WorksPhysics and Astronomy
Modeling Multi-band Active Galactic Nuclei Lightcurves Using a Self-supervised Attentive Neural Process Network
Abstract
With the the Legacy Survey of Space and Time (LSST), observations of tens of millions of quasars are expected. We present a machine-learning framework based on an attentive neural process (ANP) that models stochastic quasar variability and jointly performs two tasks: (i) probabilistic reconstruction of quasar light curves in all six LSST bands, and (ii) inference of physical parameters of the AGN system — black hole mass, Eddington ratio, and intrinsic variability timescale — with joint, full-covariance posteriors. The model processes all six bands within a single masked attention mechanism, so that observations in any band constrain the shared driving process, and natively handles irregularly and asynchronously sampled light curves with realistic heteroscedastic noise. We train and evaluate on full ten-year light curves simulated with a relativistic accretion-disk reprocessing model, observed through the fielded Rubin baseline cadence with per-visit photometric noise. The light-curve model is trained with self-supervised, learning from the sparse noisy observations alone and never accessing the underlying noiseless light curves. On a held-out test set, we show that our ANP model reach comparable performances to that of the near-optimal baseline Gaussian process model, while exceeding it in power spectrum reconstruction and damped-random walk timescale estimation. The compact learned latent representation supports future downstream applications, and the fully observation-driven training objective enables direct fine-tuning on real LSST light curves.
Introduction
Quasars, powered by accretion onto supermassive black holes (SMBHs) at the centers of galaxies, are among the most luminous sources in the Universe [15, 19]. They exhibit temporal variability across the electromagnetic spectrum (e.g. [25, 8, 17, 24, 6]), with the UV and optical brightness of quasars varying by at least 10% on weeks to years timescales [26, 31]. The Legacy Survey of Space and Time (LSST) by the Vera C. Rubin Observatory has officially begun its 10 year observations; it is expected to observe quasars across a wide range of masses () and up to high redshifts () ([14, 32, 27]). The smooth and non-periodic temporal spectrum of quasars suggests that their variability is stochastic in nature, with correlated, frequency-dependent continuum and emission-line variations [11, 16, 31, 21]. The UV and optical flux variations are believed to be either directly driven by the accretion onto the SMBH via a geometrically thin and optically thick accretion disk [23, 21, 13] or by reprocessing the X-ray radiation from the hot corona [1, 28, 2, 29]. These variations provide a time-domain probe of spatial scales that are generally too small to resolve directly.
Using a large sample of SDSS Stripe 82 quasars, C. L. MacLeod et al. [16] have confirmed these trends and further shown that quasar variability depends on black hole mass, Eddington ratio, radio loudness, rest-frame wavelength, among other source properties. They modeled the optical variability of active galactic nuclei (AGNs)1 using the damped random walk (DRW) process, a framework originally proposed by B. C. Kelly et al. [11]. The DRW is a stationary stochastic process characterized by a damping timescale, (), and a long-term asymptotic variability amplitude, (). Its power spectral density has the form of a broken power law, which is approximately flat at low frequencies () below the characteristic break and drops as () at higher frequencies. Equivalently, the structure function, corresponding to the long-term DRW variability amplitude, increases with time lag and approaches at lags much longer than . On these long timescales, pairs of measurements become effectively uncorrelated, and the variability amplitude no longer increases with temporal separation.
Despite its success, the DRW model is incomplete in describing the full diversity of AGN light curves. For example, high-cadence observations in Kepler have revealed a steeper power-spectral density at short timescales [18]. More flexible stochastic models have since then been developed, e.g. the damped harmonic oscillator model (W. Yu et al. 2022), and such models seem to better capture population-level AGN behavior (W. Yu et al. 2025).
Recent studies of AGN variability have increasingly incorporated simulation-based machine learning techniques. In this approach, physically motivated models of X-ray variability and accretion-disk reprocessing are used to generate large sets of simulated AGN light curves, which are then used to train neural-network models. For example, J. Fagin et al. [7] used a latent stochastic differential equation (SDE) framework to model AGN light-curve variability, reconstruct missing seasonal gaps, and infer accretion-disk physical parameters. While this approach provides a physically motivated and self-consistent framework, it also requires a complex model architecture and substantial computational resources for training. A complementary line of work builds on the neural process (NP) family, which combines the flexibility of deep learning with the amortized probabilistic inference of Gaussian processes. I. Čvorović-Hajdinjak et al. (2022) first applied conditional neural processes to AGN light curves, demonstrating flexible non-parametric interpolation of survey data, though without inferring physical parameters or assessing the calibration of the predictive uncertainties. J. W. Park et al. [20] extended the family to black hole parameter inference with an attentive neural process (ANP); their model is compact, but its architecture assumes synchronous multi-band observations, limiting its direct applicability to realistic LSST light curves, which are irregularly sampled and asynchronous across bands. Most recently, A. N. Raju et al. [22] combined attentive latent neural processes with self-organizing-map clustering and mixture density networks to recover accretion-disk transfer functions and SMBH parameters; their framework treats each photometric band independently, trains on noiseless synthetic light curves with short, burst-like cadences, and validates parameter recovery at the level of population distributions rather than individual objects.
In this work, we propose an extended ANP model based on the framework of J. W. Park et al. [20]. Our model retains the compact latent-representation structure of the ANP while adapting it to realistic LSST-like observations. Compared with the latent-SDE framework, the ANP provides a simpler and more modular architecture for inference from observed light curves, and its compact learned latent representation can be used not only for light-curve reconstruction and AGN parameter inference but also for further downstream tasks; for example, the latent space allows simple clustering, potentially identifying subclasses of AGN. Compared with the existing NP-based approaches, our model advances the family along four axes: it processes all six LSST bands jointly within a single masked attention mechanism, so that observations in any band constrain the shared driving process; it is trained and evaluated on visit lists drawn from the fielded Rubin observing strategy with per-visit photometric noise, rather than on synthetic cadences or noiseless targets; its light-curve model is trained in a fully self-supervised manner, learning exclusively from the sparse, noisy observations themselves; and it produces joint, full-covariance posteriors over four physical parameters whose calibration we verify per object, through the empirical coverage of individual credible intervals.
The self-supervised character of the reconstruction training deserves emphasis, as it determines the framework’s applicability beyond simulations. The model is never shown the underlying noiseless light curves. Instead, we design a tokenization pipeline that handles irregularly and asynchronously sampled multi-band observations (§ 3.1), and the loss function is evaluated only against held-out observed photometry, in contrast to approaches that train against noiseless simulated targets. Only the parameter-inference decoder is trained in a supervised approach, using the known simulation parameters. Because real LSST light curves provide exactly the same information the reconstruction model trains on — noisy, irregular photometry with error bars, and no ground truth — the framework can be fine-tuned on the millions of real AGN light curves LSST will deliver without any change to the training objective.
We start our analysis by simulating quasar lightcurves that will be observable by LSST in its 10-year survey, in § 2, starting from generating an accretion disk model that reprocesses X-ray emissions into UV/optical variability as well as simulating LSST cadences and noise. We then describe the neural network framework used to model the AGN parameters in § 3, where we describe tokenization pipeline applied on the simulated multi-band LSST lightcurves and our attentive neural processing model. We present our results of the performance of the neural process modeling on inferring AGN variability parameters and AGN physical parameters. We find high effectiveness on measuring a wide range of DRW timescales and amplitudes, as well as predicting the BH mass, Eddington fraction and redshift. We additionally find that the trained model is highly effective in reconstructing unseen multi-band LSST lightcurves in our test set. Finally, we discuss future application of this work in § 5.
Light-curve Simulation
We follow the basic approach of (J. Fagin et al. [7]) to generate simulated AGN light curves as they would be observed by LSST with slight modification. These simulated light curves are used to train and evaluate our ANP model.
The simulated AGN light curves are assumed to arise from geometrically thin, optically thick accretion disks irradiated by a central “lamp-post” X-ray source, commonly interpreted as a compact hot corona near the center of the quasar (J. H.-H. Chan et al. [3]). Variability in this central X-ray source is reprocessed by the accretion disk. Emission at different wavelengths originates predominantly from unique disk radii; in turn, different observational bands respond to the driving X-ray variability with different time delays and amplitudes. This produces a wavelength-dependent lag spectrum, , and flux spectrum, . The reprocessing can be described by a wavelength-dependent transfer function. The light curve in each photometric band is derived by convolving the driving X-ray variability with the corresponding transfer function. To simulate realistic observations, we apply realistic LSST observing cadence and properties to the reprocessed light curves. The details of each component of the simulation are described below.
X-ray Variability Simulation
A common model for describing the central X-ray variability of an AGN is the DRW process, also known as the Ornstein–Uhlenbeck process (B. C. Kelly et al. [11]). This stochastic process can be written as
where is the increment of a Wiener process over some increment of time . Equivalently, in a time-discretized form that we use in this work, can be written as , with . Here, is the characteristic damping timescale, determines the long-term mean of the variability signal through , and controls the stochastic variability amplitude.
Transfer Function Simulation
We model the reprocessing of photons by the accretion disk using reverberation mapping in the lamp-post framework (J. H.-H. Chan et al. [3]). Again, in this picture, fluctuations in the central X-ray source heat the disk and induce delayed, wavelength-dependent variations in the disk emission.
Starting from the standard thin-disk temperature profile, the disk temperature can be expanded to first order around its mean value as a function of radius and time :
Here, is the driving variability signal, modeled by the DRW process (Eqn. (1)) with a time delay . The quantity describes the contribution to the disk temperature from irradiation by the central lamp-post source, while is the mean effective disk temperature. The mean temperature includes both viscous heating and external irradiation:
Then, assuming blackbody radiation with a color-correction factor , the specific intensity at frequency is
Integrating over the solid angle gives the observed flux spectrum of the reprocessed emission. The time delay arises from the photon travel time between the central X-ray source and a given disk radius. Since each radius has a characteristic temperature, and therefore contributes preferentially to emission at particular wavelengths, the disk naturally produces wavelength-dependent time delays.
After integrating over solid angle, the time-dependent flux spectrum can be written as
Here, is the wavelength-dependent transfer function, shown in Figure 1. It describes the delayed response of the disk emission at a given wavelength to variability in the central X-ray source. The transfer function is calculated as
Figure 1. Example wavelength-dependent transfer functions, , for the six LSST bands. The black hole parameters in this example are: , , , , , , . Shorter-wavelength emission, such as the band, originates predominantly from the hotter inner regions of the accretion disk and therefore responds to the central driving variability at shorter time delays. Longer-wavelength emission arises from cooler, larger disk radii, producing broader transfer functions with longer characteristic lags. This wavelength-dependent response encodes the radial temperature structure of the accretion disk.
The mean flux and the variability amplitude are given by
The transfer function , the mean flux , and the variability amplitude all encode information about the geometry and thermal structure of the accretion disk. Therefore, by modeling the wavelength-dependent variability of AGN light curves, one can infer physical properties of the SMBH accretion disks.
To generate the simulated spectra, we sample the accretion-disk parameters uniformly within the ranges listed in Table 1. For each parameter set, we compute the corresponding transfer function, mean flux , and flux-variability amplitude . We then convolve the DRW driving signal with the transfer function and combine it with and to obtain the time-dependent flux spectrum, as described in Eq. (5).
| Min. | -0.5 | -1 | 0.3 | 10.0 | 0.0 | 0.1 | 0.01 | 7.0 |
| Max | 3.5 | 1 | 1.0 | 50.0 | 6.0 | 0.3 | 10.0 |
Table 1. Ranges of the black hole parameters used for simulation. First row is the minimum and the second row is the maximum. We set the prior distribution for all the parameters to be uniform.
One of the main goals of our attentive neural process model is to perform Bayesian inference on four parameters that characterize the AGN system and its variability: the black hole mass, ; the Eddington ratio, ; the redshift, ; and the DRW timescale, of the driving variability. In the following section, we describe how the latent representations learned from the simulated light curves are used to infer these parameters.
LSST Survey Simulation
We convolve the spectra simulated in § 2.2 with the LSST filter response curves to obtain light curves in the six observed optical bands, . To construct realistic mock observations, we further sample the simulated light curves according to the LSST observing cadence from baseline_v3.4_10yrs.db. For each simulated observation, we assign a photometric uncertainty following the LSST signal-to-noise prescription described in SMTN-002 ([10]). In this formalism, the signal-to-noise ratio is computed from the source counts, sky background, instrumental noise, gain, and the effective number of pixels contributing to the source measurement:
where is the total source counts, is the sky background counts per pixel, is the gain, is the instrumental noise per pixel, and is the effective number of pixels in the source aperture or PSF footprint. The magnitude uncertainty is then obtained by propagating the flux uncertainty into magnitude space. Since
and , the corresponding magnitude error is
We then generate the mock observed magnitude by adding Gaussian observational noise:
where is the noiseless simulated magnitude evaluated at the LSST observing epoch. This procedure assigns each observation a noise level determined by the source brightness and the observing conditions of the corresponding LSST visit.
For model training and validation, we apply an additional post-selection criterion based on the signal-to-noise ratioSpecifically, we retain only light curves with signal-to-noise ratio . This selection removes very low-quality light curves whose variability signal is dominated by observational noise, improving numerical stability and enabling more efficient model training.
Since the model is trained on LSST cadence-sampled light curves, real observational data can be incorporated into the training set in the future without requiring changes to the model architecture. This provides a natural pathway to fine-tune our model on real observational data.
Model Architecture
Inspired by the work of J. W. Park et al. [20], we introduce an ANP network for modeling AGN light curves. In their study, the model was trained and validated only on clean (i.e., noiseless) DRW signals. We extend this approach by applying the model to more realistic simulated LSST light curves, which incorporate wavelength-dependent transfer functions, the LSST observing cadence, and photometric noise. To handle the irregular sampling of LSST light curves and improve numerical stability during training, we tokenize the light-curve data before feeding them into the ANP model, and upgrade their ANP network structure as described in § 3.1 and § 3.2.
Data Tokenization
Figure 2 summarizes the tokenization pipeline applied to each simulated multi-band light curve before it is passed to the ANP model. Each light curve is tokenized into:

Figure 2. Preprocessing pipeline for simulated multi-band LSST light curves. (a) Simulated light curves in six LSST bands. Blue points denote simulated LSST observations used as model inputs during training and inference, while blue dashed curves show the simulated underlying light curves, which are used only as ground truth for testing or model evaluation and are not visible to the model. (b) Time-feature embedding. Each time step is mapped to using random Fourier features, and , together with companion features, and . A learnable linear projection with weight matrix then maps to the time embedding . (c) Band identities are encoded as six-dimensional one-hot vectors and concatenated with time to construct the input tensor . (d) For each band, observed magnitudes and observational uncertainties are normalized using the band-specific magnitude mean and standard deviation . The normalized magnitudes and uncertainties form , while the normalization statistics are stored in . (e) The preprocessed tensors , , and , together with a binary mask indicating observed entries, are used for model training and inference.
, where , , , and .
We first construct a time embedding for the observation times. For each light curve, the time tensor is denoted by , where is the number of LSST photometric bands and is the maximum sequence length. Each time value is mapped to a feature vector
so that . This embedding combines sinusoidal Fourier features with linear and logarithmic time features.
The sinusoidal Fourier features are constructed by concatenating and over a set of frequencies . We tuned the frequency range to cover the relevant temporal scales of the simulated LSST survey. The lowest frequency corresponds to the full observing baseline, , while the highest frequency corresponds to the shortest cadence-resolvable timescale, . These Fourier features allow the model to represent temporal variations across multiple characteristic timescales. This is particularly important for irregularly sampled LSST light curves, because the model receives a richer representation of relative phase and temporal separation rather than relying only on the raw observation times.
The linear time feature, , preserves the absolute temporal ordering of the observations across the survey baseline. This helps remove the ambiguity introduced by purely periodic features, which can assign similar phases to widely separated times. The logarithmic feature, , compresses large time values while retaining sensitivity to early-time variations. This improves numerical conditioning and helps the model represent variability over the broad range of timescales expected in AGN light curves, from days to years.
After constructing , we apply a learnable linear projection to map the time features into the time-embedding space to get token , which will later be used by the ANP model:
where is a trainable weight matrix and is the dimension of the final time embedding. This projection allows the model to learn the most informative combinations of Fourier, linear, and logarithmic time features directly from the training data. The resulting time embedding is then used as the temporal representation of the light curve in the downstream ANP model.
Since the LSST observing cadence samples different photometric bands irregularly, the representation used by J. W. Park et al. [20], in which the magnitudes from all six bands are combined into a single six-dimensional vector at each time step, is not directly applicable here. For sparsely and asynchronously sampled LSST light curves, this representation would introduce many empty entries. If these missing observations were filled with zeros, the model could incorrectly interpret them as true magnitude measurements rather than unobserved data.
Instead, we represent each observation as an individual token and encode the corresponding photometric band using a one-hot vector. The six LSST bands, , are each mapped to a six-dimensional one-hot band vector, as illustrated in Figure 2(c). This band-identity vector is concatenated with the observation time to form the input token , where, again, is the number of LSST bands and is the maximum sequence length after padding.
For each observation in band at time , we denote the observed magnitude and observational uncertainty as and , respectively. To improve numerical stability during training, we normalize the magnitudes and uncertainties in each band independently for all training and testing data:
Here, and are the mean and standard deviation of the observed magnitudes in band of a sample AGN:
where is the number of observations in band .
The normalized magnitude and uncertainty are concatenated to form the output token
Since this normalization removes information about the absolute brightness and variability amplitude of each band, we also store the normalization statistics,
as an additional input to the model. This allows the model to retain information about both the absolute magnitude scale and the band-dependent variability strength for each light curves.
Finally, to enable batch training, all light curves are padded along the sequence dimension to the same length . We use binary masks to distinguish real observations from padded entries, ensuring that the model does not treat padded zeros as physical measurements.
Attentive Neural Processing Model
Here we present the ANP model, following the frameworks of J. W. Park et al. [20] and Y. Dubois et al. [5].
As shown in Figure 3, the model consists of five main components: a preliminary encoder, a deterministic encoder, a latent encoder, a decoder for light-curve reconstruction, and a decoder for black hole parameter inference. The modular structure of the model also allows additional downstream tasks, such as anomaly detection or binary AGN classification, to be incorporated in future work.

Figure 3. The architecture of the attentive neural process. The preliminary encoder maps each observation through an MLP to an embedding , which is summed with a learned time embedding to form the point representation . In the latent path, masked multi-head self-attention over produces , which are mean-pooled into a set-level summary . Together with — the per-band mean and variance statistics of the observed light curve — is passed through an MLP that outputs the parameters of a Gaussian over the latent variable, from which is sampled. This latent representation serves the downstream heads: black-hole parameter inference and light-curve reconstruction. In the deterministic path, a second masked multi-head self-attention yields point representations , which act as values in a masked multi-head cross-attention with the observed times as keys and the prediction times as queries; the attention weights thus form a learned, data-adaptive interpolation kernel over the light curve’s temporal correlation structure, producing a target-specific representation for each prediction time. The light-curve reconstruction decoder consists of a shared trunk and two output branches: the trunk MLP combines and yields the mean prediction , while the standard-deviation branch additionally receives distance features — the temporal separation between and the nearest observation — and outputs , allowing the predictive uncertainty to widen inside observing gaps and contract near data. The black-hole parameter decoder maps to the mean vector and Cholesky factor of a full-rank four-dimensional Gaussian posterior over . All attention operations are masked so that padded entries never contribute.
During training, each simulated LSST light curve is divided into a context set and a target set, denoted by
where contains the time and band information, contains the normalized magnitude and observational uncertainty, and is the learned time embedding described in § 3.1. Following J. W. Park et al. [20], we choose the context set to be a subset of the target set, . In this setup, the context set provides the observed conditioning points used by the model, while the target set defines the points at which the model is trained to reconstruct the light curve and evaluate the predictive distribution. Details on the context and target set selection can be found in § 3.4.
Both the context and target observations are first passed through a preliminary encoder. This encoder maps the tokens and into an initial feature representation using a multilayer perceptron (MLP). The resulting feature vector is then combined with the learned time embedding through element-wise addition:
This operation produces a sequence of preliminary representations , which encode the photometric information, observational uncertainty, band identity, and temporal position of each data point. In Figure 3, the direct addition of the time embedding is analogous to the use of positional encoding in sequence models. It allows temporal information to be incorporated into the representation at an early stage, while keeping the dimensionality fixed for the subsequent attention layers. The resulting representations are then used as inputs to the deterministic and latent encoders, which learn the temporal correlation information of the light curve conditioned on the available context observation and form the latent representations .
In the latent encoder block, we use the sequence of preliminary representations to infer a global latent variable that summarizes the overall structure of the light curve. We first apply a masked multi-head self-attention
layer to :
where denotes the parameters of the latent masked multi-head self-attention module. The masked multi-head self-attention operation allows each observation to attend to all other observations in the light curve, thereby capturing temporal dependencies and correlations across different time and photometric bands. In this sense, the learned attention weights provide a flexible, data-driven analogue of an autocorrelation structure, but with the advantage of capturing longer correlations and cross-band photometric correlation information. The output representations are then aggregated by taking their mean over the context/target sequence dimension:
where is the length of context/target set. This aggregation produces a permutation-invariant summary of the context light curve. We then concatenate with the normalization statistics , which contain the band-dependent magnitude means and standard deviations. Including allows the latent encoder to retain information about the absolute brightness scale and variability amplitude that is removed during normalization.
The concatenated representation is passed through an MLP to parameterize a Gaussian latent distribution for both context and target set:
where
are output from the MLP given the context or target set. Latent representation and are then sampled from these distributions using the reparameterization trick. This latent variable provides a compact probabilistic summary of the input light curve and is shared by the downstream decoders for light-curve reconstruction and black hole parameter inference.
The latent representation inferred from the context set, , is used for black hole parameter inference and light-curve reconstruction. For black hole parameter inference, we concatenate with the normalization statistics . The combined representation is then passed to a black hole parameter decoder, implemented as an MLP:
The decoder outputs the mean vector and covariance matrix of a multivariate Gaussian distribution over the target physical parameters,
where denotes the set of black hole and variability parameters to be inferred, including , , , and in this research. This probabilistic output allows the model to estimate both the most likely parameter values and their associated uncertainties.
In the deterministic encoder, the model learns a target-specific representation of the light curve conditioned on the context observations. We first apply a masked multi-head self-attention layer to the preliminary context representations :
where denotes the parameters of the deterministic self-attention module. This masked multi-head self-attention operation allows the model to capture temporal correlations and cross-band dependencies among the observed context points.
We then apply a multi-head cross-attention layer to transfer information from the context set to the target locations. In this layer, the self-attention output from context set serves as the value, the context inputs serve as the keys, and the target inputs serve as the queries:
The resulting representation provides a deterministic summary of the context light curve at each target location (i.e., for a specific time and filter). Because the queries are defined at the target points, the cross-attention layer allows the model to infer which context observations are most relevant for predicting each target observation.
The deterministic representation , together with the latent variable inferred from Eq. (28), is passed to the light-curve reconstruction decoder, which predicts the light-curve magnitude distribution at the target times and filters. The decoder consists of a shared trunk and two output branches. The trunk MLP combines the target-specific deterministic representation with the latent variable, the target input location, its time embedding, and the per-band summary statistics ,
from which the mean prediction is obtained through a linear head, . The standard deviation is produced by a dedicated branch that additionally receives distance features — the temporal separation between and the nearest observation, in linear and logarithmic form:
with positivity enforced by a bounded transform. Decoupling the uncertainty branch from the mean prediction, and conditioning it explicitly on the distance to the data, allows the predictive width to contract near observations and grow inside seasonal gaps without perturbing the mean reconstruction.
The decoder thus defines a Gaussian predictive distribution for the underlying light curve at each target location,
During training, the likelihood is evaluated only against observed photometry, whose reported measurement uncertainty is added in quadrature,
so that learns to represent the model’s intrinsic predictive uncertainty about the underlying light curve rather than the photometric noise. This probabilistic formulation allows the model to reconstruct the light curve while providing calibrated uncertainties at every target location.
Loss Function
Following J. W. Park et al. [20], we train the model by maximizing the evidence lower bound (ELBO), or equivalently by minimizing the corresponding negative ELBO loss, augmented with a regularization term that anchors the model’s behavior inside observing gaps. The total loss contains four terms: the negative log likelihood of the light curve , the negative log likelihood of the black hole parameters , the KL divergence between the latent distributions inferred from the target and context sets , and the gap-anchor regularizer :
The hyperparameter controls the relative weight of the light-curve reconstruction term. Increasing places greater emphasis on accurately reconstructing the target light curves, while decreasing it gives relatively more weight to black hole parameter inference and latent-space regularization. The weight controls the strength of the latent regularization relative to the reconstruction terms, in the spirit of the -VAE framework ([9]).
For the light-curve reconstruction term, the decoder predicts a Gaussian distribution at each target point with mean and model variance . Since the target observations also contain LSST photometric noise, we add the observational variance in quadrature. Thus, for each target point ,
The corresponding negative log likelihood is
where the sum is taken over the observed target points after masking padded entries. Here, denotes the normalized LSST observational uncertainty at the -th target point.
For black hole parameter inference, the parameter decoder predicts a multivariate Gaussian distribution with mean and covariance matrix :
where is the true parameter vector, containing the black hole mass, Eddington ratio, redshift, and damped-random-walk timescale. The negative log likelihood is therefore
where is the dimension of the inferred parameter vector.
The KL term regularizes the latent space. Writing and for the latent distributions inferred from the context and the full target set,
where the KL divergence is accumulated per latent dimension . Here we clamp the KL divergence of each dimension to some value . The clamping ensures the minimum information capacity per dimension and prevents the posterior collapse to the prior ([12]). The weight for the KL divergence is increased linearly from 0 to 1 over the first 20 training epochs; this is a standard technical (”annealing”) that allows the latent variable to encode information about the light curve before the regularization toward the context distribution takes full effect. Throughout training the decoders only condition on context latent , so the target-set encoding enters the objective only through this KL term. At deployment, where the full observed light curve constitutes the context, the two distributions coincide by construction. Here we adopt per dimension.
The final term, , addresses a failure mode specific to seasonally gapped cadences. At epochs far from any observation the likelihood provides no meaningful constraint. We find that, without explicit regularization, the predictive mean can develop spurious structure with unreliable uncertainties. We therefore introduce anchor times or query times drawn uniformly at random over the observing baseline. At said anchor times, the decoder is asked to predict flux measurements even where no observed measurements exist. These predictions are never used in the reconstruction likelihood. Instead, at each training step we draw random anchor times per light curve, evaluate the decoder there in every band to obtain , and penalize departures from the AGN’s mean flux and an uncertainty prediction that significantly deviates from the intrinsic variability amplitude :
where is the separation between anchor and the nearest observation, and estimates the intrinsic variability variance of the normalized band from its observed scatter and mean photometric variance. The gates and confine the penalty to genuinely unconstrained epochs, so that anchors falling near observations are unaffected by construction. The width gate with days activates smoothly away from data, while the mean term uses a more conservative gate that is identically zero within 60 days of an observation and ramps to unity over the following 90 days, ensuring that intra-season predictions are never pulled toward the band mean. is thus a prior-consistency penalty only, independent of any data at the anchor locations. We adopt . The gap regularization term basically assumed that the process itself is stationary and thus it’s power spectrum is asymptotically constant. So it’s variability behaves the same in the seasonal gaps.
Training Details
Our final model uses a hidden and latent dimension of 64, with two layers in each of the latent encoder, the deterministic encoder, and the decoder, for a total of 367,953 trainable parameters. All attention modules — the self-attention in both encoders and the cross-attention in the deterministic encoder — use masked multi-head scaled dot-product attention with four heads; masking guarantees that padded entries never contribute to any attention weight. Observation times are encoded with a fixed random Fourier feature embedding of frequencies, log-spaced in period from day to the full observational baseline, supplemented by linear and logarithmic time terms and linearly projected into the 64-dimensional hidden space.
All MLP blocks use Monte Carlo (MC) dropout, which remains active at inference time and, combined with sampling of the latent variable, provides the epistemic contribution to the predictive uncertainty. We use a channel-dropout rate of 0.05 in the encoder and decoder trunks and 0.15 in the black hole parameter decoder. Dropout on the attention weights themselves was evaluated in an ablation and found unnecessary; the final model uses none.
We set in the loss function (Eq. (38)), chosen together with the latent dimension and number of layers via a grid search on the validation loss. We train with the Adam optimizer at an initial learning rate of and a batch size of 20 for 200 epochs (30 h on a single NVIDIA A100), clipping gradients to a global norm of 2. The black hole parameter decoder is a distinct optimizer parameter group with a reduced learning rate of and weight decay of , which we find necessary to balance its convergence against the shared encoders. The learning rate is halved whenever the validation reconstruction loss fails to improve for ten consecutive epochs.
During training, each light curve is turned into a prediction episode by sampling from its observed points. In other words, a random fraction, drawn uniformly between 30% and 95%, forms the context set, and with probability 0.7 up to three contiguous windows, each spanning 5–25% of the baseline, are additionally removed from the context to imitate extended data gaps. The full set of observed points serves as the target. The training, validation, and test sets contain 50,018, 20,029, and 8,000 simulated light curves, respectively, generated with disjoint random seeds.
Results
Black Hole Parameter Predictions
To demonstrate the effectiveness of the learned latent representation, we infer AGN physical and variability parameters from the latent representation using an MLP decoder. Figure 4 compares the predicted posterior mean, , with the ground-truth parameter values for light curves in the testing set. The white dashed diagonal line indicates the one-to-one relation corresponding to perfect recovery. The predictions for log(), log(), and follow the diagonal, indicating that the latent representation preserves information relevant to the underlying AGN parameters.

Figure 4. Model predictions compared with ground-truth values for log(), log(), , and log() on the testing data set. Each panel shows a two-dimensional histogram, with color indicating the number of objects in each bin. The dashed white line marks the one-to-one relation. The alignment of the distributions with the diagonal indicates that the model captures the dependence of the light curves on the AGN physical and variability parameters. The deviation in log() at low log(), corresponding to short-timescale or high-frequency variability, is likely caused by the finite LSST cadence. Since LSST samples the light curves on timescales of order a few days, variability on shorter or comparable timescales is difficult to resolve and may be affected by aliasing.
The prediction of log() shows larger deviations at both small and large values of log(). At log() days, corresponding to short-timescale or high-frequency variability, the finite LSST cadence limits the ability to resolve rapid variations and can introduce aliasing effects. At large log() days, the variability timescale becomes comparable to, or longer than, the survey baseline, making it difficult to constrain the full characteristic timescale from the observed light curve.
Figure 5 shows the posterior distribution of the inferred parameters for a representative testing light curve. The diagonal panels show the one-dimensional marginal posterior distributions, while the lower off-diagonal panels

Figure 5. Example posterior distribution inferred by the ANP model for a single testing light curve. The diagonal panels show the one-dimensional marginal posterior distributions for log(), log(), , and log(), while the off-diagonal panels show the corresponding two-dimensional joint posterior distributions. Dashed black lines denote the ground-truth parameter values, solid gray lines denote the predicted posterior means, the black cross on at the center of the distribution indicates the predicted posterior mean and blue contours indicate the posterior probability density. This corner plot evaluates the performance of the Bayesian inference by comparing the inferred posterior distributions with the ground-truth parameter values.
show the two-dimensional joint posterior distributions between parameter pairs. All physical parameters are well-constrained with our model.
We further evaluate the uncertainty calibration of the inferred SMBH parameters using the coverage calibration curves shown in Figure 6. For each parameter and each nominal credible level , we compute the fraction of test objects whose ground-truth value falls within the central credible interval of the marginal posterior. A perfectly calibrated posterior follows the one-to-one relation shown by the dashed line, curves above it indicate conservative (over-wide) uncertainties, and curves below it indicate over-confidence. At the 68% (95%) nominal level the empirical coverages are 82.5% (98.7%) for log(), 72.2% (96.9%) for log(), 77.6% (99.5%) for , and 75.1% (98.5%) for log(): the credible intervals are modestly wider than necessary, and no credible level for any parameter is over-confident.

Figure 6. Calibration curve for the posterior distributions of the inferred AGN host SMBH parameters on the testing set. The horizontal axis shows the fraction of posterior probability volume, and the vertical axis shows the fraction of true parameter values contained within that volume. The dashed black line indicates perfect calibration, for which the credible regions contain the true values at the expected frequency.
We quantify accuracy with the normalized RMSE of the posterior mean, defined as the RMSE divided by the width of the corresponding uniform prior. We obtain 8.2% for , 12.0% for , 11.8% for , and 16.8% for . For reference, an estimator that always returns the prior mean yields for a uniform prior, so all four parameters are recovered with substantially more information than the prior provides.
We next investigate the DRW time scale, which is the only parameter that can be directly inferred by both our model and the GP baseline. To obtain the DRW time scale from GP model, we take the median of the six per-band maximum a posteriori kernel timescales and convert to the rest frame using the true redshift. We note that this comparison gives an advantage to the GP model, since our network must actually infer the redshift itself. On a common subset of test objects, the GP recovers with a normalized RMSE of 20.2% and a systematic bias of dex — the characteristic underestimate of maximum-likelihood timescale fits on baselines comparable to — whereas our ANP model achieves 16.0% with a bias of dex and is closer to the truth for 61% of individual objects. We attribute the success of our model over the baseline to the amortized population prior. In other words, because our is trained across light curves, the network implicitly performs the hierarchical shrinkage that individual object fits cannot.
Light-curve Reconstruction
We next evaluate the ability of the attentive neural process model to reconstruct LSST-like AGN light curves.
Figure 7 shows an example reconstruction for a light curve from the testing set. Across all six LSST bands, the calibrated posterior mean predicted by the model closely follows the underlying simulated light curves, indicating that the model can recover the temporal variability structure from irregularly sampled and noisy observations. We also find that the predictive uncertainty behaves consistently with the available observational information. In contrast to the model of J. W. Park et al. [20], which predicts a relatively uniform and broad uncertainty, our model produces uncertainties that decrease near epochs with observed data by introducing the distance feature into MLP (Eq. (35)). This local reduction in predictive uncertainty reflects the expected posterior uncertainty shrinkage when the model is conditioned on nearby measurements. Similarly, the uncertainty increases away from observed epochs.

Figure 7. Example reconstruction of a simulated multi-band LSST AGN light curve from the testing set. Each panel shows one LSST photometric band ( through ) as a function of time. Gray points with error bars represent the simulated LSST observations, and dashed gray curves show the underlying simulated light curves prior to observational sampling. Solid blue curves show the ANP posterior mean prediction, with shaded blue regions indicating the and (lighter blue) predictive uncertainty. The ANP reconstruction closely follows the underlying light curves across all bands, demonstrating that the model can recover the temporal variability structure from irregularly sampled, noisy multi-band observations. The predictive uncertainty is reduced near epochs with observations, consistent with expected posterior contraction when the model is conditioned on nearby measurements.
We compare the ANP model against a baseline Gaussian process (GP) across several metrics. For the GP we adopt the DRW kernel and fit each object and band individually by maximum a posteriori estimation. Since the simulated driving signals are themselves DRWs, this baseline is a nearly optimal reference for light curve reconstruction.
Table 2 shows point-wise prediction accuracy for the light curve shown in Figure 7. The two models perform comparably (RMSE 0.174 versus 0.178). For this example the ANP produces a mean residual of 0.061 and standardized residuals with Std() = 1.01 and Mean = 0.837, close to the values of 1 and expected under , indicating essentially unbiased and well-scaled uncertainties for this object. Figure 8 extends the residual analysis. Both models produce similar residual and standardized-residual distributions, consistent with near-Gaussian behavior. In the normal Q–Q plots the ANP attains
| Metric | ANP | GP |
| MAE | 0.122 | 0.115 |
| RMSE | 0.178 | 0.174 |
| Mean residual | 0.0605 | 0.0637 |
| Std() | 1.01 | 1.08 |
| Mean | 0.837 | 0.880 |
Table 2. Summary of error and calibration metrics for the ANP and baseline GP models. Both models exhibit small mean residuals, indicating negligible systematic bias. While the GP achieves lower MAE and RMSE, the ANP model produces standardized residuals closer to a standard normal distribution (Std() , Mean ), indicating better-calibrated uncertainty estimates.

Figure 8. Residual analysis for the ANP model (left column) and the baseline GP model (right column). The top row shows histograms of raw residuals overlaid with fitted normal distributions (black dashed curves). The middle row shows standardized residual histograms overlaid with the standard normal . The bottom row shows normal Q–Q plots of the standardized residuals against theoretical standard normal quantiles; points lying along the diagonal (dashed line) indicate perfect Gaussian calibration. The ANP model achieves a Q–Q compared to 0.9900 for the GP, confirming better uncertainty calibration in this particular object.
against for the GP, implying that both are Gaussian-like. We next examine the power spectral density of the -band reconstructions (Figure 9), estimated with the Welch method [30], to assess how much of the temporal variability structure each reconstruction preserves. In general, the power spectra lose power at frequencies unconstrained by the data; the relevant comparison is how much constrained power each model retains. The ANP reconstruction traces the true spectrum substantially further into the high-frequency regime, whereas the GP reconstruction exhibits suppressed power beginning at intermediate frequencies ().

Figure 9. Power spectral density of the -band for the true underlying light curve (gray), the ANP reconstruction (blue), and the baseline Gaussian Process (GP) reconstruction (orange). The power spectrum is calculated based on mean reconstruction using Welch method with frequency bin size of . The ANP model more faithfully reproduces the power spectrum of the true light curve at high frequencies, indicating better recovery of short-timescale variability.
We also report the NLL comparison between our ANP model and the GP model across the entire testing data set in Tab. 5. We see that overall our ANP model’s light curve reconstruction performance can approach per-object DRW Gaussian-process fits — the near-optimal baseline on DRW data.
Finally, Figure 10 assesses the calibration of the predictive uncertainties over the full testing set, evaluated
Figure 10. Calibration curve for light curve reconstruction over the full validation set, comparing the ANP (blue) and the baseline GP model (orange) against perfect calibration (gray dashed diagonal). The -axis shows the nominal posterior probability volume , and the -axis shows the empirical fraction of true underlying light curve values falling within the central credible interval of the predictive posterior. A perfectly calibrated model should lie along the diagonal, where the stated credible intervals contain the true values at exactly the corresponding rate.
against the true underlying light curves on a dense time grid spanning observed and unobserved epochs. For a nominal credible level , the curve shows the empirical fraction of true values contained within the central credible interval. Both models lie on or above the diagonal at all credible levels — neither is over-confident anywhere — with the GP, as expected for the matched kernel, closest to perfect calibration, and the ANP somewhat more conservative: its 68% (95%) intervals contain 76.38% (95.77%) of the true values, versus 72.05% (94.82%) for the GP.
The Q–Q plots, standardized-residual statistics, the NLL calculation and calibration curves show that the ANP matches the near-oracle GP in reconstruction accuracy and residual Gaussianity while providing modestly conservative, never over-confident, uncertainty estimates across epochs, bands, and objects.
Discussion and Future Work
In this work, we present an upgraded attentive neural process (ANP) model based on the framework of J. W. Park et al. [20]. The model is designed to reconstruct AGN light curves across the six LSST bands and to infer physical and variability parameters of the central black hole system. Compared with the original ANP framework, our model is trained on more realistic simulated LSST light curves that include wavelength-dependent transfer functions, the fielded LSST observing cadence, and per-visit photometric noise. Most importantly, our data representation explicitly accounts for the asynchronous sampling of different photometric bands, which is a key feature of realistic LSST observations.
The light-curve reconstruction component of the model is trained in a self-supervised manner, using realistic simulated observations that include LSST-like cadence and noise properties, rather than the noiseless continuous light curves themselves. As a result, the model learns to infer the latent temporal structure of the source solely from sparse, noisy, and irregularly sampled light curves. In contrast, the parameter-inference decoder is trained in a supervised manner using the known simulation parameters. This hybrid training strategy allows the model to pre-learn a compact latent representation that supports both light-curve reconstruction and Bayesian parameter inference. In future work, this framework can be further trained or fine-tuned using the millions of real AGN light curves expected from LSST, as well as the smaller subset of objects with independently inferred physical properties, for example from spectroscopic measurements.
Our work belongs to a growing family of neural-process approaches to AGN variability. I. Čvorović-Hajdinjak et al. (2022) first applied conditional neural processes to non-parametric AGN light-curve modeling, demonstrating flexible interpolation of survey light curves, though without parameter inference or an assessment of predictive-uncertainty calibration. A. N. Raju et al. [22] recently combined attentive latent neural processes with self-organizing-map clustering and mixture density networks to recover accretion-disk transfer functions and SMBH parameters, reporting distribution-level agreement between predicted and true parameter populations. Our approach is complementary along several axes: we model the six bands jointly within a single attention mechanism rather than per band, train on visit lists drawn from the fielded Rubin observing strategy with per-visit noise rather than on synthetic burst cadences with noiseless targets, and evaluate per-object posterior quality — normalized RMSE and empirical coverage of individual credible intervals — which is a stricter criterion than population-level distributional agreement. Conversely, A. N. Raju et al. [22] recover the transfer function itself, which our model treats as part of the simulator, and their clustering-based framework is designed to operate in the few-hundred-light-curve regime, whereas ours assumes abundant training simulations.
Compared with the latent-SDE framework of J. Fagin et al. [7], our ANP model provides a complementary approach that emphasizes architectural compactness, efficient amortized inference, and reusable latent representations. The latent-SDE model embeds a physically motivated continuous-time stochastic process and an auto-differentiable accretion-disk reprocessing model within the neural-network architecture, providing a highly self-consistent physical framework. In contrast, our model uses the physical reprocessing model to generate realistic training data, but does not require solving a latent SDE or evaluating transfer-function convolutions during inference. As a result, the ANP model has fewer trainable parameters and is faster to train in our implementation: while the latent-SDE model was reported to require approximately three weeks of training (J. Fagin et al. [7]), our ANP model reaches convergence in approximately thirty hours, providing a computationally efficient pathway for future fine-tuning on real LSST data. The parameter-inference and reconstruction performance we obtain is qualitatively comparable to that reported by J. Fagin et al. [7].
The compact latent representation learned by the ANP also provides a useful summary of the information contained in the light curves. As demonstrated by the black hole parameter-inference results, this latent representation encodes information about the host AGN system and its variability. This makes the representation useful not only for light-curve reconstruction, but also for additional downstream tasks. In future work, we plan to explore applications such as anomaly detection, binary AGN classification, and identifying unusual variability patterns in large-scale time-domain surveys.
| Metric | ANP | GP |
| Mean NLL | −2.35 | −2.36 |
| Std NLL | 0.983 | 0.984 |
| Median NLL | −2.47 | −2.40 |
| Meadian absolute deviation | 0.810 | 0.835 |
| Min NLL | −3.95 | −4.10 |
| Max NLL | −0.315 | −0.427 |
Table 3. Negative Log-Likelihood (NLL) per Light curve object (Averaged over time and bands) for the ANP and baseline GP models evaluated on the whole testing set. Lower NLL indicates better probabilistic predictive performance.
Overall, our results suggest that the upgraded ANP model provides a flexible and efficient framework for simulation-based AGN light-curve analysis. Its ability to handle irregular, asynchronous multi-band observations makes it particularly well suited to LSST-like data. A natural next step is to apply the model to real AGN light curves and to investigate simulation-to-real transfer, including fine-tuning strategies in which model selection relies only on observable quantities, as no noiseless ground truth exists for real data. This provides a pathway toward using simulation-trained neural-process models for large-scale inference on real time-domain survey data.
Acknowledgments
This work made use of Claude Code (Anthropic) for assistance with code development, data-validation tooling, and manuscript polishing. All results were verified by the authors, who take full responsibility for the content of this paper.
The Villar Astro Time Lab acknowledges support through the David and Lucile Packard Foundation, the Research Corporation for Scientific Advancement (through a Cottrell Fellowship and Bridge Award) and the National Science Foundation under AST-2433718, AST-2407922 and AST-2406110. This work is supported by the National Science Foundation under Cooperative Agreement PHY-2019786 (the NSF AI Institute for Artificial Intelligence and Fundamental Interactions). T.S. Thanks the Ph.D. fellowship support from Think Global Education Trust. C.X. thanks the support by the Black Hole Initiative, which is funded by grants from the John Templeton Foundation (Grant #63445) and the Gordon and Betty Moore Foundation (Grant #13526), although the opinions expressed in this work are those of the author(s) and do not necessarily reflect the views of these Foundations. KESF is supported by NSF AST-2206096, NSF AST-1831415 and Simons Foundation Grant 533845. Support was provided by Schmidt Sciences, LLC. for J.F.
References
- [1]Blandford, R. D., & McKee, C. F. 1982, The Astrophysical Journal, 255, 419, doi: 10.1086/159843
- [2]Cackett, E. M., Bentz, M. C., & Kara, E. 2021, iScience, 24, 102557, doi: 10.1016/j.isci.2021.102557
- [3]Chan, J. H.-H., Fagin, J., Best, H., & O’Dowd, M. J. 2024, The Astrophysical Journal, 978, 54, doi: 10.3847/1538-4357/ad93af
- [5]Dubois, Y., Gordon, J., & Foong, A. Y. 2020, Neural Process Family,, http://yanndubs.github.io/Neural-Process-Family/
- [6]Edelson, R., Gelbord, J. M., Horne, K., et al. 2015, The Astrophysical Journal, 806, 129, doi: 10.1088/0004-637X/806/1/129
- [7]Fagin, J., Chan, J. H.-H., Best, H., et al. 2025, The Astrophysical Journal, 988, 59, doi: 10.3847/1538-4357/addabc
- [8]Giveon, U., Maoz, D., Kaspi, S., Netzer, H., & Smith, P. S. 1999, Monthly Notices of the Royal Astronomical Society, 306, 637, doi: 10.1046/j.1365-8711.1999.02556.x
- [9]Higgins, I., Matthey, L., Pal, A., et al. 2017, in International Conference on Learning Representations. https://openreview.net/forum?id=Sy2fzU9gl
- [10]Jones, R. L. 2022, Calculating LSST Limiting Magnitudes and SNR,, Rubin Observatory Simulations Technical Note SMTN-002 https://smtn-002.lsst.io/
- [11]Kelly, B. C., Bechtold, J., & Siemiginowska, A. 2009, The Astrophysical Journal, 698, 895, doi: 10.1088/0004-637X/698/1/895
- [12]Kingma, D. P., Salimans, T., Jozefowicz, R., et al. 2017, Improving Variational Inference with Inverse Autoregressive Flow, https://arxiv.org/abs/1606.04934
- [13]Liu, H. T., Bai, J. M., Zhao, X. H., & Ma, L. 2008, The Astrophysical Journal, 677, 884, doi: 10.1086/529361
- [14]LSST Science Collaboration, Abell, P. A., Allison, J., et al. 2009, e-print arXiv:0912.0201. https://arxiv.org/abs/0912.0201
- [15]Lynden-Bell, D. 1969, Nature, 223, 690, doi: 10.1038/223690a0
- [16]MacLeod, C. L., Ivezić, Ž., Kochanek, C. S., et al. 2010, The Astrophysical Journal, 721, 1014, doi: 10.1088/0004-637X/721/2/1014
- [17]McHardy, I. M., Koerding, E., Knigge, C., Uttley, P., & Fender, R. P. 2006, Nature, 444, 730, doi: 10.1038/nature05389
- [18]Mushotzky, R. F., Edelson, R., Baumgartner, W., & Gandhi, P. 2011, The Astrophysical Journal Letters, 743, L12, doi: 10.1088/2041-8205/743/1/L12
- [19]Netzer, H. 2015, Annual Review of Astronomy and Astrophysics, 53, 365, doi: 10.1146/annurev-astro-082214-122302
- [20]Park, J. W., Villar, A., Li, Y., et al. 2021, Inferring Black Hole Properties from Astronomical Multivariate Time Series with Bayesian Attentive Neural Processes, https://arxiv.org/abs/2106.01450
- [21]Pereyra, N. A., Vanden Berk, D. E., Turnshek, D. A., et al. 2006, The Astrophysical Journal, 642, 87, doi: 10.1086/500919
- [22]Raju, A. N., Kovačević, A. B., Ilić, D., et al. 2026, arXiv e-prints, arXiv:2606.09665. https://arxiv.org/abs/2606.09665
- [23]Shakura, N. I., & Sunyaev, R. A. 1973, Astronomy & Astrophysics, 24, 337
- [24]Soldi, S., Türler, M., Paltani, S., et al. 2008, Astronomy & Astrophysics, 486, 411, doi: 10.1051/0004-6361:200809947
- [25]Ulrich, M.-H., Maraschi, L., & Urry, C. M. 1997, Annual Review of Astronomy and Astrophysics, 35, 445, doi: 10.1146/annurev.astro.35.1.445
- [26]Vanden Berk, D. E., Wilhite, B. C., Kron, R. G., et al. 2004, The Astrophysical Journal, 601, 692, doi: 10.1086/380563
- [27]Viitanen, A., Bongiorno, A., Saccheo, I., et al. 2026, Astronomy & Astrophysics, 710, A56, doi: 10.1051/0004-6361/202558627
- [28]Wanders, I., Peterson, B. M., Alloin, D., et al. 1997, The Astrophysical Journal Supplement Series, 113, 69, doi: 10.1086/313054
- [29]Wang, S., Guo, H., & Woo, J.-H. 2023, The Astrophysical Journal Letters, 948, L23, doi: 10.3847/2041-8213/accf96
- [30]Welch, P. D. 1967, IEEE Transactions on Audio and Electroacoustics, 15, 70, doi: 10.1109/TAU.1967.1161901
- [31]Wilhite, B. C., Vanden Berk, D. E., Kron, R. G., et al. 2005, The Astrophysical Journal, 633, 638, doi: 10.1086/430821
- [32]Xin, C., & Haiman, Z. 2021, Monthly Notices of the Royal Astronomical Society, 506, 2408, doi: 10.1093/mnras/stab1856

