Glossary

ABCDEFGHIJLMNOPRSTVWZ

Aperiodic component

Also: 1/f background, aperiodic exponent

The broadband, power-law-like background of a spectrum, described by an offset and an exponent, separate from any peaks.

First used in L1.7 · Aperiodic and periodic components

Artifact subspace reconstruction

Also: ASR

Sliding-window PCA method that rejects or reconstructs components whose variance exceeds a calibration-data threshold.

First used in L2.7 · ASR, SSP and alternatives

Average reference

Also: common average reference

Re-referencing each channel to the mean of all channels; reduces data rank by one.

First used in L2.3 · Re-referencing

B

Baseline correction

Also: baseline

Subtracting the mean of a pre-stimulus window from each epoch, or modelling it as a covariate.

First used in L3.1 · Epoching and baseline

BIDS

Also: EEG-BIDS, Brain Imaging Data Structure

A standard folder, file-naming and sidecar layout for neuroimaging datasets, with an EEG extension.

First used in L0.6 · Data hygiene and metadata

C

Chance band

Also: chance interval, binomial chance interval

The interval of accuracies a classifier with no information reaches by luck in n trials; an accuracy is evidence only when it clears it.

The interval of accuracies a classifier with no information would reach by luck in n trials, at a stated confidence — the binomial interval around 1/k for k classes. With twenty trials and two classes, 95 % of chance classifiers score between 0.30 and 0.70, so a reported 0.65 is not evidence of decoding; the band shrinks as the trial count grows. Every accuracy this site reports for a decoder is printed beside its band.

First used in L6.5 · Decoding as inference

Cluster-based permutation test

Also: cluster permutation, cluster test

Nonparametric test that controls family-wise error by comparing cluster-level statistics against a permutation distribution.

First used in L3.7 · Statistics for ERPs

Complex number

Also: complex coefficient, imaginary unit, complex exponential, Euler's formula

A number with a length and an angle; a Fourier coefficient is one, and its length and angle are its amplitude and phase.

A number with two parts, written a + bi, where i is defined by i² = −1. For this curriculum the algebra matters far less than the picture: a complex number is an arrow in a plane. Its length is its magnitude, its angle is its phase, and multiplying by e^{iθ} rotates it by the angle θ without changing its length.

That is why the Fourier transform is written with exp(−2πi k n / N). The exponential is a point marching around a circle at frequency k; multiplying the signal by it and adding up asks “how much of the signal keeps step with this rotation?” The answer is one arrow per frequency — the coefficient X[k] — whose length is how much of that frequency is present and whose angle is where in its cycle it starts.

How much you need here. That the coefficient has a magnitude and a phase, that |X[k]| means the magnitude, and that a rotation leaves length alone. You do not need to manipulate complex numbers algebraically, prove Euler’s formula, or know anything about complex analysis to follow this curriculum. Where a later lesson writes a(t)·e^{iφ(t)} (L4.3) or |(1/N)·Σ e^{iθ_k}| (L4.5), it is averaging arrows: the result is short when the angles disagree and long when they agree.

First used in L1.2 · Time and frequency domains

Convolution

Also: convolve, impulse response, filter kernel

Sliding one signal over another, multiplying and summing at each shift; filtering in time is convolution, and in frequency it is multiplication.

Take a short set of weights — the filter’s impulse response, or kernel — slide it along the signal, and at each position multiply the overlapping values and add them up. The result is the filtered signal. That is all a linear filter does: every output sample is a weighted combination of nearby input samples.

The fact the curriculum uses constantly is the convolution theorem: convolving in time is the same as multiplying in frequency. So a filter can be described two ways — by its kernel, or by the gain it applies to each frequency — and those two descriptions are a Fourier pair. The practical consequence is a trade-off you cannot escape: a sharp cut in frequency requires a long kernel, and a long kernel means ringing spread over time (ringing, L1.5). A short kernel is gentle in frequency but compact in time.

Morlet wavelet analysis (L4.2) is the same operation with a kernel that is a windowed oscillation, which is why time-frequency resolution trades off the same way.

How much you need here. That a filter is a weighted sum of nearby samples, that time-convolution equals frequency-multiplication, and that sharpness in one domain costs spread in the other. You are never asked to compute a convolution by hand.

First used in L1.5 · Filters: FIR, IIR, and what they do to your data

Correlation

Also: Pearson correlation, correlation coefficient, r

How closely two signals rise and fall together, from −1 to +1; near zero between neighbouring EEG channels usually means a bad electrode.

The Pearson correlation coefficient r runs from −1 (perfectly opposite) through 0 (unrelated) to +1 (perfectly together). It measures only linear agreement, and it is insensitive to scale: multiplying one signal by ten does not change it.

Its diagnostic use in EEG runs opposite to intuition. Because of volume conduction, neighbouring scalp electrodes must look alike — they are seeing overlapping mixtures of the same sources. So a channel that correlates poorly with its neighbours is the suspicious one: it is probably disconnected or dominated by its own noise, not uniquely informative. L2.2 builds a detector on exactly this.

Correlation is not causation, and in time series it is not even association in the way it looks: two signals that both drift slowly will correlate strongly for that reason alone, which is why correlations are computed in short windows or after filtering.

How much you need here. The −1 to +1 range, that it measures linear agreement only, and that low correlation with neighbours flags a bad channel rather than an interesting one.

First used in L2.2 · Channel locations and bad channels

Current source density

Also: CSD, surface Laplacian

Second spatial derivative of the scalp potential; reference-free and spatially sharpened.

First used in L5.3 · Surface Laplacian and CSD

Dipole

Also: equivalent current dipole

A point source–sink pair used to model the net current of a synchronously active cortical patch.

First used in L0.1 · What EEG measures

DPSS taper

Also: Slepian taper, discrete prolate spheroidal sequence

Orthogonal windows with maximal spectral concentration inside a chosen bandwidth, used in multitaper estimation.

First used in L4.3 · Multitaper and filter-Hilbert

ERP image

Single trials stacked as rows of a colour image, often sorted by a variable such as reaction time.

First used in L3.6 · Single-trial approaches

Evoked vs induced

Also: evoked power, induced power, total power

Evoked activity is phase-locked to an event and survives averaging; induced activity is not and only appears in time-frequency power.

First used in L4.1 · Why time-frequency

Forward model

Also: head model

The physics mapping from source currents to sensor measurements for a given head geometry and conductivity.

First used in L0.1 · What EEG measures

Frequency resolution

The spacing between independent spectral estimates, equal to 1 divided by the analysis window length.

First used in L1.2 · Time and frequency domains

H

Hilbert transform

Also: analytic signal

Transform producing the analytic signal, from which instantaneous amplitude and phase of a narrowband signal follow.

First used in L4.3 · Multitaper and filter-Hilbert

I

ICA

Also: independent component analysis

Blind source separation that unmixes channels into statistically independent components.

First used in L2.6 · EOG regression and ICA

Interpolation (spherical spline)

Also: interpolation, spherical spline

Estimating a bad channel's signal from its neighbours using spherical-spline weights.

First used in L2.2 · Channel locations and bad channels

ITC

Also: intertrial coherence, intertrial phase coherence, phase-locking factor

Consistency of phase across trials at a time-frequency point, from 0 (random) to 1 (identical).

First used in L4.5 · Phase, ITC and cross-frequency coupling

J

Jackknife

Leave-one-out resampling used, for example, to measure latencies on subsampled grand averages.

First used in L3.3 · Measuring ERPs

L

Leadfield

Also: gain matrix

The matrix that maps unit source currents at each location to sensor potentials.

First used in L5.4 · Forward models

Leakage (source)

Also: source leakage, spatial leakage

Spread of an estimated source's activity into neighbouring source locations due to limited inverse resolution.

First used in L5.6 · Source-space connectivity and leakage

Logarithm

Also: log, log scale, log axis, log-log, decade

The exponent a base must be raised to. On a log axis equal distances mean equal ratios, so a power law becomes a straight line.

log10(x) is the power of ten that gives x: log10(1000) = 3. The reason EEG plots use it constantly is that a log axis turns equal ratios into equal distances — 1 to 10 occupies the same space as 10 to 100. Since EEG power falls by orders of magnitude between 1 Hz and 40 Hz, a linear axis would show the low frequencies and nothing else.

Two consequences the curriculum depends on:

  • A power law is a straight line on log-log axes. If P(f) = k · f^(−x) then log P = log k − x · log f, which is a line with slope −x. This is the entire basis of the aperiodic fit in L1.7: the exponent is the slope.
  • The mean of the logs is not the log of the mean. Averaging in dB and averaging in power give different answers, which is why L1.3 insists you say which you did.

A decade is a factor of ten; an octave is a factor of two.

How much you need here. That log compresses ratios, that a straight line on log-log means a power law and its slope is the exponent, and that averaging before or after taking the log gives different numbers. No log identities beyond log(ab) = log a + log b are used.

First used in L1.3 · Power spectral density

O

Octave

Also: per octave, dB per octave

A doubling of frequency. A filter roll-off "6 dB per octave" loses half its amplitude each time frequency doubles.

One octave is a factor of two in frequency — 10 to 20 Hz is one octave, 10 to 40 Hz is two. It is borrowed from music, where doubling the frequency gives the same note an octave higher.

Filter and window roll-offs are quoted per octave because the decay is a constant ratio per doubling rather than a constant amount per hertz. “Falling at 6 dB per octave” means the amplitude halves each time the frequency doubles (6 dB ≈ a factor of 2 in amplitude); 18 dB per octave means it falls to an eighth. A steeper number is a faster decay, and in L1.4 it is the difference between a window whose side lobes hide a neighbouring rhythm and one whose side lobes do not.

How much you need here. Octave = doubling, and more dB per octave = steeper. See decibel for the dB scale itself.

First used in L1.4 · Windowing, leakage and zero-padding

P

p-value

Also: significance, null hypothesis, null-hypothesis test

The probability of seeing a result at least this extreme if nothing were going on; not the probability that your hypothesis is true.

The probability of obtaining a result at least as extreme as the one you got, assuming the null hypothesis is true — that is, assuming the effect you are looking for does not exist. A small p-value means the data would be surprising under that assumption.

What it is not, and these are the confusions the curriculum keeps returning to:

  • It is not the probability that the null hypothesis is true, nor that your hypothesis is false.
  • It is not a measure of effect size. With enough trials a trivial difference reaches any threshold you like, which is why this site asks for an effect size and an interval beside every p.
  • It is not meaningful once you have tested many times without accounting for it. Testing every electrode at every time point is thousands of tests, and the smallest p among them is small by construction — the subject of L6.1.
  • A threshold of 0.05 is a convention, not a fact about nature, and “p = 0.049 versus p = 0.051” is not a difference in kind.

How much you need here. What the number is conditional on, and the four things above that it does not mean. The distribution theory behind any particular test is not required by this curriculum; what is required is the judgment about when a p-value has been earned.

First used in L3.7 · Statistics for ERPs

PLI

Also: phase lag index, PLV / PLI / wPLI

Asymmetry of the phase-difference distribution around zero; discards zero-lag interactions.

First used in L5.2 · Sensor-space connectivity

PLV

Also: phase-locking value, PLV / PLI / wPLI

Mean resultant length of the phase difference between two signals across time or trials.

First used in L5.2 · Sensor-space connectivity

Power spectral density

Also: PSD

Power per unit frequency, typically in µV²/Hz, as a function of frequency.

First used in L1.3 · Power spectral density

Pre-registration

Also: preregistration, prereg

The outcome, test and control condition written down and committed before any data exist, so the analysis cannot be chosen after seeing the result.

The outcome, the test and the control condition written down and committed before any data exist, so the analysis cannot be chosen after seeing the result. On this site a pre-registration for a feedback protocol is a small file whose hash precedes every session log in version control; whether it did is a fact about the repository, not about the file, and the CRED-nf report says so rather than ticking the box.

First used in L6.4 · Circularity and analytic flexibility

R

Radian

Also: rad, 2 pi, phase units

The angle unit phase is measured in; a full cycle is 2π radians, not 360 degrees, and every formula and library here uses radians.

An angle measured as arc length on a unit circle. A full turn is 2π ≈ 6.283 radians, so:

cycleradiansdegrees
quarterπ/2 ≈ 1.57190°
halfπ ≈ 3.142180°
full2π ≈ 6.283360°

Every phase in this curriculum, and every phase returned by NumPy, SciPy and MNE, is in radians — np.angle returns a value in (−π, π]. A notebook line like np.pi / 2 is a quarter-cycle shift, and the comment sin = cos shifted by −pi/2 is the statement that a sine lags a cosine by a quarter cycle. Degrees appear nowhere, so a reader who learned angles in degrees should convert once and then stop thinking in them.

How much you need here. 2π is one cycle, π is half, π/2 is a quarter; phase differences are read as fractions of a cycle. Converting to degrees is never required.

First used in L1.2 · Time and frequency domains

S

Sham feedback

Also: sham condition, yoked sham, control feedback

Feedback that looks and moves like the real thing but carries no information about the participant's target signal — the control a neurofeedback claim needs.

Feedback that looks and moves like the real thing but carries no information about the participant’s own target signal. Four forms are named on this site: a yoked sham replays another recording’s feedback on the same schedule, controlling for the display, the task, time on task and expectation; a control-band sham computes real-time feedback from the same participant in a band the protocol does not target; an inverted sham reverses the contingency; and veridical is the experimental condition. Each controls for something different and fails to control for something else, and the sham mode is sealed into the session log so the participant can be blind to it.

First used in L7.3 · Real-time processing and neurofeedback

Sinusoid

Also: sine wave, cosine, oscillation

A pure oscillation described by exactly three numbers — amplitude, frequency and phase; the building block every spectrum decomposes a signal into.

A wave of the form x(t) = A · cos(2πf t + φ), fixed by three numbers and no others:

  • Amplitude A — its height, in microvolts for EEG.
  • Frequency f — cycles per second (hertz); 2πf t converts that to an angle that advances with time.
  • Phase φ — where in its cycle it starts, in radians.

The whole of Fourier analysis rests on one claim: any finite signal can be written as a sum of sinusoids, and the sum can be undone. A spectrum is the list of amplitudes, one per frequency; a complex coefficient carries the phase as well.

Worth keeping separate: a sinusoid is a mathematical object, while a brain rhythm is a transient, variable, non-sinusoidal thing that a sum of sinusoids can describe. L4.6 is about not confusing the description with the phenomenon.

How much you need here. The three parameters and what each does to the picture — which the L1.2 widget lets you drag directly.

First used in L1.2 · Time and frequency domains

Standard deviation

Also: SD, sigma, σ, robust SD

The square root of the variance, in the signal's own units; the usual yardstick for "how far from typical" a sample or a channel is.

The square root of the variance, so for EEG it is in microvolts. It answers “how far from the mean is a typical sample?” and is the unit in which thresholds are usually expressed — a rejection rule of “±5 SD” means five times this distance.

Two cautions the curriculum relies on:

  • It is not robust. One large artifact inflates the SD, which then raises the threshold that was supposed to catch the artifact. This is why several lessons use a robust SD, estimated from the median absolute deviation (MAD × 1.4826) instead, which a few extreme samples cannot drag.
  • It assumes a single population. An SD computed over a recording that is half clean and half artifact describes neither half.

How much you need here. That SD is spread in microvolts, that thresholds are often quoted in SDs, and that one big artifact corrupts an ordinary SD but not a robust one.

First used in L0.5 · The artifact atlas

Standardized measurement error

Also: SME

The standard error of an ERP measurement for one participant, estimated by bootstrapping trials.

First used in L3.5 · SNR, trial counts and design

T

TFCE

Also: threshold-free cluster enhancement

Cluster-style enhancement of a statistic map that integrates over all thresholds instead of one.

First used in L4.7 · Statistics for time-frequency

V

Variance

Also: var, spread

The mean squared deviation from the mean; the square of the standard deviation, and the usual first check that a channel carries signal at all.

The average of the squared distances between each sample and the mean. It measures how much a signal moves: a flat trace has variance near zero, a noisy one has large variance. Its square root, the standard deviation, is in the same units as the signal (microvolts) and is therefore the one usually quoted.

In EEG the first use is diagnostic rather than statistical. A channel whose variance is zero or near zero over a whole file is dead, disconnected, or is the reference itself — it is not a quiet brain. A channel whose variance is far larger than its neighbours’ is usually a bad electrode rather than an interesting one. L0.5 and L0.6 use it that way, and L2.2 builds a bad-channel detector on it.

How much you need here. That variance measures spread, that zero variance means no signal, and that comparing a channel’s variance with its neighbours’ is how you spot a bad one. The formula is never asked for.

First used in L0.5 · The artifact atlas

Volume conduction

Instantaneous spread of source currents through head tissues so that every sensor sees a mixture of sources.

First used in L0.1 · What EEG measures

W

Welch's method

Also: Welch

PSD estimation by averaging periodograms of overlapping, windowed segments.

First used in L1.3 · Power spectral density

wPLI

Also: weighted phase lag index, PLV / PLI / wPLI

PLI weighted by the magnitude of the imaginary cross-spectrum, reducing sensitivity to noise near zero lag.

First used in L5.2 · Sensor-space connectivity

Z

z-score

Also: standard score, robust z-score, standardised value

How many standard deviations a value sits from the mean; it makes channels or trials with different scales comparable.

z = (x − mean) / SD. It re-expresses a value as “how unusual is this, in units of the spread”, which lets quantities with different scales be compared on one ruler — the point of it in bad-channel detection, where each channel’s amplitude is z-scored against the distribution across all channels so that one threshold works for every channel.

A robust z-score replaces the mean with the median and the SD with a robust spread (MAD × 1.4826), for the reason given under standard deviation: the outliers you are hunting would otherwise inflate the yardstick you are hunting them with.

In feedback and decoding lessons a z-score also turns a raw band power into something comparable across sessions, which is why the baseline that defines “mean” and “SD” has to be stated — a z-score is only as meaningful as the baseline behind it.

How much you need here. That z counts standard deviations from typical, that ±2 to ±3 is the usual “unusual” range, and that a robust z uses the median and MAD so outliers cannot hide themselves.

First used in L2.2 · Channel locations and bad channels