Modulation Index for Phase-Amplitude Coupling

• Modulation Index (MI) is a quantitative measure for phase–amplitude coupling defined as the normalized Kullback-Leibler divergence between empirical and uniform amplitude-phase distributions.
• MI is computed through signal filtering, Hilbert transforms, and binning of phases to create an empirical distribution that is normalized to the unit interval [0,1].
• A closed-form Beta distribution serves as the null model, providing exact p-values and significance thresholds that bypass the need for computationally intensive surrogate simulations.

Searching arXiv for the specified paper and closely related phase–amplitude coupling / modulation index work to ground the article. Modulation Index (MI) is a quantitative measure for phase-amplitude coupling, a phenomenon in which the phase of one signal modulates the amplitude of another signal with a distinct frequency. In the framework of Pinto Orellana et al., MI is defined through the Kullback-Leibler divergence between a uniform distribution and the empirical conditional distribution of amplitudes with respect to phases, with the uniform distribution representing the ideal case expected under the absence of coupling. The same framework also provides exact statistical inference under a null hypothesis that an observed MI value can be entirely explained by chance, yielding a closed-form reference distribution and eliminating the need for surrogate-data sampling procedures (Pinto-Orellana et al., 2023).

1. Phase-amplitude coupling and the role of MI

Phase-amplitude coupling is observed in several neurological processes. Its operational form in this setting uses two real-valued time series: a low-frequency component from which an instantaneous phase is extracted, and a high-frequency component from which an instantaneous amplitude is extracted. The quantity of interest is whether the amplitude distribution depends on phase in a way that departs from uniformity.

Within this formulation, MI is a common technique used to quantify the interaction by assessing the Kullback-Leibler divergence between a uniform distribution and the empirical conditional distribution of amplitudes with respect to the phases of the observed signals. The uniform distribution is an ideal representation that is expected to appear under the absence of coupling. However, it does not reflect the statistical properties of coupling values caused by random chance. This distinction is central: the framework separates descriptive non-uniformity from inferential significance by introducing a null hypothesis of independence between signals.

A common source of confusion is to treat the raw MI value as sufficient evidence of coupling. The framework explicitly rejects that interpretation by requiring significance to be obtained through comparison with a reference distribution derived under the null hypothesis of independence, that is, no coupling between signals (Pinto-Orellana et al., 2023).

2. Mathematical definition

The construction begins with two real-valued time series, $x_{LF}(n)$ and $x_{HF}(n)$, from which one extracts the instantaneous phase $\phi_L(n)$ and instantaneous amplitude $A_H(n)$, respectively. The phase is discretized into $K$ bins, $\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$.

For each bin $k$, the empirical “pseudo-probability” is formed in two steps. First, one computes the mean amplitude within the bin,

$m_k = E[ A_H(n) \mid \phi_L(n)\in \text{bin } k ].$

Second, one normalizes these quantities so that they sum to one,

$P(\phi_k)=\frac{m_k}{\sum_{j=1}^K m_j}.$

Under a uniform (“no coupling”) distribution $U$, each bin has

$x_{HF}(n)$0

The Shannon entropy of $x_{HF}(n)$1, with logarithm base $x_{HF}(n)$2, is

$x_{HF}(n)$3

Because the uniform entropy is

$x_{HF}(n)$4

the normalized Kullback-Leibler divergence becomes

$x_{HF}(n)$5

An equivalent compact form is

$x_{HF}(n)$6

This definition fixes the range of MI to $x_{HF}(n)$7 through entropy normalization. In this representation, larger values correspond to greater divergence from the uniform phase-amplitude distribution, while MI equal to zero corresponds to maximal entropy of the empirical distribution.

3. Construction of the empirical phase-amplitude distribution

The empirical phase-amplitude distribution is built through a standard sequence of signal-processing operations. First, one filters or bandpass filters $x_{HF}(n)$8 around the low frequency band, computes its analytic signal via Hilbert transform, and extracts

$x_{HF}(n)$9

Second, one filters $\phi_L(n)$0 around the high frequency band, computes its analytic signal, and extracts

$\phi_L(n)$1

The phase circle $\phi_L(n)$2 is then partitioned into $\phi_L(n)$3 equal bins. For each bin $\phi_L(n)$4, one gathers those times $\phi_L(n)$5 for which $\phi_L(n)$6 falls in bin $\phi_L(n)$7. Writing

$\phi_L(n)$8

the mean amplitude in that bin is computed as

$\phi_L(n)$9

Normalization yields

$A_H(n)$0

and thus the discrete empirical amplitude-phase distribution

$A_H(n)$1

This construction is descriptive rather than inferential. It produces the empirical distribution on which MI is based, but by itself it does not determine whether the observed departure from uniformity is statistically significant. That inferential step requires a null model for MI itself.

4. Closed-form null model under no coupling

The inferential contribution of the framework is a closed-form null distribution for MI under the hypothesis that $A_H(n)$2 and $A_H(n)$3 are independent and each time series is white Gaussian noise. Under this null, the vector

$A_H(n)$4

can be approximated as a sample from a Dirichlet distribution with all parameters equal, so that each marginal

$A_H(n)$5

Consequently, the random entropy $A_H(n)$6 has an approximate Beta-distribution on $A_H(n)$7, and since $A_H(n)$8, MI itself is Beta-distributed on $A_H(n)$9 with swapped parameters. In particular, Theorem 2 gives

$K$0

with \begin{align*} a_{MI} &= \bigl(1-\mu_H\bigr)\,D_H,\ b_{MI} &= \mu_H\,D_H,\ D_H &= \frac{\mu_H - \mu_H^2}{\,\sigma_H^2 - 1\,}, \end{align*} where $K$1 are the mean and variance of the normalized entropy $K$2, themselves functions of $K$3 via auxiliary digamma-function expressions (Pinto-Orellana et al., 2023).

A more explicit form of the shape-parameter construction uses the intermediate quantities $K$4 and $K$5 with an $K$6-dependent correction, followed by \begin{align*} \alpha_p&=\mu_p(\mu_p-\mu_p^2)/(\sigma_p^2-\mu_p^2),\ \beta_p&=(1-\mu_p)(\mu_p-\mu_p^2)/(\sigma_p^2-\mu_p^2). \end{align*} Closed-form digamma-based expressions for $K$7 and $K$8 appear in Equations (15)–(18). In practice, one simply plugs in $K$9 and $\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$0 into the provided formulas to obtain $\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$1 and $\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$2.

The resulting null density of MI is

$\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$3

where $\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$4 is the Beta function.

5. Hypothesis testing, p-values, and critical values

Once the null model

$\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$5

has been specified, significance testing is one-sided. For an observed value $\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$6, the p-value is

$\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$7

where $\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$8 is the Beta cumulative distribution function.

For a desired significance level $\phi_k=(2\pi k/K,(2\pi(k+1)/K]]$9, the critical value $k$0 is defined by

$k$1

hence

$k$2

In code, this is a single call to the inverse-Beta-CDF with parameters $k$3 at the point $k$4. No Monte Carlo or surrogate generation is needed.

This inferential pipeline directly addresses the limitation that the uniform distribution alone does not reflect the statistical properties of coupling values caused by random chance. A raw MI value can therefore be interpreted only relative to the null distribution determined by the sample length $k$5 and bin count $k$6.

6. Surrogate methods, validation, and implications

The framework is presented in explicit contrast to traditional surrogate approaches. Such methods generate hundreds or thousands of artificial datasets, for example by permuting phases, shifting blocks, or resampling, and recompute MI each time to build an empirical null. This typically requires $k$7 Hilbert transforms and histogram computations, where $k$8 may be $k$9–$m_k = E[ A_H(n) \mid \phi_L(n)\in \text{bin } k ].$0. By contrast, the closed-form Beta approximation needs only the two inputs $m_k = E[ A_H(n) \mid \phi_L(n)\in \text{bin } k ].$1 to yield exact p-values and thresholds in $m_k = E[ A_H(n) \mid \phi_L(n)\in \text{bin } k ].$2 time, plus special-function evaluations for digamma and Beta inverses (Pinto-Orellana et al., 2023).

The paper identifies several consequences of this closed-form specification. Because the null is fully specified, different studies using different bin counts $m_k = E[ A_H(n) \mid \phi_L(n)\in \text{bin } k ].$3 and sample lengths $m_k = E[ A_H(n) \mid \phi_L(n)\in \text{bin } k ].$4 can compare MI values on common significance thresholds, facilitating meta-analysis. The method is also described as simple and computationally efficient, and it provides p-values and significance levels without resorting to generating surrogate data through sampling procedures. In addition, Monte Carlo surrogates may introduce biases, including edge artefacts and “phase-jumps,” and are sensitive to the surrogate-generation scheme; the closed-form null is free of these sampling artefacts.

Validation is reported through comprehensive Monte Carlo simulations assessing the significance of MI values under various experimental scenarios, including amplitude modulation, trains of spikes, and sequences of high-frequency oscillations. The framework is also corroborated by comparing its statistical significance thresholds with reported values from other research studies conducted under different experimental settings. This suggests that the method is intended not only as an analytical approximation, but as a practical inferential standardization scheme for phase-amplitude coupling studies.

A concise summary of the full pipeline is therefore: compute MI via the normalized KL divergence, derive its approximate null $m_k = E[ A_H(n) \mid \phi_L(n)\in \text{bin } k ].$5 from $m_k = E[ A_H(n) \mid \phi_L(n)\in \text{bin } k ].$6, obtain p-values or critical values by standard Beta-CDF or inverse-CDF calls, and avoid large surrogate simulations. Within the stated assumptions of independence and white Gaussian noise, MI is thus elevated from a descriptive coupling index to a statistically calibrated test quantity.