Time-Augmented Lasso Methods

• Time-augmented lasso is a family of methods that embeds temporal structure directly into lasso estimation via lag expansion, coefficient modulation, dependence-aware inference, and continuous-time optimization.
• The approach enhances prediction accuracy and network inference by imposing ordered lag constraints, adaptive penalties, and heteroscedastic reweighting on time series data.
• Applications include rolling prediction, gene regulatory network reconstruction, and dynamic system optimization, illustrating its practical value in time-dependent data analysis.

Time-augmented lasso denotes a family of lasso-based constructions in which temporal structure is built directly into estimation, regularization, inference, or optimization. In the cited literature, the expression is used in several closely related senses: as a lag-expanded lasso with monotone decay across past coefficients, as a varying-coefficient lasso in which time acts as a modifier, as a lasso adapted to dependent and heteroskedastic time series, and as a continuous-time dynamical representation in which the lasso solution is reached along an explicit temporal trajectory (Suo et al., 2014, Tibshirani et al., 2017, Wong et al., 2017, Wu et al., 2 Apr 2026).

1. Conceptual scope

At its narrowest, time-augmented lasso refers to the ordered, time-lagged lasso developed by Suo and Tibshirani, where the feature space is expanded to include lagged predictors and the coefficient sequence within each predictor is constrained to decay with lag (Suo et al., 2014). In a broader sense, the same phrase has been used for models in which coefficients vary as a function of time or time-basis variables, rather than remaining constant across observations (Tibshirani et al., 2017).

A second extension treats time not as a covariate index but as a stochastic dependence structure. In that setting, the lasso is applied to regressions built from lagged observations, but the theory is reformulated under strict stationarity, mixing, Near-Epoch Dependence, heavy tails, and heteroskedasticity, so that estimation and inference remain valid for dependent data rather than only for i.i.d. samples (Wong et al., 2017, Adamek et al., 2020, Ziel, 2015). A third extension makes time explicit in the optimization process itself: the lasso estimator is realized as the equilibrium of a continuous-time ODE, or more generally as a generalized lasso solved by augmented-Lagrangian dynamics in which temporal coupling can be encoded into the regularization operator (Wu et al., 2 Apr 2026, Aleotti et al., 28 Oct 2025).

This suggests that “time augmentation” is not a single estimator but a structural motif. Time can enter through lagged design expansion, coefficient modulation, dependence-aware asymptotics, heteroscedastic reweighting, or continuous-time solver dynamics.

2. Ordered lag structures and sparse time-lagged regression

The canonical time-lagged formulation begins with regression on lagged predictors,

$$y_i = \beta_0 + \sum_{j=1}^p \sum_{k=1}^K x_{ikj}\beta_{kj} + \epsilon_i,$$

where $j$ indexes predictors and $k$ indexes lags. The defining temporal prior is that more recent lags should be at least as influential as older lags, encoded as

$$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$

A direct absolute-value constraint is nonconvex, so the coefficients are split as

$$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$

with separate monotone constraints

$$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$

The resulting objective is a convex lasso-type problem with squared-error loss and $\ell_1$-penalty

$$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$

applied on a design of $pK$ lagged variables (Suo et al., 2014).

In rolling prediction, a single multivariate time series $\{y_t,x_{t1},\dots,x_{tp}\}$ is embedded into a lagged design matrix $j$0 whose $j$1-th row contains blocks

$j$2

The ordered lasso is then applied blockwise, one monotone lag sequence per predictor. Because the monotonicity is imposed on each lag block, the model automatically chooses an effective lag length: once a sequence reaches $j$3, later lags remain $j$4 under the non-increasing constraint (Suo et al., 2014).

The optimization exploits the fact that the proximal operator of the combined $j$5 term and monotone cone can be reduced to isotonic regression. For a block constrained by

$j$6

the proximal step is obtained by isotonic regression of $j$7, followed by truncation at zero. The isotonic subproblem is solved by the Pool Adjacent Violators Algorithm, and each proximal step is $j$8 for a block of length $j$9; in the time-lagged case the corresponding block cost is $k$0 (Suo et al., 2014). The same framework extends to logistic regression through IRLS, and to near-monotone formulations that replace hard order constraints with penalties on positive differences (Suo et al., 2014).

3. Network reconstruction and modifier-based time variation

A direct biological specialization is the time-lagged Ordered Lasso for gene regulatory network inference. For each target gene $k$1, the model regresses current expression on lagged expression of all genes,

$k$2

with the de novo reconstruction problem

$k$3

subject to

$k$4

Because $k$5 forces all later lags to zero, edge existence can be read off from the first lag coefficient, while the full lag block determines how influence decays over time (Nguyen et al., 2018).

The same paper introduces a semi-supervised variant in which prior network information is embedded by assigning different penalties to prior edges and non-edges: $k$6 The ordered lag constraint remains unchanged, but prior edges are favored through smaller penalty weights. In that formulation the method is simultaneously temporal, sparse, and prior-weighted (Nguyen et al., 2018).

A different route to time augmentation is provided by the pliable lasso, where time is not a lag index but a modifying variable. Its core model is

$k$7

so the effective coefficient on predictor $k$8 at modifier value $k$9 is

$$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$0

When $$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$1 is time, or a basis expansion of time, each lasso coefficient becomes a sparse linear function of time. The penalty

$$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$2

induces predictor sparsity, additional sparsity in the modifier effects, and a weak hierarchy in which $$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$3 is effectively nonzero only when $$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$4 is nonzero (Tibshirani et al., 2017).

For scalar time $$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$5, the model reduces to

$$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$6

with time-varying coefficient

$$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$7

For basis-expanded time $$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$8, one obtains

$$|\beta_{1j}| \ge |\beta_{2j}| \ge \cdots \ge |\beta_{Kj}|.$$9

so time dependence can be linear, spline-based, polynomial, Fourier, or regime-coded, while remaining inside a convex optimization problem (Tibshirani et al., 2017).

4. Estimation under dependence, heavy tails, and misspecification

When the lasso is applied to time series, lag expansion alone does not provide valid theory. One influential formulation defines the target as the best linear predictor

$$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$0

where $$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$1 is built from past lags and $$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$2 is the current response. The corresponding estimator is

$$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$3

The associated theory does not assume that the data-generating mechanism is a finite-order Gaussian VAR. Instead, it assumes strict stationarity, sparsity of $$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$4, and either Gaussian $$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$5-mixing or geometrically decaying $$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$6-mixing together with subweibull tails (Wong et al., 2017).

The subweibull framework is broad enough to include subgaussian and subexponential variables, but also variables with tails heavier than an exponential. Under these conditions, the paper derives non-asymptotic inequalities for estimation error and prediction error of the lasso estimate of the best linear predictor. In the Gaussian case, summability of $$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$7-mixing coefficients is sufficient; in the heavy-tailed case, the analysis uses geometrically decaying $$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$8-mixing coefficients and a dependence-tail compatibility condition summarized by

$$\beta_{kj} = \beta_{kj}^+ - \beta_{kj}^-, \qquad \beta_{kj}^+,\beta_{kj}^- \ge 0,$$9

A central implication is that the lasso remains valid under non-Gaussian and non-linear time series models, and under model misspecification, because it targets the best sparse linear predictor rather than a presumed true VAR parameter (Wong et al., 2017).

This perspective is particularly relevant when omitted variables or nonlinear state evolution make the observed process fail to satisfy a finite-order VAR representation. The paper’s examples include Gaussian VAR, a VAR with omitted variables whose observed marginal is no longer finite-order VAR, and a multivariate ARCH-like model. In each case, the time-augmented lasso estimates the sparse linear predictor associated with the chosen lag embedding, rather than requiring exact structural correctness (Wong et al., 2017).

5. Desparsified inference and heteroscedastic reweighting

High-dimensional inference in time series requires an additional layer beyond estimation. One construction extends the desparsified lasso to

$$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$0

under Near-Epoch Dependence assumptions that allow non-Gaussian, serially correlated, and heteroskedastic processes. The debiased estimator is

$$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$1

where $$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$2 is obtained from nodewise lasso regressions. The asymptotic variance is not a simple residual variance but a long-run covariance of score-type products $$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$3, estimated by a Bartlett-kernel HAC estimator

$$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$4

Under the paper’s NED, weak sparsity, and growth conditions, the desparsified lasso is uniformly asymptotically normal, and the long-run variance estimator is consistent, enabling confidence intervals, Wald tests, and simultaneous inference for high-dimensional time-series parameters (Adamek et al., 2020).

A complementary adaptation addresses conditional heteroscedasticity directly at the estimation stage. The iteratively reweighted adaptive lasso solves

$$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$5

with heteroscedasticity weights

$$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$6

and adaptive penalty weights

$$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$7

The algorithm alternates between weighted lasso estimation of the conditional mean and estimation of a conditional variance model from the residuals, updating the weights iteratively (Ziel, 2015).

This reweighted scheme is designed for AR-ARCH type processes and their extensions. The paper states sign consistency and asymptotic normality for the active coefficients, and it further states that any iteration $$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$8 achieves the same optimal asymptotic variance. The framework is extended to multivariate AR-ARCH, periodic AR-ARCH, threshold AR-ARCH, and ARMA-GARCH type models, so temporal augmentation is present both in the mean equation through many lags and in the weighting scheme through time-varying volatility (Ziel, 2015).

6. Continuous-time dynamical and generalized-lasso formulations

A conceptually different notion of time augmentation treats the lasso itself as a continuous-time dynamical system. One paper considers the $$\beta_{1j}^+ \ge \beta_{2j}^+ \ge \cdots \ge \beta_{Kj}^+ \ge 0, \qquad \beta_{1j}^- \ge \beta_{2j}^- \ge \cdots \ge \beta_{Kj}^- \ge 0.$$9-regularized lasso

$\ell_1$0

splits $\ell_1$1 with $\ell_1$2, and rewrites the problem as a smooth nonnegative quadratic program in

$\ell_1$3

The KKT system is

$\ell_1$4

and the solver is a projection-free Newton-based ODE that drives the residual

$\ell_1$5

according to a fixed-time-stable flow. The settling time is independent of the problem data and can be prescribed by choosing

$\ell_1$6

so that the worst-case settling time satisfies $\ell_1$7 (Wu et al., 2 Apr 2026).

In this representation, time is not merely an iteration counter but the physical variable along which the optimizer evolves. The paper proves that with strictly positive initialization, nonnegativity is preserved without projection, equilibria coincide with KKT points, and the recovered

$\ell_1$8

is the unique optimal solution of the $\ell_1$9-regularized lasso. The construction is explicitly motivated by analog and neuromorphic implementation, where prescribed-time convergence and projection-free dynamics are advantageous (Wu et al., 2 Apr 2026).

A further generalization appears in variable projected augmented Lagrangian methods for the generalized nonlinear lasso

$$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$0

Introducing a split variable $$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$1 with constraint $$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$2, the augmented Lagrangian is reduced by eliminating $$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$3 through soft-thresholding,

$$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$4

The resulting projected objective is smooth in $$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$5, with gradient

$$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$6

For linear models, a preconditioned variant uses a Hessian-like operator and mimics Newton-type updates (Aleotti et al., 28 Oct 2025).

In that framework, any temporal coupling can be encoded into the linear operator $$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$7. This suggests a direct route to space-time or purely temporal generalized lasso models in which $$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$8 contains temporal finite differences, fused-lasso penalties across time, or other structured lag couplings. The paper itself treats generalized nonlinear inverse problems rather than a dedicated temporal benchmark, but its formulation makes temporal regularization an operator-design choice rather than a change in optimization principle (Aleotti et al., 28 Oct 2025).

7. Empirical behavior, interpretability, and limitations

In the ordered time-lagged setting, empirical behavior is closely tied to the validity of monotone decay. In a rolling-prediction simulation with $$\frac{1}{2}\sum (y_i-\hat y_i)^2 + \lambda \sum_{j=1}^p \sum_{k=1}^K (\beta_{kj}^+ + \beta_{kj}^-),$$9 predictors, maximum lag $pK$0, and $pK$1, the ordered lasso achieved average MSE $pK$2, whereas the standard lasso achieved average MSE $pK$3. In a larger example with $pK$4, the ordered lasso clearly outperformed the lasso when the true coefficients were monotone in lag, whereas the standard lasso could perform better when the true coefficients were randomly permuted. On Los Angeles ozone data, both time-lagged lasso and ordered lasso improved prediction error over cross-sectional lasso, and the ordered lasso achieved its minimum error with fewer degrees of freedom while producing interpretable lag windows such as nonzero wind coefficients up to about $pK$5 days and humidity coefficients up to about $pK$6 days. On yearly sunspot data, classical AR with AIC chose order $pK$7, ordered lasso suggested order about $pK$8 with monotone coefficients, and all three compared methods had similar validation error (Suo et al., 2014).

For network inference, the same monotone-lag constraint yields robustness to the choice of maximum lag. On repressilator simulations, time-lagged Ordered Lasso achieved AUC $pK$9 when temporal coverage was sufficient. On DREAM and HeLa datasets, the paper reports that Ordered Lasso AUCs increase monotonically with $\{y_t,x_{t1},\dots,x_{tp}\}$0 and then plateau, while Granger and Lasso-Granger peak at intermediate lags and then degrade. The semi-supervised variant also ranked highly edges absent in the older BioGRID-based prior but present in newer BioGRID versions (Nguyen et al., 2018).

Modifier-based time augmentation produces a different empirical pattern: not lag selection, but sparse selection of predictors whose effects change over time or regime. In simulations with modifying variables, pliable lasso consistently performed best in test error among the compared methods. In a stock-returns forecasting example based on time regimes, it improved the correlation between predicted and actual returns by about $\{y_t,x_{t1},\dots,x_{tp}\}$1 in the test period. In crime data with regional modifiers, it produced better test error than a plain lasso and showed that only a few predictors had modifier-varying coefficients (Tibshirani et al., 2017).

The continuous-time ODE formulation exhibits a separate empirical criterion: convergence within a prescribed physical time. In numerical experiments with $\{y_t,x_{t1},\dots,x_{tp}\}$2, $\{y_t,x_{t1},\dots,x_{tp}\}$3, $\{y_t,x_{t1},\dots,x_{tp}\}$4, $\{y_t,x_{t1},\dots,x_{tp}\}$5, and $\{y_t,x_{t1},\dots,x_{tp}\}$6 generated instances, trajectories reached the optimal solution within prescribed times $\{y_t,x_{t1},\dots,x_{tp}\}$7, and convergence by $\{y_t,x_{t1},\dots,x_{tp}\}$8 was maintained across initial conditions $\{y_t,x_{t1},\dots,x_{tp}\}$9, $j$00 (Wu et al., 2 Apr 2026).

The limitations are correspondingly heterogeneous. Ordered lag models assume monotone decay in $j$01 with lag, which can be wrong in the presence of delayed responses, complex oscillatory dynamics, or non-monotone impulse responses (Nguyen et al., 2018). Modifier-based formulations can overfit if the time basis is too rich, and standard random cross-validation can be optimistic when observations are temporally correlated (Tibshirani et al., 2017). Dependence-aware estimation and inference require mixing, NED, weak sparsity, and growth-rate conditions that are substantially stronger than the classical i.i.d. setting (Wong et al., 2017, Adamek et al., 2020). Iterative heteroscedastic reweighting depends on consistent volatility estimation and is developed under stationarity assumptions (Ziel, 2015). The fixed-time ODE work reports only “soft verification” via digital simulation, while “hard verification” on actual analog hardware is deferred to future work (Wu et al., 2 Apr 2026).

Taken together, these strands define time-augmented lasso as a technically diverse but coherent class of methods: sparse regression becomes temporally structured by lag-order constraints, modifier-driven coefficient variation, dependence-aware estimation and inference, volatility-adaptive weighting, or explicit continuous-time optimization dynamics.