Star-Diff: Differential Stellar Methods

• Star-Diff is an umbrella term representing differential approaches in stellar studies, encompassing diffusion measurements, wave optics, and Bayesian photometry.
• The framework unifies diverse methods such as direct measurement of stellar diffusion in clusters, diffractive microlensing, and generative modeling to capture differential stellar phenomena.
• Applications include tracing gravitational relaxation in globular clusters, enhancing photometric precision with beam-shaping diffusers, and enabling robust star–galaxy separation using probabilistic algorithms.

Searching arXiv for the papers and term usage relevant to “Star-Diff.” “Star-Diff” is not a single standardized term in the literature represented here. It is better understood as an informal umbrella for several technically distinct ideas that combine stars with difference, differential, or diffusion operations. In the cited work, those ideas include direct measurement of stellar diffusion in a globular cluster core, wave-optics diffraction in stellar microlensing, reddening-free differential photometric estimators of stellar diameter, harmonic-ratio methods for stellar differential rotation, stabilized differential photometry with beam-shaping diffusers, Bayesian separation of point-like stars from diffuse backgrounds, map-to-map residual statistics between stellar and gas kinematics, and the “star-shaped” diffusion formalism in generative modeling [(Heyl et al., 2015); (Heyl, 2013); (Chelli et al., 2016); (Reinhold et al., 2015); (Stefansson et al., 2017); (Knollmüller et al., 2018); (Powley et al., 13 Apr 2026); (Okhotin et al., 2023)].

1. Terminological scope and major usages

A common misconception is that “Star-Diff” names one established astrophysical framework. The literature instead supports a disambiguated reading: the label attaches to different objects in different subfields, and in some cases the underlying paper does not use “Star-Diff” as a formal term at all.

Domain	Formal construct	Representative paper
Globular-cluster dynamics	Diffusion of young white dwarfs through the core of 47 Tucanae	(Heyl et al., 2015)
Gravitational lensing	Diffractive microlensing of background stars by nearby compact lenses	(Heyl, 2013)
Stellar parameter inference	Differential surface brightness and pseudomagnitudes for angular diameters	(Chelli et al., 2016)
Stellar variability	Harmonic-ratio method for the sign of surface differential rotation	(Reinhold et al., 2015)
High-precision photometry	Beam-shaping diffusers for differential photometry	(Stefansson et al., 2017)
Source separation	Bayesian star-versus-diffuse decomposition (“starblade”)	(Knollmüller et al., 2018)
Kinematic residuals	$\Delta V_{\star-g}$ as a stellar–gas map mismatch statistic	(Powley et al., 13 Apr 2026)
Generative modeling	Star-Shaped DDPM (SS-DDPM)	(Okhotin et al., 2023)

This suggests that “Star-Diff” is best treated encyclopedically as a family resemblance term rather than a canonical noun. The recurring motifs are direct residual comparison, differential observables, or diffusion-like evolution, but the underlying mathematics ranges from Poisson likelihoods and Bayesian model comparison to exponential-family variational objectives.

2. Direct stellar diffusion in 47 Tucanae

The most literal “Star-Diff” usage is the direct measurement of diffusion of stars through the core of the globular cluster 47 Tucanae. The observable is the radial distribution of young white dwarfs as a function of cooling age. In the 47 Tuc core, the core relaxation time is about $t_r \sim 70$ Myr, while stars evolve from $\sim 0.9\,M_\odot$ turnoff stars to $\sim 0.53\,M_\odot$ white dwarfs over $\sim 150$ Myr. That coincidence of timescales makes newly formed white dwarfs effective tracers of gravitational relaxation: they are born with the progenitors’ centrally concentrated distribution and then random-walk outward through two-body encounters (Heyl et al., 2015).

The observational basis is Hubble Space Telescope WFC3 ultraviolet imaging of the 47 Tuc core in F225W and F336W over ten epochs between November 2012 and September 2013. The ultraviolet regime is advantageous because young white dwarfs are comparatively bright there and because the HST PSF is narrower in the UV, reducing crowding. A central technical requirement was the completeness model: the analysis used about $10^8$ artificial stars to map completeness and photometric errors as functions of radius and magnitude, because incompleteness is strongly radius-dependent for faint white dwarfs and could otherwise mimic outward diffusion (Heyl et al., 2015).

The physical model treats the dense core potential as approximately harmonic, $\phi \approx \frac{2}{3}\pi G\rho_c r^2$, and adopts a diffusion equation for the spatial density,

$$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$

with

$\kappa \sim \frac{r_c^2}{t_r}.$

For the projected cumulative distribution under a point-like initial condition, the paper uses

$C(<R) = 1 - e^{-R^2/(4\kappa t)}.$

For finite initial width, it uses a Gaussian initial condition and folds the resulting $t_r \sim 70$0 through a white-dwarf cooling curve $t_r \sim 70$1, producing an expected observed density

$t_r \sim 70$2

Inference is performed with an unbinned Poisson likelihood, supplemented by a more exact convolution with the measured non-Gaussian magnitude-error distribution; the paper reports that including or omitting that convolution does not materially change the recovered diffusion parameters (Heyl et al., 2015).

The principal result is a diffusion coefficient of order

$t_r \sim 70$3

with representative best fits near $t_r \sim 70$4, $t_r \sim 70$5, and $t_r \sim 70$6. Using $t_r \sim 70$7, the inferred relaxation time is

$t_r \sim 70$8

which the authors describe as consistent with the traditional estimate of about $t_r \sim 70$9 Myr, given the model’s simplicity and the ambiguity in the effective diffusion scale. The no-diffusion model $\sim 0.9\,M_\odot$0 is excluded at high confidence. The paper emphasizes that this is the first direct measurement of diffusion due to gravitational relaxation in a globular cluster, as distinct from inference based on static mass segregation or velocity-distribution modeling (Heyl et al., 2015).

Important approximations remain explicit: constant $\sim 0.9\,M_\odot$1, spherical symmetry, a roughly constant white-dwarf birthrate over $\sim 0.9\,M_\odot$2 Myr, and a cooling curve derived from MESA models for $\sim 0.9\,M_\odot$3, $\sim 0.9\,M_\odot$4 progenitors. The data already hint that real diffusion slows at larger radii and saturates as the white dwarfs approach the equilibrium distribution appropriate to their mass. Within a few core radii and a few core relaxation times, however, the simple diffusion picture captures the observed trend well (Heyl et al., 2015).

3. Differential optical signatures: diffraction, intensity interferometry, and turbulence

A second usage concerns stellar light as a probe of wave-optics and atmospheric differential effects. In diffractive microlensing, a compact gravitational lens produces not only geometric magnification but also interference fringes in the observed stellar signal. The control parameter is

$\sim 0.9\,M_\odot$5

If $\sim 0.9\,M_\odot$6, diffraction dominates and suppresses the peak magnification; if $\sim 0.9\,M_\odot$7, many fringes occur but finite source size or finite bandwidth can wash them out, recovering the geometric-optics limit. The wave-optics magnification is obtained by coherently summing amplitudes over paths, and the observable consequence is oscillatory magnification rather than a smooth light curve. For $\sim 0.9\,M_\odot$8, the paper describes about three peaks per Einstein radius. For large $\sim 0.9\,M_\odot$9, the oscillation amplitude scales as $\sim 0.53\,M_\odot$0. The proposed lenses are nearby substellar objects, including Earth-mass planets, asteroids, interstellar planets, and other compact low-mass lenses; the target sources are giant stars in the Galactic bulge. The paper argues that the SKA may have sufficient sensitivity, gives a 10 GHz example for an Earth-mass planet at 1 pc lensing a bulge source, estimates that about $\sim 0.53\,M_\odot$1 of OGLE-II bulge sources would have a detectable diffractive signal with the SKA, and states that the astrometric signature of Earth-mass lenses could be detected out to about $\sim 0.53\,M_\odot$2 pc (Heyl, 2013).

Differential Hanbury Brown–Twiss interferometry for binary stars uses an entirely different observable: second-order intensity correlations. For two detectors $\sim 0.53\,M_\odot$3 and $\sim 0.53\,M_\odot$4, the normalized correlator is

$\sim 0.53\,M_\odot$5

For a binary source, the correlation acquires a modulation depending on the apparent stellar separation $\sim 0.53\,M_\odot$6, so that the single-star HBT envelope is multiplied by a binary term containing $\sim 0.53\,M_\odot$7. In the paper’s simplified forms, the time dependence of $\sim 0.53\,M_\odot$8 allows recovery of semi-major axis, orbital phase, orbital period, eccentricity, and orientation. The “differential HBT” step consists of comparing correlators at different orbital phases, especially between periastron and apastron, so that the difference vanishes for $\sim 0.53\,M_\odot$9 and becomes nonzero for eccentric systems. The method is attractive because it uses intensity correlations rather than phase-resolved imaging and is therefore robust against atmospheric phase fluctuations, although the paper stresses the practical requirement of high photon flux and good coincidence statistics (Csernai et al., 2015).

The Multi-star Turbulence Monitor extends stellar differential measurement to atmospheric profiling. It uses short-exposure images of a star field and computes pairwise differential centroid motions. For each star pair, the relative motion is decomposed into longitudinal and transverse components with respect to the separation vector, and structure functions $\sim 150$0 and $\sim 150$1 are measured as functions of angular separation. Those structure functions are then fit, by Markov-Chain Monte-Carlo, to parameterized turbulence profiles and outer-scale models. In the AST3-2 implementation, the telescope is a 0.5 m modified Schmidt with a 2.9° field of view, a 10k × 5k CCD, 1 arcsec/pixel scale, 10 ms exposures, and 30-frame sequences. The method estimates the lower-atmosphere turbulence profile, total seeing, free-atmosphere seeing, and outer scale. In the illustrative Dome A examples, the inferred free-atmosphere seeing is about $\sim 150$2–$\sim 150$3. The trade-off is altitude resolution: the method provides low-resolution vertical structure rather than fine tomography (Hickson et al., 2019).

4. Differential photometric inference and stellar observables

In stellar parameter estimation, the closest formal construction to “Star-Diff” is the differential surface brightness framework for angular diameters. The method introduces a reddening-free pseudomagnitude,

$\sim 150$4

and a reddening-free, distance-independent differential surface brightness,

$\sim 150$5

The paper calibrates DSB as a function of spectral type number $\sim 150$6, mapping O0 to M9 as 0 to 69, using direct diameter measurements and photometric pairs $\sim 150$7, $\sim 150$8, and $\sim 150$9 with $10^8$0. The final filtered calibration sample contains 573 measurements of 404 distinct stars, spanning spectral types from about O5 to M7 and angular diameters from 0.23 to 44 mas. The reported median statistical diameter error is about $10^8$1, with possible astrophysical or systematic biases at about $10^8$2. Two catalogs are central to the framework: the JMMC Measured Diameters Catalog (JMDC) and the second version of the JMMC Stellar Diameter Catalog (JSDC), expanded to about 453,000 stars (Chelli et al., 2016).

Another strictly differential photometric method infers the sign of stellar surface differential rotation from light-curve morphology. The rotation law is

$10^8$3

with $10^8$4 for solar-like differential rotation and $10^8$5 for antisolar differential rotation. The algorithm computes a generalized Lomb–Scargle periodogram, identifies significant periods $10^8$6 near the dominant period, seeks first harmonics $10^8$7 near $10^8$8, and defines a peak-height ratio

$10^8$9

Lower-latitude spots are inferred to have larger $\phi \approx \frac{2}{3}\pi G\rho_c r^2$0, because they produce less sinusoidal rotational modulation and hence stronger harmonic power. Comparing the assigned lower- and higher-latitude periods then yields the sign of differential rotation. In synthetic tests with solar-like differential rotation, the reported false-positive rate is less than $\phi \approx \frac{2}{3}\pi G\rho_c r^2$1 under the stated conditions and rises to about $\phi \approx \frac{2}{3}\pi G\rho_c r^2$2 in the most permissive setup. Applied to 50 Kepler G stars, the method yields corrected counts of 21–34 solar-like rotators and 5–10 antisolar rotators, depending on the minimum peak-separation criterion (Reinhold et al., 2015).

Beam-shaping diffusers address an instrumental differential-photometry problem rather than a stellar-physics inversion. The diffuser reshapes the stellar PSF into a broad, stable top-hat, reducing sensitivity to pixel-response non-uniformity, guiding drifts, seeing-driven PSF changes, and telescope-induced aberrations. The paper explicitly states that this is not ordinary defocus: the PSF is deterministic and repeatable, and diffusers work in both converging and collimated beams. In a converging beam, the PSF size is approximated by

$\phi \approx \frac{2}{3}\pi G\rho_c r^2$3

Observed precisions include $\phi \approx \frac{2}{3}\pi G\rho_c r^2$4 ppm in 30-minute bins for 16 Cyg A on the ARC 3.5 m telescope, $\phi \approx \frac{2}{3}\pi G\rho_c r^2$5 ppm for WASP-85A b, a conservative 101 ppm for TRES-3b, $\phi \approx \frac{2}{3}\pi G\rho_c r^2$6 ppm for 55 Cnc, and $\phi \approx \frac{2}{3}\pi G\rho_c r^2$7 ppm in the NIR for a $\phi \approx \frac{2}{3}\pi G\rho_c r^2$8 star on the 200-inch Hale Telescope. The paper states that these precisions match or surpass the expected TESS precision in the same magnitude range (Stefansson et al., 2017).

5. Star–diffuse and star–galaxy separation

In imaging inverse problems, “Star-Diff” naturally denotes the separation of point-like stars from diffuse emission. The starblade algorithm poses the decomposition as

$\phi \approx \frac{2}{3}\pi G\rho_c r^2$9

where $$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$0 is the diffuse component and $$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$1 is the point-like component. The point-source prior is independent across pixels and follows an inverse-gamma form encoding a power-law brightness distribution. The diffuse term is modeled as a log-normal field, with Gaussian statistics in log-space and an a priori unknown correlation structure inferred non-parametrically. To stabilize the numerics, the paper introduces a separation field $$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$2 such that $$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$3, $$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$4, and

$$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$5

Inference proceeds through a variational Gaussian approximation to the posterior over $$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$6, with joint estimation of the diffuse power spectrum. On synthetic data, starblade with $$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$7 gives the best reported RMS error on the log-scale separation, $$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$8, compared with $$\frac{\partial \rho(r,t)}{\partial t} = \kappa \nabla^2 \rho(r,t),$$9 for the denoising auto-encoder and $\kappa \sim \frac{r_c^2}{t_r}.$0 or $\kappa \sim \frac{r_c^2}{t_r}.$1 for SExtractor background estimation, depending on the window size. On an HST WFPC2 image of M100, the cosine similarity between the separated diffuse and point-like components is $\kappa \sim \frac{r_c^2}{t_r}.$2 for starblade, compared with $\kappa \sim \frac{r_c^2}{t_r}.$3 for the denoising auto-encoder and $\kappa \sim \frac{r_c^2}{t_r}.$4 for SExtractor. The point-like component naturally absorbs stars, cosmic rays, sharp edges, and other compact artifacts (Knollmüller et al., 2018).

Morphological star–galaxy separation is a related but distinct classification problem: decide whether a pixelized object is better described by a PSF-convolved point-source model or by an extended-source model. The statistical foundation is either a hypothesis test or a Bayesian model-selection problem. With Gaussian pixel noise, the paper writes down the star likelihood for $\kappa \sim \frac{r_c^2}{t_r}.$5 and the galaxy likelihood for $\kappa \sim \frac{r_c^2}{t_r}.$6, then shows that common metrics such as the SDSS PSF–model magnitude difference and SExtractor’s spread_model are closely related to the same underlying model odds ratio. One headline survey result is that 10% worse seeing can be compensated by approximately 0.4 magnitudes deeper data to achieve the same star–galaxy classification performance. The paper also gives an explicit Bayesian prescription for combining multiple morphological or heterogeneous measurements and argues that LSST should continue to improve star–galaxy separation as it pushes to fainter magnitudes, despite the increasing number density of small galaxies (Slater et al., 2019).

The conceptual link between starblade and morphological star–galaxy separation is the move from heuristic scalar cuts to explicit probabilistic modeling. In one case the competition is between point-like and diffuse flux superposed in a single image; in the other it is between unresolved and extended object classes. In both, the decisive quantities are residual structure, priors, and the treatment of uncertainty rather than any uniquely privileged hand-crafted discriminator.

6. Differential residual maps in stellar–gas kinematics

A more recent “Star-Diff-style” construction appears in the kinematic disturbance parameter

$\kappa \sim \frac{r_c^2}{t_r}.$7

which measures the mean absolute mismatch between co-spatial stellar and gas velocity maps after independent min–max normalization to $\kappa \sim \frac{r_c^2}{t_r}.$8. The computation begins from MaNGA DAP stellar and gas velocity maps, keeps only bins unmasked in both maps, requires mean $\kappa \sim \frac{r_c^2}{t_r}.$9-band weighted $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$0, applies a $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$1 clip, normalizes each map separately, and averages the binwise absolute residuals. The metric is deliberately unweighted: there is no flux weighting, no radius weighting, and no velocity-dispersion term. The explicit motivation is to avoid over-weighting galaxy centers and to avoid dependence on MaNGA’s limited uniform sensitivity to low dispersions below the BOSS instrumental width of $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$2 (Powley et al., 13 Apr 2026).

The resulting statistic is intended to be universal in the sense that it does not require a well-defined kinematic position angle, a bi-antisymmetric velocity field, or a parametric rotation model. Artificial rotation tests show that aligned rotation-supported systems move down to $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$3–0.3 when aligned and up to $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$4 when anti-aligned, while disturbed systems remain elevated at roughly $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$5–0.6. The full-sample distribution is bi-modal, with peaks near $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$6 and $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$7 and a minimum near $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$8, which the authors use as an empirical dividing line between “kinematically undisturbed” and “kinematically disturbed,” while cautioning that the threshold can move by $C(<R) = 1 - e^{-R^2/(4\kappa t)}.$9 with preprocessing choices. For seven regular galaxies with clear asymmetric drift, the reported values are $t_r \sim 70$00–0.08 with mean $t_r \sim 70$01, suggesting that asymmetric drift is not a major contaminant (Powley et al., 13 Apr 2026).

In the MaNGA obscured-AGN application, the final AGN sample contains $t_r \sim 70$02 objects, and the control sample is matched in stellar mass and redshift by the Hungarian Method. The $t_r \sim 70$03 distributions of AGN and inactive controls are statistically indistinguishable, with Anderson–Darling statistic $t_r \sim 70$04 and $t_r \sim 70$05. The paper interprets this as evidence against any uniquely preferred kinematic disturbance channel for AGN triggering in that sample. A plausible implication is that “Star-Diff” has expanded, in some contexts, from star-centered photometric or dynamical observables to direct field-to-field mismatch operators between stellar and gaseous components (Powley et al., 13 Apr 2026).

7. Extensions into generative modeling and star formation simulations

Outside observational astronomy, the closest formal analogue is Star-Shaped Denoising Diffusion Probabilistic Models (SS-DDPM). The forward process is not Markovian in the DDPM sense. Instead, every noisy variable is sampled directly from the clean datum,

$t_r \sim 70$06

The reverse process is therefore non-Markovian in the original $t_r \sim 70$07, but for a class of exponential-family forward marginals with linear natural parameters, the dependence on the noisy tail $t_r \sim 70$08 can be compressed into a sufficient statistic,

$t_r \sim 70$09

The practical reverse model then predicts $t_r \sim 70$10 from $t_r \sim 70$11 and plugs that prediction back into the chosen forward family. In the Gaussian case, SS-DDPM is equivalent to DDPM. The paper also gives non-Gaussian constructions for Beta, Dirichlet, von Mises–Fisher, Wishart, Gamma, and categorical data, which are useful when the data lie on constrained manifolds. Reported results include KL $t_r \sim 70$12 versus $t_r \sim 70$13 for Dirichlet SS-DDPM against Gaussian DDPM on simplex-valued data, KL $t_r \sim 70$14 versus $t_r \sim 70$15 for Wishart SS-DDPM on positive definite matrices, text8 NLL $t_r \sim 70$16 bits/char versus $t_r \sim 70$17, and best FID $t_r \sim 70$18 for Beta SS-DDPM on CIFAR-10, comparable to a Gaussian DDPM (Okhotin et al., 2023).

A broader editorial extension would place the semi-deterministic model for individual-star sampling in star-formation simulations alongside these “Star-Diff” usages, although its formal term is SDT rather than Star-Diff. The scheme combines deterministic regulation of the highest-mass star in a cluster with stochastic sampling of lower-mass stars. It uses reservoir particles, on-the-fly friends-of-friends grouping, and the instantaneous cluster mass $t_r \sim 70$19 to define the mass budget for the next sampling step. Relative to purely stochastic schemes, SDT reproduces the observed $t_r \sim 70$20–$t_r \sim 70$21 relation, yields numbers of massive stars consistent with optimal sampling theory, produces the smallest run-to-run variation, delays the first $t_r \sim 70$22 star by about $t_r \sim 70$23 Myr in the tested cloud, and predicts a steeper high-mass galaxy-wide IMF at low star formation rates, with the high-mass slope reaching about $t_r \sim 70$24 and in some fits $t_r \sim 70$25 at SFR $t_r \sim 70$26 (Deng et al., 8 Jun 2026).

Taken together, these extensions reinforce the central terminological point. “Star-Diff” does not identify a unique object but a recurrent methodological pattern: a star-anchored quantity is inferred by tracking diffusion, comparing fields, extracting differential signatures, or summarizing a structured family of stochastic transformations. The specific mathematical content depends entirely on the subfield—stellar dynamics, photometric inversion, wave optics, Bayesian imaging, kinematic comparison, or generative modeling—so any precise use of the term requires immediate disambiguation by its governing formalism and cited source.