Does EEG Lempel-Ziv complexity just reflect temporal autocorrelation?

Two metrics that both track temporal predictability

Lempel-Ziv complexity counts how many distinct new patterns appear when a symbolic sequence is parsed from left to right. A signal with many repeated patterns can be compressed into few blocks and gets a low complexity value; a signal with few repeats gets a high one. In EEG work the raw signal is first turned into a binary string, typically by comparing each sample with a threshold such as the median, and the block count is divided by a maximum-complexity term so that the normalized value falls between 0 and 1.

The autocorrelation function measures how similar a signal is to a shifted copy of itself. The first-lag coefficient, AC1, captures the similarity between each sample and the next one, so a slowly decaying autocorrelation means strong short-term memory and a rapid decay means noisy, broadband activity. The paper's argument is that these two quantities are not independent descriptions of a signal: any process that raises temporal predictability also raises pattern repetition, which is exactly what makes the Lempel-Ziv parser compress more efficiently.

Simulated signals show a robust inverse relation

The authors first avoided the confound of binarizing continuous data by generating binary sequences directly. A two-state Markov chain was tuned through a transition parameter from weakly to strongly persistent dynamics, and the measured AC1 matched the prescribed parameter across the explored range. As persistence increased, LZC fell systematically, and plotting 1 − LZC against AC1 produced a monotonic, weakly nonlinear curve.

A second generator, a one-dimensional random Cantor set, allowed long-range correlations to be manipulated through a retention probability. Here too the relation was monotonic, though LZC behaved somewhat nonlinearly with respect to AC1. A third procedure progressively coarse-grained a correlated sequence by replacing windows with their rounded mean, driving the data toward an uncorrelated state; AC1 and LZC declined together at each step. Because these three generators rest on different theoretical grounds, the authors read the shared pattern as evidence that the link is not an artifact of one model class.

Finite-length estimates sit above the entropy rate

For the Markov chain the paper can go further than a correlation plot. The probability of switching state at each step is tied to AC1, which yields an entropy rate for the binary process. Ziv's theorem says normalized Lempel-Ziv complexity converges to that entropy rate as sequence length grows without bound, so a monotonic relation between LZC and the entropy rate implied by AC1 is expected for long sequences.

The simulations matched this expectation in direction but not in level: estimated LZC stayed systematically above the asymptotic reference line. The authors attribute the offset to the known slow finite-size convergence of the normalized estimator. This is a useful boundary for readers, because it means the simulated LZC values are not direct readings of entropy rate; the agreement is about ordering and trend, not about exact numerical equality.

EEG recordings reproduce the same trade-off

The relationship was then tested on physiological data. In an event-related potential dataset from an emotion-attention task, the authors filtered for the alpha band (8–13 Hz), binarized with a median threshold, and computed LZC alongside AC1 at a first lag of 0.5 ms. Scatterplots for individual participants showed a stable, approximately linear AC1/LZC relation, both in the pre-stimulus interval and around the evoked-potential maximum, and the same tendency held when all participants were pooled. The authors state explicitly that their aim was not to interpret the ERP responses but to show that LZC and AC1 fluctuate in parallel in a typical ERP experiment.

Resting-state recordings from a public dataset gave a second test. With 53 channels from one subject sampled at 512 Hz, five-minute eyes-open and eyes-closed windows formed two well-separated clusters: eyes open sat at relatively low AC1 and high LZC, eyes closed at the opposite corner. Recomputing both metrics in sliding, non-overlapping 75-second windows traced the same quasi-linear dependency over time. The eyes-open/closed maneuver is useful here because it shifts cortical activity between more and less synchronous states, so the two clusters are a physiological version of the manipulation performed in simulation.

Why lower complexity follows stronger correlation

The mechanism the paper offers is compression, not information processing. When a signal has stronger short-term memory, successive samples resemble one another more often, so recurrent symbolic patterns become more probable. The Lempel-Ziv algorithm then encounters fewer novel substrings and compresses the sequence more efficiently, which registers as lower complexity. In this account, a large share of the variability that EEG studies attribute to changes in complexity may instead reflect differences in the temporal correlation of ongoing activity.

The authors are careful about the limits of this claim. They do not argue that LZC and autocorrelation are mathematically equivalent: a single autocorrelation coefficient does not represent higher-order correlations, and signals can differ in other aspects of temporal organization. Their claim is that temporal dependence is a major determinant of LZC and that a substantial proportion of the variability attributed to complexity can be explained by simple temporal dependencies.

Autocorrelation offers a more direct physiological reading

The interpretive payoff concerns what each metric can be tied to. Changes in autocorrelation map onto mechanisms that electrophysiology already describes: stronger rhythmic activity, greater oscillatory stability, enhanced synchronization among neuronal populations, or slower recovery of cortical networks after spontaneous fluctuations. A change in LZC, by contrast, is often read as altered neural complexity, information richness, dynamical repertoire, or conscious content, but the algorithm itself quantifies only the compressibility of a symbolic sequence. Similar LZC changes could therefore arise from shifts in oscillatory amplitude, spectral composition, synchronization, signal-to-noise ratio, or temporal dependence.

The paper also notes a practical asymmetry. Estimating autocorrelation is computationally trivial, needs no symbolization or thresholding, and has an intuitive reading as temporal memory. Computing LZC requires discretizing the signal and depends on choices such as sequence length and symbolization strategy. The authors do not dismiss complexity metrics, which they still treat as compact descriptors of signal organization that may capture structure beyond conventional linear statistics; their point is that the extra machinery needs justification against a simpler reference.

Where the correlation argument connects to criticality

The paper links its result to the idea that brain function operates near a phase transition. As a complex system approaches the edge of synchronization, its average activity becomes more correlated in time, a phenomenon the authors describe as critical slowing down, and the same rise in AC1 appears in systems as different as flocks of birds, magnetic spin collectives, and spiking neural networks. Under those circumstances, increases in AC1 are expected to be accompanied by increases in pairwise cross-correlations, or functional connectivity.

This gives AC1 and LZC a shared interpretation as trackers of spontaneous arousal fluctuations, provided those fluctuations reflect distance from a critical point. The authors cite recent evidence in that direction but present it as a suggestion rather than a demonstrated result of this study. The criticality discussion is therefore a proposed interpretive frame, not an additional experiment reported here.

What to check before attributing meaning to LZC

The paper's practical priority is a comparison, not a ban. Before assigning functional significance to an observed change in EEG Lempel-Ziv complexity, the authors argue, studies should routinely assess how that change relates to autocorrelation and other basic descriptors of temporal structure. If a complexity difference tracks AC1 closely, as it did across the simulated generators, the ERP trials, and the eyes-open/closed resting-state windows, then the simpler and more physiologically transparent measure is the natural reference against which any added value of LZC must be shown.

The scope of that recommendation follows the evidence: binary stochastic processes, alpha-band event-related potentials, and resting-state recordings from healthy adults, including a single subject in the 53-channel resting-state analysis. The relation was robust but not exact, varying in shape across signal classes, and the authors' own framing leaves room for complexity metrics to capture temporal structure that linear autocorrelation misses. The decision threshold the paper proposes is therefore comparative rather than absolute: a complexity finding earns a mechanistic interpretation only after autocorrelation has been ruled out as its explanation.

The narrower claim worth carrying forward

The strongest takeaway is not that LZC is uninformative but that its information content in typical EEG overlaps heavily with first-lag autocorrelation. The most decision-relevant comparison in the paper is between a metric that requires symbolization, thresholding, and length choices and one that is computed directly and maps onto oscillatory persistence and synchronization. Where the two move together, as they did in the eyes-open versus eyes-closed clusters and in the trial-by-trial ERP scatterplots, the autocorrelation reading is the more transparent one. Where they diverge, that divergence is the specific evidence a complexity claim would need to present.

Sources used in this answer