betaFeedback

WorksPhysics and Astronomy

WISE-VAULT: Decade-long Mid-Infrared Variability of WISE AGN

Abstract

Mid-infrared (MIR) variability provides a powerful probe of accretion processes and dust reprocessing in active galactic nuclei (AGNs), as it is less affected by extinction than optical flux. Using multi-epoch photometry from WISE and NEOWISE spanning over a decade, we investigate the MIR variability of ∼\sim4.5 million color-selected AGNs. Variability is quantified using maximum–minimum amplitudes (ΔW1ΔW1), excess variance, reduced χ2χ^2, and structure function (SF) analysis over timescales of 0.5–10 years. We identify 632,939 variable AGNs, including 54,897 sources that exhibit large-amplitude variability (ΔW1>0.5ΔW1 > 0.5 mag) and 624 with extreme intrinsic variability (Avar>0.5A_{\rm var} > 0.5 mag). A subset of these sources shows flare-like behavior, identifying them as candidate changing-look AGNs or tidal disruption events that warrant further spectroscopic and multi-wavelength follow-up. The SF analysis shows radio-detected AGNs exhibit the highest variability amplitudes, followed by X-ray detected and the overall AGN sample. The variability amplitude is strongly anti-correlated with the Eddington ratio, indicating enhanced variability in lower accretion rate systems. A weak positive dependence on radio-loudness further suggests a contribution from non-thermal emission in radio sources. These results demonstrate that MIR variability provides a robust diagnostic of accretion physics and an efficient means of identifying extreme and rapidly evolving AGN phenomena. The full catalog of variable AGNs is made publicly available.

Introduction

Accretion of matter onto a supermassive black hole is widely accepted as the primary mechanism powering active galactic nuclei (AGNs), producing luminous emission across the electromagnetic spectrum and giving rise to pronounced variability on timescales ranging from hours to decades ([30, 50]). This variability is a fundamental consequence of the dynamical nature of the accretion flow and its interaction with surrounding structures, including the X-ray–emitting corona and circumnuclear material. Because the characteristic spatial scales of the accretion disk and reprocessing regions are far below the resolution limits of direct imaging, temporal variability serves as a powerful indirect probe of the physical conditions, geometry, and size scales of AGN central engines (e.g., [35, 59]).

The physical origin of AGN variability is commonly attributed to stochastic processes within the accretion disk, including fluctuations in the mass accretion rate, magneto-rotational turbulence, thermal instabilities, and variable irradiation by a compact X-ray–emitting corona ([51, 6, 8, 46]). These processes produce correlated variability over a wide range of timescales and motivate the use of statistical descriptions such as damped random walk (DRW) and related stochastic models. Such models have been successfully applied to optical AGN light curves from large surveys and monitoring campaigns, revealing empirical correlations between variability properties and fundamental AGN parameters, including black hole mass, accretion rate, and luminosity (see [20, 25, 32, 23, 44]).

While optical variability directly traces emission from the accretion disk, infrared variability probes a physically distinct and spatially extended component of the AGN structure. In the MIR, the emission is dominated by thermal radiation from dust in the circumnuclear torus that absorbs ultraviolet and optical photons and

Email: [email protected] re-emits them at longer wavelengths [2, 39]. Variability in the accretion disk luminosity is therefore expected to drive infrared variability through dust reprocessing, with characteristic time delays and smoothing determined by the light-travel time across the torus and the thermal response of the dust grains [4, 22]. Consequently, the amplitude and timescale dependence of infrared variability provide direct constraints on the geometry, size, and radial distribution of the dusty environment.

Early observational studies established that AGNs exhibit measurable variability at infrared wavelengths and that such variability is broadly correlated with optical emission, consistent with a reprocessing origin (e.g., [15]; [40]; [19]; [22]). However, these studies were limited by small sample sizes, heterogeneous cadences, and relatively short temporal baselines, restricting their ability to probe long-term variability and population-level trends. The advent of wide-field, space-based infrared surveys has fundamentally transformed this field. Recent studies have demonstrated that infrared variability is common among AGNs and exhibits systematic dependencies on luminosity, wavelength, and rest-frame timescale [24, 49, 31, 43, 53, 41, 21]. Long-term WISE monitoring has also enabled detailed studies of MIR variability and dust reverberation in AGNs. J. Lyu et al. [31] developed methods to construct WISE MIR light curves and measure dust-reverberation lags in PG quasars, accounting for sparse sampling and photometric limitations. Subsequent studies combining long-term optical and WISE light curves measured MIR–optical lags and established the dust-lag–luminosity relation over a broad luminosity range (e.g., [70]; [34]; [58]), further demonstrating the connection between optical variability and delayed MIR emission from circumnuclear dust. J. Li & Y. Shen (2023) showed that ensemble optical and MIR structure functions can be used to characterize the smoothing and geometric dilution of variability by the dusty torus and to constrain its effective size.

Beyond characterizing typical variability, MIR time-domain studies have proven particularly powerful for identifying extreme and rapidly evolving nuclear phenomena. Changing-look AGNs (CL-AGNs), which undergo dramatic transitions in their optical spectral type on timescales of years, are increasingly interpreted as manifestations of large-amplitude changes in accretion rate rather than obscuration alone (e.g., [26]; [47]; [33]). In such events, the MIR emission often responds with delayed but significant brightening or fading, reflecting changes in dust heating and, in some cases, dust destruction or reformation [52, 69]. Similarly, tidal disruption events (TDEs), in which a star is tidally disrupted by a dormant or weakly active supermassive black hole, can produce luminous infrared echoes as the intense flare irradiates surrounding dust (e.g., [10]; S. van Velzen et al. 2016; [36]; [18]). In both cases, MIR variability provides a unique and often less obscured diagnostic compared to optical emission, which can be significantly affected by extinction and host-galaxy contamination. Moreover, using mid-IR data, S. Kim et al. [21] demonstrate that AGN candidates can be identified even in optically inactive hosts, with a low contamination rate (~1%) from transients such as TDEs and supernovae.

Despite these advances, a comprehensive characterization of long-term MIR variability across large AGN samples remains incomplete. In particular, uniform statistical analyses that explicitly quantify variability as a function of timescale—such as structure function measurements—are still relatively scarce in the infrared regime. Such analyses are essential for distinguishing between stochastic accretion-driven variability, dust reprocessing effects, and rare transient phenomena, and for enabling direct comparisons with variability properties observed at shorter wavelengths.

In this paper, we present a comprehensive study of long-term MIR variability in a large sample of AGNs selected from the MIR color-selected AGN catalog of R. J. Assef et al. [3] using their R90 reliability criterion, corresponding to a 90% reliability threshold for AGN selection based on WISE colors. Utilizing multi-epoch photometry from WISE and NEOWISE, we perform a suite of variability analyses, including statistical variability diagnostics and ensemble structure function measurements, to characterize the amplitude, timescale dependence, and population-level behavior of infrared variability. In addition, our analysis enables the identification of a significant population of highly variable sources, which are strong candidates for extreme accretion-driven events such as CL-AGN and TDEs. We compile the key variability properties derived in this work into the WISE Variability of Active Galaxies over Long Timescales (“WISE-VAULT”) catalog, which provides a valuable resource for studying AGN variability and identifying extreme accretion events. The paper is organized as follows: Section 2 describes the data and sample selection, § 3 outlines the variability analysis methods, § 4 presents the main results, and § 5 discusses their implications with a conclusion in § 6.

Sample and Data

Our AGN sample is drawn from the Wide-field Infrared Survey Explorer (WISE; E. L. Wright et al. [68]) MIR color-selected AGN catalog constructed by R. J. Assef et al. [3] using the R90 reliability criterion. The R90 selection is designed to provide a 90% reliability AGN sample based on W1 (3.4 µm) and W2 (4.6 µm) colors, effectively separating luminous AGNs from stars and normal galaxies over a wide redshift range. This reliability was established through a detailed analysis of UV-to-near-IR spectral energy distributions of sources in the 9 deg2^2 Boötes field. The resulting R90 catalog contains 4,543,530 AGN candidates over the extragalactic sky, with an AGN candidate surface density of 130 ± 4 deg−2^{-2}, providing one of the largest uniformly selected MIR AGN samples currently available. This selection is particularly well-suited for infrared variability studies, as it identifies sources whose MIR emission is dominated by hot dust heated by the central engine and is therefore sensitive to variability driven by changes in accretion activity. In Figure 1, we show the distribution of the parent AGN sample in the WISE color–color diagram, together with the color distributions of other astronomical source populations for comparison. The figure illustrates the location of the parent sample in WISE color space relative to these populations, while the solid contours show the distribution of color-selected AGNs from the R90 catalog of R. J. Assef et al. [3].

Figure 1. WISE color–color diagram for the AGN sample constructed from the WISE Single Exposure Source Table. The dotted contours indicate the 1σ\sigma (red), 2σ2\sigma (blue), and 3σ3\sigma (black) levels for different classes of astronomical objects; their locations are marked in the diagram. For reference, the distribution of color-selected AGNs from the R90 catalog of R. J. Assef et al. [3] is represented by solid contours.

We cross-matched the parent R90 AGN catalog with the WISE database using a matching radius of 3″. This radius is comparable to the WISE point-spread function (PSF), whose full width at half maximum (FWHM) is approximately 6.1″ in W1 and 6.4″ in W2 (E. L. Wright et al. [68]). The adopted radius therefore provides a balance between recovering genuine optical counterparts and minimizing chance associations. The WISE mission provides multi-epoch photometry from several operational phases, including the cryogenic mission (2010), the post-cryogenic phase, and the subsequent NEOWISE reactivation survey. After the primary mission concluded in 2011, WISE was reactivated as NEOWISE in December 2013 and has continued surveying the sky until August 2024. In this work, we combine photometric measurements from the AllWISE data release with the NEOWISE-R release (2023), which includes observations obtained between December 2013 and December 2022 in the W1 (3.4 µm) and W2 (4.6 µm) bands. Together, these data provide a temporal baseline exceeding a decade. Owing to WISE’s scanning strategy, most sky locations are observed in discrete six-month intervals (hereafter “epochs”), with individual exposures separated by approximately 1.5 hours within each epoch. By combining AllWISE and NEOWISE-R photometry, we construct decade-long light curves suitable for robust statistical analysis of MIR variability of AGNs.

Several factors can introduce spurious variability into WISE photometry, including scattered moonlight, contamination from nearby sources, image artifacts, and reduced image quality. To ensure robust variability measurements, we therefore restrict our analysis to high-quality single-exposure detections that are free from known contamination flags. Specifically, we apply the following criteria: saa_sep > 0 to exclude exposures affected by the South Atlantic Anomaly; moon_masked = 0000 to avoid moon contamination; qi_fact > 1 to retain frames with acceptable image quality; nb < 3 and na = 0 to remove blended or artifact-affected sources; cc_flags = 0000 to exclude known image artifacts; w1rchi2_ep < 3 to eliminate poor PSF fits; and w1sigmpro_ep < 0.2 mag to ensure reliable photometric uncertainties. We note that blending and host-galaxy emission may still affect the measured WISE fluxes. Unresolved nearby sources can dilute or add scatter to the measured variability, while host-galaxy emission can dilute the intrinsic AGN variability, particularly for lower-luminosity sources. These effects are not explicitly corrected for and may affect the measured variability amplitudes.

To construct long-term light curves, we group all qualified single-exposure measurements within each six-month observing interval into discrete epochs, requiring at least two measurements per epoch, and compute their weighted mean magnitudes. This epoch-averaging process reduces short-term stochastic noise and emphasizes variability on timescales of years. The uncertainty of the epoch-averaged magnitude is estimated following J. Lyu et al. [31], combining the photometric errors and the observed scatter within each epoch as

σint2=1N−1∑j=1N(fj−fˉepoch)2+1N2∑j=1Nσj,phot2+1Nσsys2(1)\sigma_{\mathrm{int}}^2 = \frac{1}{N-1}\sum_{j=1}^{N}\left(f_j-\bar{f}_{\mathrm{epoch}}\right)^2 + \frac{1}{N^2}\sum_{j=1}^{N}\sigma_{j,\mathrm{phot}}^2 + \frac{1}{N}\sigma_{\mathrm{sys}}^2 \tag*{(1)}

Here, NN is the number of data points within a 6-month window. fjf_j, σj,phot\sigma_{j,\mathrm{phot}}, and fˉepoch\bar{f}_{\mathrm{epoch}} are the flux, its uncertainty for the jj-th individual exposure, and the average flux within an epoch, respectively. σsys\sigma_{\mathrm{sys}} is the systematic uncertainty due to the instability of the instrument. J. Lyu et al. [31] estimated the σsys\sigma_{\mathrm{sys}} for ALLWISE and NEOWISE surveys using mid-IR calibration stars and found σsys\sigma_{\mathrm{sys}} to be ∼0.029\sim0.029 mag and ∼0.016\sim0.016 mag for ALLWISE and NEOWISE measurements. S. Son et al. [55] conducted an extensive investigation of the structure function using non-variable sources and found that the uncertainty estimated from Equation (1) is overestimated. They showed that the true uncertainty is better represented by σtrue=σint/N\sigma_{\mathrm{true}}=\sigma_{\mathrm{int}}/\sqrt{N}.

To empirically assess the adopted epoch uncertainty, we constructed a control sample from the MPA-JHU DR8 catalog1 by selecting sources classified as SPECTROTYPE=GALAXY that are also present in our WISE sample. We applied the same WISE quality cuts and epoch construction procedure as for the AGN sample and calculated the ratio

Rerr,true=S2⟨σtrue2⟩.(2)R_{\mathrm{err,true}}=\frac{S^2}{\left\langle\sigma_{\mathrm{true}}^2\right\rangle}. \tag*{(2)}

where

S2=1N−1∑i=1N(mi−mˉ)2,(3)S^2=\frac{1}{N-1}\sum_{i=1}^{N}(m_i-\bar{m})^2, \tag*{(3)}

is the sample variance of the epoch magnitudes, with NN being the number of epochs, mim_i the magnitude measured in the ii-th epoch, and mˉ\bar{m} the mean magnitude. For a sample with correctly estimated uncertainties and negligible intrinsic variability, Rerr,trueR_{\mathrm{err,true}} is expected to be approximately unity. We find that Rerr,trueR_{\mathrm{err,true}} decreases systematically with increasing W1 magnitude, with values of 2.20–0.99 (median 1.5) across the magnitude bins (from brightest to faintest). This indicates that the corrected uncertainty (σtrue\sigma_{\mathrm{true}}) prescription provides a substantially better description of the observed scatter than the original σint\sigma_{\mathrm{int}}, although a residual dependence on source brightness remains, particularly for the brighter sources. For comparison, using the uncorrected σint\sigma_{\mathrm{int}} gives Rerr≃0.18R_{\mathrm{err}}\simeq0.18–0.12 (median 0.14) across the same magnitude range, indicating that σint\sigma_{\mathrm{int}} substantially overestimates the observed scatter. Accordingly, we adopt σtrue\sigma_{\mathrm{true}} as the uncertainty at each epoch in the final light curve.

To ensure reliable characterization of long-term variability, we further restrict our analysis to sources with at least 16 well-sampled epochs. This requirement guarantees sufficient temporal coverage over the full WISE baseline and allows for robust statistical measurements of variability amplitude and structure function properties. After applying this criterion, the final sample used for subsequent variability analysis consists of 1,898,589 AGNs.

Variability Analysis

To characterize the MIR variability properties of our AGN sample, we adopt a two-step approach. First, we identify intrinsically variable sources using statistical variability diagnostics that account for photometric uncertainties. Second, we quantify the variability amplitude and its dependence on timescale using the structure function formalism.

Detection of Intrinsic Variability

Chi-square Variability Test

For each light curve, we evaluate deviations from a constant-flux hypothesis using the chi-square statistic (e.g., [37, 48]),

χ2=∑i=1N(mi−⟨m⟩)2σi2,(4)\chi^2=\sum_{i=1}^{N}\frac{(m_i-\langle m\rangle)^2}{\sigma_i^2}, \tag*{(4)}

where mim_i and σi\sigma_i are the observed magnitude and its uncertainty at epoch ii, ⟨m⟩\langle m\rangle is the weighted mean magnitude, and NN is the total number of epochs in a light curve. Note that all variability analyses in this work are performed consistently in the magnitude scale.

Assuming no intrinsic variability, χ2\chi^2 follows a chi-square distribution with N−1N-1 degrees of freedom. Therefore, we compute the cumulative probability

Pvar=Fχ2(χ2;N−1),(5)P_{\mathrm{var}}=F_{\chi^2}(\chi^2;N-1), \tag*{(5)}

where Fχ2F_{\chi^2} is the cumulative chi-square distribution. We then define the variability parameter as

Vpar=−log⁡10(1−Pvar).(6)V_{\mathrm{par}}=-\log_{10}\left(1-P_{\mathrm{var}}\right). \tag*{(6)}

Following P. Sánchez et al. [48], sources are classified as variable if Vpar>1.3V_{\mathrm{par}}>1.3, corresponding to a ∼95%\sim95\% probability that the observed variability is inconsistent with a constant-flux model. Hence, a larger VparV_{\mathrm{par}} indicates stronger variability.

Intrinsic Variability Amplitude

To quantify the intrinsic variability amplitude, we adopt the excess variance formalism commonly used in AGN variability studies (e.g., S. Vaughan et al. [63]; S. Rakshit & C. S. Stalin [44]).

The mean square photometric uncertainty is

⟨σi2⟩=1N∑i=1Nσi2.(7)\left\langle\sigma_i^2 \right\rangle=\frac{1}{N}\sum_{i=1}^{N}\sigma_i^2. \tag*{(7)}

The intrinsic (noise-corrected) variance is then

σXS2=S2−⟨σi2⟩.(8)\sigma_{\mathrm{XS}}^2=S^2-\left\langle\sigma_i^2 \right\rangle. \tag*{(8)}

If σXS2>0\sigma_{\mathrm{XS}}^2>0, the intrinsic variability amplitude is defined as

Avar=σXS2.(9)A_{\mathrm{var}}=\sqrt{\sigma_{\mathrm{XS}}^2}. \tag*{(9)}

Otherwise, AvarA_{\mathrm{var}} is set to zero.

Structure Function Analysis

To characterize how variability amplitude evolves with timescale, we compute the first-order structure function (SF), a widely used diagnostic in AGN time-domain studies (J. H. Simonetti et al. [54]). The SF provides a non-parametric description of stochastic variability by measuring the typical magnitude difference between pairs of observations separated by a given time lag. It has been extensively applied to quasar variability analysis (e.g., D. E. Vanden Berk et al. [62]; S. Kozłowski et al. [24]; P. Sánchez et al. [48]; S. Son et al. [55]).

The noise-corrected structure function is computed as

SF2(Δt)=1NΔt∑k=1NΔt[(m(tk)−m(tk+Δt))2−(σk2+σk+Δt2)].(10)\mathrm{SF}^2(\Delta t)=\frac{1}{N_{\Delta t}}\sum_{k=1}^{N_{\Delta t}}\left[\left(m(t_k)-m(t_k+\Delta t)\right)^2-\left(\sigma_k^2+\sigma_{k+\Delta t}^2\right)\right]. \tag*{(10)}

where NΔtN_{\Delta t} is the number of magnitude pairs within a given time-lag bin Δt\Delta t, and the second term removes the contribution of photometric uncertainties. Since redshift information is unavailable for a significant fraction

of our sample, all time lags are computed in the observed frame, except for the SF analysis presented in § 4.3 for SDSS AGNs with redshift measurements. While rest-frame corrections are commonly applied in optical quasar variability studies, the observed-frame analysis remains appropriate for characterizing ensemble variability trends, particularly given the broad redshift range (as large as z=6.51z=6.51; G. C. K. Leung et al. [27]) expected for WISE-selected AGNs. We therefore interpret the derived timescales as observed-frame quantities. Alternative robust estimators based on the interquartile range have been proposed for SF computation to reduce sensitivity to outliers (S. Kozłowski et al. [24]). However, the variance-based definition adopted here enables direct comparison with stochastic variability models and previous MIR variability studies (e.g., P. Sánchez et al. [48]; S. Son et al. [55]).

To investigate population-level variability trends, we construct ensemble structure functions for different AGN subsamples based on source classification and physical properties. Ensemble SFs are obtained by combining all valid magnitude pairs within each time-lag bin across the subsample. Although ensemble structure functions are often approximated by a single power-law relation, particularly for sparsely sampled light curves (e.g., S. Kozłowski et al. [24]), recent studies have demonstrated that a single slope does not always provide an adequate description over the full range of timescales (e.g., S. Son et al. [55]). We therefore fit the ensemble SFs using both a single power-law and a broken power-law (BPL) model. The single power-law model is expressed as

SF(Δt)=APL(Δt365)γPL.(11)\mathrm{SF}(\Delta t)=A_{\mathrm{PL}}\left(\frac{\Delta t}{365}\right)^{\gamma_{\mathrm{PL}}}. \tag*{(11)}

whereas the BPL model is given by

SF(Δt)={A(Δt365)γ1,Δt≤tbreak,A(tbreak365)γ1(Δttbreak)γ2,Δt>tbreak,(12)\mathrm{SF}(\Delta t)= \begin{cases} A\left(\dfrac{\Delta t}{365}\right)^{\gamma_1}, & \Delta t\le t_{\mathrm{break}},\\ A\left(\dfrac{t_{\mathrm{break}}}{365}\right)^{\gamma_1}\left(\dfrac{\Delta t}{t_{\mathrm{break}}}\right)^{\gamma_2}, & \Delta t>t_{\mathrm{break}}, \end{cases} \tag*{(12)}

where Δt\Delta t is the observed-frame time lag in days, AA represents the variability amplitude at a timescale of one year, tbreakt_{\mathrm{break}} denotes the characteristic break timescale, γ1\gamma_1 describes the slope at short time lags, and γ2\gamma_2 characterizes the slope beyond the break. We calculate the ensemble structure function using time-lag bins of 180 days, with bin edges spanning 90–3700 days, requiring a minimum of 10 pairs in each time-lag bin to ensure a reliable estimate of the SF. No clipping or additional outlier rejection is applied to the pairwise measurements. The two SF models are compared using the Bayesian Information Criterion (BIC), defined as:

BIC=kln⁡(n)−2ln⁡(L^),(13)\mathrm{BIC} = k \ln(n) - 2\ln(\hat{L}), \tag*{(13)}

where kk is the number of free parameters, nn is the number of SF data points used in the fit, and L^\hat{L} is the maximum likelihood. A lower BIC indicates a preferred model after accounting for model complexity. Our comparison shows that the broken power-law model is preferred over the single power-law model in most of the cases considered. We therefore use the broken power-law fits to characterize the ensemble SFs and investigate the dependence of their amplitudes and slopes on AGN classification and physical properties.

For comparative analysis between different subsamples, we focus primarily on the variability amplitude at one year (AA) and the short-timescale slope (γ1\gamma_1), as these parameters are less sensitive to long-timescale sampling limitations and provide a robust characterization of the stochastic variability growth. The uncertainties on the fitted parameters are estimated using a bootstrap re-sampling procedure. For each subsample, we generate multiple realizations of the ensemble SF within its 1σ1\sigma uncertainties and refit the model. The dispersion of the resulting parameter distributions is adopted as the 1σ1\sigma confidence interval.

Results

Variable AGNs selection

We considered a source to be variable if Pvar>0.95P_{\mathrm{var}} > 0.95 and Avar>0A_{\mathrm{var}} > 0, yielding a sample of 632,939 AGNs. In Figure 2, we plotted the AvarA_{\mathrm{var}} against the maximum to minimum magnitude (ΔW1\Delta W1) in the W1 band, along with density contours and the 3σ3\sigma outliers representing highly variable sources. We identify 3008 sources with Avar>0.3A_{\mathrm{var}} > 0.3 mag and 624 sources with Avar>0.5A_{\mathrm{var}} > 0.5 mag variation, respectively. There are 54,897 sources showing ΔW1>0.5\Delta W1 > 0.5 mag in their W1 band lightcurve. The highly variable (Avar>0.5A_{\mathrm{var}} > 0.5 mag) sources show ΔW1\Delta W1 variation of 1 to 5 mag within a time scale of a few years. A few examples of such highly variable sources are shown in Figure 3. We note that many of our sources have no spectroscopic information, and some have very little information in different databases. The object type as found in SIMBAD is also mentioned. The lightcurves show strong variability in W1 and W2 bands, including a source showing a huge flare and classified as a radio galaxy (top-left panel), a flare-like lightcurve with a fast rise and slow decay (top-right), a QSO-type light curve which is in a bright phase (bottom-left), and a galaxy

Variability amplitude in W1 plotted against maximum to minimum magnitude variation, with density contours and marginal histograms

Figure 2. Variability amplitude (AvarA_{\mathrm{var}}) in W1 is plotted against the maximum to minimum magnitude variation. The 1σ1\sigma, 2σ2\sigma, and 3σ3\sigma density contours are shown along with the outliers, which represent the objects with high variability. The corresponding histograms of the variability amplitudes are shown in the upper-left and lower-right in log-scale.

Figure 3. Example lightcurves of some variable objects in our sample in W1 (blue) and W2 (red) bands. The Object type from SIMBAD (’otype’) is mentioned for each object.

showing gradual dimming since the start of WISE observations (bottom-right). A detailed investigation of extreme MIR variables will be presented in a forthcoming paper. In this work, we focus on the variability analysis of the full sample and compile the derived variability properties into Table 1, which is made publicly available to facilitate a wide range of studies. The catalog includes a notable source, J034947.08+163415.7, which has recently been reported as a highly variable quasar at cosmic dawn (z=6.51z = 6.51; [27]), consistent with our findings.

Column NameDescription
Object NameUnique source identifier
RARight ascension (J2000)
DECDeclination (J2000)
np_W1Number of data points in the light curve (W1 band)
np_W2Number of data points in the light curve (W2 band)
mean_mag_W1Mean magnitude in the W1 band
err_mag_W1mean square uncertainty in W1 band lightcurve (eq. (1))
del_mag_W1maximum–minimum magnitude variation in W1
var_amp_W1Intrinsic variability amplitude (Avar,W1A_{\mathrm{var,W1}}) in W1 (eq. (9))
chi2_W1χ2\chi^2 in W1 band lightcurve (eq. (4))
Pvar_W1cumulative probability of variation in W1 (eq. (5))
Vpar_W1Variability parameter in W1 (eq. (6))
mean_mag_W2Mean magnitude in the W2 band
err_mag_W2mean square uncertainty in W2 band lightcurve (eq. (1))
del_mag_W2maximum–minimum magnitude variation in W2
var_amp_W2Intrinsic variability amplitude (Avar,W2A_{\mathrm{var,W2}}) in W2 (eq. (9))
chi2_W2χ2\chi^2 in W2 band lightcurve (eq. (4))
Pvar_W2cumulative probability of variation in W2 (eq. (5))
Vpar_W2Variability parameter in W2 (eq. (6))

Table 1. Description of the columns in the WISE-VAULT catalog. The full catalog is available in FITS format on Zenodo under an open-source Creative Commons Attribution license: doi: 10.5281/zenodo.23682572.

We observe a strong correlation in variability between the W1 and W2 bands, as shown in Figure 4 with a mean difference of 0.005 mag (a dispersion of 0.03 mag) in AvarA_{\mathrm{var}}. The scatter in AvarA_{\mathrm{var}} in two bands is larger for low variable sources and reduces significantly as AvarA_{\mathrm{var}} increases, as expected. The mean W1-W2 color of the variable AGNs is 1 mag, and the majority of the sources have W1 ∼\sim15 mag

Comparison of variability amplitude in W1 and W2 bands

Figure 4. Comparison of variability amplitude in W1 and W2 bands showing similar correlated variation in both bands. The inset plot shows the histogram of the difference in variability amplitude in the W1 and W2 bands.

It is important to note that the majority of the variable AGNs in our sample lack spectroscopic measurements. For instance, cross-matching with the SDSS DR14 quasar spectral property catalog ([45]) within 3″ search radius yields matches for only ∼\sim10% of the sources. Expanding this comparison,

cross-match our sample with the DESI quasar catalog2 using a 3″ search radius and identify 134,956 counterparts, corresponding to ∼20% of the sample. However, the DESI catalog does not provide physical parameters such as bolometric luminosity or black hole mass. Consequently, DESI-based measurements are not included in our analysis to study the dependency of variability amplitude on AGN physical properties in § 4.3.

To estimate the false-positive rate associated with our variability-selection criteria, we performed Monte Carlo simulations using a representative sample of 10000 objects. For each source, we retained the actual WISE sampling and the adopted epoch uncertainties, but removed any intrinsic variability by assigning a constant magnitude to the source. We then generated 500 mock light curves by adding Gaussian noise according to the corresponding epoch uncertainties. Each simulated light curve was analyzed using the same procedure as the real data, and sources satisfying Vpar>1.3V_{\mathrm{par}} > 1.3 and Avar>0A_{\mathrm{var}} > 0 were classified as variable. We find that 7.5% of the simulated light curves satisfy these criteria, increasing from ∼5% for brighter sources to ∼9% for fainter sources, providing an empirical estimate of the false-positive rate expected from statistical fluctuations and the adopted uncertainty model.

Variability of different AGN samples

In this section, we investigate the variability properties of our AGN sample in the context of their multiwavelength counterparts across different surveys.

Variability of Radio-detected Sample

We cross-matched our variable AGN sample with the Faint Images of the Radio Sky at Twenty-Centimeters survey [67] using a matching radius of 5″, yielding 16,406 radio-detected sources. To characterize their radio properties, we computed the radio-loudness parameter, defined as the logarithmic ratio of the radio flux density at 1.4 GHz to the optical gg-band flux. The gg-band magnitudes were obtained from the SIMBAD database using a matching radius of 3″ from the WISE positions. The resulting radio-loudness parameter spans the range log⁡R=−0.6\log R = -0.6 to 6. Note that since the optical gg-band fluxes are taken from SIMBAD, they may originate from heterogeneous surveys and epochs. Therefore, the resulting radio-loudness estimates should be regarded as approximate, and differences in optical coverage and photometric measurements may introduce additional scatter.

Following standard conventions, we classify sources with log⁡R<1\log R < 1 as radio-quiet (RQ) and those with log⁡R>1\log R > 1 as radio-loud (RL), yielding 7528 RL AGNs in our sample. We further estimate the radio power at 1.4 GHz assuming a power-law spectrum of the form fν∝ν−αf_{\nu} \propto\nu^{-\alpha} with a fixed spectral index α=0.8\alpha= 0.8. The radio luminosity is calculated as (e.g., [7])

P1.4 GHz=4πDL2(1+z)α−1fν,1.4 GHz,(14)P_{1.4\,\mathrm{GHz}} = 4\pi D_L^2 (1+z)^{\alpha-1} f_{\nu,1.4\,\mathrm{GHz}}, \tag*{(14)}

where DLD_L is the luminosity distance and fν,1.4 GHzf_{\nu,1.4\,\mathrm{GHz}} is the observed flux density at 1.4 GHz. Note that P1.4 GHzP_{1.4\,\mathrm{GHz}} was calculated for sources having redshift.

In Figure 5, we plotted the correlation of AvarA_{\mathrm{var}} with radio-loudness and radio-power and found a weak but statistically significant positive correlation between the variability amplitude and radio-loudness (RR), with a Spearman rank coefficient of ρ≈0.20\rho\approx0.20 and a very small pp-value (p∼4×10−39p \sim4 \times10^{-39}). In contrast, the correlation between variability amplitude and radio power is negligible (ρ≈0.04\rho\approx0.04). This indicates a modest contribution from jet-related non-thermal emission, though not the dominant driver of MIR variability (see also J. Lyu et al. [31]; S. Rakshit et al. [43]). This is consistent with the expectation that, for predominantly non-blazar AGNs, MIR variability arises from thermal dust reprocessing of accretion disk fluctuations rather than non-thermal jet emission. We note that the radio-loud and radio-quiet samples may differ in redshift and luminosity, and the observed variability differences should therefore not be attributed solely to jet or non-thermal emission.

Correlation between variability amplitude and radio properties

Figure 5. The correlation between variability amplitude (W1) against radio loudness (left-panel) and radio-power at 1.4GHz (right-panel) are shown. The 1σ and 2σ density contours are shown in thick lines along with 2σ outliers in dots.

Variability of X-ray-detected Sample

To identify X-ray counterparts of our variable AGN sample, we cross-matched our sources with several major X-ray catalogs using a search radius of 5′′5^{\prime\prime}, following the convention commonly used in previous studies (e.g., V. S. Paliya et al. [42]). Specifically, we identified 5641 matches in the Chandra Source Catalog v2.0 (I. N. Evans et al. [12]), 59,135 matches in the eROSITA Final Equatorial Depth Survey AGN catalog (eFEDs; T. Liu et al. [29]), 13,224 matches in the XMM-Newton serendipitous source catalog (4XMM-DR12; N. A. Webb et al. [65]), and 18,837 matches in the Swift X-ray Telescope Point Source Catalog (P. A. Evans et al. [13]). In cases where a source is detected in multiple X-ray surveys, we retain only a single counterpart to avoid duplication. After accounting for overlapping detections, we obtain a total of 86,253 unique AGNs with X-ray counterparts. The X-ray-detected subsample is drawn from surveys with different depths and sky coverage, and therefore does not represent a uniform X-ray selection of the full AGN sample.

In Figure 6, we plotted the ensemble structure function in the W1 band of AGN samples, namely, radio-detected, X-ray-detected, and variable-AGN samples (which include all variable sources found in this work). The best-fit parameters for the broken power-law fit of the SF are given in Table 2. The structure function (SF) analysis reveals clear and systematic differences in variability properties among the different AGN populations. The variability amplitude at 1 yr (AA) shows a strong dependence on AGN type, with radio-detected sources exhibiting the highest amplitudes (A=0.126±0.001A = 0.126 \pm0.001), followed by X-ray-detected sources (A=0.096±0.001A = 0.096 \pm0.001), and the variable-AGN sample (A=0.073±0.001A = 0.073 \pm0.001). This trend indicates that radio-detected AGNs are the most variable population in the MIR, consistent with the presence of additional non-thermal components, such as jet emission, contributing to the variability.

Figure 6. The ensemble structure function in the W1 band of different samples of AGN is shown. The best-fit broken power-law model is shown by solid lines. The same for W2 is shown on the right. The lower panels show the SF when samples are subdivided into smaller bins of W1 band magnitude (14.5-15.0 & 15.0-15.5 mag) for comparison.

SampleBandAAtbreakt_{\mathrm{break}} (days)γ1\gamma_1γ2\gamma_2APLA_{\mathrm{PL}}γPL\gamma_{\mathrm{PL}}BICBPL\mathrm{BIC}_{\mathrm{BPL}}BICPL\mathrm{BIC}_{\mathrm{PL}}
RadioW10.126 ± 0.001319.4 ± 5.40.271 ± 0.0170.221 ± 0.0030.124 ± 0.0010.225 ± 0.0034949
RadioW20.126 ± 0.001358.9 ± 94.40.229 ± 0.0190.251 ± 0.0040.126 ± 0.0010.249 ± 0.0032419
X-rayW10.096 ± 0.0011258.4 ± 78.30.375 ± 0.0050.256 ± 0.0050.099 ± 0.0010.315 ± 0.001651304
X-rayW20.101 ± 0.002349.2 ± 490.40.462 ± 0.0330.346 ± 0.0180.099 ± 0.0010.353 ± 0.001357465
AGNW10.073 ± 0.001343.0 ± 118.50.434 ± 0.0040.363 ± 0.0050.072 ± 0.0010.367 ± 0.00181388767
AGNW20.064 ± 0.0011382.0 ± 11.40.515 ± 0.0010.394 ± 0.0010.068 ± 0.0010.449 ± 0.0013906360

Table 2. Structure Function (SF) fitting parameters in W1 and W2 bands for the different AGN samples. The broken-power-law (BPL) parameters are given by AA, tbreakt_{\mathrm{break}}, γ1\gamma_1, and γ2\gamma_2, while APLA_{\mathrm{PL}} and γPL\gamma_{\mathrm{PL}} denote the parameters of the single-power-law (PL) model. The final two columns give the Bayesian Information Criterion (BIC) values for the BPL and PL models, respectively.

To investigate whether the differences in MIR variability among the radio-detected, X-ray-detected, and general AGN samples are driven by differences in brightness, we further constructed ensemble structure functions in two W1 magnitude intervals, 14.5≤W1<15.014.5 \leq\mathrm{W1} < 15.0 and 15.0≤W1<15.515.0 \leq\mathrm{W1} < 15.5, as shown in the lower panels of Figure 6. In both magnitude intervals, we recover the same relative trend, with the radio-detected AGNs exhibiting the highest variability amplitude, followed by the X-ray-detected AGNs and the general AGN sample. This persistence of the relative ordering within narrow W1 magnitude intervals indicates that the observed differences in variability are not solely driven by differences in the W1 magnitude distributions of the samples. We also find that the SF amplitude is systematically lower in the fainter W1 magnitude interval, indicating a dependence of the measured MIR variability amplitude on source brightness.

The short-timescale slopes (γ1\gamma_1) also differ among the samples. X-ray detected AGNs show the steepest slopes (γ1∼0.38−0.46\gamma_1 \sim0.38 - 0.46), implying rapid growth of variability at short timescales, consistent with strong coupling to the inner accretion disk. Radio-detected sources exhibit significantly flatter slopes (γ1∼0.23\gamma_1 \sim0.23–0.270.27), indicating slower variability growth at short timescales. This suggests that radio AGNs have more gradual and smoother variability evolution, possibly due to the contribution of extended or jet-related emission. At longer timescales, the SF slopes (γ2\gamma_2) flatten for all samples, indicating saturation of variability. However, X-ray and variable-AGN samples maintain relatively higher γ2\gamma_2 values compared to radio sources, suggesting that their variability continues to evolve over longer timescales, whereas radio sources reach saturation earlier. The value of γ1\gamma_1 found in our variable-AGN sample is consistent with the finding of γ1=0.47\gamma_1 = 0.47–0.510.51 obtained in local Type 1 and Type II AGNs by S. Son et al. [55] and the γ∼0.45\gamma\sim0.45 derived from the MIR lightcurves of quasars by S. Kozłowski et al. [24].

The fitted break timescales vary across the AGN sub-samples and between the W1 and W2 bands. These values should be interpreted cautiously because they may be affected by the approximately six-month cadence of WISE observations, source brightness, and photometric uncertainties. Furthermore, reliable redshifts are unavailable for a substantial fraction of the sample, so the SFs in this section are measured in the observed frame. We therefore treat (tbreak)(t_{\mathrm{break}}) as a phenomenological parameter describing the shape of the ensemble SFs rather than as a direct measurement of a physical dust-reprocessing or accretion timescale. Previous studies have shown that MIR variability and SF shapes can contain information on dust reverberation and torus structure (e.g., J. Lyu et al. [31]; J. Li & Y. Shen 2023); however, extracting such physical scales requires well-characterized redshifts and complementary well-sampled optical/MIR data.

Correlation of variability amplitude with AGN parameters

To investigate the dependence of infrared variability on fundamental AGN physical parameters, we examine correlations between the variability amplitude (AvarA_{\mathrm{var}}) and the bolometric luminosity (LbolL_{\mathrm{bol}}), black hole mass (MBHM_{\mathrm{BH}}), and Eddington ratio (REddR_{\mathrm{Edd}}) adopted from the catalog of S. Rakshit et al. [45]. In this catalog, bolometric luminosities are estimated from monochromatic continuum luminosities at 5100 Å, 3000 Å, and 1350 Å depending on the redshift, using appropriate bolometric corrections. Black hole masses are derived using single-epoch virial estimators based on broad emission lines, specifically Hβ, Mg II, and C IV, depending on the redshift of the source. The Eddington ratio is defined as REdd=Lbol/LEddR_{\mathrm{Edd}} = L_{\mathrm{bol}}/L_{\mathrm{Edd}}, where the Eddington luminosity LEddL_{\mathrm{Edd}} is computed from the black hole mass. Further details of these measurements are provided in S. Rakshit et al. [45].

In Figure 7, we present the correlations between AvarA_{\mathrm{var}} and LbolL_{\mathrm{bol}}, MBHM_{\mathrm{BH}}, and REddR_{\mathrm{Edd}}. We find that the variability amplitude exhibits a systematic anti-correlation with all three parameters, with Spearman rank coefficients of ρ≈−0.38\rho\approx-0.38, −0.27-0.27, and −0.20-0.20 for LbolL_{\mathrm{bol}}, MBHM_{\mathrm{BH}}, and REddR_{\mathrm{Edd}}, respectively, all with highly significant (very small) pp-values. As a robustness check, we examined the correlations between the variability amplitude and the AGN physical parameters within bins of the other relevant parameters (see Table 3 for bin size). The anti-correlation between AvarA_{\mathrm{var}} and LbolL_{\mathrm{bol}} persists when the sample is divided into four MBHM_{\mathrm{BH}} bins, with ρ\rho ranging from ρ=−0.15\rho= -0.15 to −0.39-0.39. Similarly, the correlation between AvarA_{\mathrm{var}} and MBHM_{\mathrm{BH}} was examined within four LbolL_{\mathrm{bol}} bins, yielding ρ=0.03\rho= 0.03–0.180.18 in three bins and ρ=−0.21\rho= -0.21 in the highest-LbolL_{\mathrm{bol}} bin, suggesting the dependence on MBHM_{\mathrm{BH}} is comparatively weak and changes across the LbolL_{\mathrm{bol}} bins. The anti-correlation between AvarA_{\mathrm{var}} and REddR_{\mathrm{Edd}} also persists within the four LbolL_{\mathrm{bol}} bins, with ρ=−0.04\rho= -0.04 to −0.17-0.17. Such trends are broadly consistent with previous optical variability studies, which have reported that AGNs with higher luminosities and accretion rates tend to exhibit lower variability amplitudes (e.g., D. E. Vanden Berk et al. [62]; C. L. MacLeod et al. [32]; S. Rakshit & C. S. Stalin [44]).

Figure 7. Correlation of variability amplitude (W1) with AGN physical parameters such as bolometric luminosity (left), black hole mass (middle), and Eddington ratio (right). The 1σ1\sigma, 2σ2\sigma, and 3σ3\sigma density contours are shown along with the median variability at different bins of these physical parameters.

Parameter binBandAAtbreakt_{\mathrm{break}} (days)γ1\gamma_1γ2\gamma_2
LbolL_{\mathrm{bol}}
Lbol<44.6L_{\mathrm{bol}} < 44.6W10.112±0.0020.112 \pm0.002640.8±139.6640.8 \pm139.60.423±0.0370.423 \pm0.0370.229±0.0160.229 \pm0.016
W20.115±0.0030.115 \pm0.003479.4±146.3479.4 \pm146.30.554±0.0520.554 \pm0.0520.335±0.0160.335 \pm0.016
44.6<Lbol<45.244.6 < L_{\mathrm{bol}} < 45.2W10.122±0.0010.122 \pm0.001501.3±16.9501.3 \pm16.90.525±0.0090.525 \pm0.0090.271±0.0030.271 \pm0.003
W20.114±0.0010.114 \pm0.001498.6±30.1498.6 \pm30.10.590±0.0130.590 \pm0.0130.348±0.0050.348 \pm0.005
45.2<Lbol<45.545.2 < L_{\mathrm{bol}} < 45.5W10.115±0.0010.115 \pm0.001508.8±34.3508.8 \pm34.30.589±0.0150.589 \pm0.0150.300±0.0060.300 \pm0.006
W20.100±0.0010.100 \pm0.001527.2±84.0527.2 \pm84.00.621±0.0250.621 \pm0.0250.378±0.0110.378 \pm0.011
Lbol>45.5L_{\mathrm{bol}} > 45.5W10.088±0.0010.088 \pm0.001872.2±108.9872.2 \pm108.90.587±0.0270.587 \pm0.0270.297±0.0210.297 \pm0.021
W20.073±0.0010.073 \pm0.0011024.3±102.51024.3 \pm102.50.587±0.0180.587 \pm0.0180.367±0.0140.367 \pm0.014
MBHM_{\mathrm{BH}}
MBH<7.8M_{\mathrm{BH}} < 7.8W10.112±0.0010.112 \pm0.001510.2±62.5510.2 \pm62.50.485±0.0200.485 \pm0.0200.246±0.0080.246 \pm0.008
W20.101±0.0010.101 \pm0.001499.5±83.2499.5 \pm83.20.525±0.0260.525 \pm0.0260.326±0.0070.326 \pm0.007
7.8<MBH<8.37.8 < M_{\mathrm{BH}} < 8.3W10.113±0.0010.113 \pm0.001507.7±27.1507.7 \pm27.10.555±0.0110.555 \pm0.0110.298±0.0050.298 \pm0.005
W20.099±0.0010.099 \pm0.001503.6±46.8503.6 \pm46.80.607±0.0170.607 \pm0.0170.379±0.0060.379 \pm0.006
8.3<MBH<8.88.3 < M_{\mathrm{BH}} < 8.8W10.113±0.0010.113 \pm0.001531.3±69.2531.3 \pm69.20.594±0.0210.594 \pm0.0210.314±0.0130.314 \pm0.013
W20.100±0.0010.100 \pm0.001711.5±116.2711.5 \pm116.20.593±0.0270.593 \pm0.0270.360±0.0180.360 \pm0.018
MBH>8.8M_{\mathrm{BH}} > 8.8W10.109±0.0010.109 \pm0.001665.9±25.5665.9 \pm25.50.641±0.0110.641 \pm0.0110.312±0.0060.312 \pm0.006
W20.097±0.0010.097 \pm0.001701.7±94.5701.7 \pm94.50.674±0.0190.674 \pm0.0190.388±0.0180.388 \pm0.018
REddR_{\mathrm{Edd}}
REdd<−1.5R_{\mathrm{Edd}} < -1.5W10.122±0.0010.122 \pm0.001649.9±18.4649.9 \pm18.40.558±0.0090.558 \pm0.0090.270±0.0040.270 \pm0.004
W20.116±0.0010.116 \pm0.001655.1±30.1655.1 \pm30.10.614±0.0120.614 \pm0.0120.345±0.0050.345 \pm0.005
−1.5<REdd<−1.0-1.5 < R_{\mathrm{Edd}} < -1.0W10.104±0.0010.104 \pm0.001547.1±65.9547.1 \pm65.90.597±0.0180.597 \pm0.0180.315±0.0120.315 \pm0.012
W20.091±0.0010.091 \pm0.001731.1±71.0731.1 \pm71.00.605±0.0180.605 \pm0.0180.364±0.0100.364 \pm0.010
−1.0<REdd<−0.5-1.0 < R_{\mathrm{Edd}} < -0.5W10.109±0.0010.109 \pm0.001516.0±52.5516.0 \pm52.50.543±0.0170.543 \pm0.0170.289±0.0080.289 \pm0.008
W20.094±0.0010.094 \pm0.001526.7±96.0526.7 \pm96.00.575±0.0230.575 \pm0.0230.371±0.0110.371 \pm0.011
REdd>−0.5R_{\mathrm{Edd}} > -0.5W10.094±0.0010.094 \pm0.001719.8±156.7719.8 \pm156.70.488±0.0320.488 \pm0.0320.271±0.0210.271 \pm0.021
W20.080±0.0010.080 \pm0.0011012.1±224.51012.1 \pm224.50.475±0.0350.475 \pm0.0350.337±0.0180.337 \pm0.018

Table 3. Broken-power-law structure-function fitting parameters for AGN subsamples at z<0.5z < 0.5, divided according to bolometric luminosity (LbolL_{\mathrm{bol}}), black-hole mass (MBHM_{\mathrm{BH}}), and Eddington ratio (REddR_{\mathrm{Edd}}). The broken power-law model is preferred over the single power-law model for all subsamples considered here.

In order to further investigate the dependence of variability on AGN physical parameters, we construct ensemble structure functions in bins of LbolL_{\mathrm{bol}}, MBHM_{\mathrm{BH}}, and REddR_{\mathrm{Edd}} for z<0.5z < 0.5. Here, the structure functions are calculated by converting the time lags to the rest frame using τrest=τobs/(1+z)\tau_{\mathrm{rest}} = \tau_{\mathrm{obs}}/(1 + z). The resulting rest-frame ensemble SFs are shown in Figure 8, and the corresponding best-fit parameters are summarized in Table 3. We find clear systematic trends of variability amplitude with AGN properties. The SF amplitude (AA) decreases with increasing bolometric luminosity, indicating that more luminous AGNs exhibit reduced variability across all timescales. A weaker but noticeable dependence is also observed with black hole mass, where higher-mass systems tend to show slightly lower variability amplitudes.

Ensemble structure functions in W1 band for different bins of bolometric luminosity, black hole mass, and Eddington ratio

Figure 8. The ensemble structure function in W1 band for the SDSS-DR14 quasars for different bins of LBOLL_{\mathrm{BOL}}, MBHM_{\mathrm{BH}}, and REDDR_{\mathrm{EDD}} at z<0.5z < 0.5.

The strongest dependence is seen with the Eddington ratio. AGNs with lower REddR_{\mathrm{Edd}} display significantly higher variability amplitudes compared to high accretion rate systems, indicating that variability is closely linked to the accretion state. This trend is consistently reflected in the fitted SF amplitudes listed in Table 3, where AA decreases systematically toward higher REddR_{\mathrm{Edd}}. The structure function slopes further support this interpretation. The short-timescale slope (γ1\gamma_1) typically ranges between ∼0.4\sim0.4–0.60.6, while the long-timescale slope (γ2\gamma_2) is flatter (∼0.2\sim0.2–0.40.4), indicating a gradual saturation of variability at longer timescales. The break timescale (tbreakt_{\mathrm{break}}) is generally of order ∼500\sim500–600600 days, with a tendency to increase for higher luminosity sources.

Despite these differences in amplitude, the overall shape of the structure function remains broadly similar across different parameter bins, implying that the underlying stochastic variability process is largely universal. Instead, the normalization of the variability appears to be primarily regulated by the physical properties of the accretion flow. These results demonstrate that MIR variability is most strongly governed by the Eddington ratio, with secondary dependence on luminosity and black hole mass.

Discussions

Highly Variable AGNs

A subset of our sample exhibits extreme MIR variability, with 54,897 AGNs showing maximum–minimum variations of ΔW1>0.5\Delta W1 > 0.5 mag. In addition, 624 sources display large intrinsic variability amplitudes with Avar>0.5A_{\mathrm{var}} > 0.5 mag. Such high-amplitude variability is uncommon for typical stochastic AGN light curves and points to substantial changes in the accretion flow or its reprocessing environment. A large fraction of these highly variable sources are radio-undetected, suggesting that their variability is unlikely to be dominated by jet-related processes. These extreme MIR-variable sources provide a promising sample for investigating rare and rapidly evolving nuclear phenomena. Some of their variability patterns may be consistent with those expected from changing-look AGNs (CL-AGNs) or tidal disruption events (TDEs), but MIR variability alone is insufficient to establish such classifications.

In the CL-AGN scenario, large-amplitude MIR variations are interpreted as dust echoes of significant changes in the accretion rate, often accompanied by the appearance or disappearance of broad emission lines ([52]). We cross-matched our catalog with the recent homogeneous catalog of known changing-look AGNs (CL-AGNs) compiled by M. Camus & S. Panda [5]. Their catalog contains 1,438 unique sources, including 987 spectroscopically confirmed CL-AGNs and 451 candidate or photometrically selected systems ([5]). We find 533 of these known CL-AGNs in our catalog. The MIR variability amplitudes of these CL-AGNs in the W1 band have a median of ∼0.15\sim0.15 mag, with the most variable source showing ∼0.67\sim0.67 mag.

TDEs can produce luminous ultraviolet/optical flares that heat the surrounding dust, leading to delayed and long-lived infrared emission ([18] (S. van Velzen et al. 2016)). To place our sample in the context of previously identified infrared transients, we cross-match with catalogs of known MIR outbursts. We identify 7 matches with the “MIRONG” sample [18], which consists of nearby galaxies exhibiting large-amplitude infrared flares associated with extreme accretion events. We also find 122 matches with the “Flaires” catalog [38], which compiles a large population of infrared flares interpreted as dust echoes of transient accretion episodes. We also cross-matched our catalog with the N. Franz et al. [14] TDE catalog from the ‘Open multiwavelength Transient Event Repository’ (OTTER), which provides a multiwavelength compilation of published TDEs. The initial OTTER release contains photometric data for 232 TDEs spanning wavelengths from radio to X-rays [14]. We identify 9 matches between the OTTER TDE catalog and ours. Their MIR variability amplitudes in the W1 band have a median of ∼0.22\sim0.22 mag, with the most variable source showing ∼0.35\sim0.35 mag. These associations demonstrate that our variability selection recovers sources previously identified as infrared transients and TDEs, highlighting the potential of long-term MIR variability searches for identifying nuclear transient phenomena. The large number of high-amplitude MIR variables identified in this work highlights the potential of WISE time-domain data for uncovering rare and rapidly evolving accretion phenomena. The sources with ΔW1>0.5\Delta W1 > 0.5 mag and Avar>0.5A_{\mathrm{var}} > 0.5 mag represent a particularly compelling set of candidates for future spectroscopic and multi-wavelength follow-up aimed at identifying changing-look behavior and transient accretion events. In this context, upcoming large time-domain surveys such as the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST; [17]) will be particularly valuable, providing dense optical monitoring that complements the MIR variability and enables a more complete understanding of the underlying

accretion-driven processes. In addition, space missions such as SPHEREx (O. Doré et al. [9]) and Euclid (Euclid Collaboration et al. [11]) will provide complementary near-infrared spectro-photometric coverage and high-resolution imaging, enabling detailed characterization of dust emission, robust redshift determination, and AGN–host decomposition, thereby offering critical context for understanding the physical origin and environment of the observed variability.

Wavelength-dependent Variability

The dependence of AGN variability on wavelength provides important insight into the physical origin of the emission regions. In Figure 9, we compare the structure function derived for SDSS DR14 quasars after correcting for redshift (see § 4.3) and all variable AGNs (redshift uncorrected) obtained in this work from WISE W1-band light curves with those obtained at shorter wavelengths, including the GALEX FUV band (B. Y. Welsh et al. [66]) and the SDSS i-band (Y. L. Ai et al. 2013). As shown in the figure, the variability amplitude exhibits a clear wavelength dependence, with the largest amplitudes observed in the ultraviolet, intermediate amplitudes in the optical, and the lowest amplitudes in the MIR.

Figure 9. Comparison of the ensemble structure function of quasars at different wavelength regions, GALEX FUV-band quasars from B. Y. Welsh et al. [66], SDSS i-band broad-line AGN from Y. L. Ai et al. (2013), SDSS DR14 quasars (SF corrected for redshift) and WISE all variable AGNs in W1 bands (SF uncorrected for redshift).

At a given timescale, the UV structure function lies significantly above the optical and infrared curves, while the WISE W1-band variability is systematically suppressed across all time lags. This trend is consistent with numerous previous studies, which have shown that AGN variability amplitude decreases with increasing wavelength (e.g., D. E. Vanden Berk et al. [62]; C. L. MacLeod et al. [32]; S. Kozłowski et al. [24]). The enhanced variability at shorter wavelengths is generally attributed to emission originating from the inner regions of the accretion disk, where temperature fluctuations and accretion rate variations are more pronounced.

In contrast, MIR emission arises predominantly from circumnuclear dust that reprocesses the variable ultraviolet and optical radiation from the accretion disk (R. Barvainis [4]; J. Lyu et al. [31]). This reprocessing introduces both temporal smoothing and dilution of variability amplitude, leading to a reduced structure function normalization in the infrared. The similarity in the overall shape of the structure function across wavelengths suggests that the underlying driving mechanism is common, while the amplitude modulation reflects the spatial extent and thermal response of the emitting region. This wavelength dependence is consistent with previous studies of WISE AGNs. J. Lyu et al. [31] found that the MIR variability amplitude generally decreases toward longer wavelengths, with blazars and flat-spectrum radio sources showing an exception to this trend, likely due to the contribution from non-thermal jet emission. Our finding of stronger variability in W1 than in W2 is broadly consistent with this wavelength-dependent behavior. These results support a picture in which AGN variability originates in the accretion disk and propagates outward, with the infrared variability tracing a delayed and smoothed response of the dusty torus. The wavelength dependence of the structure function is consistent with the expected connection between accretion-disk variability and its reprocessed emission from circumnuclear dust.

Conclusion

We present a comprehensive study of MIR variability in a large sample of 4,543,530 WISE color-selected AGNs, aimed at characterizing their variability properties and identifying extreme MIR-variable sources that provide promising targets for follow-up studies of changing-look AGNs (CL-AGNs) and tidal disruption events (TDEs). Using multi-epoch photometry from WISE and NEOWISE spanning over a decade, we quantify variability through maximum–minimum amplitudes, excess variance, reduced χ2\chi^2, and structure function (SF) analysis over timescales of 0.5–10 years. Our main findings are summarized below:

  • We identify 632,939 variable AGNs, including 54,897 sources exhibiting large maximum-to-minimum variability (ΔW1>0.5\Delta W1 > 0.5 mag) and 624 sources with extreme intrinsic variability (Avar>0.5A_{\mathrm{var}} > 0.5 mag). A subset of these sources shows flare-like behavior, making them promising candidates for CL-AGN or TDE events and warranting spectroscopic and multi-wavelength follow-up. The variability in the W1 and W2 bands is strongly correlated.

  • The SF analysis shows that radio-detected AGNs exhibit the highest variability amplitudes (A=0.127±0.001A=0.127\pm0.001 mag), followed by X-ray-detected sources (A=0.096±0.001A=0.096 \pm0.001 mag) and the overall AGN sample (A=0.073±0.001A=0.073 \pm0.001 mag). Within the radio-detected subsample, the variability amplitude shows a positive correlation with radio-loudness, suggesting a contribution from non-thermal synchrotron emission in addition to thermal emission from the accretion disk.

  • The broken power-law fits to the SF further reveal that X-ray-detected AGNs exhibit steeper short-timescale slopes (γ1≈0.38−0.46\gamma_1 \approx0.38 - 0.46), indicating rapid growth of variability, whereas radio-detected sources show flatter slopes (γ1∼0.23−0.27\gamma_1 \sim0.23-0.27), consistent with smoother variability evolution. At longer timescales, the SF slopes flatten for all samples, indicating saturation of variability. These differences suggest that the ensemble MIR variability properties depend on AGN population and emission characteristics.

  • By examining the dependence of variability on physical parameters, we find that MIR variability is strongly anti-correlated with Eddington ratio, such that high-accretion-rate sources exhibit lower variability amplitudes compared to low-accretion-rate systems. This trend is consistent with results from optical studies and supports a scenario in which MIR variability primarily arises from dust reprocessing of the UV/optical emission from the accretion disk. The ensemble SFs generally exhibit steeper slopes at shorter time lags (γ1\gamma_1) than at longer time lags (γ2\gamma_2), indicating a change in the temporal scaling of MIR variability across timescales.

Acknowledgments

We thank the Reviewer for their constructive comments and suggestions, which have helped improve the quality and clarity of the manuscript. This publication makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration.

Facilities: IRSA

Software: TOPCAT ([56]), NumPy ([60]), SciPy ([64]), Matplotlib ([16]), Astropy ([57]).

References

  1. [2]Antonucci, R. 1993, ARA&A, 31, 473, doi: 10.1146/annurev.aa.31.090193.002353DOI
  2. [3]Assef, R. J., Stern, D., Noirot, G., et al. 2018, ApJS, 234, 23, doi: 10.3847/1538-4365/aaa00aDOI
  3. [4]Barvainis, R. 1987, ApJ, 320, 537, doi: 10.1086/165571DOI
  4. [5]Camus, M., & Panda, S. 2026, Compiling the largest multiwavelength database of known changing-look AGN: Prospects in the Rubin era, https://arxiv.org/abs/2607.25839
  5. [6]Chen, X., & Taam, R. E. 1993, ApJ, 412, 254, doi: 10.1086/172916DOI
  6. [7]Condon, J. J. 1992, ARA&A, 30, 575, doi: 10.1146/annurev.aa.30.090192.003043DOI
  7. [8]Dexter, J., & Agol, E. 2011, ApJL, 727, L24, doi: 10.1088/2041-8205/727/1/L24DOI
  8. [9]Doré, O., Bock, J., Ashby, M., et al. 2014, arXiv e-prints, arXiv:1412.4872, doi: 10.48550/arXiv.1412.4872DOI
  9. [10]Dou, L., Wang, T.-g., Jiang, N., et al. 2016, ApJ, 832, 188, doi: 10.3847/0004-637X/832/2/188DOI
  10. [11]Euclid Collaboration, Scaramella, R., Amiaux, J., et al. 2022, A&A, 662, A112, doi: 10.1051/0004-6361/202141938
  11. [12]Evans, I. N., Primini, F. A., Glotfelty, K. J., et al. 2010, ApJS, 189, 37, doi: 10.1088/0067-0049/189/1/37DOI
  12. [13]Evans, P. A., Page, K. L., Beardmore, A. P., et al. 2023, MNRAS, 518, 174, doi: 10.1093/mnras/stac2937DOI
  13. [14]Franz, N., Alexander, K. D., Gomez, S., et al. 2026, ApJ, 999, 243, doi: 10.3847/1538-4357/ae346eDOI
  14. [15]Glass, I. S. 1992, MNRAS, 256, 23P, doi: 10.1093/mnras/256.1.23PDOI
  15. [16]Hunter, J. D. 2007, Computing in Science Engineering, 9, 90
  16. [17]Ivezić, Ž., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111, doi: 10.3847/1538-4357/ab042cDOI
  17. [18]Jiang, N., Wang, T., Dou, L., et al. 2021, ApJS, 252, 32, doi: 10.3847/1538-4365/abd1dcDOI
  18. [19]Kawaguchi, T., & Mori, M. 2011, ApJ, 737, 105, doi: 10.1088/0004-637X/737/2/105DOI
  19. [20]Kelly, B. C., Bechtold, J., & Siemiginowska, A. 2009, ApJ, 698, 895, doi: 10.1088/0004-637X/698/1/895
  20. [21]Kim, S., Kim, M., Son, S., & Ho, L. C. 2026, ApJ, 1000, 245, doi: 10.3847/1538-4357/ae4ecdDOI
  21. [22]Koshida, S., Minezaki, T., Yoshii, Y., et al. 2014, ApJ, 788, 159, doi: 10.1088/0004-637X/788/2/159DOI
  22. [23]Kozłowski, S. 2016, ApJ, 826, 118, doi: 10.3847/0004-637X/826/2/118DOI
  23. [24]Kozłowski, S., Kochanek, C. S., Ashby, M. L. N., et al. 2016, ApJ, 817, 119, doi: 10.3847/0004-637X/817/2/119DOI
  24. [25]Kozłowski, S., Kochanek, C. S., Strn, D., et al. 2010, ApJ, 716, 530, doi: 10.1088/0004-637X/716/1/530DOI
  25. [26]LaMassa, S. M., Cales, S., Moran, E. C., et al. 2015, ApJ, 800, 144, doi: 10.1088/0004-637X/800/2/144DOI
  26. [27]Leung, G. C. K., Eilers, A.-C., Panagiotou, C., et al. 2026, Discovery of Quasar Variability and Early Accretion Disk Signatures at Cosmic Dawn, https://arxiv.org/abs/2605.00978
  27. [29]Liu, T., Buchner, J., Nandra, K., et al. 2022, A&A, 661, A5, doi: 10.1051/0004-6361/202141643DOI
  28. [30]Lynden-Bell, D. 1969, Nature, 223, 690, doi: 10.1038/223690a0
  29. [31]Lyu, J., Rieke, G. H., & Smith, P. S. 2019, ApJ, 886, 33, doi: 10.3847/1538-4357/ab481dDOI
  30. [32]MacLeod, C. L., Ivezić, Ž., Kochanek, C. S., et al. 2010, ApJ, 721, 1014, doi: 10.1088/0004-637X/721/2/1014DOI
  31. [33]MacLeod, C. L., Green, P. J., Anderson, S. F., et al. 2019, ApJ, 874, 8, doi: 10.3847/1538-4357/ab05e2DOI
  32. [34]Mandal, A. K., Woo, J.-H., Wang, S., et al. 2024, ApJ, 968, 59, doi: 10.3847/1538-4357/ad414dDOI
  33. [35]Matthews, T. A., & Sandage, A. R. 1963, ApJ, 138, 30, doi: 10.1086/147615DOI
  34. [36]Mattila, S., Pérez-Torres, M., Efstathiou, A., et al. 2018, Science, 361, 482, doi: 10.1126/science.aao4669DOI
  35. [37]McLaughlin, M. A., Mattox, J. R., Cordes, J. M., & Thompson, D. J. 1996, ApJ, 473, 763, doi: 10.1086/178188DOI
  36. [38]Necker, J., Graikou, E., Kowalski, M., et al. 2025, A&A, 695, A228, doi: 10.1051/0004-6361/202451340DOI
  37. [39]Netzer, H. 2015, ARA&A, 53, 365, doi: 10.1146/annurev-astro-082214-122302DOI
  38. [40]Neugebauer, G., & Matthews, K. 1999, AJ, 118, 35, doi: 10.1086/300945DOI
  39. [41]Pai, A., Blanton, M. R., & Moustakas, J. 2024, ApJ, 977, 102, doi: 10.3847/1538-4357/ad89b8DOI
  40. [42]Paliya, V. S., Rakshit, S., Domínguez, A., Saikia, D. J., & Stalin, C. S. 2026, ApJS, 285, 20, doi: 10.3847/1538-4365/ae7c6fDOI
  41. [43]Rakshit, S., Johnson, A., Stalin, C. S., Gandhi, P., & Hoenig, S. 2019, MNRAS, 483, 2362, doi: 10.1093/mnras/sty3261DOI
  42. [44]Rakshit, S., & Stalin, C. S. 2017, ApJ, 842, 96, doi: 10.3847/1538-4357/aa72f4DOI
  43. [45]Rakshit, S., Stalin, C. S., & Kotilainen, J. 2020, ApJS, 249, 17, doi: 10.3847/1538-4365/ab99c5DOI
  44. [46]Ruan, J. J., Anderson, S. F., Dexter, J., & Agol, E. 2014, ApJ, 783, 105, doi: 10.1088/0004-637X/783/2/105DOI
  45. [47]Ruan, J. J., Anderson, S. F., Cales, S. L., et al. 2016, ApJ, 826, 188, doi: 10.3847/0004-637X/826/2/188DOI
  46. [48]Sánchez, P., Lira, P., Cartier, R., et al. 2017, ApJ, 849, 110, doi: 10.3847/1538-4357/aa9188DOI
  47. [49]Sánchez-Sáez, P., Lira, P., Mejía-Restrepo, J., et al. 2018, ApJ, 864, 87, doi: 10.3847/1538-4357/aad7f9
  48. [50]Shakura, N. I., & Sunyaev, R. A. 1973, A&A, 24, 337
  49. [51]Shakura, N. I., & Sunyaev, R. A. 1976, MNRAS, 175, 613
  50. [52]Sheng, Z., Wang, T., Jiang, N., et al. 2017, ApJL, 846, L7, doi: 10.3847/2041-8213/aa85deDOI
  51. [53]Sheng, Z., Wang, T., Jiang, N., et al. 2020, ApJ, 889, 46, doi: 10.3847/1538-4357/ab5af9DOI
  52. [54]Simonetti, J. H., Cordes, J. M., & Heeschen, D. S. 1985, ApJ, 296, 46, doi: 10.1086/163418DOI
  53. [55]Son, S., Kim, M., & Ho, L. C. 2023, ApJ, 958, 135, doi: 10.3847/1538-4357/ad01bcDOI
  54. [56]Taylor, M. B. 2005, in Astronomical Society of the Pacific Conference Series, Vol. 347, Astronomical Data Analysis Software and Systems XIV, ed. P. Shoppbell, M. Britton, & R. Ebert, 29
  55. [57]The Astropy Collaboration, Robitaille, Thomas P., Tollerud, Erik J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068
  56. [58]Tomar, A., Rakshit, S., Mandal, A. K., & Pandey, S. 2025, ApJ, 993, 203, doi: 10.3847/1538-4357/ae0f96DOI
  57. [59]Ulrich, M.-H., Maraschi, L., & Urry, C. M. 1997, A&A, 35, 445, doi: 10.1146/annurev.astro.35.1.445DOI
  58. [60]van der Walt, S., Colbert, S. C., & Varoquaux, G. 2011, Computing in Science Engineering, 13, 22
  59. [62]Vanden Berk, D. E., Wilhite, B. C., Kron, R. G., et al. 2004, ApJ, 601, 692, doi: 10.1086/380563DOI
  60. [63]Vaughan, S., Edelson, R., Warwick, R. S., & Uttley, P. 2003, MNRAS, 345, 1271, doi: 10.1046/j.1365-2966.2003.07042.xDOI
  61. [64]Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Medicine, 17, 261, doi: 10.1038/s41592-019-0686-2DOI
  62. [65]Webb, N. A., Coriat, M., Traulsen, I., et al. 2020, A&A, 641, A136, doi: 10.1051/0004-6361/201937353DOI
  63. [66]Welsh, B. Y., Wheatley, J. M., & Neil, J. D. 2011, A&A, 527, A15, doi: 10.1051/0004-6361/201015865DOI
  64. [67]White, R. L., Becker, R. H., Helfand, D. J., & Gregg, M. D. 1997, ApJ, 475, 479, doi: 10.1086/303564DOI
  65. [68]Wright, E. L., Eisenhardt, P. R. M., Mainzer, A. K., et al. 2010, AJ, 140, 1868, doi: 10.1088/0004-6256/140/6/1868DOI
  66. [69]Yang, Q., Wu, X.-B., Fan, X., et al. 2018, ApJ, 862, 109, doi: 10.3847/1538-4357/aaca3aDOI
  67. [70]Yang, Q., Shen, Y., Liu, X., et al. 2020, ApJ, 900, 58, doi: 10.3847/1538-4357/aba59bDOI

Paper details

vlnc:W_7z6ak78h
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents