Fourier-domain analysis + geodesic-coherence refinement

A Fourier-Domain Analysis of BHEX for Photon-Ring Inference

Jonathan R. Landers

Independent researcher. This project is inspired by published BHEX work and is not affiliated with the BHEX collaboration or mission team.

BHEX is motivated by the idea that sufficiently long interferometric baselines may isolate a thin, more universal photon-ring signal beneath brighter, broader plasma emission around a black hole.[4][5][6] That turns the inference problem into a separation problem: the ring carries the target geometric information, but the surrounding plasma can dominate the image and distort the visibility pattern that a direct amplitude heuristic tries to read.[1][5]

At a high level, the project presented here turns the BHEX-motivated ring-versus-plasma question into a runnable synthetic Fourier-domain inference prototype, then sharpens its interpretation with later geometric theory. This page starts from the implemented Fourier-domain source-separation model in the first note and then adds the mathematical extensions from the geodesic-provenance note and the summary note. The simulations, estimator pipeline, and held-out results shown below come entirely from the first Fourier-domain paper.[1][2][3]

Open Fourier Note Open Geodesic Note Open Summary Note

Executive picture

The Fourier-domain analysis shows that a long-baseline amplitude heuristic of the kind emphasized in the BHEX program can become hard to read cleanly before the photon ring itself becomes unrecoverable. The geodesic-coherence refinement asks what physically controls the key coherence term in that story and answers by lifting the nuisance model to photon initial-condition space, where ring-background overlap is bounded by the correlation of escaping geodesic families.[1][2][3]

The scripts implement only the first paper's abstract Fourier-domain smooth-nuisance prototype. They do not implement the geodesic-provenance model from the second paper or any additional construction from the summary note.[1][2][3]
  • Start from images, not abstract vectors.
  • Push the images through a full Fourier visibility representation.
  • Implement the abstract estimator from the first paper.
  • Discuss the later mathematical notes separately as theory.
  • Use the implemented tuning, inference, and ring-emphasized reconstructions as shown.
Synthetic sample

Representative synthetic dataset

Synthetic preview panel
A representative synthetic image panel from the prototype dataset. The composite image contains a thin photon ring plus broader plasma emission; the true components are shown separately for intuition.
Best tuned $\lambda$
63.1
smooth-nuisance regularization
Best template width
4.0px
input image pixels
Tuning radius MAE
0.24px
median 0.23px
Held-out radius MAE
0.27px
mean confidence 0.74

Physical intuition

At the image level, the source contains a broad plasma structure and a thinner photon ring. The plasma dominates visual appearance, whereas the photon ring carries the target geometric information.[1][5][6]

In Fourier space, this geometric distinction becomes a separation problem. Broad image features and thin oscillatory ring structure concentrate differently, so the first inference question is: what fraction of the visibility is attributable to ring structure, and what fraction is attributable to structured nuisance? The refinement then asks where that nuisance comes from physically and answers: from ordinary escaping geodesics rather than an arbitrary visibility-space contaminant.[2][3]

The important conceptual shift is that this is not blind source separation. First it is a structured inverse problem in visibility space; then, in the refinement, it becomes a provenance-constrained separation problem in which the nuisance is restricted to the background escaping sector.

Why the heuristic can lose readability early

A long-baseline amplitude heuristic, in the same spirit as the BHEX ring-shape methodology, is effectively trying to read the ring directly from the oscillatory pattern of $|y|$. That works when contamination is mild. But the modulus is nonlinear, so structured nuisance can shift peaks and spacing before the underlying ring is truly gone.[1][5]

A fuller estimator is complementary in what it demands. It does not require the ring to be visually obvious in amplitude space. It only requires the ring to remain identifiable under the joint model. The geodesic-coherence extension sharpens this by proposing that identifiability should improve when ring-generating and background-generating escaping geodesics are dynamically well separated.[2][3]

Core mathematics

Observation model

$$ y = \alpha g_{\theta} + q + \varepsilon, $$

$g_{\theta}$ is the photon-ring family, $\alpha$ is ring amplitude, $q$ is structured nuisance, and $\varepsilon$ is noise.[1][5][6]

Structured estimator

$$ (\hat{\alpha},\hat{\theta},\hat{q}) \in \arg\min_{\alpha,\theta,q \in \mathcal{Q}} \|y-\alpha g_{\theta}-q\|_{\mathcal{H}}^2 + \lambda R(q). $$

This is the layer implemented in code: fit the full visibility model and regularize the nuisance rather than assuming the ring can be read directly from amplitude alone.[1]

Mismatch bound

$$ \bigl|\hat{\alpha}_{\mathrm{heur}} - \hat{\alpha}_{\mathrm{model}}\bigr| \le \mu(g,\mathcal{P})\frac{\|p\|}{\|g\|} + \frac{\|\varepsilon\|}{\|g\|} + \frac{\|(I-\Pi_{\mathcal{P}})\varepsilon\|}{\|(I-\Pi_{\mathcal{P}})g\|}. $$

The heuristic-model gap increases with ring-plasma coherence, nuisance magnitude, and noise. The geodesic-provenance refinement interprets this coherence term more physically.[1][2]

Recoverability statement

$$ \{\text{cleanly amplitude-readable}\} $$ $$ \subsetneq $$ $$ \{\text{structured-separation recoverable}\} $$

Clean amplitude readability is stricter than joint-model identifiability. The ring may cease to be visually obvious while remaining stable to estimate under the structured model.[1]

Provenance lift

$$ z=(p,\xi), \qquad E = E_{\mathrm{ring}} \,\dot\cup\, E_{\mathrm{bg}}, \qquad g_{\theta}=A_r f_{\theta}, \quad q=A_b h. $$

The geodesic-coherence note lifts the nuisance model back to photon initial-condition space. Ring and background are the visibility-space images of distinct escaping geodesic families, not just abstract components in $\mathcal{H}$.[2][3]

Provenance-constrained coherence

$$ \mu(g,Q_{\mathrm{prov}}) := \sup_{q \in Q_{\mathrm{prov}}\setminus\{0\}} \frac{|\langle g,q\rangle|}{\|g\|\,\|q\|}, \qquad Q_{\mathrm{prov}}=\{A_b h : h \in \mathcal{A}\}. $$

The overlap term is redefined over a physically admissible nuisance family generated by ordinary escaping geodesics.[2][3]

Main new theorem

$$ \mu(g,Q_{\mathrm{prov}}) \le \frac{1}{\sigma_r\sigma_b} \left( \int_{E_{\mathrm{ring}}}\!\int_{E_{\mathrm{bg}}} |G(z,z')|^2\,d\nu(z)\,d\nu(z') \right)^{1/2}. $$

The coherence term in the mismatch bound is itself bounded by a cross-Gram integral over escaping geodesic families.[2][3]

Geometric separation corollary

$$ \mu(g,Q_{\mathrm{prov}}) \le \frac{M}{\sigma_r\sigma_b} e^{-\beta\Delta} \sqrt{\nu(E_{\mathrm{ring}})\,\nu(E_{\mathrm{bg}})}. $$

If near-critical and ordinary escaping geodesics are separated by a criticality gap $\Delta$, the ring-background coherence decays exponentially. The refinement is therefore geometric, not just "smaller nuisance."[2][3]

Mathematically, the page tells a three-step story: Fourier-domain mismatch and recoverability, provenance lift to photon initial conditions, and a geometric corollary that links recoverability margin to separation of escaping photon families. The code shown below implements only the first paper's structured estimator with a smooth nuisance regularizer. The second note and the summary note appear here as mathematical extensions only.[1][2][3]

Prototype estimator used in the scripts

Prototype scope

The implementation is not a mission-grade astrophysical forward model; it is an inference prototype designed for methodological clarity:

  • synthetic images: thin ring + broad crescent + blur + noise
  • circular template bank over ring radius and width
  • Fourier-domain nuisance regularization
  • image-domain reconstruction for interpretation

This scoped formulation operationalizes the first note's estimator while isolating estimation behavior before adding full astrophysical realism.

Working resolution during tuning and inference: 64 x 64 after downsampling by a factor of 4.

1
Generate image

Create a synthetic black-hole-like image with a thin photon ring and broader plasma component.

2
Move to Fourier space

Apply the FFT-based image-to-visibility map used in the prototype.

3
Search ring templates

Fit ring radius and amplitude over a candidate bank while regularizing the nuisance.

4
Estimate plasma

Absorb smoother contamination into $\hat q$ instead of letting it bias the ring directly.

5
Reconstruct image

Bring the recovered ring and plasma back into image space and create a ring-emphasized visualization.

Concrete nuisance model

How the scripts operationalize the estimator

The scripts instantiate the abstract nuisance regularizer with a concrete model: use a circular Gaussian ring template bank on the full FFT visibility plane and penalize high-frequency structure in the nuisance term. This drives the plasma estimate toward smoother Fourier behavior while preserving thin oscillatory ring structure in the explicit template family.

For fixed $(\alpha,\theta)$, this yields a closed-form nuisance estimate, so a practical grid search over $\theta$ and $\lambda$ is feasible in the prototype.

This estimator implements only the first paper's abstract Fourier-domain model with a smooth Fourier nuisance term. No script estimates $A_r$, $A_b$, the family $Q_{\mathrm{prov}}$, a criticality index, or the cross-Gram kernel, and nothing from the second or summary paper is implemented numerically.

Prototype objective

$$ \min_{\alpha,\theta,q} \|y-\alpha g_{\theta}-q\|^2 + \lambda \|H q\|^2, $$

Here $H$ is a frequency-weighting operator that penalizes high-frequency content in $q$.

Closed-form nuisance update

$$ \hat q(\alpha,\theta) = (I + \lambda H^*H)^{-1}(y-\alpha g_{\theta}). $$

This update is what makes the template and regularization sweep numerically practical in the scripts.

Tuning / calibration results

The calibration loop begins with synthetic images, pushes them into Fourier space, runs the estimator, compares the recovered radius to ground truth, and then chooses the hyperparameters that minimize a simple score balancing radius error, amplitude error, and confidence.

Tuning heatmap radius MAE
The tuning heatmap for mean absolute radius error, with x-axis = regularization strength $\lambda$ and y-axis = template width. Several larger-$\lambda$, wider-template settings perform similarly well; the final selected setting is chosen by a score that also includes amplitude error and confidence.
Tuning heatmap confidence
Confidence heatmap over the same search grid, with x-axis = regularization strength $\lambda$ and y-axis = template width. The best region is not merely accurate; it is also numerically confident under the prototype score.
Tuning scatter true vs estimated radius
True versus estimated radius on the tuning set under the best hyperparameters, with x-axis = true ring radius and y-axis = estimated ring radius.
Tuning histogram radius error
Radius-error distribution on the tuning data, with x-axis = estimated radius minus true radius and y-axis = count. In this run the average absolute error is 0.24 pixels with median 0.23 pixels.
Tuning example best
Example reconstruction panel from the tuning stage.
Tuning example mid
Another tuning reconstruction, illustrating ring / plasma / residual decomposition.
Tuning example hard
A harder example. Even here the estimated ring remains visually clean.

Held-out inference and ring-emphasized outputs

On new images, the calibrated estimator is applied without retuning. It estimates ring parameters in Fourier space, reconstructs ring and plasma components in image space, and produces a diagnostic image in which the ring is brightened with a confidence-controlled boost while plasma remains visible as context.

Held-out scatter true vs estimated radius
Held-out true versus estimated radius, with x-axis = true ring radius and y-axis = estimated ring radius. The average held-out absolute radius error in this run is 0.27 pixels.
Held-out histogram radius error
Held-out radius-error distribution, with x-axis = estimated radius minus true radius and y-axis = count. The median held-out absolute error is 0.31 pixels.
Held-out confidence vs error
Confidence versus absolute radius error on held-out examples, with x-axis = confidence score and y-axis = absolute radius error. Confidence is tracked as a diagnostic, though in this simple synthetic prototype the errors are already very small across the board.
Held-out example

holdout_0153

true radius = 51.99, estimated radius = 52.00, absolute error = 0.008 px

confidence 0.69
holdout_0153 observed composite
Observed composite image. Broad plasma emission and thin ring are mixed together.
holdout_0153 ring emphasized
Ring-emphasized diagnostic. The recovered ring is brightened while the plasma stays visible as context.
Held-out example

holdout_0139

true radius = 65.31, estimated radius = 65.00, absolute error = 0.315 px

confidence 0.92
holdout_0139 observed composite
Observed composite image. Broad plasma emission and thin ring are mixed together.
holdout_0139 ring emphasized
Ring-emphasized diagnostic. The recovered ring is brightened while the plasma stays visible as context.
Held-out example

holdout_0129

true radius = 73.54, estimated radius = 73.00, absolute error = 0.536 px

confidence 0.89
holdout_0129 observed composite
Observed composite image. Broad plasma emission and thin ring are mixed together.
holdout_0129 ring emphasized
Ring-emphasized diagnostic. The recovered ring is brightened while the plasma stays visible as context.

Limitations and next steps

What this prototype already shows

  • The abstract source-separation story can be instantiated operationally.
  • A ring-aware estimator can be tuned directly from image-generated synthetic data.
  • The output can be rendered back into image space for direct visual interpretation.
  • The prototype supports the central conceptual argument.

Natural upgrades

  • elliptical or spin-informed ring families instead of purely circular templates
  • explicit baseline masks and more realistic visibility sampling
  • provenance-constrained nuisance classes instead of purely smooth Fourier penalties
  • cross-Gram or $A_b^*A_r$ estimates from ray tracing / simulation to evaluate coherence directly
  • criticality indices that turn geometric separation into a computable recoverability diagnostic
  • uncertainty bands or posterior summaries instead of point estimates only
  • mission-style diagnostics that tie confidence to actual recoverability margins
Conceptual arc: heuristic readability can fail before recoverability, and the geodesic-coherence refinement says the size of that gap should be governed by how strongly ring-generating and background-generating escaping geodesics correlate in visibility space.[1][2][3][5]

References

  1. Jonathan R. Landers, A Fourier-Domain Analysis of BHEX for Photon-Ring Inference, 2026. manuscript/fourier-domain-analysis-bhex.pdf
  2. Jonathan R. Landers, Geodesic-Provenance Coherence Bounds for Fourier-Domain Photon-Ring Inference, 2026. manuscript/geodesic_coherence_bhex_note.pdf
  3. Jonathan R. Landers, A Summary of the Geodesic-Coherence Refinement for Fourier-Domain Photon-Ring Inference, 2026. manuscript/geodesic_coherence_summary_note_updated.pdf
  4. Black Hole Explorer Collaboration, Black Hole Explorer: Motivation and Vision, arXiv:2406.12917. https://arxiv.org/abs/2406.12917
  5. Black Hole Explorer Collaboration, The Black Hole Explorer: Photon Ring Science, Detection and Shape Measurement, arXiv:2406.09498. https://arxiv.org/abs/2406.09498
  6. Black Hole Explorer Collaboration, Interferometric Inference of Black Hole Spin from Photon Ring Size and Brightness, arXiv:2509.23628. https://arxiv.org/abs/2509.23628