Resonant inelastic x-ray scattering (RIXS) shines x-rays tuned to an atomic absorption edge onto a material and measures how much energy the scattered photons lose. Those losses reveal the material's collective excitations: spin waves, lattice vibrations, and the electron-hole continuum. The quantity physicists usually report is the spectral weight of a mode, for example a paramagnon (a damped spin wave), obtained by adding up counts inside a fixed energy-loss window or by fitting a few peak shapes. Computer vision learned the hard way that a fixed rectangle drawn around a target also captures whatever sits behind it. Box-pooled features absorb background, trackers drift, identities get swapped, and box-based metrics overstate how good a method is. A fixed energy-loss window has the same problem. It also collects fluorescence (photons whose loss shifts with the incoming photon energy rather than staying fixed) and the tail of the elastic line, which is the intense peak of photons that lose no energy. The share of contamination inside the window changes with doping, temperature and incident energy, which are exactly the variables whose trends RIXS papers try to measure.
We propose to replace the window with a mask, meaning a per-pixel assignment of measured counts to physical components. The mask is estimated from a kinematic fact RIXS already provides. When the incident energy is detuned from the absorption peak, collective (Raman-like) features stay at a fixed energy loss, while fluorescence stays at a fixed emitted energy and so moves along a diagonal in the plane of incident energy versus energy loss. Crossover from Collective to Incoherent Spin Excitations in Superconducting Cuprates Probed by Detuned Resonant Inelastic X-Ray Scattering used this signature qualitatively, tracking peak positions against detuning to show that overdoped cuprates cross over from collective paramagnons near 0.3 eV to an incoherent continuum. Our claim is that the same signature, written as a constrained two-coordinate decomposition of the full detuning map, gives a quantitative and less biased Raman-like spectral weight. We also claim that the size of the bias from window integration can be measured, in the same way BUAA-PRO: A Tracking Dataset with Pixel-Level Annotation measured the gap between boxes and true object shapes.
Choosing the idea
The research file's most direct bridge (its Gap 1) is to apply learned mask-pooling, in the style of Learning Multi-Object Tracking and Segmentation from Automatic Annotations, to the curved lines that spectrometer optics project onto the 2D detector. We decline it. The curvature is set by the instrument and can be calibrated by fitting a low-order polynomial to the elastic line of a reference sample. A skeptical reviewer would rightly expect that calibration to match any learned mask. The place where a box actually hides physics is one level up, in energy-loss space. There, the contamination is not a fixed geometric artifact; it is a competing physical process whose weight changes along the very axis being studied. No paper found on Lacuna connects RIXS to bounding-box contamination, so the contribution has to stand on whether it changes a physical conclusion, not on being the first to combine the two topics.
Core Claim
Index a detuning series at fixed momentum by incident energy $\omega$ and energy loss $\Delta$. The emitted energy is then $\omega - \Delta$. We model the expected photon count as
$$ \mu(\omega,\Delta) = \sum_k a_k(\omega)\, r_k(\Delta) \;+\; \sum_j b_j(\omega)\, f_j(\omega-\Delta) \;+\; e(\omega)\, g(\Delta), $$
where the symbols are as follows. The $r_k$ are nonnegative lineshapes fixed in energy loss (Raman-like components), with amplitudes $a_k(\omega)$ that may vary with incident energy. The $f_j$ are lineshapes fixed in emitted energy (fluorescence-like components), with amplitudes $b_j(\omega)$. The term $g$ is the measured elastic resolution function, and $e(\omega)$ is its amplitude. We fit this model with a Poisson likelihood, because the data are photon counts.
The Raman mask is $m(\omega,\Delta)=\sum_k a_k r_k/\mu$, the fraction of expected counts at each pixel attributed to Raman-like components. The quantity we report is the mask-weighted spectral weight in place of the window sum. A real continuum may be neither purely Raman nor purely fluorescent, and that intermediate character is the physics of the crossover. The model therefore lets the mask take any value between 0 and 1 rather than forcing a hard assignment, so the mask itself becomes an observable.
Research Program
The work has three parts. Each one borrows a specific lesson from the bounding-box literature.
The first part is the explicit decomposition described above: a nonnegative factorization in the two coordinate systems (energy loss and emitted energy), with smoothness priors on the amplitudes and the elastic shape fixed from calibration. This is the method itself. It needs no neural network. It takes the idea from Scene-Adaptive Person Search via Bilateral Modulations of treating context as a separable nuisance to be modeled, rather than something subtracted after the fact. The difference is that our separation comes from physics, where SEAS learned it through a loss on feature norms.
The second part is an evaluation protocol in the spirit of BUAA-PRO, which showed that boxes overlap true object shapes by only 0.48 to 0.56 on average and that rankings of trackers reverse when they are scored against true masks. We build a simulated benchmark of detuning maps with known ground-truth decompositions. In these maps the paramagnon weight, the continuum's Raman fraction, the fluorescence onset, the elastic intensity, the resolution and the count level are all controlled. The ground truth comes from parametric lineshapes and from small-cluster spectra of the kind used to interpret cuprate data. We then measure how far window integration and per-spectrum peak fitting misreport the doping trend.
The third part tests whether an amortized estimator adds anything. This is a small network trained on simulated maps that outputs the components directly. It is the "silver-standard" bet of MOTSNet. We test it against a sparse-calibration alternative modeled on the finding in Segmentation Mask Guided End-to-End Person Search that masks on roughly 15% of training images sufficed. Here, the sparse set would be a few experimental maps whose decomposition is physically well constrained. This part is conditional. If the explicit factorization already reaches the noise floor, the network is dropped.
The real-data targets are detuning series on overdoped Bi2201 and Tl2201, where the crossover was reported, and on Sr3Ir2O7. In Sr3Ir2O7, Antiferromagnetic excitonic insulator state in Sr3Ir2O7 found a longitudinal mode at $q_c=0.5$ that softens toward the Néel temperature and overlaps an electronic continuum. That softening is a trend whose size could depend on how weight is assigned between mode and continuum. Obtaining these series requires either collaborating with the original groups or getting new beamtime. This is a real dependency, and it is not resolved in the file.
Baselines
The naive reference point is fixed-window integration. Counts are summed inside an energy-loss window around the mode, after subtracting the elastic line, separately for each incident energy. This is what "paramagnon weight versus doping" often means in practice. We run it with window widths chosen on simulated data to minimize error. That choice makes it a tuned baseline, not a straw man, and it beats the proposal if its error on the simulated doping trend is within noise of ours.
The strongest established method is per-spectrum multi-component fitting with Gaussian, Lorentzian, damped-oscillator or Fano lineshapes, the standard practice the file describes for cuprate and iridate RIXS. Each incident energy is fitted independently. Afterwards, peak positions are plotted against detuning, as in the cuprate crossover paper. We give it the same lineshape families, the same elastic resolution function, the same Poisson likelihood and the same number of random restarts as our method, on identical maps. There is no reported number on Lacuna for the decomposition accuracy of RIXS peak fitting on any benchmark, so its error has to be measured by us on the simulated ground truth. We count it as beaten only if our method reduces bias in the recovered Raman weight and its trend with doping, and also reduces disagreement across choices of detuning range, beyond bootstrap error bars.
Two ablations isolate the key ingredient. In the first, the decomposition is fitted without the emitted-energy coordinate, so both component types are fixed in energy loss. In the second, a generic nonnegative factorization is run on the map with no kinematic structure. If either ablation matches the full model, then the coupled two-coordinate structure is not what produces the gain. For the amortized network, the matching ablation is the same architecture trained without the structured output head.
Why not just…?
Why not just correct the detector curvature and keep rectangular binning? We agree, and it is the reason we dropped the detector-mask idea. Curvature correction handles the instrument's geometry. It does nothing about fluorescence and continuum weight that physically sits inside the energy-loss window and changes with doping. The simulated benchmark tests this directly by applying perfect curvature correction to every baseline.
Why not just fit more peaks per spectrum? Adding components to independent 1D fits makes them less identifiable. A broad continuum and a damped mode trade weight freely, and the cuprate crossover paper's central point was that the continuum's behavior under detuning is what distinguishes the two. Per-spectrum fitting discards the cross-spectrum constraint our model imposes. This is a real threat only when the detuning range is so narrow that the diagonal and vertical directions barely differ. We will report the detuning range below which the advantage disappears.
Why not just read peak position against detuning, as Minola et al. did? That method answers whether a feature is Raman-like. It does not say how much weight is Raman-like when the two kinds of feature overlap. Our mask is the quantitative version of the same idea. On real data, the test is whether the mask-weighted trend with doping agrees with their qualitative crossover and narrows its location.
Why not just train a big denoising network on the maps? Generalized Relation Modeling for Transformer Tracking found that unconstrained attention lets background contaminate the target's features, and that blocking target-background interaction improved GOT-10k from 71.0% to 73.4% average overlap. That points to structure, not capacity, as what matters. Our ablation, the structured head versus the generic head, tests this in the RIXS setting.
Proposed Experiments
The decisive check on simulated data is a sweep over doping-like parameter series in which the true Raman fraction of the continuum falls monotonically. For each method we report the bias and variance of the recovered paramagnon weight, the error in the recovered trend, and the calibration of the mask against the true per-pixel fraction. These are reported across count levels, energy resolutions and detuning ranges. The expected result is that window integration overstates collective weight at heavy overdoping, where fluorescence moves into the window, and so flattens the apparent crossover. If it does not, the main motivation fails.
The decisive check on real data does not need ground truth. A good decomposition should give the same Raman weight no matter which subset of incident energies is used, while window and per-spectrum estimates should drift with that subset. This is a held-out consistency test. Applied to the Sr3Ir2O7 longitudinal mode, it tells us whether the reported softening toward the Néel temperature keeps its magnitude when weight is assigned by the mask.
Finally, we compare the silver-standard and sparse-calibration regimes for the amortized estimator. We vary the number of calibrated real maps from one upward and record where performance plateaus. This tests whether the finding that a small fraction of masks suffices carries over to a field where no human can label the ground truth.
References
- Crossover from Collective to Incoherent Spin Excitations in Superconducting Cuprates Probed by Detuned Resonant Inelastic X-Ray Scattering
- BUAA-PRO: A Tracking Dataset with Pixel-Level Annotation
- Learning Multi-Object Tracking and Segmentation From Automatic Annotations
- Scene-Adaptive Person Search via Bilateral Modulations
- Segmentation mask guided end-to-end person search
- Antiferromagnetic excitonic insulator state in Sr3Ir2O7
- Generalized Relation Modeling for Transformer Tracking