In this paper
AbstractIntroductionThe AGN sampleMethodsMutual Information Lag EstimationMutual Information as a lag estimatorKSG estimatorLight-curve pre-processingRandom Subset Selection bootstrapResults and DiscussionNGC 5548NGC 4151NGC 4593Mrk 509Band behaviour across the sampleComparison with JAVELIN and ICCFComparison with the multifractal characterisationConclusionsAcknowledgementsData AvailabilityTime Series for each AGNReferencesWorksPhysics and Astronomy
Time-Delay Recovery in AGN Reverberation Mapping via Mutual Information
Abstract
Active galactic nuclei exhibit complex, nonlinear multiwavelength variability that standard cross-correlation techniques, sensitive only to linear covariance, may fail to capture fully. We apply time-lagged Mutual Information (MI) lag estimation, via the Kraskov-Stoegbauer-Grassberger -nearest-neighbour estimator () with a 150-iteration Random Subset Selection bootstrap, to the light curves of four Seyfert galaxies (Mrk 509, NGC 5548, NGC 4151, and NGC 4593) monitored across eight energy bands by Swift. In the UV/optical regime the MI estimator successfully reproduces the inter-band disc reprocessing lags established by the interpolated cross-correlation function, with performance degrading toward the optical where low variability amplitude produces poorly defined MI peaks. In the X-ray bands, where the physical X-ray/UV coupling is known to be weak, the estimator fails to recover consistent lags across the sample, with the notable exception of NGC 4593. Contextualising these outcomes within the multifractal characterisation of our previous work, we discuss how MI performance relates to the multifractal spectral width of each band, and we propose as a candidate diagnostic for anticipating where MI-based reverberation mapping is likely to succeed or fail in multiband datasets of active galactic nuclei.
Introduction
Active galactic nuclei (AGNs) are among the most energetically extreme objects in the Universe, powered by accretion of matter onto supermassive black holes at the centres of their host galaxies (e.g. Peterson 1997; Netzer 2013). Their emission spans the entire electromagnetic spectrum, from radio waves to hard X-rays (Ulrich et al. 1997; Shang et al. 2011), and varies on timescales ranging from hours to years (Vaughan et al. 2003). This variability encodes fundamental information about the geometry, kinematics, and physical conditions of the emitting regions surrounding the central engine (Ulrich et al. 1997). Understanding these regions, in particular the accretion disc and the broad-line region (BLR), is crucial for constraining models of black hole growth and AGN feedback across cosmic time (Kormendy & Ho 2013; Heckman & Best 2014).
Reverberation mapping (Blandford & McKee 1982; Peterson 1993) is one of the most powerful observational techniques for probing the inner structure of AGNs. Variations in the ionising continuum emission propagate outward and are reprocessed by surrounding gas and dust, producing delayed echoed signals arising from the finite travel time of light across the system, at longer wavelengths. By measuring the time lag, , between a reference and a subsidiary band, one can infer the light-travel distance to the reprocessing region, mapping the spatial structure of the AGN on scales inaccessible to direct imaging (Cackett et al. 2021). In the scenario of thermal reprocessing of extreme UV/X-ray photons to optical/UV photons by the accretion disc (Cackett et al. 2007), the observed lags follow the characteristic wavelength-dependent scaling , a direct prediction of the standard thin-disc model (Shakura & Sunyaev 1973). This scenario is supported by X-ray reverberation lags observed in the Fourier domain of individual Seyfert galaxies (Uttley et al. 2014; Kara et al. 2016; Fabian et al. 2009), and by dedicated multiwavelength monitoring campaigns spanning X-ray, UV, and optical bands simultaneously (Edelson et al. 2015, 2017, 2019; McHardy et al. 2018a). These campaigns have provided tight constraints on the accretion disc structure and the geometry of the corona, while also revealing discrepancies between observed lag amplitudes and standard thin-disc predictions, pointing towards a contribution from the BLR and diffuse continuum emission to the observed lag amplitudes (Cackett et al. 2018).
The standard approach to measuring inter-band lags relies on the interpolated cross-correlation function (ICCF; Gaskell & Peterson 1987; Peterson et al. 1998) or its discrete counterpart (DCF; Edelson & Krolik 1988). More sophisticated methods, such as JAVELIN (Zu et al. 2011), model the light curves as damped random walk processes and infer lags through maximum-likelihood fitting. While these approaches have yielded significant results, they share a common limitation: they are fundamentally based on second-order statistics, i.e., the linear covariance between bands. AGN Light Curves, however, are inherently non-Gaussian and nonlinear (Vio et al. 1992; Longo et al. 1996; Provenzale et al. 1994; Green et al. 1999), and exhibit complex temporal behaviour that includes intermittency (Chian [12]) and non-linear variability signatures incompatible with simple stochastic models (Vaughan et al. 2003).
In particular, their variability has been shown to exhibit multifractal scaling in lensed quasars (Belete et al. 2018; Bewketu Belete et al. 2019b,a; Souza et al. [49]; Souza et al. [50]) and in multiband Seyfert emission (Souza et al. [51]). The central measure in multifractal analysis is the width of the Hölder exponent distribution, , a positive quantity that is related to monofractality when , or to the presence of multiple fractal structures, i.e., multifractality, when deviates from zero. Each Hölder exponent is derived from the local scaling relation between the logarithm of the fluctuation function , analogous to a generalised variance, and the logarithm of the corresponding time scale . By varying the order , which controls the relative weight given to large and small fluctuations, one obtains a family of scaling exponents that, through a Legendre transform, map onto the corresponding values composing the multifractal spectrum. When this scaling exponent is approximately independent of , the spectrum collapses to a narrow range and tends to, or equals, zero, indicating monofractal behaviour; when the scaling exponent instead varies with , multiple Hölder exponents emerge, characterising a multifractal behaviour (Kantelhardt et al. 2002; Souza et al. [51]; Souza et al. [52]). For , is dominated by large fluctuations, whereas for it is dominated by small fluctuations. A left-skewed spectrum indicates dominance of small fluctuations, a right-skewed spectrum indicates dominance of large fluctuations, and a symmetric spectrum indicates comparable contributions from both. This asymmetry is quantified by the parameter , with , , and corresponding respectively to left-skewed, right-skewed, and symmetric spectra (Ihlen [25]).
In parallel, information theory, originally formulated by Shannon (1948) in the context of communication systems, provides a natural framework for characterising statistical dependence in a model-free, nonlinear fashion. Unlike second-order statistics, information-theoretic measures capture any form of dependence between two variables, linear or nonlinear, without assuming a specific functional form or a particular noise model (Cover & Thomas 2006). Practical estimation from finite, irregularly sampled data is made tractable by nonparametric density estimators that require no binning or bandwidth choice. Using such a measure as a lag estimator, by maximising it as a function of a trial time shift , yields a natural generalisation of the Cross-Correlation Function (CCF) that imposes no linearity assumption.
In the context of astronomical time-delay estimation, Chakraborty & van Leeuwen (2022) introduced a Discrete Mutual Information Function (DMIF) as a nonlinear alternative to the discrete correlation function, and used toy models of accreting sources, including X-ray binaries and AGNs, to demonstrate that the DMIF recovers time lags in regimes where the linear DCF is biased or fails outright. The present work can be regarded as a direct extension of that proof-of-concept to real, multi-band AGNs light curves: rather than synthetic data, we apply a Kraskov-Stoegbauer-Grassberger (KSG) -nearest-neighbour MI estimator to the Swift Infrared Detected Reverberation Mapping (IDRM) light curves of four Seyfert AGNs, benchmarking the recovered lags against the established ICCF and JAVELIN estimates of Edelson et al. (2019), providing an observational test of the DMIF approach advocated by Chakraborty & van Leeuwen (2022). Chelouche et al. (2017) explored randomness measures as a further alternative reverberation mapping estimator. More recently, Gaussian-process-based cross-correlation (Pozo Nunez et al. 2023) and hierarchical Bayesian frameworks (Brewer & Elliott 2014) have further expanded the methodological landscape. Despite its broad applicability, MI-based lag estimation in AGN reverberation mapping remains largely unexplored.
The present work addresses this gap directly. We apply time-lagged MI estimation, using the Kraskov-Stoegbauer-Grassberger (KSG; Kraskov et al. 2004) -nearest-neighbour estimator, as implemented in the scikit-learn library, to the multiwavelength light curves of four well-studied Seyfert galaxies monitored simultaneously across eight energy bands as part of the first Swift intensive AGN accretion disc reverberation mapping survey (Edelson et al. 2019), which extended earlier pilot campaigns on NGC 5548 (Edelson et al. 2015) and NGC 4151 (Edelson et al. 2017). Black hole masses are drawn from the AGN Black Hole Mass Database (Bentz & Katz 2015). The estimated lags are compared against the ICCF and JAVELIN values of Edelson et al. (2019).
Our analysis is further contextualised within the multifractal and complexity characterisation of the same dataset performed by Souza et al. [51], who applied Multifractal Detrended Moving Average analysis (Gu & Zhou 2010) and the Fisher-Shannon information plane (Vignat & Bercher 2003) to the identical light curves, finding that all bands exhibit persistent long-range correlations, meaning that an increase (decrease) in the signal is more likely to be followed by another increase (decrease), whereas an antipersistent signal would tend to alternate between increases and decreases (Hurst exponent , where corresponds to uncorrelated noise, to persistence, and to anti-persistence), alongside a superlinear scaling between multifractal width and fractional variability across all four sources.
This paper is organised as follows. § 2 describes the observational dataset and the fundamental properties of the four AGNs. § 3 presents the methodology, covering the time-lagged MI estimator, the KSG -nearest-neighbour scheme, the light-curve pre-processing pipeline, and the Random Subset Selection bootstrap (Efron 1979). The results and their comparison to the results of other methods are presented in § 4, including the MI lag curves for all sources, a comparison with the ICCF and JAVELIN values of Edelson et al. (2019), and a discussion contextualised within the multifractal signatures of Souza et al. [51]. § 5 summarises our main conclusions.
The AGN sample
This study investigates the variability properties of four well-studied AGNs, Mrk 509, NGC 5548, NGC 4151, and NGC 4593, each observed by Swift in eight energy bands within a single monitoring epoch as part of the intensive disc reverberation mapping survey Edelson et al. (2019). An example of these light curves is shown for NGC 5548 in Figure 1. The sample consists of light curves with a minimum of 150 data points, meeting the minimum requirement for the method used here. It is restricted to sources observed in the same eight photometric bands to ensure consistent cross-source comparison.

Figure 1. Light curves of NGC 5548 across all eight observational bands, ordered by wavelength from top to bottom. The top two panels show XRT observations in the HX (1.5,10 keV) and SX (0.3,1.5 keV) bands, while the lower six panels correspond to the Swift/UVOT filters W2 ( Å), M2 ( Å), W1 ( Å), U ( Å), B ( Å), and V ( Å). Each panel displays flux as a function of MJD, with a shared time axis and independent flux scales. X-ray fluxes are given in counts s, and UVOT fluxes in units of erg cm s Å.
The multiwavelength variability of these AGNs reflects distinct physical mechanisms. In the Seyfert 1 galaxy Mrk 509 (Phillips et al. 1983), a direct correlation between soft X-ray and UV flux variations has been reported, interpreted as variability intrinsic to the AGN, driven by changes in the ionising continuum illuminating the surrounding gas (Kaastra [26]). A companion multifractal analysis of the same light curves (Souza et al. [51]) finds that all bands of Mrk 509 except B and V are multifractal, with the complexity most pronounced in the X-ray bands ( for Soft X-ray (SX) and for Hard X-ray (HX)), suggesting that the innermost regions of the accretion flow drive the most complex variability in this source. NGC 5548, among the most thoroughly studied Seyfert 1 galaxies since its original classification by Seyfert (1943), has been monitored since the 1960s (Mehdipour [35]); rapid X-ray spectral variations have been linked to changes in the ionisation state of intervening absorbers, while infrared coronal line variability points to an X-ray-heated wind as a relevant energy-transport mechanism ([35, 31]). Souza et al. [51] report that NGC 5548 shows the broadest multifractal spectrum in the full sample in SX (), the only band with a near-antipersistent Hurst exponent (), alongside strong right-skewed multifractality across its UV bands ().
For the Seyfert 1 galaxy NGC 4151 (Zoghbi et al. 2019), long-baseline observations have revealed asymmetric Balmer line profiles, indicating a BLR with multiple kinematically distinct components (Shapovalova [48]), while the historical B-band light curve shows nonlinear, intermittent behaviour consistent with complex underlying variability mechanisms (Longo et al. 1996). Notably, Souza et al. [51] finds that NGC 4151 is the only source in the sample for which a single band, HX, exhibits clear multifractal behaviour, with and a left-skewed spectrum (), indicating that the variability is dominated by large-amplitude fluctuations corresponding to stronger singularities with smaller Hölder exponents, while all other bands are consistent with near-monofractal behaviour (). The Seyfert 1 galaxy NGC 4593 (Cackett et al. 2018) showed UV variability over a ten-month baseline, attributed to fluctuations in the intrinsic AGN continuum (Ward [60]); simultaneous Swift monitoring revealed correlated variability across the X-ray, UV, and optical bands, in agreement with X-ray reprocessing by extended regions of the accretion disc or BLR (McHardy et al. 2018b). For NGC 4593, Souza et al. [51] report multifractality concentrated in the X-ray bands ( for HX and for SX), with a strongly right-skewed W2 spectrum () and near-monofractal B and V bands.
The IDRM strategy improves on earlier reverberation-mapping campaigns by incorporating six UV/optical filters, a higher sampling cadence, and substantially more observations, enhancing sensitivity to short inter-band lags. It was first applied to NGC 5548 and later extended to additional targets (Edelson et al. 2015; McHardy et al. 2018a; Edelson et al. 2019), with simultaneous HST and ground-based coverage.
The light curves show strong variability in all UVOT bands and in HX, with more moderate variability in SX. The IDRM cadence adequately tracks UV/optical variations, whereas X-ray fluctuations occur on shorter timescales and may be partially undersampled (Edelson et al. 2019). The fundamental properties of the four AGNs, redshift, black hole mass, Eddington ratio, and cadence are summarised in Table 1.
| Name | Visits | Duration (days) | Cadence (days) | |||
| Mrk 509 | 0.0341 | 8.05 | 5% | 257 | 272.6 | 1.065 |
| NGC 5548 | 0.0163 | 7.72 | 5% | 291 | 127.6 | 0.440 |
| NGC 4151 | 0.0032 | 7.56 | 1% | 322 | 69.3 | 0.216 |
| NGC 4593 | 0.0083 | 6.88 | 8% | 194 | 22.6 | 0.117 |
Table 1. Table 1. Fundamental properties of the selected AGNs, including redshift (), black hole mass (), Eddington ratio (), campaign time span, total duration, number of visits, and mean sampling cadence. Redshifts are drawn from the SIMBAD database. Black hole masses are adopted from the AGN Black Hole Mass Database (Bentz & Katz 2015). The Eddington ratio is defined as . Campaign duration is expressed in days, the number of visits corresponds to all observations with usable data in at least one band, and the mean cadence is derived from the ratio of campaign duration to number of visits.
Methods
Information theory, originally formulated by Shannon (1948) in the context of communication systems, provides a natural framework for characterising statistical dependence in a model-free, nonlinear fashion. The central quantity is the Shannon entropy,
where is the probability density function of , and measures the average uncertainty in the random variable (Cover & Thomas 2006). From this foundation, one constructs the Mutual Information (MI) between two random variables and (Eq. (2)). Crucially, MI vanishes if and only if and are statistically independent, and captures any form of dependence, linear or nonlinear, without assuming a specific functional form (Cover & Thomas 2006). MI has been widely applied in time-series analysis across diverse fields, including neuroscience (Kraskov et al. 2004; Wibral et al. 2014), climatology and geophysical signals (Telesca et al. 2011; Telesca & Lovallo 2011), and financial markets (Fiedor 2014). The -nearest-neighbour method of Kraskov et al. (2004) is widely adopted for this purpose and is available via the scikit-learn library. Using MI as a lag estimator, by maximising the MI as a function of a trial time shift, , is a natural generalisation of the CCF that requires no linearity assumption.
To characterise temporal dependencies between the multi-band light curves of our four Seyfert galaxies, we employ a sequential, self-consistent analysis implemented as a standalone Python pipeline. We estimate the inter-band time lag by maximising the time-lagged MI as a function of trial lag, with uncertainties derived from a Random Subset Selection bootstrap. The pipeline relies on numpy, scipy, and matplotlib, with the MI estimation step performed by the -nearest-neighbour mutual information estimator (Kraskov et al. 2004) implemented in the scikit-learn library.
Mutual Information Lag Estimation
Mutual Information as a lag estimator
The MI between two continuous random variables and is defined as (Cover & Thomas 2006):
where is the joint probability density function (PDF) and , are the marginal PDFs. Practical computation from finite data requires a non-parametric estimator of the densities entering Eq. (2). Higher MI values indicate a stronger dependence between the two variables, whereas MI becomes zero if and only if and are statistically independent. Unlike the Pearson correlation coefficient, MI captures any statistical dependence, linear or non-linear, making it sensitive to the full structure of the joint variability between two AGN Light Curves.
For a pair of light curves, we define the time-lagged MI curve as
where is the reference band and is the subsidiary band. In this convention, a positive lag corresponds to the reference band leading the subsidiary band, which is the expected sign for disc reprocessing, in which the UV continuum leads the longer-wavelength optical response (Cackett et al. 2007). The lag that maximises is taken as the estimated inter-band time delay, as that displacement maximises the statistical information shared between the two series.
The shift is implemented by linear interpolation: for each trial lag , both light curves are first resampled onto a common uniform time grid spanning their temporal overlap, and the subsidiary series is then evaluated on that grid after an additional shift of using scipy.interpolate.interp1d. This produces the pair on the common grid, which is then passed to the MI estimator. This manual resampling is required because the -nearest-neighbour MI estimator has no native lag parameter and must be evaluated independently for each trial .
KSG estimator
The PDFs in Eq. (2) are unknown and must be estimated from the data. We adopt the Kraskov-Stoegbauer-Grassberger (KSG) -nearest-neighbour estimator ((Kraskov et al. 2004)), as implemented in the mutual_info_regression function of the scikit-learn library. For each observation , the distance to its -th nearest neighbour is computed in the joint space using the maximum norm. The marginal neighbour counts and are then defined as the number of points whose distance from the -th point is strictly smaller than in the one-dimensional and subspaces, respectively. MI is estimated as:
where is the digamma function, , and is the number of data points. We write explicitly as a function of the time lag , since is shifted by relative to before the joint and marginal neighbour counts are computed, and of , the number of nearest neighbours used in the estimator. We use nearest neighbours, the value recommended by Kraskov et al. (2004) for moderate sample sizes. The KSG estimator is asymptotically unbiased, requires no bandwidth or binning choice, and is well suited to the moderately sized, normalised continuous series produced by the light-curve pre-processing described below.
Light-curve pre-processing
Each light curve is sorted by observation time and standardised to zero mean and unit variance.
Following Edelson et al. (2019), we adopt the W2 band ( Å) as the reference for all inter-band lag measurements, as it provides the shortest-wavelength UVOT continuum reference with high variability amplitude and dense temporal coverage across all four campaigns. For each object, the reference band and subsidiary light curves are linearly interpolated onto a common uniform time grid with spacing covering their temporal overlap; grid points at which either interpolant is undefined are flagged as NaN and excluded via a finite-value mask. Trial lags for which fewer than 20 valid common points remain are excluded from the MI() curve entirely, as the KSG estimator is unreliable at such sample sizes.
Because the four campaigns differ substantially in total duration and mean cadence (Table 1, spanning from 23 d for NGC 4593 to 273 d for Mrk 509), the trial-lag range, lag step, and grid spacing are tuned for each object to match the relevant timescales while keeping the computational cost of the bootstrap (§ 3.1.4) tractable. The grid spacing is matched to the mean sampling cadence of each campaign (Table 1), while is scaled to between 5% and 18% of the total campaign duration (Table 2), reflecting the substantially different baselines and expected lag scales across the four objects, so that the lag-scan parameters reflect the temporal sampling and baseline available for each object.
| Object | (d) | (d) | (d) | ||
| NGC 4151 | 6 | 0.15 | 0.20 | 81 | 0.087 |
| NGC 5548 | 10 | 0.15 | 0.25 | 134 | 0.078 |
| NGC 4593 | 4 | 0.10 | 0.15 | 81 | 0.177 |
| Mrk 509 | 15 | 0.25 | 0.50 | 121 | 0.055 |
Table 2. Table 2. Lag-scan parameters adopted for each object: maximum trial lag , lag step , common-grid spacing , number of trial lags , and the fraction of the total campaign duration covered by the lag scan.
For each object, the trial-lag grid spans in steps of . The peak of MI() is located by fitting a second-order polynomial to all points with MI :
providing a sub-step peak estimate.
Random Subset Selection bootstrap
To quantify the uncertainty on , we apply a Random Subset Selection (RSS) bootstrap (Efron 1979): in each of iterations, a fraction of the data points is drawn without replacement from each light curve independently, and the full analysis (interpolation, normalisation, MI curve computation, and peak detection) is repeated on the subset. Sampling without replacement preserves the temporal ordering of the selected points and avoids artificially duplicated observations that could bias the density estimate.
The 150 bootstrap peak values characterise the lag uncertainty. We report:
where and denote the 16th and 84th percentiles of the bootstrap distribution, corresponding to the interval of a Gaussian distribution. The estimated lag and its asymmetric uncertainty are then reported as \hat{\tau}_{\mathrm{med}}^{+\sigma^{+}}_{-\sigma^{-}}.
The output of the full procedure consists of two diagnostic plots per band pair: the normalised MI lag curve MI() (Figure 2) and the histogram of bootstrap peak values (Figure 3). In a well-behaved case such as NGC 5548/B ( Å), the MI curve displays a single, well-isolated peak and the bootstrap distribution is approximately Gaussian and narrow ( d). Conversely, in bands with weak physical coupling (e.g. the X-ray bands HX and SX), the MI curve is broad and multi-peaked, and the bootstrap histogram is correspondingly wide and skewed, signalling that no reliable lag can be extracted.
Figure 2. Normalised MI lag curves for all seven bands of NGC 5548, relative to the W2 reference band. The solid blue vertical line marks the bootstrap median lag ; the blue shaded region spans the 16th to 84th percentile interval. The red dashed and green dotted lines show the ICCF and JAVELIN reference lags of Edelson et al. (2019), respectively, where available. The UV bands (M2 through U) display a single well-defined peak consistent with disc reprocessing; the X-ray bands (HX, SX) show broad, multi-peaked curves indicative of weak X-ray - UV coupling.
Figure 3. Bootstrap lag distributions (, RSS ) for each band of NGC 5548. The solid blue vertical line marks the bootstrap median ; dashed blue lines indicate the 16th and 84th percentiles. The red dashed and green dotted lines show the ICCF and JAVELIN reference lags of Edelson et al. (2019), respectively, where available. UV bands yield compact, approximately Gaussian distributions; the X-ray bands HX and SX produce broad, skewed histograms consistent with an unresolvable lag.
Results and Discussion
The KSG MI inter-band time lags for all four AGNs are listed in Table 3. Below we discuss the results source by source (§ 4.1,§ 4.4) and then synthesise the band-level behaviour across the full sample (§ 4.5). The MI lag curves and bootstrap distributions for NGC 5548 are shown as representative examples in Figure 2 and Figure 3; equivalent figures for the remaining sources are provided as online supplementary material. The comparison between MI lags and the ICCF/JAVELIN reference values of Edelson et al. (2019) for all four AGNs is shown in Figure 4.

Figure 4. MI lag versus ICCF (left) and JAVELIN (right) reference lags for all four AGNs. Each panel shows the MI bootstrap median lag (with RSS uncertainties) against the reference lag of Edelson et al. (2019), relative to the W2 reference band. The dashed line marks the 1:1 relation.
| Object | Band | (d) | (d) | (d) | (d) | (d) | |
| NGC 5548 | M2 | −0.030 | 0.052 | 0.052 | −0.007 | +0.012 | 0.244 |
| NGC 5548 | W1 | +0.226 | 0.076 | 0.076 | +0.301 | +0.102 | 0.186 |
| NGC 5548 | U | +0.855 | 0.097 | 0.103 | +1.146 | +0.974 | 0.153 |
| NGC 5548 | B | +1.078 | 0.141 | 0.170 | +1.108 | +0.980 | 0.085 |
| NGC 5548 | V | +1.325 | 0.169 | 0.225 | +1.410 | +1.266 | 0.038 |
| NGC 5548 | HX | −0.782 | 1.157 | 1.250 | −4.550 | - | 0.293 |
| NGC 5548 | SX | −0.113 | 1.094 | 1.283 | −2.008 | - | 0.681 |
| NGC 4151 | M2 | −0.025 | 0.088 | 0.091 | +0.055 | +0.045 | 0.017 |
| NGC 4151 | W1 | −0.066 | 0.109 | 0.124 | −0.011 | +0.064 | 0.017 |
| NGC 4151 | U | +0.480 | 0.114 | 0.113 | +0.679 | +0.443 | 0.018 |
| NGC 4151 | B | +0.423 | 0.120 | 0.139 | +0.877 | +0.475 | 0.008 |
| NGC 4151 | V | +0.658 | 0.318 | 0.294 | +0.960 | +0.714 | 0.003 |
| NGC 4151 | HX | +0.508 | 0.872 | 1.163 | −3.324 | - | 0.542 |
| NGC 4151 | SX | −2.715 | 1.928 | 1.653 | −2.408 | - | 0.040 |
| NGC 4593 | M2 | +0.030 | 0.034 | 0.050 | +0.048 | +0.009 | 0.120 |
| NGC 4593 | W1 | −0.020 | 0.043 | 0.059 | +0.077 | −0.010 | 0.077 |
| NGC 4593 | U | +0.163 | 0.063 | 0.111 | +0.337 | +0.337 | 0.054 |
| NGC 4593 | B | +0.000 | 0.200 | 0.231 | +0.182 | +0.041 | 0.013 |
| NGC 4593 | V | +0.054 | 0.379 | 0.275 | +0.351 | +0.182 | 0.004 |
| NGC 4593 | HX | −0.531 | 0.490 | 0.350 | −0.602 | - | 0.274 |
| NGC 4593 | SX | −0.405 | 0.702 | 0.273 | −0.538 | - | 0.348 |
| Mrk 509 | M2 | −0.398 | 0.178 | 0.143 | −0.047 | −0.172 | 0.194 |
| Mrk 509 | W1 | +0.046 | 0.148 | 0.129 | −0.047 | +0.005 | 0.144 |
| Mrk 509 | U | +1.705 | 0.230 | 0.200 | +2.626 | +1.959 | 0.123 |
| Mrk 509 | B | +1.263 | 0.245 | 0.310 | +1.937 | +1.548 | 0.089 |
| Mrk 509 | V | +1.727 | 0.517 | 0.310 | +2.469 | +2.776 | 0.056 |
| Mrk 509 | HX | −0.915 | 1.535 | 1.183 | +4.941 | - | 0.216 |
| Mrk 509 | SX | +0.824 | 0.751 | 0.848 | +2.332 | - | 0.321 |
Table 3. KSG MI inter-band time lags relative to the W2 (UVW2) reference band for the four AGNs in the sample. is the bootstrap median lag, with and denoting the asymmetric uncertainties derived from the 16th and 84th percentiles of 150 RSS bootstrap realisations. and are the ICCF and JAVELIN literature lags from Edelson et al. (2019) (JAVELIN not available for X-ray bands). is the multifractal spectrum width for each band, taken from Souza et al. [51].
NGC 5548
The MI lag curves and bootstrap distributions for NGC 5548 are shown in Figure 2 and Figure 3. In the UV bands M2, W1, and U the MI returns , , and d, respectively, against ICCF values of , , and d (Edelson et al. (2019)), giving , , and d respectively. The JAVELIN lags (, , d) bracket these estimates from above in the near-UV and are consistent within the bootstrap uncertainties. In the optical, B returns d against d (ICCF, d; JAVELIN d), and V returns d against d (ICCF, d; JAVELIN d). The bootstrap histograms for these bands are compact and approximately Gaussian, confirming a well-constrained peak.
In the X-ray bands the MI lags depart from the ICCF reference. For HX the MI returns d against d (ICCF, d); for SX, d against d ( d). Both bootstrap distributions are wide and skewed (Figure 3), consistent with the broad, multi-peaked MI curves in Figure 2, and the ICCF lags lie well outside the RSS uncertainty intervals in both cases.
Physically, the close UV/optical agreement between MI, ICCF, and JAVELIN in the near-UV (M2, W1) supports the disc-reprocessing picture, in which continuum reprocessing by the accretion disc drives a wavelength-ordered lag sequence (Cackett et al. (2007)). The growing toward U indicates an increasing diffuse continuum/BLR contribution to the observed lag (Cackett et al. (2018)), while the failure in HX/SX reflects the intrinsically weak X-ray/UV coupling reported for this source by Edelson et al. (2019), rather than a limitation of the MI estimator itself.
NGC 4151
For NGC 4151 (Table 3), the MI lags in M2, W1, and U are , , and d, against ICCF values of , , and d (Edelson et al. (2019)), giving , , and d. The corresponding JAVELIN lags (, , d) are compatible within the bootstrap uncertainties. The U band provides the tightest three-way agreement in this source. For B the MI returns d against d (ICCF, d; JAVELIN d), and for V d against d (ICCF, d; JAVELIN d).
In the X-ray bands, HX returns d against d (ICCF, d, opposite sign). The SX band gives d against d ( d). Both X-ray bootstrap distributions are broad, with the HX histogram spanning several days (Table 3).
The tight three-way agreement in U, alongside the shorter MI lags in B and V relative to ICCF, is consistent with a disc-dominated UV/optical response with a modest diffuse continuum contribution superposed at longer wavelengths (Cackett et al. (2018)). The large, sign-discordant X-ray offsets echo the weak physical X-ray/UV connection found by Edelson et al. (2019) for the full sample, suggesting the corona-to-disc reprocessing channel is not the dominant driver of the observed X-ray variability in this source.
NGC 4593
For NGC 4593 (Table 3), the MI lags in the UV/optical bands are (M2), (W1), (U), (B), and d (V). The corresponding ICCF values are , , , , and d (Edelson et al. (2019)). The M2 band shows excellent agreement ( d), and W1, B, and V also agree within the bootstrap uncertainties. The JAVELIN lags (, , , , d) are broadly consistent with the MI estimates. The U band shows the largest UV/optical discrepancy in this source ( d).
For the X-ray bands the MI returns d (HX) and d (SX), against ICCF values of and d ( and d). This is the only source in the sample for which the MI X-ray lags fall within one of the ICCF, although the bootstrap distributions remain broad relative to those of the UV bands.
NGC 4593 is the only source for which the X-ray lags also agree with the ICCF, alongside the tightest UV/optical baseline consistency. This is compatible with the correlated X-ray/UV/optical variability reported by Cackett et al. (2018) for this object, pointing to a cleaner, more centrally concentrated reprocessing geometry with comparatively little diffuse BLR contamination, in contrast to the more extended reprocessing regions inferred for the other three AGNs.
Mrk 509
For Mrk 509 (Table 3), the MI lags in the UV/optical bands are (M2), (W1), (U), (B), and d (V), against ICCF values of , , , , and d (Edelson et al. (2019)). The MI recovers the correct sign for all five bands. W1 shows the closest agreement ( d; JAVELIN d). The U, B, and V lags are systematically shorter than both ICCF and JAVELIN reference values (, , and d relative to ICCF), suggesting that for this source the MI peak identifies the leading edge of the reprocessing response rather than the centroid.
In the X-ray bands, HX returns d against d (ICCF, d, opposite sign), the largest discrepancy in the full sample. SX gives d against d ( d; see Table 3).
The systematically shorter MI lags in U, B, and V relative to ICCF and JAVELIN mirror the diffuse-continuum excess identified by Edelson et al. (2019), who report U-band lags a factor of above thin-disc predictions due to BLR diffuse continuum emission. Because MI traces the peak rather than the lag-weighted centroid, it is comparatively insensitive to this extended tail, recovering a delay closer to the direct disc response than to the diffuse-continuum-inflated ICCF/JAVELIN values.
Band behaviour across the sample
Considering all four AGNs jointly, the M2 band shows , , , and d for NGC 5548, NGC 4151, NGC 4593, and Mrk 509, respectively, with the MI recovering the correct sign in all cases and the JAVELIN lags consistent within uncertainties. The W1 band yields , , , and d, the most uniformly small discrepancies in the sample and consistently the best-constrained bootstrap distributions. The U band shows good agreement for NGC 5548, NGC 4151, and NGC 4593 ( d) but the MI is shorter than the ICCF for Mrk 509 ( d). The B band discrepancy grows across sources: , , , and d. The V band yields , , , and d, ranging from close agreement in NGC 5548 to the largest V-band discrepancy in Mrk 509.
The X-ray bands HX and SX show the largest and least consistent discrepancies across all sources. For HX: , , , and d; for SX: , , , and d. NGC 4593 is the single exception where the MI X-ray lags agree with the ICCF within the bootstrap uncertainties. In two of the four AGNs (NGC 5548 and Mrk 509/HX) the MI and ICCF X-ray lags differ in sign. Edelson et al. (2019) report for all eight X-ray/UV pairs across the sample, versus for all twenty UV/optical pairs.
Across the full sample, the MI lag is smaller than the ICCF lag in 19 of the 20 UV/optical band pairs (the only exception being Mrk 509/W1; see left panel of Figure 4), pointing to a structural rather than source-specific effect. ICCF reports the centroid of the cross-correlation function, which averages over the full width of the response; if the transfer function has an extended tail toward longer delays, as expected from diffuse BLR/continuum reprocessing (Cackett et al. 2021), this tail pulls the centroid toward larger . The MI lag, instead, marks the peak of MI(), the single delay with the most shared information, so it is not pulled by that tail and stays closer to the shortest, dominant lag. This difference between an averaging estimator and a peak-seeking one naturally makes MI lags shorter
The near-uniform excellence of M2 and W1 across all sources reflects their proximity to the disc continuum with minimal diffuse contamination, while the growing B/V discrepancy traces the increasing diffuse-continuum/BLR contribution at longer wavelengths (Cackett et al. 2018). The X-ray bands, showing the largest and most inconsistent offsets, corroborate the weak X-ray/UV coupling (, the peak value of the cross-correlation coefficient as defined by Edelson et al. 2019) reported by Edelson et al. (2019), confirming that disc reprocessing, not coronal X-ray heating, governs the UV/optical lag structure recovered here.
Comparison with JAVELIN and ICCF
Although the primary goal of this work is to assess the ability of the KSG mutual information estimator to recover inter-band reverberation lags, it is also instructive to compare its performance against the two most widely adopted techniques in AGN reverberation mapping, namely the Interpolated Cross-Correlation Function (ICCF) and JAVELIN.
For each UV/optical band pair, we computed the absolute discrepancy between the MI lag estimate and the JAVELIN lag and the corresponding discrepancy between ICCF and JAVELIN, Eq. (9) and Eq. (10), respectively.
The mean absolute discrepancy was calculated, as shown in Eq. (11)
where is the number of UV/optical lag measurements. The median discrepancy was obtained from the ordered set of values, while the root-mean-square (RMS) discrepancy was computed as shown in Eq. (12)
The resulting statistics are presented in Table 4.
| Method | Mean (d) | Median (d) | RMS (d) |
| MI - JAVELIN | 0.151 | 0.084 | 0.267 |
| ICCF - JAVELIN | 0.180 | 0.143 | 0.240 |
Table 4. Comparison between MI and ICCF relative to the JAVELIN lag estimates for the 20 UV/optical band pairs.
The MI estimator exhibits smaller mean and median discrepancies relative to JAVELIN than the ICCF. The mean absolute difference
decreases from 0.180 d for ICCF to 0.151 d for MI, corresponding to a reduction of approximately 16%. Likewise, the median discrepancy decreases from 0.143 d to 0.084 d, corresponding to a reduction of approximately 41%, indicating that the typical MI lag estimate lies substantially closer to the JAVELIN solution.
A complementary comparison was performed on a case-by-case basis. Among the 20 UV/optical lag measurements analysed in this work, the MI estimator produced lag values closer to the JAVELIN estimates in 14 cases (70%), while ICCF was closer in only 6 cases (30%). This tendency suggests that the MI estimator generally reproduces the JAVELIN solutions more faithfully than ICCF.
To assess whether this apparent advantage is statistically significant, we applied a paired Wilcoxon signed-rank test to the two discrepancy distributions. The resulting -value () does not allow rejection of the null hypothesis that both methods perform equivalently at the conventional 95% confidence level. Nevertheless, the systematic reduction in both the mean and median discrepancies, together with the 70% agreement rate, indicates a consistent trend favouring MI.
This result is particularly relevant because JAVELIN is generally regarded as one of the most sophisticated lag-recovery methods currently available, relying on Bayesian inference and stochastic process modelling of AGN variability. While JAVELIN often provides highly robust lag estimates, its application can become computationally expensive when analysing large samples, densely sampled light curves, or extensive Monte Carlo experiments.
In contrast, the KSG mutual information estimator is comparatively simple to implement, computationally efficient, and free from assumptions regarding linear correlations between the underlying signals. The present results therefore suggest that MI may represent a practical alternative to ICCF when a computationally inexpensive estimator is desired, while still maintaining close agreement with the lag solutions obtained by JAVELIN.
The closer agreement between MI and JAVELIN is physically consistent with both being less sensitive to the extended, diffuse-continuum-driven tail of the transfer function that biases the ICCF centroid (Cackett et al. 2018). This suggests that peak/model-based lag estimators recover a delay closer to the direct disc-reprocessing signal, while the ICCF centroid is more strongly weighted by BLR diffuse continuum contamination, particularly in high-luminosity sources such as Mrk 509.
Although the current sample is too small to establish a statistically significant superiority of MI over ICCF, the observed tendency motivates future investigations using larger AGN samples. If confirmed, this behaviour would support the use of mutual-information-based lag recovery as a competitive and computationally efficient tool for reverberation mapping studies.
Comparison with the multifractal characterisation
Souza et al. [51] characterised the same eight light curves of these four sources using Multifractal Detrended Moving Average (MFDMA), reporting for each band the multifractal spectral width , the asymmetry parameter , the Hurst exponent , and the fractional variability, , together with a superlinear scaling across the full sample. Their band-by-band picture is markedly source-dependent. In Mrk 509, every band except B and V is multifractal, most strongly HR X and SX ( and 0.321); in NGC 5548, multifractality is again strongest in HX and SX, with SX reaching the highest in the whole sample and the only near-antipersistent Hurst exponent (), while B and V are near-monofractal. NGC 4593 follows a similar pattern, HX and SX more multifractal than the optical bands, with W2 and M2 strongly right-skewed ( and 2.836) and B, V essentially monofractal. NGC 4151 stands apart: only HX is clearly multifractal (, left-skewed with ), while every other band, including SX, is near-monofractal ().
Comparing this with the MI-ICCF discrepancies in Table 3, the clearest qualitative link is in the UV/optical regime: among the five subsidiary bands compared to W2 in this work (M2, W1, U, B, V), M2 and W1 carry the highest and second-highest in three of the four sources [51], and these are also the bands with the smallest mean MI-ICCF offsets ( and 0.06 d; § 4.5), consistent with their comparatively structured, higher-amplitude variability producing a better-defined MI() peak. The X-ray bands break this correspondence: NGC 5548/SX and NGC 4151/HX have the two highest values in the entire sample [51], yet their MI-ICCF discrepancies are among the largest found here ( 1.90 and 3.83 d). The exception is NGC 4593, whose HX and SX bands are simultaneously among the most multifractal in the sample [51] and the only X-ray bands for which the MI lag agrees with the ICCF to within 0.16 d (§ 4.3). Taken together, this suggests that alone does not determine MI performance: in the UV/optical regime higher multifractality appears to coincide with more reliable lag recovery, but in the X-ray bands the outcome instead seems to depend on source-specific properties, such as the X-ray/UV coupling strength (Edelson et al. 2019), that a single complexity measure does not capture. A direct, band-by-band statistical test of this relation across the full 28-entry table is left for future work.
The UV/optical correspondence between higher and better MI recovery suggests that stronger, more structured disc-driven variability yields a sharper, more identifiable MI peak. The breakdown of this relation in the X-ray bands indicates that multifractality alone does not encode reprocessing efficiency; instead, source-specific X-ray/UV coupling strength (Edelson et al. 2019) remains the dominant factor, implying is a UV/optical-specific, not universal, diagnostic for MI-based lag recovery.
Conclusions
We have applied KSG MI lag estimation (, 150-iteration RSS bootstrap) to the multiwavelength light curves of four Seyfert AGNs from the Swift IDRM survey of Edelson et al. (2019), providing a systematic comparison of MI-based reverberation mapping against both the ICCF and JAVELIN lags on a multi-source, multi-band dataset. In the UV/optical regime the method recovers inter-band lags consistent with the ICCF and JAVELIN across all four sources, with the closest agreements in the M2 and W1 bands ( d on average) and degrading performance toward the optical B and V bands, where low variability amplitude reduces the discriminating power of the MI peak. The best single-band agreement in the full sample is NGC 5548/W2 ( d), and the MI recovery shows the correct lag sign for all UV/optical bands in Mrk 509. These results provide a consistent basis for the continuation on the next page. observational validation, on real multi-band AGN data, of the proof-of-concept put forward by Chakraborty & van Leeuwen (2022) using toy models, and extend the scope of alternative randomness-based estimators explored by Chelouche et al. (2017) to the multiband AGN disc-reprocessing regime.
A complementary comparison with JAVELIN further supports the potential of the MI approach. Across the 20 UV/optical band pairs analysed, MI lag estimates were closer to the JAVELIN solutions in 14 cases (70%), compared to 6 cases for ICCF. Moreover, the mean and median MI-JAVELIN discrepancies (0.151 d and 0.084 d, respectively) are smaller than the corresponding ICCF-JAVELIN values (0.180 d and 0.143 d). Although the current sample does not provide statistically significant evidence for the superiority of MI over ICCF, these results indicate that MI is capable of reproducing JAVELIN-like lag estimates while retaining a comparatively simple and computationally efficient implementation.
In the X-ray bands the MI lags deviate substantially from the ICCF reference for three of the four sources, in several cases with opposite sign. The discrepancies reach d for Mrk 509/HX, 3.83 d for NGC 4151/HX, 3.77 d for NGC 5548/HX, and 1.90 d for NGC 5548/SX. The exception is NGC 4593, for which both X-ray lags agree with the ICCF within the bootstrap uncertainty ( d). These failures are consistent with the weak X-ray/UV physical coupling reported by Edelson et al. (2019), who find for all eight X-ray/UV pairs versus for all twenty UV/optical pairs, indicating that the X-ray variability is not straightforwardly connected to the UV through the disc-reprocessing channel that governs the UV/optical lags.
The multifractal characterisation of the same light curves by Souza et al. [51] provides a band-resolved framework for understanding these outcomes. The MI-ICCF agreement is best in bands with moderate to high multifractal spectral width and degrades as approaches zero in near-monofractal optical bands with low . In the X-ray bands, high , reaching 0.681 for NGC 5548/SX, the largest in the sample, which also carries the only near-antipersistent Hurst exponent () reported by Souza et al. [51], does not on its own improve MI performance, with NGC 4593 as the notable exception. Taken together, these results suggest as a candidate diagnostic for anticipating where MI-based lag estimation is likely to succeed in AGN multiband datasets, particularly in the UV/optical regime, while source-specific factors remain important for the X-ray bands.
Physically, these results reinforce the standard disc-reprocessing picture (Shakura & Sunyaev 1973; Cackett et al. 2007): MI reliably recovers UV/optical lags where the disc continuum dominates, while previous work has shown that lags in some bands, most notably the U band, are longer than predicted by the thin-disc model, implying either a larger disc size than expected or an additional contribution from the BLR and diffuse continuum emission (Cackett et al. 2018; Edelson et al. 2019). The pervasive X-ray failure confirms the weak physical X-ray/UV coupling reported by Edelson et al. (2019), indicating that coronal X-ray variability is not simply reprocessed into the UV/optical bands through the same channel governing the disc response.
The main current limitation of the MI-based approach presented here is not the size of the sample but its homogeneity. Because requires both light curves to be evaluated on a common time grid, irregular sampling must be bridged by interpolation, and the fraction of interpolated points needs to remain minimal to avoid introducing spurious structure into the curve. The sample analysed here satisfies this requirement reasonably well, which allowed a first successful application of the method to reverberation-mapping data.
Applying this technique to other AGN samples, spanning a wider range of sampling cadences and data quality, will allow us to systematically characterise the regime in which the MI estimator remains reliable.
Acknowledgements
RS acknowledges the Universidade de Pernambuco (UPE) for its support and resources. This work was carried out within the activities of the Grupo de Ensino das Ciências e Matemáticas, UPE. We also acknowledge the Graduate Program in Applied Physics (PPGF) at the Federal Rural University of Pernambuco (UFRPE) for its support and resources throughout this research. DOF acknowledges the Universidade Federal do Rio Grande do Norte (UFRN) for its support and resources, and CNPq for financial support. LBTS acknowledges the Group of Orbital & Rotational Research in Irregular Objects & Observational astroNomY (ORION), UNESP, and the PostGrad Program in Systems Engineering (PPGES), Universidade de Pernambuco (UPE), for their support and resources. MVDS acknowledges the Departamento de Química Fundamental, Universidade Federal de Pernambuco (UFPE), for its support and resources. TS acknowledges the Departamento de Estatística e Informática, Universidade Federal Rural de Pernambuco (UFRPE), for its support and resources. EH has received funding from the European Union’s Horizon Europe research and innovation programme under grant agreement No. 101188037 (AtLAST2). The IAC is an accredited Severo Ochoa Centre of Excellence supported by grant CEX2025-001609-S, funded by MICIU/AEI/10.13039/501100011033. This research has made use of the VizieR catalogue access tool, CDS, Strasbourg, France (DOI: 10.26093/cds/vizier). The original description of the VizieR service was published in Ochsenbein et al. (2000), A&A, 143, 23.
Data Availability
The data used in this study are available in the VizieR catalogue at https://doi.org/10.26093/cds/vizier.18700123. The mutual information estimator used in this work is the mutual_info_regression implementation of scikit-learn, documented at https://scikit-learn.org/stable/modules/generated/sklearn.feature_selection.mutual_info_regression.html.
Time Series for each AGN

Figure A2. Same as Figure A1 for NGC 4151.

Figure A1. The original light curve of Mrk 509 for all eight observational bands, ordered by wavelength from top to bottom. The top two panels correspond to XRT observations in the HX (1.5,10 keV) and SX (0.3,1.5 keV) bands, while the lower six panels show the UVOT bands (W2, M2, W1, U, B, and V). Each panel displays the flux as a function of MJD, sharing a common time axis but with independent flux scales. X-ray fluxes are given in units of counts , and UVOT fluxes are expressed in units of erg .

Figure A3. Same as Figure A1 for NGC 4593.

Figure A4. Same as Figure A1 for NGC 5548.
References
- [12]Chian A. C. L., 1996, pp 249–295, doi:10.1007/BF00645115DOI
- [25]Ihlen E. A. F. E., 2012, Frontiers in physiology, 3, 141
- [26]Kaastra J. S. e. a., 2011, Astronomy & Astrophysics, 534, A36
- [31]Landt H. e. a., 2015, The Astrophysical Journal, 806, 115
- [35]Mehdipour M. e. a., 2015, Astronomy & Astrophysics, 575, A22
- [48]Shapovalova A. I. e. a., 2010, Astronomy & Astrophysics, 509, A106
- [49]Souza R. A. A., Bewketu Belete A., Canto Martins B. L., de Azevedo L. M. C., Campelo J. P. S., Leão I. C., De Medeiros J. R., 2024, Monthly Notices of the Royal Astronomical Society, 535, 2009
- [50]Souza R. A. A., Bewketu Belete A., Canto Martins B. L., Leão I. C., Azevedo L. M. C., Gomes R. L., De Medeiros J. R., 2026a, Monthly Notices of the Royal Astronomical Society, p. stag905
- [51]Souza R. A. A., Hatziminaoglou E., de Pádua Santos A., Stosic T., 2026b, Monthly Notices of the Royal Astronomical Society, 549, sg016
- [52]Souza R. A. A., Bewketu Belete A., Canto Martins B. L., Leão I. C., Azevedo L. M. C., Gomes R. L., De Medeiros J. R., 2026c, Monthly Notices of the Royal Astronomical Society, 549, stag905
- [60]Ward M. J. e. a., 1983, Monthly Notices of the Royal Astronomical Society, 202, 85

