Pearson correlation has long been the workhorse of time-series comparison in neuroscience. As a model-based measure, it operates under the assumption that the two input signals are jointly Gaussian and largely free of outliers. When those assumptions hold, the sample Pearson correlation coefficient is easy to compute from experimental recordings and offers a transparent interpretation: a value near one signals a strong linear relationship, a value near zero signals the absence of a consistent linear one. But the brain rarely cooperates so neatly. Neural evoked responses can be noisy, non-Gaussian, and, crucially, linked by dependencies that go beyond simple linearity. A zero correlation, the authors note, does not even imply the absence of structure, since alternating stretches of positive and negative correlation can cancel out. This is precisely where mutual information, a model-free measure rooted in information theory dating back to Claude Shannon’s foundational 1948 work, offers an attractive alternative capable of capturing both linear and higher-order relationships between signals.

The catch, and the reason mutual information has not displaced correlation in everyday neuroimaging practice, is estimation. For continuous data, there is no simple, universally accepted estimator of mutual information, and accurate estimation from the limited samples typical of neurophysiological recordings is notoriously challenging. Decades of methodological development have produced a crowded toolbox: adaptive binning schemes, kernel density estimators, the Kraskov–Stögbauer–Grassberger nearest-neighbor estimator, Gaussian copula approaches, and even neural network–based estimators. Each carries its own advantages, biases, and tuning parameters. The underlying probability distribution of the data, the choice of estimator-specific parameters, and normalization factors can all dramatically shift the resulting estimates. The authors of the new study, Anni Hukari, Silvia Federica Cotroneo, and Riitta Salmelin of Aalto University’s Department of Neuroscience and Biomedical Engineering, set out to bring order to this landscape for the specific case of neural evoked responses.

The study’s methodological core is a carefully controlled simulation framework. The researchers generated base signals as Morlet wavelets, resembling the oscillatory waveforms common in evoked brain activity, and then subjected pairs of copies to a battery of realistic transformations. They varied the signal-to-noise ratio, injected sparse high-amplitude outliers resembling muscle artifacts, appended noise-only segments to mimic signal cropping, imposed small time shifts, altered response duration, and changed relative magnitude. Each configuration was repeated one thousand times across different noise iterations, allowing the team to map how each estimator’s output behaved across the full parametric space. In parallel, they anchored their interpretation with a novel statistical device: adaptive lower bounds, constructed by comparing each reference signal against one thousand Gaussian noise realizations matched in mean and standard deviation, yielding empirical ninety-nine percent confidence thresholds against which the true comparisons could be judged.

The parameter tuning phase alone yielded practical insights that researchers can apply immediately. For the kernel density estimator, the team adopted Scott’s rule for bandwidth selection, which produced stable performance in both simple and nonlinear comparison scenarios. For the Kraskov estimator, the authors found that while the original developers recommend choosing between two and four nearest neighbors, setting k equal to five widened the gap between genuine signal similarity and noise comparisons, improving separability. For adaptive binning, Doane’s rule correctly identified the optimal number of bins, thirteen in their configuration, corresponding to the peak of the similarity estimate curve. Crucially, the lower bounds proved instrumental in these decisions: small kernel bandwidths, for instance, inflated the noise floor so much that true similarity became indistinguishable from chance, a failure mode that would be invisible without such a reference.

The simulation results revealed a consistent behavioral split. Pearson correlation and the kernel density estimator formed one pair, while the Kraskov and adaptive binning estimators formed another. When signals were strictly shape-identical, Pearson correlation performed reliably, even tolerating remarkably small sample sizes, and it separated signal from noise comparisons slightly earlier, at zero decibels of signal-to-noise ratio, where the mutual information estimators still struggled. Yet when signals shared information without sharing shape, Pearson correlation and the kernel estimator faltered. Under small time shifts, Pearson correlation and the kernel estimator dropped below their noise thresholds almost immediately, while the Kraskov and binning estimators remained above their bounds, detecting the underlying relationship. Similarly, with sparse high-amplitude outliers, Pearson correlation and the kernel estimator degraded rapidly, whereas the nearest-neighbor and binning approaches declined more gracefully. The two estimation families, in short, are sensitive to different properties of the data.

The study also delivered concrete numerical guidance. All estimators required positive signal-to-noise ratios to function meaningfully, but their performance stabilized around twenty decibels, a level at which any added noise causes negligible distortion. Sample sizes of roughly one hundred began to separate genuine similarity from noise, and by two hundred and ten samples, corresponding to a six-hundred-hertz sampling rate for their signals, all estimators converged to stable estimates. Sampling frequencies above two hundred hertz proved sufficient, improvements plateaued near five hundred hertz, and oversampling beyond that added computational cost and redundancy-driven bias without benefit. Perhaps counterintuitively, the researchers recommend deliberately adding a small amount of Gaussian noise to signals, even after preprocessing, when using the Kraskov or kernel density estimators, since near-identical samples can cause these methods to break down entirely.

To demonstrate that these findings survive contact with real data, the team applied all four estimators to magnetoencephalography recordings from a picture-naming task, comparing a reference channel near a known brain activation source with every other sensor in the whole-head helmet array. The behavioral pairing observed in simulations reappeared: Pearson correlation and the kernel estimator produced variable estimates with tight noise bounds, while the Kraskov and binning estimators yielded broader bounds and more uniform similarity across sensors. When the researchers ranked all channels by similarity rather than comparing absolute values, sensors closest to the reference ranked highest for every estimator, consistent with the spatial spread of magnetic sources. But the mutual information estimators additionally assigned high similarity to sensors capturing delayed or slightly distorted versions of the response, which Pearson correlation systematically under-ranked. The ranking-based approach proved essential, because all channels in a single recording share so much information, from common stimulus drive to shared preprocessing artifacts, that comparison against random noise alone provided limited insight.

The study is candid about the interpretive traps that remain. Mutual information is unbounded, and normalization strategies, including the widely used one adopted here, introduce their own upward biases at low values. The choice of reference signal matters enormously: because the reference channel in the case study contained a mixture of neural activity, noise, and artifacts, the mutual information estimators recognized similarities in other channels based on all of these factors, making it difficult to judge whether high similarity reflected genuine neural activation or merely shared residual artifacts such as eye blinks. The authors’ prescription is methodological humility: never interpret similarity values in absolute terms, always establish adaptive lower bounds tailored to the signal’s sample size and variance, and reason in terms of rankings. Their conclusion offers a practical decision rule for the field. When the goal is recognizing the exact same waveform across signals, Pearson correlation remains the simpler, efficient choice. When signals are expected to share features without being identical, the Kraskov estimator, which proved the easiest of the mutual information methods to tune, emerges as the recommended tool, potentially reshaping how connectivity, artifact removal, and stimulus-response analyses are conducted across EEG and MEG laboratories worldwide.

Subject of Research: Estimating mutual information and Pearson correlation for comparing neural evoked responses in EEG and MEG signals

Article Title: Estimating Mutual Information and Pearson Correlation on Neural Evoked Responses

Article References: Hukari, A., Cotroneo, S. F., & Salmelin, R. (2026). Estimating Mutual Information and Pearson Correlation on Neural Evoked Responses. Neuroinformatics, 24(4), Article 61. https://doi.org/10.1007/s12021-026-09784-3

Image Credits: AI Generated

DOI: 10.1007/s12021-026-09784-3

Keywords: mutual information, Pearson correlation, neural evoked responses, magnetoencephalography, electroencephalography, similarity estimators, signal processing, information theory, neuroimaging, time-series analysis, nonlinear dependence, statistics