Reachable-Set Certificates for Blazar Flare Mechanisms

Blazars are galaxies whose central black holes shoot jets of plasma almost straight at us. Electrons in the jet are pushed to enormous energies, and we see the result as light spanning radio waves through very-high-energy gamma rays. The central open question is how the electrons are pushed. Candidate mechanisms include repeated bouncing across a shock front, magnetic reconnection (the sudden snapping and rejoining of magnetic field lines), and various multi-zone geometries.

Today these mechanisms are compared by fitting a forward model to the observed spectrum. The analyst picks a handful of parameters and a simple shape for how particles are injected over time, then tunes until the curve passes through the data. When the fit fails or needs absurd values, nobody can tell whether the physics is wrong or the chosen injection shape was too narrow. The 2011 outburst of 1ES 1215+303 is the textbook case. A one-zone model needed a Doppler beaming factor near 60 or an unusually high minimum electron energy (Aleksić et al.).

We propose to ask a different question, one borrowed from safety verification in robotics. Instead of "which parameters fit best?", we ask "can any acceleration history allowed by this mechanism produce what we observed?" That is a reachability question: it asks for the set of all states a system can reach under every admissible input. Answering it gives a checkable certificate. Either the observed flare lies inside the mechanism's reachable set, or there is a proof that no allowed history could produce it.

Core Claim

The electron population in a blazar emission zone is described by a kinetic equation for how many particles sit at each energy. Discretized into $n$ energy bins, it becomes an ODE with $n$ in the range of 50 to 200. That size is far beyond grid-based Hamilton–Jacobi methods; Gillula and Tomlin note these are practically limited to 2–3 states.

The equation has a structure that makes high-dimensional reachability tractable without a grid:

  • With the magnetic field fixed, synchrotron cooling and escape act linearly on the particle spectrum.
  • Particle injection enters linearly.
  • Every observed flux is a linear functional of the spectrum.

So we write the core dynamics as

$$\dot N = A\,N + G\,u(t) + w(t), \qquad y = C\,N,$$

where:

  • $N \in \mathbb{R}^n$ is the binned electron spectrum.
  • $A$ encodes cooling and escape at a nominal field strength.
  • $u(t)$ is the time-varying injection or acceleration input, restricted to an admissible set $U$ that defines the mechanism hypothesis (for example, the injection-index range and power budget a shock model allows).
  • $w(t)$ is a bounded term absorbing everything nonlinear: self-Compton cooling, field fluctuations, and the bilinear energy-gain term of in-situ acceleration.
  • $y$ is the vector of observable band fluxes.

Following Sharpless, Chow and Herbert, who linearize nonlinear dynamics and treat the bounded linearization error as an adversarial player so that a Hopf-type formula yields guaranteed conservative reachable sets without grids, the reachable set of observables at time $T$ has a support function. A support function gives, for any direction $\ell$ in observable space, the furthest the reachable set extends along $\ell$:

$$h_T(\ell) = \ell^\top C e^{AT} N_0 + \int_0^T \Big[\max_{u \in U} \ell^\top C e^{A(T-s)} G u + \max_{|w| \le \bar w(s)} \ell^\top C e^{A(T-s)} w\Big]\, ds .$$

Here $N_0$ is the initial spectrum, and $\bar w(s)$ is the bound on the nonlinear residual.

Because the inner maximizations are over the input set, not over trajectories, $h_T(\ell)$ costs one backward integration per direction regardless of $n$. If some direction satisfies $h_T(\ell) < \ell^\top y_{\text{obs,lo}}$ (the observed flux interval's lower bound projected on $\ell$), the observation is certifiably unreachable. The pair $(\ell, h_T(\ell))$ is a compact certificate that anyone can recheck.

The new thing is not "machine learning for SED fitting." It is turning mechanism discrimination into a for-all-input-histories infeasibility question, and exploiting the near-linear kinetic structure so that this question stays cheap in 100+ dimensions.

Research Program

Building the model. The first component is the mechanism-as-input-set formulation. We encode three hypothesis classes as admissible sets $U$, each paired with explicit bounds $\bar w$ derived from energy-density limits:

  • shock injection with a bounded spectral index and rate;
  • harder, faster reconnection-style injection;
  • in-situ Fermi acceleration with a bounded acceleration time.

These are inputs the analyst states openly, not fitted parameters. We start with leptonic, synchrotron-dominated sources, where the linear part dominates and the residual is small. The lepto-hadronic cascades used for TXS 0506+056 (MAGIC and multiwavelength study) have strongly nonlinear coupling, so they are deferred to a stress test.

Two-zone systems. The second component handles stratified geometries of the kind IXPE motivated for Mrk 501: a compact zone behind a shock feeding an extended downstream region (Abe et al.). Here we borrow the decomposition idea of FaSTrack, which splits a 6D quadrotor into small subsystems and bounds the coupling error. In our case, the downstream zone sees the shock zone's escaping particles as a bounded input. This keeps each zone's reachable set computable separately while the composition stays sound.

Measuring conservativeness. The third component pairs every outer bound with an inner bound. We optimize explicit input histories to push the simulated observables as far as possible along each $\ell$. The gap between the inner value and $h_T(\ell)$ quantifies how much the certificate gives away.

Software. Everything is built on a standard kinetic forward solver for validation, and we will release the certificate checker.

Baselines

Naive reference: timescale arguments. The first reference point is what astrophysicists already do implicitly. Causality limits the size of the emitting region, $R \le c\,\Delta t\,\delta$, and an observed rise or decay is compared with the cooling and acceleration times. A flare violating these is "ruled out"; otherwise it is "allowed." These bounds are effectively a crude reachability argument. Beating them means producing certified infeasibility on observations they declare allowed, while never ruling out a case the forward solver can actually reproduce.

Strongest astrophysical competitor: Bayesian sampling. This is MCMC or nested sampling over a parametric model of the kind behind the fits in Aleksić et al. and Tagliaferri et al.. We give the sampler a flexible injection history with a piecewise-constant rate over $K$ segments, increasing $K$ until the wall-clock budget matches ours. Both methods use the same kinetic model, the same bins and the same observational intervals.

The reported reference value is the single-zone requirement of $\delta \approx 60$ for 1ES 1215+303 (Aleksić et al.). We found no reported result on Lacuna for nested-sampling posteriors with time-varying injection on these sources, so we will run that baseline ourselves. Beating it means one of two things. Either we certify a lower minimum Doppler factor, by exhibiting an inner-bound history the sampler missed, or we certify that no admissible history with lower $\delta$ exists, which sampling cannot claim.

Strongest reachability competitor: general-purpose tools. These are CORA and JuliaReach run directly on the discretized ODE, the two tools that led the nonlinear category of ARCH-COMP24. CORA verified the 10D TRAF22 traffic model in 40 seconds, against 51 seconds for JuliaReach. We will first reproduce those TRAF22 timings on our hardware for calibration. We will then run both tools on our models at $n = 20, 50, 100, 200$ with the same input sets and residual bounds. Beating them means tighter projected bounds at equal time, or finishing at $n$ where they time out.

Ablation of the key ingredient. This ablation removes the adversarial residual: we set $\bar w = 0$ and use the plain linear model. If certificates barely change, the nonlinear terms were irrelevant and the method is just linear systems theory. If unsound "certificates" appear on cases the forward solver can reach, the residual player is doing necessary work.

Why not just…?

Why not just run MCMC with a richer injection model? This is the most serious alternative, and on fitting it may match us. Its weakness is the negative case. A sampler that finds no acceptable fit has shown only that it did not find one, and the space of time-varying histories is infinite-dimensional. The synthetic experiment below settles the comparison. We construct flares that are reachable only through unusual injection histories, and flares that are provably unreachable. We then measure whether the sampler, at matched compute, separates the two classes as reliably as the certificate. If it does, the proposal's value shrinks to producing auditable negative results, and we will say so.

Why not just use CORA or JuliaReach off the shelf? Zonotope and Taylor-model propagation is mature and could genuinely suffice on a linear-dominant model. Our bet is that these tools bound the full $n$-dimensional set, while we only need its projection onto a few observables. Support-function evaluation along those directions should scale far better. The scaling sweep over $n$ is designed to test exactly that claim. If the tools keep pace at $n = 200$, we adopt them as the back end and keep only the mechanism-as-input-set formulation.

Why not just learn a neural barrier function or surrogate? Data-driven certificates trade the grid curse for a sample-density curse. The guarantees in Lindemann et al. rely on Lipschitz coverage of the state space, which is hopeless in 100 dimensions. Exact ReLU verification as in SEEV reached 8D in about 22 seconds, where baseline solvers timed out after 3 hours, but that is still an order of magnitude short of our state size. We therefore do not compete on this axis. A neural surrogate could still serve to propose directions $\ell$, while the certificate itself remains analytic.

Why not just simulate with particle-in-cell codes? These codes simulate the plasma from first principles and answer which particle spectrum a given shock or reconnection layer produces. They do not answer which observations are impossible for any history. We treat their outputs as the source of physically motivated input sets $U$, not as a competitor.

Proposed Experiments

Soundness and power on synthetic flares. We generate several thousand synthetic flares with the forward kinetic solver from randomly drawn admissible and inadmissible histories, and add noise at the levels of the Swift-XRT and MAGIC intervals reported in the cited campaigns. Soundness requires zero certified-infeasible verdicts on truly admissible flares; any violation is a bug or a bad residual bound. Power is the fraction of inadmissible flares certified, reported against the timescale baseline and against matched-compute sampling. The inner–outer gap is reported per observable direction.

Decisive check on 1ES 1215+303. On the published 2011 SED, we compute the smallest Doppler factor for which the observation remains inside the shock-class reachable set, bracketed by inner and outer bounds. Two results would each be informative. A certified minimum well below 60 shows that the extreme-parameter conclusion of Aleksić et al. was an artifact of the injection parameterization. A lower bound near 60 converts that conclusion from a fitting outcome into a theorem about the model class.

Spectral-peak dynamics in 1ES 1959+650. For this source, Tagliaferri et al. report a synchrotron peak that moved to higher energy as the X-ray flux changed by about a factor of two. They read this as support for standing-shock acceleration. We test whether that joint flux–peak trajectory is reachable under the shock class but unreachable under a softer injection class, which would turn a qualitative argument into a quantitative one.

Stratified two-zone composition on Mrk 501. On the Mrk 501 IXPE-era campaign, we test whether the composed two-zone reachable set contains the simultaneous X-ray and TeV states while a single-zone set does not (Abe et al.). Polarization is excluded at this stage, because it requires a field-geometry model outside the linear structure.

Interpretation. The project succeeds if the certificates are sound in simulation, stay informative where timescale arguments are not, and change or harden at least one published mechanism conclusion. It fails cleanly if the residual bounds needed for soundness make every reachable set too large to exclude anything. That outcome is itself a measurable statement about how much current blazar data can actually discriminate between mechanisms.