Harder, or Just Later? Blind Deconvolution of Blazar Flare Hysteresis

When a blazar (a galaxy whose central jet points almost at us) flares, its spectrum usually shifts toward higher-energy photons as it brightens. This "harder-when-brighter" behavior is read as evidence that the jet's particle accelerator changed: newly injected electrons carry a flatter energy distribution. There is a second way to get the same signature. High-energy electrons lose energy faster than low-energy ones, so the X-ray and gamma-ray bands respond to an injection quickly while lower-energy bands respond slowly and with a long tail. Shared brightening seen through band-dependent delays will trace hardness–flux loops even if the injected spectrum never changes. We propose to treat this as a blind deconvolution problem. Each band's light curve is a common injection history smeared in time by an unknown band-specific response kernel, and the question "is the hardening intrinsic or caused by lags?" is exactly the ambiguity between signal and kernel that blind deconvolution is known for.

The claim is that this ambiguity can be quantified and often broken with multi-band time series that already exist. Two things make that possible: all bands share one injection history, and the kernels obey physical constraints (they are causal, decay monotonically, and have response times ordered by photon energy). We solve the resulting non-convex problem with half-quadratic splitting, which alternates cheap closed-form steps and replaces the slow derivative-free searches now used for flare spectral inversion. The output is a decomposition of each observed hysteresis loop into an intrinsic-hardening part and a lag part, together with a profile showing how strongly the data actually constrain that split.

Why this intersection, and what is already known

The phenomenology is well established. Hardening during flux increases has been documented in the X-ray and very-high-energy (VHE) gamma-ray spectra of Markarian 421 and 1ES 0647+250. The closest and most important prior work is Time-dependent Modeling of the Subhour Spectral Evolution during the 2013 Outburst of Mrk 421. That paper fit a two-zone synchrotron self-Compton model (one slow steady zone plus one compact fast zone) to 240 broadband spectra binned at 15 minutes. It found clockwise hysteresis loops in X-rays and VHE gamma-rays and attributed them to changes in both the injected electron luminosity and the injected slope, consistent with the shock compression ratio varying by a factor of about 2.

That paper motivates our work and is also our strongest competitor. Its answer to "intrinsic versus lag" comes from inside one physical model. It does not separate what the data demand from what the model's parameterization allows. Two other results constrain what we can assume. X-ray intraday variability of OJ 287 shows that the same source can be softer-when-brighter over long times and can produce a major burst in 2020 with no spectral evolution at all. Harder-when-brighter therefore cannot be imposed as a prior; it has to be an output. This is also why we reject the obvious combination the literature seems to invite, which is using harder-when-brighter as a regularizer for multi-band image deblurring. The relation fails across cooling phases and thermal events, and using it as a prior would assume the very thing astronomers want to measure.

The deblurring side supplies both the diagnosis and the method. The classic failure of naive maximum-a-posteriori blind deconvolution is the "no-blur" solution, in which the kernel collapses to an impulse and the observation is returned as the answer. The file describes this trap, and Noise-Blind Image Deblurring documents the analogous collapse for noise variance. Our problem has the same attractor. Kernels that collapse to impulses force all spectral evolution into the intrinsic term. Vision methods break the attractor with image statistics: $L_0$ sparsity of intensity and gradients, and the dark channel and extreme channels priors. All three optimize with half-quadratic splitting. We keep their optimizer and discard their priors, since light curves have no dark pixels or two-toned text.

Constraining kernels to a low-dimensional family follows the spirit of kernel manifold constraints, but here the family comes from physics rather than learning. Multi-frame astronomical blind deconvolution, as in Hirsch et al., handles time-varying spatial blur but treats channels independently. The shared-source coupling across bands is precisely what we exploit.

Core Claim and Model

Let $Y_\nu(t)$ be the observed light curve in energy band $\nu$ on a regular grid (15-minute bins for Mrk 421), with a mask marking gaps. We model it as

$$ Y_\nu(t) = \big(K_\nu * [\,a\, f_0(\nu) + b\, f_1(\nu)\,]\big)(t) + c_\nu + N_\nu(t). $$

The terms are defined as follows. $a(t)\ge 0$ is the shared injection luminosity history, and $b(t)$ is the shared change in injected spectral slope. The band weights $f_0(\nu)$ and $f_1(\nu)$ come from a first-order expansion of a power-law injection around a reference slope. $K_\nu$ is a causal, normalized, monotonically decaying response kernel with characteristic time $\tau_\nu$. $c_\nu$ is a steady per-band baseline, playing the role of the slow zone. $N_\nu$ is noise.

For synchrotron bands, cooling time scaling as $E^{-1}$ and emitted frequency scaling as $E^2$ suggest $\tau_\nu \propto \nu^{-1/2}$. The VHE inverse-Compton band does not follow this scaling. We therefore impose only an ordering constraint, with $\tau_\nu$ non-increasing in photon energy within each emission component, and test the $-1/2$ exponent as an output.

In this model, apparent hysteresis with $b\equiv 0$ is pure lag. Hysteresis with $\tau_\nu\to 0$ is pure intrinsic hardening, which is the no-blur attractor. The estimate minimizes

$$ \sum_{\nu,t} m_\nu(t)\,\rho\big(Y_\nu(t) - \hat Y_\nu(t)\big) + \lambda_a \|\nabla a\|_1 + \lambda_b \|\nabla b\|_1 , $$

where $\hat Y_\nu$ is the model prediction and $m_\nu(t)\in\{0,1\}$ masks missing bins. The loss $\rho$ is a robust Welsch loss, which down-weights residuals from instrument glitches and unmodeled micro-flares. The two $\ell_1$ terms favor piecewise-smooth injection histories.

The penalty on $b$ could bias the answer toward "pure lag." We therefore choose $\lambda_b$ by held-out-band prediction and never by fit quality. That choice matters most when the two explanations fit equally well, which is exactly when the bias would otherwise decide the answer.

Research Program

The optimizer is a half-quadratic scheme with three alternating blocks. First, the robust loss becomes iteratively reweighted least squares: auxiliary weights are updated in closed form from the current residuals, following the half-quadratic treatment of correntropy-type losses in Zhang et al.. Second, splitting variables for $\nabla a$ and $\nabla b$ are updated by soft-thresholding. Third, with kernels and weights fixed, $(a,b,c)$ solves a weighted linear least-squares problem, which we handle by conjugate gradient with FFT-based convolutions under the non-negativity constraint. With the signal fixed, the kernel parameters $\{\tau_\nu\}$ are only a handful of scalars. We update them by projected Newton steps under the ordering constraint, and restart from a coarse grid to avoid poor basins.

The relaxation parameter is increased on a continuation schedule, which is the standard remedy for the $\beta$-sensitivity noted for splitting methods. Convergence of the augmented objective is monotone by construction. We do not claim global optimality. Instead, we compute an ambiguity profile: we fix a global lag scale $\tau_0$ on a grid from 0 to large values, re-optimize everything else, and report the resulting objective and the intrinsic fraction. A flat profile is itself a scientific result, because it means the data cannot distinguish lags from hardening.

The main derived quantity is the intrinsic loop fraction. We recompute the model hardness–flux loop with $b$ set to zero and with all kernels set to impulses. The fraction of the observed loop area that survives with lags removed is the fraction attributable to intrinsic hardening. A bootstrap over time blocks gives its uncertainty.

Baselines

The naive reference point is the raw hardness–flux relation: a linear regression slope of hardness on flux, plus the signed loop area, computed from the Mrk 421 and OJ 287 data with no modeling. This is what most observational papers report. It cannot separate the two mechanisms, and on simulations it serves as the "do nothing" error level.

The strongest existing method is the time-dependent two-zone synchrotron self-Compton fit of the Mrk 421 2013 study. It is strongest because it already encodes cooling physically and returns time histories of injected luminosity and slope, which map directly onto our $a$ and $b$. We will reimplement a two-zone kinetic model and fit it to exactly the same 240 binned spectra and the same simulated data. Both methods get equal wall-clock budgets and an equal number of restarts, and hyperparameters for both are chosen on the same held-out bands.

The file reports that model's qualitative outputs: clockwise loops in both X-ray and VHE bands, and a compression ratio change of a factor of about 2. No quantitative goodness-of-fit or parameter-recovery metric was found on Lacuna for it. On real data, beating it therefore means better held-out-band predictive likelihood. On simulations, it means lower error in recovering the true slope history when the generating physics departs from two zones.

A lag-corrected cross-correlation baseline estimates one pure time shift per band from cross-correlation peaks, shifts the light curves into alignment, and then recomputes the hardness–flux loop. This is the simplest way an astronomer would try to remove lags. Our method beats it if kernel shape, not just delay, measurably changes the inferred intrinsic fraction.

Two ablations isolate the key ingredients. One removes the physical kernel constraints and allows free non-negative kernels per band. This isolates whether causal, ordered kernels are what break the no-blur attractor. The other keeps the model and replaces half-quadratic splitting with a genetic algorithm and with Markov chain Monte Carlo sampling, mirroring the derivative-free approach of the Siberian Radioheliograph flare fits. This isolates the optimizer's contribution to speed and to final objective value.

Why not just…?

Why not just fit the two-zone kinetic model, as the Mrk 421 study did? This is the real threat. If the true emitting geometry is two zones with the assumed cooling, a correctly specified physical fit should match or beat us, and we will report that honestly. Our bet is about misspecification. Simulations with three zones, adiabatic losses, or an evolving magnetic field will show whether the kinetic fit misattributes lag-driven loops to slope changes, the same kind of confusion the solar fits face between intrinsic change and source motion. Our method needs only causality and energy ordering. If the two-zone fit is equally robust under these mismatched simulations, our contribution shrinks to the ambiguity profile and the speed.

Why not just shift light curves by their cross-correlation lags? A pure shift cannot reproduce the long, energy-dependent tails that cooling produces. Tails, not offsets, are what generate much of the loop area. The lag-corrected baseline above tests this directly on simulations where the true kernels are known exponentials. If shifting recovers the intrinsic fraction as well as full deconvolution does, we will conclude that the simpler tool suffices.

Why not just use a generic blind deconvolution method with non-negativity and sparsity? Without physical kernel structure, the problem keeps the no-blur attractor and its opposite: kernels can absorb arbitrary spectral change, or collapse to impulses. The vision priors that break the attractor in images (dark channel, extreme channels) rely on pixel statistics with no counterpart in light curves. The free-kernel ablation measures how much ambiguity remains when the constraints are dropped. We expect a flat ambiguity profile there, and a peaked one with the constraints.

Why not use harder-when-brighter itself as the prior? Because OJ 287 shows flares with no spectral change and long-term softer-when-brighter behavior. A monotonic prior would manufacture the effect. Our design leaves $b$ unconstrained in sign.

Proposed Experiments

The first test is on simulated data where the truth is known. We generate multi-band light curves from a kinetic electron model under three regimes: pure injection-slope change with no lags, pure cooling lags with a constant slope, and mixtures. Each regime is generated with two-zone and misspecified multi-zone geometries, sampled at the Mrk 421 band set and 15-minute cadence with realistic gaps and noise. Metrics are the absolute error of the intrinsic loop fraction, correlation between recovered and true $b(t)$, held-out-band log-likelihood, and runtime.

The decisive check comes from the pure-lag and pure-intrinsic regimes. The method must return near-zero and near-one intrinsic fractions respectively, with bootstrap intervals that exclude the wrong extreme. Otherwise the ambiguity is not being broken.

The second test uses the Mrk 421 2013 dataset of 240 spectra. We report the intrinsic loop fraction for X-ray and VHE loops separately, compare the recovered $b(t)$ against the slope history inferred in the original modeling, and test whether the fitted $\tau_\nu$ follow the $\nu^{-1/2}$ scaling within the synchrotron bands.

The third test is a negative control on the XMM-Newton OJ 287 pointings, including the burst with no spectral evolution. There the method should return $b$ consistent with zero while still fitting the flux. If it instead hallucinates hardening, the anti-lag bias concern is confirmed and $\lambda_b$ selection must be revisited.

A stretch application uses the Siberian Radioheliograph flares, where the electron index hardens at peak. There the frequency-dependent delay model would test how much of that hardening survives once lags are removed.