Many measured systems, such as brain signals, battery cells and traffic flows, remember their past through slowly decaying power laws rather than through the fast exponential forgetting that standard time-series models assume. When such a system has hundreds of interacting variables and we want to learn which variable drives which, the usual tool fits a sparse vector autoregression: each variable is regressed on a few lags of all variables, with a penalty that zeroes out most coefficients. This note argues that the tool fails in a specific way on power-law memory. The long tail of each variable's own history leaks into the lag coefficients, which forces a choice between many dense nuisance coefficients and a large truncation bias. The long memory also makes unrelated variables look correlated. The damage is worst when the system sits close to its stability boundary, which is where these systems often operate.
We propose to separate the two jobs. Each variable's self-memory gets a single parameter, its fractional order: a number between 0 and 1 that sets how slowly the influence of its past decays. The sparse cross-variable network is then selected on the "fractionally differenced" response, meaning the signal after that memory has been filtered out. The central claim has two parts. First, this approach recovers the network at nearly the rate available for short-memory systems. Second, the quantity that governs how hard recovery is becomes a fractional stability margin, which depends jointly on the interaction matrix and the fractional orders, and not the spectral radius alone. We test both parts with theory, controlled simulation and real multichannel recordings.
Core Claim
Consider a discrete-time fractional-order network in the Grünwald–Letnikov form:
$$\sum_{j=0}^{t+1} c_j(\alpha_i)\, x_{i,t+1-j} \;=\; \sum_{k=1}^{p} A_{ik}\, x_{k,t} + e_{i,t+1}.$$
Here $x_{i,t}$ is variable $i$ at time $t$. $\alpha_i \in (0,1]$ is its fractional order. $c_j(\alpha)$ are the fixed binomial weights of the fractional difference, with $c_0=1$ and $|c_j| \sim j^{-1-\alpha}$. $A$ is a sparse $p \times p$ interaction matrix, and $e$ is noise. Setting $\alpha_i = 1$ recovers an ordinary first-order difference model.
Rewritten as an ordinary autoregression, this system has one lag of cross-effects, driven by $A$, plus an infinite, dense, diagonal self-history whose weights decay as a power law. A lag-$d$ vector autoregression fit with the Lasso, as in Estimating Structured Vector Autoregressive Models, must therefore do one of two things. It can carry about $pd$ nonzero self-lag coefficients on top of the $s$ true network edges. Or it can pay a truncation bias of order $\sum_{j>d} j^{-1-\alpha} \sim d^{-\alpha}$, which for $\alpha = 0.3$ is still about 0.25 at $d = 100$. Plugging the inflated sparsity into that paper's $\sqrt{s\log(pd)/N}$ rate suggests an error of roughly
$$\sqrt{\frac{(s + pd)\log(pd)}{N}} + d^{-\alpha},$$
where $N$ is the number of time samples. Balancing the two terms gives a rate far slower than $\sqrt{s \log p / N}$.
Our estimator instead profiles each $\alpha_i$ and runs sparse selection on the filtered response. We conjecture it attains the $\sqrt{s\log p/N}$ rate, up to a constant $C(\Delta_\alpha)$, where $\Delta_\alpha$ is the distance of the system from its fractional stability boundary. The second part of the claim is that $C$ blows up like $1/\Delta_\alpha^{2}$. That is the analogue of the inverse-square margin laws proven for integer-order systems by On the Approximation of Toeplitz Operators for Nonparametric H∞-norm Estimation and Finite-sample-based Spectral Radius Estimation and Stabilizability Test for Networked Control Systems. The contrast we care about is that two systems with identical spectral radius $\rho(A)$ but different orders $\alpha$ can differ by orders of magnitude in how many samples they need.
A simple mechanism supports the first part. Filtering only the response, and not the regressors, turns a long-memory response into a short-memory one. Consider the sample cross-covariance between that short-memory response and an unrelated long-memory regressor. Its variance is a sum of products of their autocovariances, and that sum converges because the response's autocovariances are summable. So the spurious correlations that inflate marginal screening under long memory, the concern raised in Guarding against Spurious Discoveries in High Dimensions, shrink back to the independent-data scale.
Relation to Prior Work
Two lines of work motivate the proposal by exposing the gap. The first is sparse autoregressive estimation with guarantees. Estimating Structured Vector Autoregressive Models bounds structured-norm estimators for stable vector autoregressions. Sampling Requirements for Stable Autoregressive Estimation ties the sample requirement to the spectral spread, meaning the ratio of largest to smallest power spectral density. Both assume exponential memory. A fractional process has a spectral density that diverges at zero frequency, so the spread is infinite and neither bound applies. We keep their machinery for restricted eigenvalue arguments (conditions ensuring the design matrix is well-conditioned on sparse directions) but apply it only after filtering the response.
The second line is fractional-order identification. A Sampling Complexity-aware Framework for Discrete-time Fractional-Order Dynamical System Identification turns nonlinear fractional identification into least squares on basis expansions, with $N = O(d/\epsilon)$. It is low-dimensional and starts from reset initial conditions. We borrow its least-squares reformulation and basis idea, and add three things it lacks: sparsity for $p \gg N$, unknown pre-sample history, and behavior near the boundary. Tutorial on system identification using fractional differentiation models supplies the parsimony argument that one order per variable replaces many rational-approximation parameters.
Two further papers constrain rather than motivate. High-dimensional structure learning of sparse vector autoregressive models using fractional marginal pseudo-likelihood looks close by name, but its "fractional" is a fractional power of the likelihood used for Bayesian scoring; its dynamics are integer-order. It is a baseline, not prior art for our idea. End-to-End Learning Framework for Solving Non-Markovian Optimal Control shows that fractional linear systems admit block-Toeplitz structure and that integer-order surrogates lose badly on them. Its setting is control of known, low-dimensional systems, not selecting structure.
For the selection step we reuse SLOPE—Adaptive variable selection via convex optimization, a penalty that sorts coefficients and penalizes them by rank so as to control the false discovery rate. For the nonlinear extension we reuse spline-based screening from Nonparametric Independence Screening in Sparse Ultra-High-Dimensional Additive Models. Both were analyzed for independent data only, and SLOPE's false discovery guarantee holds only for orthogonal designs. Whether they survive under long memory is part of what we test, not something we assume.
Research Program
The estimator is node-wise, so it runs in parallel across variables. For each variable $i$ and each candidate order on a fine grid, we compute the fractionally differenced response using the full observed history. With a fast Fourier transform this costs $O(N\log N)$ per candidate. We then fit a SLOPE or Lasso regression of that response on the lagged state $x_t$. We select $\alpha_i$ by residual error on held-out contiguous time blocks. Profiling over the order, instead of fitting it jointly by gradient methods, is meant to address the noise sensitivity of joint coefficient-and-order estimation noted in the fractional-identification literature. Because the pre-sample history is unknown, its effect decays like $t^{-\alpha}$. We handle this with a burn-in period and a small set of nuisance regressors for that transient, and we measure how much each option matters.
The theory proceeds in three steps.
- Restricted eigenvalue condition. We prove that the unfiltered long-memory design $x_t$ still satisfies a restricted eigenvalue condition, with high probability, once $N$ exceeds a threshold. The smallest eigenvalue of the spectral density stays bounded below even though the largest diverges, so we expect this to hold with a slower concentration exponent.
- Error bound. We derive an error bound for the estimated $A$ that has the $\sqrt{s\log p/N}$ form, plus a term for errors in $\hat\alpha$.
- Fractional margin. We define $\Delta_\alpha$ as the distance from the eigenvalues of $A$ to the region where the characteristic function $(1-z)^{\alpha}-\lambda z$ has no roots inside the unit disk. We compute this region numerically, because we do not rely on a closed form. We then show how $C$ depends on $\Delta_\alpha$.
If the $1/\Delta_\alpha^2$ law fails to hold, we will report the scaling we actually observe instead.
The nonlinear extension replaces $A_{ik}x_{k,t}$ with sparse additive functions $f_{ik}(x_{k,t})$ expanded in B-splines and selected by group penalties. Screening is performed on the filtered response, which is the dynamic analogue of nonparametric independence screening. This part is empirical, not theoretical.
Baselines
The naive reference is a first-order vector autoregression with the Lasso that ignores memory. It shows how much damage unmodeled power-law history does.
The strongest existing method is the lag-$d$ sparse vector autoregression with the Lasso and with ordered weighted $\ell_1$ penalties, from Estimating Structured Vector Autoregressive Models. It is the strongest because any finite power-law tail is, in principle, representable at a large enough lag. We give it an unfair advantage: $d$ tuned over a range up to $N/2$ using the same blocked cross-validation, the same data and the same number of penalty values as our method. It is counted as beaten only if our method has higher edge-recovery F1 and lower error on $A$ at every tested $N$ below the point where both saturate.
We also run a stronger variant of this baseline: a lag-$d$ model whose own-lag coefficients are left unpenalized, so only cross-lags are sparsified. This removes the inflated-sparsity argument and leaves only truncation bias.
We run PLVAR from High-dimensional structure learning of sparse vector autoregressive models using fractional marginal pseudo-likelihood at the lag order it selects itself. That paper reports faster and more accurate structure recovery than cross-validated Lasso and SCAD on 80-channel EEG, but the file gives no numerical precision or recall, so no reported result was found on Lacuna for PLVAR's edge-recovery accuracy.
In the low-dimensional regime we run the least-squares estimator of A Sampling Complexity-aware Framework for Discrete-time Fractional-Order Dynamical System Identification. That paper reports $\hat\alpha \approx 0.807$ for true $\alpha = 0.8$ and $\hat\alpha \approx 0.407$ for true $\alpha = 0.4$ over 100 synthetic trials. Matching that order accuracy when $p < N$ is a sanity requirement, not a win. For the nonlinear extension, INIS-penGAM from Nonparametric Independence Screening in Sparse Ultra-High-Dimensional Additive Models, applied to raw lagged data, is the screening baseline.
Two ablations isolate the key ingredient. One fixes every $\alpha_i = 1$, which is the integer-order version of our pipeline. The other gives our pipeline the true $\alpha$, which bounds how much error comes from estimating the orders.
Why not just…?
Why not just fit a high-order vector autoregression with the Lasso? The Lasso rate grows only logarithmically in the lag $d$, so this looks cheap. But in a power-law system the true self-lag coefficients are nonzero at every lag, so the effective sparsity grows linearly in $d$ while the bias shrinks only like $d^{-\alpha}$. For small $\alpha$ no single $d$ makes both small at realistic $N$. The unpenalized-own-lag variant is the honest version of this objection, and it is a real threat when $\alpha$ is close to 1 and memory is short. The simulation sweep over $\alpha$ is designed to show where, if anywhere, the gap closes. We will report that crossover rather than hide it. Consistent evidence that integer-order surrogates lose on power-law systems comes from SOC fusion estimation for lithium battery using dual OCV-SOC relationships combined with fractional order extended Kalman filter. There, an integer-order Kalman filter has 2.65% state-of-charge error versus 0.70% for a fractional model, on a problem that is low-dimensional and does not involve selection.
Why not just fractionally difference each series on its own with a standard univariate long-memory estimator, then run a sparse first-order regression? This is the most serious alternative, because it is our pipeline with the order estimated marginally. The difference is that a univariate order estimate absorbs memory that is actually transmitted through the network $A$. Near the stability boundary, the network itself generates slow decay that a univariate estimator will attribute to $\alpha_i$, so it will over-filter. We include this two-stage method as a named baseline and predict it matches us far from the boundary and degrades as $\Delta_\alpha \to 0$. If it matches us everywhere, the contribution shrinks to the theory and the margin law, and we will say so.
Why not just use a neural sequence model such as a recurrent network or Neural Laplace? Neural Laplace: Learning diverse classes of differential equations in the Laplace domain shows that non-Markovian dynamics can be learned by flexible models. But the goal here is a sparse, interpretable network with false-discovery control from $p \gg N$ data, and these models give neither. We include a recurrent network with input-gradient attribution as a secondary comparison on edge recovery only, and do not expect it to be competitive when $N$ is small.
Why not just collect more data? Active Learning for Identification of Linear Dynamical Systems and the margin laws cited above show that, near the boundary, passive data requirements grow without bound. Our margin experiment measures exactly how fast this happens for fractional systems.
Proposed Experiments
Synthetic study. We extend the fractional-system benchmark of A Sampling Complexity-aware Framework for Discrete-time Fractional-Order Dynamical System Identification to networks with $p$ from 50 to 500 and $s$ around $2p$ edges. Orders are drawn from $\alpha \in [0.2, 1]$, and $N$ ranges from $p/4$ to $4p$. Both Gaussian and heavy-tailed noise are used, and there are 100 trials per cell.
The decisive check for the second claim is a matched-radius design. We build pairs of systems with equal $\rho(A)$ but different $\alpha$, and therefore different $\Delta_\alpha$. We then plot the samples needed to reach edge-recovery F1 of 0.9 against both $\rho(A)$ and $\Delta_\alpha$. The claim holds if $\Delta_\alpha$ collapses the curves and $\rho(A)$ does not, and if the slope on log–log axes is near $-2$. For the first claim, the check is that our error curve, plotted against $N$, tracks the oracle-order curve while the tuned lag-$d$ Lasso plateaus at its truncation bias for small $\alpha$. We also measure empirical false discovery rate under SLOPE at nominal $q = 0.1$, with and without response filtering. The screening claim predicts that filtering is what restores nominal control.
Real data. On the CHB-MIT scalp EEG database, used by High-dimensional structure learning of sparse vector autoregressive models using fractional marginal pseudo-likelihood, there is no ground-truth network. We therefore evaluate multi-step forecast error on held-out segments, the stability of selected edges across recordings from the same patient, and the number of edges needed to reach a given forecast error. On the NASA flight data from Estimating Structured Vector Autoregressive Models, we check whether the known physical links, such as vertical speed and angle of attack, are recovered with fewer spurious edges. On the traffic speed data from Sampling Requirements for Stable Autoregressive Estimation, we test whether the fractional model explains the slow component without the dense lag structure the integer-order fit needs. On the CALCE battery profiles, which are low-dimensional, we only verify that the estimated orders fall in a physically plausible range and that the voltage fit does not degrade relative to the integer model. We make no selection claim there.
The note's contribution fails if either of two things happens: the tuned, unpenalized-own-lag vector autoregression matches our method across the $\alpha$ sweep, or the matched-radius experiment shows $\rho(A)$ predicts difficulty as well as $\Delta_\alpha$.
References
- Estimating Structured Vector Autoregressive Models
- On the Approximation of Toeplitz Operators for Nonparametric H∞-norm Estimation
- Finite-sample-based Spectral Radius Estimation and Stabilizability Test for Networked Control Systems
- Guarding against Spurious Discoveries in High Dimensions
- Sampling Requirements for Stable Autoregressive Estimation
- A Sampling Complexity-aware Framework for Discrete-time Fractional-Order Dynamical System Identification
- Tutorial on system identification using fractional differentiation models
- High-dimensional structure learning of sparse vector autoregressive models using fractional marginal pseudo-likelihood
- End-to-End Learning Framework for Solving Non-Markovian Optimal Control
- SLOPE—Adaptive variable selection via convex optimization
- Nonparametric Independence Screening in Sparse Ultra-High-Dimensional Additive Models
- SOC fusion estimation for lithium battery using dual OCV-SOC relationships combined with fractional order extended Kalman filter
- Neural Laplace: Learning diverse classes of differential equations in the Laplace domain
- Active Learning for Identification of Linear Dynamical Systems