing and communication but intensifies V2V interference; and residual adapters mitigate GCN oversmoothing.

6. Empirical behavior, baselines, and terminological distinctions

The two VariSAC usages are evaluated in entirely different empirical regimes. The optimization VariSAC is analyzed through convergence theory and complexity reductions to SAGA- and SEGA-type limits (Hanzely et al., 2019). The V2X VariSAC is evaluated in simulation and on real-world trajectories (Tang et al., 8 Sep 2025). In its reported setup, the environment is an urban road grid of size $F$08 with 12 vehicles, $F$09 sensing targets, one BS at $F$10, and one RIS at $F$11. Vehicle speeds lie in $F$12, the V2V payload is $F$13 bits, the RIS has $F$14 elements and $F$15 phase levels, V2I and sensing power is $F$16, V2V power ranges over $F$17, the noise powers are $F$18, and the bandwidth is $F$19. The GNNRA module consists of three NodeEncoder networks and a two-layer GCN with one residual adapter, while the SAC module has an actor, two critics, and two targets, each with 5 fully-connected layers and 3 hidden layers of 512 units. Training uses 1,800 episodes, $F$20, Adam optimizer learning rate $F$21, replay buffer size $F$22, batch size $F$23, and target update $F$24.

The baseline set is also specific: VariSAC-GNNRA-DDPG replaces SAC with DDPG; VariSAC-SAC omits GNNRA and feeds the raw state to SAC; VariSAC-RandomRIS uses random RIS phases; and VariSAC-Greedy performs myopic maximization of immediate reward. The reported metrics are $F$25, $F$26, and total CCR. The reported performance claims include the following: VariSAC achieves the fastest and most stable reward growth across learning rates $F$27; when varying the V2V payload $F$28, total CCR gains over VariSAC-DDPG, RandomRIS, SAC-only, and Greedy are $F$29, $F$30, $F$31, and $F$32, respectively; under V2I transmit-power sweeps, total CCR outperforms baselines by $F$33; with longer window length $F$34, VariSAC still leads by up to $F$35 in V2I CCR; and on Didi GAIA trajectories, V2V CCR gains over the best baselines are $F$36.

A frequent misconception is to treat these results as informing the optimization-theoretic VariSAC of regularized empirical risk minimization. They do not. The two methods share a name but not a problem statement, objective, state representation, convergence analysis, or evaluation protocol. A second misconception is to identify VariSAC with the vision method VSAC. That is likewise incorrect: VSAC is a RANSAC-type estimator for homography and fundamental-matrix estimation, with independent inliers, adaptive SPRT, lightweight local optimization, Gaussian-elimination solvers, and DEGENSAC$F$37, and is experimentally reported to run in $F$38 ms on a CPU across EVD, HPatches, PhotoTourism, and Kusvod2 (Ivashechkin et al., 2021). This suggests that “VariSAC” is best understood not as a single algorithmic lineage, but as a reused acronym whose meaning is determined entirely by domain context.