First-hand information for everyone
This paper gives a clear method for answering a simple question: when several time series move together, which one moves first and how strongly? The authors’ organizing idea is physical and easy to state. If you record a system for a long time and then play that record backwards, anything that changes is the system’s direction. If the forward and reversed records are statistically the same, there is no arrow of time to measure; if they differ, the difference is the arrow.
Formally the paper shows that the directional part of a multivariate linear system sits in a specific matrix they call the circulation Q. It is carried by the lagged covariance — that is, the covariance between a variable now and the same set of variables a short time earlier. The usual tools built only on same-time covariance (correlation matrices, principal components, undirected network measures) cannot tell direction, because two systems with opposite direction can have the same same-time covariance. Under a Gaussian benchmark the size of Q can be read as an entropy‑production rate, a physical measure of how much the forward record diverges from the reversed one.
Turning this estimand into a practical tool required attention to finite-sample problems. The natural magnitude of direction is quadratic in Q, and quadratic functions of noisy estimates are biased upwards. The paper shows that this bias grows like n^2/T for n series and T observations, and gives a concrete example: a naive estimator returned 0.845 while a bias‑corrected one returned −0.009. To fix the problem they build a cross‑fitted estimator that removes first‑order bias, use delete‑block jackknife standard errors that avoid reusing observations, and supply a randomisation test that is exact under the null that the record and its reverse have the same law. The testing approach also allows a familywise correction at no extra cost, and a nonlinear extension using feature maps that preserves the exact reference distribution.
The paper also studies when an arrow can be seen at all. If the underlying rotations or signals happen much faster than the sampling interval, or if the data are sampled too coarsely, the measured arrow collapses to zero — an effect known as aliasing. The authors present a sampling theory that quantifies this window of timescales where direction is visible. They also add a design layer to separate a measured arrow from non‑transmission sources of ordering, such as markets in different time zones, where apparent lead–lag can be driven by clock timing rather than causal transmission.
As a reality check the authors run a laboratory of four systems with known answers and compare their method to common tools (correlation networks, Granger causality, transfer entropy, and connectedness indices). The comparison shows standard tools can be misleading if read as measures of direction: for example, in an exactly reversible system a Granger causality test rejected in 1.00 of samples and a connectedness index reported a net directional flow of 10% — correct answers for those methods’ own questions but wrongly read as an arrow. The authors also report that their linear method misses purely nonlinear transmission, and that the feature‑map extension recovers such cases. The paper supplies complete algorithms and code and notes the main limitations: the need for sufficient sampling relative to the system’s time scales, and the extra care required for nonlinear dynamics and for distinguishing clock effects from genuine transmission.