Diffusion Entropy Analysis (DEA)

• Diffusion Entropy Analysis (DEA) is an entropy‐based method that transforms a time series into a diffusion process to derive scaling exponents.
• The approach estimates the full probability density of diffusion displacements, avoiding misleading variance measures in anomalous regimes.
• DEA is widely applied to detect non-Gaussian transport phenomena, as demonstrated in studies like Voyager-I magnetic field fluctuation analysis.

Diffusion Entropy Analysis (DEA) is an entropy-based method for detecting scaling in time-series data by converting a sequence into a diffusion process, estimating the probability density of diffusion displacements, and inferring the scaling exponent from the growth of Shannon entropy with $\ln t$. In the DEA framework, scaling is not read from a single moment such as the variance, but from the full diffusion-generated probability density $p(x,t)$. This is the central reason DEA is used in contexts where Gaussian assumptions fail, where second moments are misleading or undefined, or where anomalous transport is of interest (Haubold et al., 2012).

1. Definition and formal construction

DEA treats the numbers in a time series as the trajectory of a diffusion process. Let the original time series be $\{\xi_i\}$. The standard construction forms moving partial sums, called sub-trajectories,

$X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$

as reproduced in the Voyager-I application, where the notation is printed with a slight inconsistency between $l$ and $t$ but clearly denotes the usual windowed accumulation of increments into diffusion displacements (Haubold et al., 2012).

From the ensemble of such sub-trajectories one estimates the diffusion probability density $p(x,t)$, where $x$ is the diffusion variable and $t$ is the window length or diffusion time. The fundamental scaling ansatz is

$$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$

with $p(x,t)$0 the DEA scaling exponent and $p(x,t)$1 a scaling function. DEA then computes the Shannon entropy

$p(x,t)$2

Under the scaling form above, one obtains

$p(x,t)$3

The scaling exponent $p(x,t)$4 is therefore determined as the slope of the fitted straight line in the linear-log plot of $p(x,t)$5 over time $p(x,t)$6 (Haubold et al., 2012).

A closely related discrete formulation is used in tutorial expositions of DEA. Starting from a time series $p(x,t)$7, one forms a cumulative trajectory

$p(x,t)$8

then windowed displacements

$p(x,t)$9

constructs the empirical distribution $\{\xi_i\}$0 of these displacements, and evaluates the discrete Shannon entropy

$\{\xi_i\}$1

The scaling relation is then written as

$\{\xi_i\}$2

with $\{\xi_i\}$3 again estimated from the slope (Culbreth et al., 2023).

2. Entropy growth as a diffusion measure

DEA rests on the proposition that entropy growth can quantify diffusion speed more robustly than variance. A general theoretical treatment distinguishes two asymptotic regimes. For systems with no stationary density, entropy grows logarithmically in time and the coefficient of $\{\xi_i\}$4 specifies the speed of diffusion. For systems with a stationary large-time density, the appropriate relative entropy approaches its limiting value exponentially fast rather than growing indefinitely (Aghamohammadi et al., 2013).

In the nonstationary case, the general logarithmic law is

$\{\xi_i\}$5

or equivalently

$\{\xi_i\}$6

when the evolution is asymptotically scale invariant under

$\{\xi_i\}$7

In one-dimensional self-similar form, this reduces exactly to the standard DEA relation

$\{\xi_i\}$8

The coefficient of $\{\xi_i\}$9 is thus the scaling exponent controlling spreading, and in ordinary diffusion in $X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$0 dimensions one gets $X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$1 as the entropy-growth coefficient (Aghamohammadi et al., 2013).

This formulation clarifies why DEA differs from width-based diagnostics. Variance can diverge, remain constant, depend strongly on the derivative definition in fractional models, or be ill-defined on compact spaces. Entropy, by contrast, is computed from the full density $X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$2. In DEA language, variance measures the growth of a single moment, whereas entropy measures the growth of the full support or scale of the distribution. This suggests that entropy scaling is often the more robust indicator in anomalous, non-Gaussian, or heavy-tailed regimes (Aghamohammadi et al., 2013).

3. Relation to variance methods, Hurst scaling, and anomalous diffusion

DEA is often paired with Standard Deviation Analysis (SDA), which tracks the scaling of the second moment of the same diffusion process: $X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$3 where $X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$4 is the Hurst exponent or standard deviation scaling exponent. The comparison between $X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$5 and $X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$6 is used diagnostically rather than descriptively (Haubold et al., 2012).

Three benchmark cases are stated explicitly. For fractional Brownian motion,

$X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$7

For random noise with finite variance, the central limit theorem implies Gaussian diffusion,

$X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$8

If $X_n(t)=\sum_{i=0}^{l}\xi_{i+n}, \;\; n=0,1,\ldots$9 differs from $l$0, the scaling is interpreted as anomalous. Within anomalous diffusion, the distinction made is the following. For Lévy flights, the scaling form

$l$1

can still hold, but the variance is infinite, so the variance scaling exponent is not defined. For Lévy walks, the second moment remains finite and the exponents satisfy

$l$2

Agreement $l$3 is thus consistent with fractional-Brownian-type scaling, while disagreement together with the Lévy-walk relation is taken as evidence of non-Gaussian anomalous transport (Haubold et al., 2012).

A theoretical treatment of fractional-derivative diffusions sharpens this point. For the class

$l$4

the entropy law becomes

$l$5

For normal diffusion, $l$6 and $l$7, yielding

$l$8

For fractional time derivative diffusion,

$l$9

while for one-dimensional fractional space derivative diffusion,

$t$0

The same paper emphasizes that in Lévy-flight-type cases the variance may diverge at every positive time while the entropy still has a clean asymptotic law, precisely the situation for which DEA was introduced (Aghamohammadi et al., 2013).

4. Empirical use: Voyager-I magnetic-field fluctuations

A concrete DEA application is provided by the analysis of Voyager-I heliospheric magnetic field strength fluctuations. The data are daily averages of Voyager-I magnetic field strength fluctuations in the heliosphere, illustrated for 1989 at about 45 AU and discussed in connection with earlier Voyager-I studies at 40 AU and 85 AU. DEA and SDA are applied jointly to determine whether the stochastic behavior is closer to Gaussian diffusion or to Lévy-type anomalous diffusion (Haubold et al., 2012).

The principal reported exponents are

$t$1

Both are markedly larger than $t$2, which is used as the first argument against ordinary Gaussian diffusion. The second argument is the mismatch between the exponents,

$t$3

which excludes the simple fractional Brownian benchmark. The third argument is that the pair satisfies the Lévy-walk relation

$t$4

Substituting $t$5 gives $t$6, matching the reported DEA estimate very closely. On this basis the fluctuations are interpreted as anomalous and Lévy-type rather than Gaussian (Haubold et al., 2012).

The same study introduces a general fractional-order spatial and temporal diffusion equation,

$t$7

together with special cases including neutral fractional diffusion, space-fractional diffusion, time-fractional diffusion, and the classical diffusion equation. The classical limit

$t$8

has the Gaussian solution

$t$9

The paper does not derive a closed-form mapping from the measured $p(x,t)$0 to $p(x,t)$1; instead, the fractional model is proposed as a possible interpretation of the observed non-Gaussian scaling. This suggests a qualitative link: DEA indicates Lévy-type anomalous transport, and fractional diffusion provides a model class consistent with that evidence (Haubold et al., 2012).

The same work situates DEA within a nonextensive statistical-mechanics context. It refers to earlier Voyager-I analyses that led to a nonextensivity $p(x,t)$2-triplet

$p(x,t)$3

with the theoretical inequalities

$p(x,t)$4

No numerical $p(x,t)$5-triplet values are reported there; rather, the DEA/SDA results are presented as supporting the earlier conclusion of nonextensivity (Haubold et al., 2012).

5. Methodological extensions: modified, multifractal, and generalized-entropy DEA

A later methodological development modifies DEA by introducing an event-detection stage before diffusion construction. In this modified DEA, the original time series is not used directly as the diffusion trajectory. Instead, the data are partitioned into stripes. With data range

$p(x,t)$6

an event is registered whenever the scaled signal crosses a stripe between times $p(x,t)$7 and $p(x,t)$8, using the checks

$p(x,t)$9

If either condition fails, one sets $x$0; otherwise $x$1. The modified diffusion trajectory is then the cumulative event count

$x$2

DEA proceeds on $x$3 in the usual way via windowed displacements, histograms, and entropy regression (Culbreth et al., 2023).

The purpose of this modification is to emphasize event timing over amplitude. The reported claim is that standard DEA is significantly affected by fractional Brownian motion contributions, whereas modified DEA filters out existing fractional Brownian motion contributions to the scaling more effectively and especially improves detection of events with power-law distributed waiting times. The paper identifies two unresolved methodological issues: rigorous selection of the fitting interval and principled choice of the number of stripes $x$4 (Culbreth et al., 2023).

DEA has also been extended to a multifractal setting by replacing Shannon entropy with the family of Rényi entropies. Starting from a stationary time series $x$5, the Fluctuation Collection Algorithm constructs overlapping sums

$x$6

which are binned into histograms $x$7. The differential Rényi entropy of the diffusion density is

$x$8

In the multiscale setting discussed there, regression of $x$9 against $t$0 yields a $t$1-indexed scaling spectrum $t$2, operationally interpreted as a multifractal $t$3-spectrum (Jizba et al., 2014).

That extension is accompanied by a practical result on optimal histogram bin widths. Since DEA and its multifractal generalization estimate densities from histograms of fluctuation sums, poor binning can distort the entropy curve and the estimated scaling spectrum. The asymptotic mean integrated squared error for the $t$4-histogram is derived as

$t$5

leading to $t$6-dependent Scott- and Freedman–Diaconis-type width rules. The S&P500 example is used to show that poor binning materially changes the estimated $t$7-spectrum, especially for large $t$8 (Jizba et al., 2014).

A distinct generalization replaces Shannon entropy with Mathai’s additive generalized entropy,

$t$9

For a scaling diffusion process,

$$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$0

the generalized DEA law becomes

$$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$1

The key point is that the coefficient of $$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$2 remains $$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$3, independent of the generalized entropy parameter $$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$4. The parameter changes only the additive constant, not the slope extracted by DEA (Sebastian, 2014).

6. Scope, assumptions, and relation to adjacent entropy-based methods

DEA assumes that the diffusion-generated density has a scaling structure of the form

$$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$5

or, in nonstationary extensions,

$$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$6

When that assumption is violated, or when the system has a stationary asymptotic density, the characteristic DEA law $$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$7 should not be expected. In stationary linear systems, the appropriate relative entropy instead approaches zero exponentially,

$$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$8

with $$p(x,t)=\frac{1}{t^\delta}F\!\left(\frac{x}{t^\delta}\right),$$9 the slowest decaying nonzero eigenvalue of the generator (Aghamohammadi et al., 2013).

Several empirical cautions recur in the literature. The Voyager-I study does not report uncertainties for $p(x,t)$00 or $p(x,t)$01, fitting intervals in $p(x,t)$02, robustness tests, explicit probability-density fits, surrogate-data comparisons, or stationarity tests (Haubold et al., 2012). The modified-DEA tutorial identifies fitting-interval selection and stripe-count choice as major unresolved problems (Culbreth et al., 2023). The multifractal extension shows that histogram choice is not a secondary implementation detail, because non-optimal widths can visibly distort $p(x,t)$03 versus $p(x,t)$04 plots and the resulting $p(x,t)$05-spectrum (Jizba et al., 2014).

DEA is also distinct from several later entropy methods that employ the word “diffusion” in different senses. Diffusion spectral entropy, for example, is not classical DEA; it computes Shannon-style entropy over normalized eigenvalues of a graph diffusion operator,

$p(x,t)$06

and is interpreted as a noise-resilient measure of intrinsic manifold complexity rather than as an estimator of transport scaling from displacement densities (Liao et al., 2023). Likewise, recent work on generative diffusion models tracks class-conditional entropy

$p(x,t)$07

as a signature of semantic uncertainty over diffusion time. That framework is conceptually related to DEA in its use of entropy trajectories, but it is not classical DEA because the entropy is over semantic labels conditioned on noisy states rather than over the diffusion-state density itself (Handke et al., 10 Feb 2026).

A plausible implication is that DEA is best understood as a specific member of a broader family of entropy-versus-diffusion-time diagnostics. What distinguishes DEA within that family is the particular object whose entropy is measured: the probability density of diffusion displacements generated from a time series, together with the use of the slope in

$p(x,t)$08

as the operational scaling estimator.