Upper Limits on Out-of-Plane Dispersion from Photon-Starved ARPES

Photoemission spectroscopy measures a crystal's electronic bands by shining light on it and recording the energy and angle of the electrons that fly out. One step of routine analysis carries unusual weight. Physicists sweep the photon energy and check whether a band moves. A band that stays put is assigned to the surface. A band that shifts is assigned to the bulk, meaning it disperses along the direction perpendicular to the surface. Claims of topological surface states, Fermi arcs and bulk Dirac nodes rest on this check. The check is almost always made by eye. It is also often made in the soft X-ray regime, where far fewer electrons are ejected per photon and spectra are dominated by counting noise.

This note proposes to treat "this band does not disperse along the surface normal" the way particle physicists treat "this decay was not seen." The answer should be an upper limit, with a stated confidence level, on how much the band moves across the Brillouin zone. The limit comes from a Poisson likelihood that models electron counts across the whole photon-energy sweep. Photon-energy-dependent intensity (the "matrix element" effects that make bands fade in and out) and the unknown final-state parameters are treated as nuisance parameters and profiled out. The same likelihood yields its expected sensitivity before any beamtime is spent. That expected sensitivity tells an experimenter how to spread a fixed count budget across photon energies so the limit becomes as tight as possible.

Core Claim

Energy-sweep ARPES can report surface-versus-bulk assignments as calibrated upper limits on out-of-plane bandwidth $W$, with correct coverage down to a few hundred electrons per energy step. Two ingredients make this work, and both can be isolated by ablation. The first is a likelihood that fits peak position jointly across all photon energies, with a free amplitude at each energy, so a matrix-element dropout cannot masquerade as dispersion or hide it. The second is a Fisher-information schedule of photon energies, which should reach a given limit with fewer total counts than uniform stepping.

The statistical template is the "median sensitivity versus observed limit" reporting of rare-event searches, as in Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset. That paper quotes a median 90% CL half-life sensitivity of 5.0 × 10²⁵ yr next to an observed limit of 3.5 × 10²⁵ yr. It also belongs to the broader practice of extended likelihood fits over Poisson data used in meson-decay analyses such as Bergfeld et al. and Achasov et al.. These papers supply the method; none of them touches band mapping.

Relation to Prior Work

The closest machine-learning work reconstructs bands but does not test hypotheses about them. A machine learning route between band mapping and band structure fits band surfaces of WSe₂ with a Markov random field, which is a smoothness prior linking neighbouring momentum points. It is warm-started from density functional theory and reports 40–50 meV reconstruction error on synthetic crossings. It works at fixed photon energy, uses a continuous intensity likelihood rather than counts, and its theory prior is exactly what should be avoided when the question is whether the data themselves show dispersion.

An autoencoder for compressing angle-resolved photoemission spectroscopy data (ARPESNet) denoises 2D slices by training on pairs of independent noisy copies of the same spectrum. It reports MSE 0.38 and PSNR 41.34 on high-quality test spectra, but it returns images, not parameters with uncertainties. Autonomous micro-focus angle-resolved photoemission spectroscopy steers sampling with Gaussian-process regression, a probabilistic interpolator that also estimates its own uncertainty. It achieves a 40–75× point reduction over position and tilt, with photon energy fixed and Gaussian noise assumed.

Photoemission Orbital Tomography Using Robust Sparse PhaseLift is the one paper here that benchmarks reconstruction under controlled electron counts, from 10³ to 10⁵ electrons. We borrow its evaluation protocol but not its method, which targets molecular orbitals rather than periodic bands.

The physics papers motivate the problem. Disentanglement of Surface and Bulk Rashba Spin Splittings in Noncentrosymmetric BiTeI separated stationary surface states in UV from a dispersive bulk band in soft X-ray. Observation of unconventional chiral fermions with long Fermi arcs in CoSi, Evidence of topological insulator state in the semimetal LaBi and Topological Dirac surface states and superconducting pairing correlations in PbTaSe2 rest on the same null-dispersion argument, stated qualitatively. None of them quotes a bound on residual out-of-plane dispersion.

Research Program

At fixed in-plane momentum, the band energy is modelled as a nearest-neighbour form along the surface normal, $E(k_z) = E_0 - \tfrac{W}{2}\cos(c\,k_z)$. Here $W$ is the out-of-plane bandwidth and $c$ the lattice spacing along the normal. The perpendicular momentum $k_z$ follows from photon energy through the free-electron final-state relation, with inner potential $V_0$ as a nuisance parameter. The expected counts in energy pixel $e$ at photon energy $h\nu_j$ are

$$\mu_{je} = A_j\,\big[\,V_{\Gamma_j} * \bar{\delta}_{\lambda_j}\big](E_e - E(k_z(h\nu_j;V_0))) + B_j .$$

In this expression, $A_j$ and $B_j$ are a free amplitude and background at each photon energy. The free amplitude absorbs matrix-element and flux variation. The term $V_{\Gamma_j}$ is a lineshape of width $\Gamma_j$ convolved with the analyser resolution. The term $\bar{\delta}_{\lambda_j}$ averages the band over the spread in $k_z$ set by the electron escape depth $\lambda_j$.

The likelihood is the product of Poisson terms over pixels. We deliberately work with the detector's native pixels rather than a truly unbinned event list. At ARPES pixel sizes, finer binning should cost little information, and we will check this claim directly.

The escape-depth term matters and is often ignored. In the VUV the escape depth is short, so the $k_z$ spread is large and genuinely bulk bands appear flattened. The reported quantity is therefore a limit on $W$ conditional on an assumed $\lambda$. Scanning $\lambda$ shows how much of the "surface" verdict is really a statement about the probe's depth resolution.

Limits on $W$ come from the profile likelihood ratio with Feldman–Cousins-style toy calibration, since $W \ge 0$ sits on a boundary. Expected sensitivity comes from Asimov datasets, which are noise-free expected-count datasets used to compute median sensitivity without running toys. A Fisher-information argument gives the scaling, because a peak position is estimated to roughly $\Gamma/\sqrt{N}$ from $N$ signal counts. The detectable bandwidth should then scale as

$$W_{90} \propto \frac{\Gamma}{\sqrt{N_{\rm eff}}}\,\frac{1}{s(\lambda)} ,$$

where $N_{\rm eff}$ is the number of counts weighted by how strongly each photon energy's $k_z$ probes the cosine, and $s(\lambda)$ is the attenuation factor from escape-depth broadening. The scheduler chooses the photon energies and dwell times that maximise $N_{\rm eff}$ under a flux-versus-energy curve and a monochromator move cost. Placing counts near $k_z = 0$ and $\pi/c$, where the cosine is at its extremes, should beat uniform steps. How much it beats them is an empirical question.

Baselines

The naive reference is standard practice, which cannot be scored for coverage. We instead implement the stricter version of it: fit a peak to each energy distribution curve (EDC, an intensity-versus-energy cut at fixed momentum) at each photon energy with least squares, then test for constancy with a weighted chi-square. It is the strongest simple competitor because it uses the same information, a peak position per energy, and needs no new modelling. It runs on identical data with the same lineshape family.

The strongest existing reconstruction method is the Markov random field of Xian et al.. Its reported figure is 40–50 meV reconstruction error on synthetic crossing benchmarks. We will run it slice by slice across photon energies, with and without its theory warm start, and derive a dispersion estimate from its output with bootstrap intervals. No reported result was found on Lacuna for any method bounding out-of-plane bandwidth, so no competitor number exists for the headline metric.

A third baseline chains ARPESNet denoising with the pointwise fit. It uses the published noise-to-noise training recipe on the same augmented training library, with no retraining on test materials.

Two ablations isolate the key ingredients. The first ties the per-energy amplitudes into a smooth curve instead of leaving them free, which tests the matrix-element robustness. The second replaces the Poisson likelihood with a Gaussian one at equal counts. Beating the baselines means two things. We must hold 90% coverage where they under-cover, and we must reach a tighter median limit, or equal limit at fewer counts, with the scheduler adding at least a further factor of about 1.5 in count efficiency over uniform stepping.

Why not just…?

Why not just count longer? Brute-force counting is the correct answer whenever it is affordable, and the sensitivity formula tells an experimenter when it is. It fails the stated case: soft X-ray cross sections are two to three orders of magnitude lower than in the VUV, and beam damage limits dwell time. More importantly, longer counting still produces no calibrated statement. The proposal reports the count budget needed for a target limit, so it complements long integration rather than replacing it.

Why not just fit each EDC and run a chi-square? This is the real threat. At high counts it should match the joint fit closely, and we expect that to happen. The joint fit should win in two places. The first is below a few hundred counts per step, where least squares on Poisson data biases peak positions and intervals under-cover. The second is at matrix-element dropouts, where individual fits fail or lock onto background. The thinning experiment below maps exactly where the two methods diverge. If they never diverge within realistic count ranges, the likelihood machinery is not worth it and we will say so; the scheduler would survive as the contribution.

Why not extend Gaussian-process active learning to photon energy? Such an extension of Ágústsson et al. would sample where spectral curvature is interesting. That rule targets the wrong quantity for a null test. A surface state has no curvature along the photon-energy axis, so curvature-seeking sampling has nothing to aim at precisely where the verdict matters. We will run it as a scheduling comparison anyway.

Why not denoise with ARPESNet and then read off dispersion? Denoising without a forward model can smear or invent small shifts. It also gives no interval. The chained baseline tests this directly.

Proposed Experiments

The first study is synthetic. We simulate energy sweeps with known $W$ from 0 to 200 meV, Poisson counts from 10² to 10⁴ per energy step, and injected matrix-element dropouts, following the controlled-count protocol of the PhaseLift paper. We measure coverage of the 90% interval, median limit versus counts against the predicted $\Gamma/\sqrt{N}$ scaling, and false "bulk" rates at $W=0$.

The second study uses real data through binomial thinning. Keeping each detected electron with probability $p$ turns a high-statistics sweep into an exact Poisson realisation at lower counts. The full-statistics fit then serves as ground truth. Sweeps would come from BiTeI-like systems with both a known surface state and a known dispersive bulk band, obtained from authors or new beamtime, with CoSi arcs and α-Sn as further targets. This also tests the pixelisation claim, by comparing information at native and coarsened binning.

The third study is prospective. Two beamtime runs at equal total dose would compare the Fisher-optimal photon-energy schedule with uniform stepping. The pass criterion is a 90% limit on the surface state and a significant nonzero $W$ for the bulk band, with the scheduled run needing fewer counts.

The decisive check is the combination of three results: coverage holds on thinned real data, the naive fit under-covers at low counts or at dropouts, and the free-amplitude ablation shows that matrix-element modulation changes conclusions when not profiled.

References

  1. Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset
  2. Observation of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">B</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mo>→</mml:mo><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">ω</mml:mi><mml:mi mathvariant="italic">K</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow></mml:math>and Search for Related<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="italic">B</mml:mi></mml:math>Decays Modes
  3. Observation of a Neutral Structure near the D(D)over-bar* Mass Threshold in e(+)e(-) -> (D(D)over-bar*)(0)pi(0) at root s=4.226 and 4.257 GeV
  4. A machine learning route between band mapping and band structure
  5. An autoencoder for compressing angle-resolved photoemission spectroscopy data
  6. Autonomous micro-focus angle-resolved photoemission spectroscopy
  7. Photoemission Orbital Tomography Using Robust Sparse PhaseLift
  8. Disentanglement of Surface and Bulk Rashba Spin Splittings in Noncentrosymmetric BiTeI
  9. Observation of unconventional chiral fermions with long Fermi arcs in CoSi
  10. Evidence of topological insulator state in the semimetal LaBi
  11. Topological Dirac surface states and superconducting pairing correlations in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi>PbTaSe</mml:mi><mml:mn>2</mml:mn></mml:msub></mml:math>