Signal Lasso for Binary Network Inference
- Signal Lasso is a convex method that adds dual ℓ1 penalties to shrink coefficients toward 0 or 1 for inferring binary network topologies.
- It adapts classical Lasso regression to target unweighted adjacency matrices, improving accuracy and reducing mean-squared errors in network reconstruction.
- Adaptive Signal Lasso minimizes tuning complexity by using adaptive weights to force coefficients exactly to 0 or 1, eliminating ambiguous classifications.
Searching arXiv for recent and foundational papers on Signal Lasso. Signal Lasso is a convex, double-shrinkage regularization method for inferring binary adjacency structures from noisy, finite data in network reconstruction problems. It was introduced for settings in which the unknown connection parameters are signals taking values $0$ or $1$, so that the absence or existence of a link is itself the target of inference rather than a continuous edge weight. In that setting, classical Lasso and compressed sensing methods are effective at shrinking coefficients toward zero, but they do not force nonzero coefficients to one; Signal Lasso adds a second -type term that attracts coefficients toward $1$, thereby adapting sparse regression machinery to unweighted topology inference (Shi et al., 2021).
1. Definition and problem class
Signal Lasso arose from the network reconstruction problem: inferring an underlying topology from observable time series or experimental data. The motivating regime is that of real networks with sparsely connected topologies and binary links, where the relevant parameter is an element of an adjacency matrix that takes value $0$ or $1$. In the linearized regression formulation used in the method, the goal is to reconstruct a vectorized adjacency representation from observed data and a design matrix , via
This formulation was used for evolutionary-game and synchronization dynamics, as well as for empirical network data (Shi et al., 2021).
The conceptual novelty lies in treating network reconstruction as a sparse estimation problem with a discrete target space. Classical Lasso and compressed sensing methods fit naturally when the aim is to identify a sparse continuous parameter vector, but unweighted topologies are elusive to those methods because they do not explicitly encode attraction to $1$0. Signal Lasso was proposed specifically to recover such binary structure, and the original study reported improved accuracy and mean square errors relative to classical approaches in synthetic and empirical network settings (Shi et al., 2021).
A common source of terminological confusion is that the phrase “signal lasso” can be conflated with broader Lasso-based signal processing. In the literature surveyed here, however, “Signal Lasso” denotes a specific network-inference penalty design for binary topology recovery rather than a generic use of Lasso on signals.
2. Penalized formulation and statistical mechanism
The original Signal Lasso estimator augments the least-squares criterion with two absolute-value penalties: $1$1 The first term is the data-fidelity term. The penalty $1$2 is the usual Lasso component that shrinks coefficients toward zero, while $1$3 is the distinctive Signal Lasso component that shrinks coefficients toward one (Shi et al., 2022).
An equivalent parameterization used in the original paper writes the penalty as
$1$4
so that $1$5 controls the balance between the attraction to $1$6 and the attraction to $1$7 (Shi et al., 2021).
Geometrically, the original paper described the Signal Lasso constraint as polygonal, with corners at both $1$8 and $1$9. This differs from the standard Lasso geometry, which is centered only at zero. In the orthogonal design case, the method admits a coordinatewise analytic solution that explicitly partitions the least-squares estimate into regions that are shrunk toward 0 or 1, thereby making the binary intent of the estimator transparent (Shi et al., 2021).
The essential statistical mechanism is therefore dual shrinkage. Classical Lasso reduces small coefficients and leaves the nonzero part continuous; Signal Lasso attempts to discriminate both non-links and links by penalizing distance to the two admissible states. This makes the method particularly natural when the underlying graph is assumed unweighted.
3. Computation and theoretical structure
Because both penalty terms are convex, the original Signal Lasso optimization problem is convex. The original paper proposed a coordinate descent algorithm with a specialized thresholding update that incorporates the two-direction shrinkage structure (Shi et al., 2021). In the notation summarized in the later adaptive paper, the coordinate update takes the form
2
where 3 is a piecewise thresholding operator and the coordinates are updated cyclically (Shi et al., 2022).
The theoretical appeal of the original construction is that it remains within convex optimization while modifying the target geometry from “sparse continuous coefficients” to “binary coefficients.” That is why the method could be developed within a familiar regularized regression framework and still be analyzed using standard convex-optimization tools (Shi et al., 2021).
The principal limitation of the original Signal Lasso is equally structural. Although it encourages coefficients to be close to 4 or 5, it does not guarantee that all fitted coefficients fall exactly at those values. The later adaptive paper describes this as an “unclassified” portion, specifically estimates in 6 that are neither close to 7 nor close to 8 and therefore cannot be properly classified. A second limitation is the need to tune two regularization parameters, 9 and $1$0, which increases computational cost and complicates application (Shi et al., 2022).
4. Adaptive Signal Lasso
Adaptive Signal Lasso (AS-lasso) was proposed to remove the unclassified region and reduce the burden of parameter tuning. Its objective function is
$1$1
with adaptive weights $1$2 and $1$3 chosen as functions of an initial estimator $1$4. The recommended choice is $1$5, $1$6 (Shi et al., 2022).
To simplify implementation, the method sets $1$7, $1$8, yielding
$1$9
Under this reparameterization, only one tuning parameter, $0$0, must be selected, usually in $0$1 (Shi et al., 2022).
For large enough $0$2, and with a given cutoff $0$3, the adaptive method has the complete-classification property
$0$4
so that all estimates are shrunk either to $0$5 or $0$6. The paper presents this as eliminating the unclassified portion in network reconstruction (Shi et al., 2022).
The adaptive method was presented as having three advantages: it can effectively uncover the network topology with high accuracy and completely shrink the signal parameter to either $0$7 or $0$8; it performs well in scenarios of both sparse and dense signals and is robust to noise contamination; and it needs only one tuning parameter versus two in Signal Lasso, which greatly reduces the computational cost and is easy to apply (Shi et al., 2022).
| Variant | Penalty structure | Tuning burden |
|---|---|---|
| Signal Lasso | $0$9 | Two parameters |
| Adaptive Signal Lasso | Weighted version with $1$0 | One parameter |
| Non-convex Signal Lasso | Product or minimum penalties | Large enough $1$1 |
5. Empirical evaluation and application domains
The original Signal Lasso paper illustrated the method on evolutionary-game and synchronization dynamics in several synthetic and empirical networks, including Erdős–Rényi, Watts–Strogatz, and Barabási–Albert graphs, and reported higher accuracy, lower mean-squared errors, higher AUROC/AUPR, and better detection rates such as SREL and SRNL than Lasso and compressed sensing, even with small sample sizes (Shi et al., 2021).
The adaptive paper broadened the empirical scope to linear regression, evolutionary games, Kuramoto models, a human behavioral experiment, and a world trade web. In the simulation studies reported there, AS-lasso consistently achieved the lowest misclassification rates measured by MCC and MCCa and almost zero unclassified rates measured by UCR, across sparse and dense signal regimes and under both low and high noise. In evolutionary games and Kuramoto synchronization, it outperformed alternative methods in classifying network structure even with noisy or limited data, and its UCR was described as near zero (Shi et al., 2022).
The real-data illustrations are especially important for understanding the intended application range. In the human behavioral experiment, AS-lasso achieved higher MCC and MCCa and lower UCR than Signal Lasso, especially when data amount was small, while classifying all links. In the world trade web, which the paper describes as a very dense real economic network, AS-lasso detected essentially all true connections with high accuracy, with SREL and SRNL approximately $1$2 and UCR approximately $1$3. The reconstructed networks also reflected empirical properties such as degree, clustering, and assortativity (Shi et al., 2022).
These results support a specific interpretation of the method’s niche. Signal Lasso and its adaptive successor are not merely sparse estimators for network regression; they are binary-topology classifiers designed for settings in which complete assignment of coefficients to “edge” or “non-edge” is itself a central performance criterion.
6. Extensions, limitations, and related terminology
A later extension introduced non-convex Signal Lasso penalties intended to remove the need for parameter tuning and to force all coefficients to the binary extremes. Two penalties were proposed: $1$4 and
$1$5
These were denoted “SL_prod” and “SL_min.” The paper states that, with sufficiently large $1$6, both methods guarantee that all $1$7 converge to exactly $1$8 or $1$9, and that they are faster because they avoid the tuning procedures required by convex Signal Lasso variants. The corresponding disadvantage is non-convexity: coordinate descent is used in practice, but multiple local minima may arise (Shi et al., 30 Aug 2025).
The development from Signal Lasso to Adaptive Signal Lasso and then to non-convex variants clarifies the central technical tension in this line of work. Convex penalties are computationally convenient and theoretically tractable, but they may leave ambiguous coefficients or require parameter search; non-convex penalties more aggressively enforce binary outputs, but they give up global convexity. This suggests a trade-off between optimization guarantees and direct binary classification, although the precise boundary of that trade-off remains method-dependent.
The method also has explicit scope conditions. The original and adaptive papers emphasize binary or unweighted topology recovery. For weighted or more general networks, extension or adjustment is needed; for very large 0, dimensionality-reduction methods may be needed in advance; and, as with other network reconstruction methods, performance degrades with very high noise or missing data, although the reported performance remains robust relative to alternatives (Shi et al., 2021).
Finally, Signal Lasso in this network-reconstruction sense should be distinguished from other Lasso-based signal methodologies that solve different inverse problems. Examples include a two-stage LASSO-ADMM signal detection algorithm for large scale MIMO (Elgabli et al., 2018), the LAD fused lasso signal approximator for sparse and blocky signals under heavy-tailed or outlier-prone noise (Gao, 2021), two-dimensional Lasso frequency analysis for dynamic spectra (Kato, 2021), and Bayesian LASSO methods for complex-valued signal recovery with uncertainty quantification (Green et al., 2024). Those methods share the broader 1-regularization lineage, but they do not implement the dual attraction to 2 and 3 that defines Signal Lasso as a method for binary network topology inference.