betaFeedback

WorksComputer Science

Inducing Fourier Geometry: A Smoothness Prior Imposes Approximate Translation Equivariance in Multimodal Stellar Embeddings

Abstract

Learned representations often contain geometric structure induced by symmetries in the underlying data. We ask whether a known continuous parameter space can instead be used to induce such structure when the corresponding symmetry is only approximate. We introduce a graph-Laplacian smoothness prior that encourages representations to vary smoothly over a physical parameter space without specifying a representation geometry; in the continuum limit it selects Laplacian eigenfunctions and therefore predicts the Fourier-like geometry associated with translation symmetry. On multimodal stellar representations spanning six astronomical surveys, the prior induces Toeplitz Gram structure and sinusoidal modes along stellar parameters, including in modalities whose native representations lack this geometry, while modality-specific residuals preserve structure outside the imposed coordinates. Of the exact-symmetry predictions, stationarity in a recoverable warped coordinate survives; diagnostics assuming exact quantization or asymptotic rates do not. The prior also improves downstream linear-probe performance and label efficiency, suggesting a general strategy: where a meaningful continuous parameter space exists, its local geometry can serve as a data-informed prior.

Introduction

Recent work on representation geometry has shown that translation-symmetric structure in data statistics leaves a characteristic fingerprint in learned embeddings: when pairwise co-occurrence statistics are approximately invariant under shifts of an underlying coordinate, the representation Gram matrix is approximately Toeplitz, its eigenvectors are Fourier modes, and one-dimensional manifolds embed as Lissajous-like curves in successive principal components ([20]). There the geometry emerges, discovered by generic objectives on data whose statistics happen to carry the symmetry. We ask the converse: can it be deliberately induced where the symmetry is only approximate, and is it useful to do so?

Scientific data provides a natural testbed. Stellar properties vary continuously with physical coordinates such as effective temperature (TeffT_{\mathrm{eff}}) and surface gravity (log⁡g\log g), making translation along them physically meaningful, yet stellar populations are not exactly stationary and different surveys carry distinct noise, resolution, and selection effects. We study the shared representation of a multimodal stellar embedding model in which six surveys are aligned into a common latent space, augmenting the alignment objective with a graph-Laplacian smoothness penalty in physical-label space. The penalty pulls nearby stars in (Teff,log⁡g)(T_{\mathrm{eff}}, \log g) together but specifies no Fourier geometry; in the continuum limit it becomes a Dirichlet energy whose fixed-variance minimizers are Laplacian eigenfunctions. Under approximate stationarity, the Fourier geometry of Karkada et al. [20] is therefore a prediction of the prior rather than an empirical surprise.

The prior is not new in isolation; what differs is where we point it. A graph-Laplacian Dirichlet penalty is manifold regularization (Belkin et al., [4]), with the density-weighted Laplacian of diffusion maps as its continuum limit (Coifman and Lafon, [8]), and the formalism has been used for data fusion and alignment ([22, 16] (Wang and Mahadevan 2009)). Here the manifold is supplied externally, as a trusted physical coordinate space, and the regularized object is a shared representation in a deep network. That only part of a multimodal representation admits a common basis is long established ([13, 12, 21] (Tucker 1958; Tsai et al. 2018)) and parallels our shared/residual split; recovering a withheld modality can also be seen as graph-regularized matrix completion (Kalofolias et al., [19]; Boyarski et al., [6]).

Together our results turn Fourier-like representation geometry from a diagnostic of symmetric data into a design target: a smoothness prior defined in physical-label space installs Toeplitz Gram structure and sinusoidal modes in all six surveys’ shared representations, including three whose raw features lack them, and improves linear probing on every label we test.

Setting: data, model, and the smoothness prior

Data. We use six large-scale stellar surveys spanning spectroscopy (LAMOST, APOGEE, GALAH), spectrophotometry (Gaia XP), and photometric time series (Kepler, TESS), together with Gaia astrometry, cross-matched through the Gaia DR3 source identifier. We probe geometry along continuous stellar parameters, primarily TeffT_{\mathrm{eff}}, log⁡g\log g, and [Fe/H]; survey statistics and preprocessing are in § 5.1.

Model. Each modality is mapped by a pretrained, frozen encoder to features ziz_i. A Perceiver-IO-style alignment module maps these into a d=64d = 64 shared stream, trained for cross-survey agreement, and a d=192d = 192 per-survey residual stream that can retain modality-specific structure. The shared stream is trained with VICReg (Bardes et al., [3]); astrometry enters as a conditioning vector.

The smoothness prior. We add a density-normalised graph-Laplacian penalty in physical-label space. Let sˉa\bar{s}_a be the mean shared token of star aa and ca=(Teff,log⁡g)ac_a = (T_{\mathrm{eff}}, \log g)_a its normalized label coordinates. With a fixed-bandwidth Gaussian kernel wab=exp⁡(−∥ca−cb∥2/2ε2)w_{ab} = \exp(-\lVert c_a - c_b \rVert^2 / 2\varepsilon^2) and degrees qa=∑bwabq_a = \sum_b w_{ab},

Lsmooth=1ds∑(a,b)∈Ew~ab∥sˉa−sˉb∥2∑(a,b)∈Ew~ab,w~ab=wabqaqb,(1)\mathcal{L}_{\mathrm{smooth}} = \frac{1}{d_s}\frac{\sum_{(a,b)\in E}\widetilde{w}_{ab}\lVert\bar{s}_a-\bar{s}_b\rVert^2}{\sum_{(a,b)\in E}\widetilde{w}_{ab}}, \qquad\widetilde{w}_{ab} = \frac{w_{ab}}{q_a q_b}, \tag*{(1)}

where EE retains edges with ∥ca−cb∥<3ε\lVert c_a-c_b\rVert< 3\varepsilon. The 1/(qaqb)1/(q_aq_b) normalisation is the α=1\alpha= 1 member of the Coifman–Lafon family (Coifman and Lafon, [8]), which makes the continuum limit the unweighted Dirichlet energy ∫∥∇cs∥2 dc\int\lVert\nabla c_s\rVert^2\,dc rather than a density-weighted one and ensures the penalty is not dominated by the most densely populated regions of the HR diagram (Figure 3); the VICReg variance term, evaluated under a matching density-free measure, fixes its scale (§ 5.4). Minimising then selects the low-lying eigenfunctions of the ordinary Laplacian, which for an approximately one-dimensional coordinate are sinusoids of increasing node count. The penalty encodes no Fourier structure explicitly: the geometry of Karkada et al. [20] is a predicted minimiser of the loss rather than a feature we impose by hand.

Heatmaps of graph degree relative to the batch median, without and with density correction

Figure 3. Where the smoothness penalty acts. Each star’s graph degree, relative to the batch median (log⁡10\log_{10}), binned in (TeffT_{\mathrm{eff}}, log⁡g\log g). Left: without the density correction (α=0\alpha=0), the weight concentrates on the densest regions, the main sequence and red clump, and giants, hot stars and subgiants are nearly unweighted. Right: with it (α=1\alpha=1, our model), the weight spreads across a wider part of the HR diagram; the few very high-degree bins at the sparse edges are stars the correction up-weights where the density estimate is thin.

Measurements and scope. Geometry is measured from bin-mean embeddings along individual coordinates (or coordinate pairs), on signature-consistent main-sequence samples to avoid survey-availability effects; we quantify Toeplitz structure by the fraction of off-diagonal Gram variance explained (RToep2R^{2}_{\mathrm{Toep}}), count sign changes in leading eigenmodes, and fit kernels to Gram-diagonal decay. Two scope conditions apply: stellar parameters enter parts of the training pipeline, so we claim induced consistency, not emergence from unlabeled data; and the physical symmetry is only approximate, so we test explicitly which exact-symmetry predictions survive. Throughout, “with Lsmooth\mathcal{L}_{\mathrm{smooth}}” denotes the full density-normalised prior—the α=1\alpha=1 kernel together with the 1/q1/q-weighted variance and covariance statistics; the ablation removes both.

Results

The prior installs the geometry. The direct test compares the shared stream with and without Lsmooth\mathcal{L}_{\mathrm{smooth}}, holding data, architecture, and seed fixed. With the term, the modality-averaged shared token’s Gram matrix along TeffT_{\mathrm{eff}} is strongly Toeplitz (RToep2=0.92R^{2}_{\mathrm{Toep}} = 0.92) with leading eigenmodes of node count 1⋅2⋅3⋅41\mathbin{\cdot}2\mathbin{\cdot}3\mathbin{\cdot}4, tracing closed Lissajous curves in successive principal-component pairs (Figure 1). Without it the fit drops to 0.63 and the node ladder disorders; log⁡g\log g, the second graph coordinate, behaves the same way (Figure 4).

Translation symmetry in the shared latent space

Figure 1. Translation symmetry in the shared latent space. Upper row: our full model, with Lsmooth\mathcal{L}_{\mathrm{smooth}}. Fourier modes in TeffT_{\mathrm{eff}} with frequency matching eigenmode degree, and principal-component manifolds matching the prediction of Karkada et al. [20]. Lower row: the same model without Lsmooth\mathcal{L}_{\mathrm{smooth}}, showing neither.

Fourier modes and principal-component projections for log $g$, with and without $L_{\mathrm{smooth}}$

Figure 4. Same as Figure 1, but for log gg.

Where the geometry lives. Figure 6 probes three layers of the with-prior model. In the raw features ziz_i the symmetry is strong in the spectroscopic surveys (0.75–0.85) and weak in the photometric and spectrophotometric ones (0.35–0.60); the shared stream makes it stronger and uniform (all six in 0.85–0.93) while the residual stays uneven (0.45–0.85). The battery along log⁡g\log g reproduces the pattern (Figure 7). Along [Fe/H], a coordinate the prior does not name, the shared stream transfers the symmetry rather than taxing it: native spectroscopic structure is approximately preserved while the surveys lacking it gain it outright (TESS 0.27 → 0.76; Figure 8). The transfer is carried by alignment, not the prior: the pooled [Fe/H] Gram is indistinguishable with and without Lsmooth\mathcal{L}_{\mathrm{smooth}} (RToep2=0.76R^{2}_{\mathrm{Toep}} = 0.76 vs. 0.75; Figure 5). The induced geometry is targeted: installed along the named coordinates, inert off-graph.

Figure 5. Same as Figure 1, but for [Fe/H].

Fourier geometry per modality and layers in the model for $T_{\mathrm{eff}}$

Figure 6. Fourier geometry per modality and layers in the model for TeffT_{\mathrm{eff}}. Each column represents a specific modality embedding. Each row represents different parts in the model - raw pretrained features, ziz_i, (upper), shared space (middle), and residual space (lower). We see that the shared space improves Fourier geometry compared to ziz_i across all modalities. The residual stream preserves the geometry of ziz_i.

Fourier geometry per modality and layers in the model for log $g$

Figure 7. same as Figure 6, for log gg

Figure 8. same as Figure 6, for [Fe/H][\mathrm{Fe}/\mathrm{H}]. Note that this is not a graph coordinate.

Which symmetry diagnostics survive approximation. Karkada et al. [20] predict three signatures of a distance-only latent coordinate: a Toeplitz Gram matrix, quantized eigenmode wavenumbers, and decoding error ∼r−1/D\sim r^{-1/D}. Only the first survives, and Toeplitz structure alone does not exclude a matched non-stationary kernel; the evidence is the co-ordinate: reparameterising by embedded arclength raises it to 0.97, and an independently fitted monotone warping agrees (r=0.92r = 0.92). Mode counting is undecidable, its quantization ladder assuming an exponential kernel where our spectrum is Gaussian (ℓ≈440\ell\approx440 K), and decoding-error scaling deviates (r−0.5r^{-0.5} vs. r−1r^{-1} along TeffT_{\mathrm{eff}}; r−0.8r^{-0.8} vs. r−1/2r^{-1/2} on the plane), consistent with a finite lattice below the asymptotic regime.

Does the induced geometry pay? Linear probes on the frozen shared latent improve with the prior on every label at every training-set size (Figure 2), most where labels are scarce: age reaches R2≈0.38±0.09R^{2} \approx0.38 \pm0.09 from 50 labelled stars, comparable within error to the without-prior probe at 20,000. We attribute these gains to the prior as a whole; whether Fourier structure specifically, rather than smoothness generally, buys the improvement remains open.

Effect of $\mathcal{L}_{\mathrm{smooth}}$ on downstream tasks

Figure 2. Effect of Lsmooth\mathcal{L}_{\mathrm{smooth}} on downstream tasks. Linear probes on the frozen shared latent, with Lsmooth\mathcal{L}_{\mathrm{smooth}} (black) and without (magenta), against the number of labelled training stars.

Physical coordinates, not generic smoothness. We retrained the model changing only the graph the Laplacian is built on, with data, seed and recipe fixed (Table 3). A permutation null (identical kernel, bandwidth and neighbour count, with (TeffT_{\mathrm{eff}}, log⁡g\log g) shuffled across stars) falls below a no-Laplacian baseline on every label and destroys the geometry (RToep2=0.39R^{2}_{\mathrm{Toep}} = 0.39). A label-free regulariser built on the model’s own embedding also falls below it. Only the physical graph improves on no-Laplacian across labels, by +0.04 to +0.11 on five of six; a graph on the input embeddings, whose neighbours are only partly physical, helps only on TeffT_{\mathrm{eff}} and log⁡g\log g. Without the density correction (α=0\alpha=0), the penalty’s weight collapses onto the main sequence and red clump (Figure 3): the geometry on our main-sequence lattice is the cleanest of any arm (RToep2=0.91R^{2}_{\mathrm{Toep}} = 0.91), but probes over the full population are weaker. A lattice score measured in one region therefore cannot rank models that weight regions differently. What the controls do show is that with the wrong neighbours, the geometry and the gain fail together.

Graph | TeffT_{\mathrm{eff}} | log gg | [Fe/H] | Age | Act. | vsin⁡iv\sin i | RToep2R^{2}_{\mathrm{Toep}} Physical (rerun) | 0.872 | 0.873 | 0.530 | 0.557 | 0.165 | 0.481 | 0.852 Physical (paper) | 0.860 | 0.875 | 0.502 | 0.562 | 0.167 | 0.509 | 0.889 No Laplacian | 0.781 | 0.762 | 0.435 | 0.490 | 0.123 | 0.474 | 0.759 Permuted | 0.548 | 0.449 | 0.180 | 0.302 | 0.005 | 0.399 | 0.394 Latent self-graph | 0.705 | 0.504 | 0.204 | 0.448 | 0.109 | 0.442 | 0.775 Input ziz_i | 0.826 | 0.869 | 0.326 | 0.478 | 0.124 | 0.486 | 0.846 Physical, α=0\alpha=0 | 0.812 | 0.813 | 0.147 | 0.508 | 0.103 | 0.468 | 0.909 Physical, warped | 0.879 | 0.889 | 0.442 | 0.555 | 0.158 | 0.427 | 0.861

Table 3. Graph controls: linear-probe R2R^{2} on the shared token (ntrain=1000n_{\mathrm{train}}=1000, 30 paired draws) and RToep2R^{2}_{\mathrm{Toep}} along TeffT_{\mathrm{eff}} on 11,347 clean main-sequence singles with LAMOST, APOGEE and Gaia, fixed so that survey coverage cannot drive the comparison. All arms share data, seed and recipe; only the Laplacian’s graph differs. The two FAME rows are the same recipe trained twice and set the noise level (≲0.03\lesssim0.03 in probe R2R^{2}, 0.04 in RToep2R^{2}_{\mathrm{Toep}}). One training seed per arm; best checkpoints fall at different epochs (3 to 79).

Implied distribution. A Fourier geometry further implies that leading directions are arcsine-valued rather than Gaussian, the opposite of what distributional priors such as SI-GReg ([2]) target. This holds in all six surveys with the prior, in four without it, and in none of a vanilla VICReg latent (§ 5.6).

Discussion and conclusions

Emergence and induction are complementary: emergence reveals what data statistics already contain, induction supplies what they do not. Here the induction succeeds without any underlying physical invariance. What stellar physics guarantees is that observables vary smoothly with TeffT_{\mathrm{eff}} and log⁡g\log g, the same assumption behind spectral-grid interpolation, and approximate stationarity follows where information content varies slowly along the coordinate. The failures obey the same logic: age, whose governing physics changes regime across the Kraft break, shows no symmetry at any layer, and rotation only a partial one. The prior installs geometry where it is pointed and nowhere else. Importantly, the prior consistently improves performance over all labels, not only on graph coordinates. This highlights the importance of such priors in scientific domains.

Two questions follow. Our controls rule out generic smoothness, but not smoothness in physical coordinates: whether the Fourier structure itself drives the downstream gains, or comes along with them, remains open, since the two fail together when the graph’s neighbours are wrong. And what replaces exact quantization and decoding-rate predictions under approximate symmetry is, to our knowledge, open. The physical sciences may be the natural place to settle both, since there the coordinates are real and the geometry can be checked against nature.

appendix

data and pre processing

Kepler. The Kepler mission [5] monitored a single ∼115 deg2\sim115\,\mathrm{deg}^{2} field in Cygnus–Lyra for four years at 30-minute (long) cadence, delivering the highest-precision, longest-baseline optical photometry available for main-sequence stars. We use the long-cadence PDCSAP light curves (STScI (2016)) on their native, uniform 48 samples day−1^{-1} grid, with gaps zero-filled and no rebinning, cut into 450-day windows (21,600 points). Each window is lightly smoothed (Savitzky–Golay, ∼0.5\sim0.5 d) and expanded into six channels: standardized flux, median-normalized flux, two activity-amplitude envelopes (a slow envelope on a 6 ProtP_{\rm rot} window and a high-passed one on a 1 ProtP_{\rm rot} window), the autocorrelation function (ACF, [25]), and the Lomb-Scargle periodogram (LS, [23] (), Scargle 1982)). The resulting sample is 193,951 stars.

TESS. TESS (Ricker et al., 2015) trades Kepler’s depth for sky coverage, observing nearly the full sky in ∼\sim27-day sectors at multiple cadences. We use light curves from the Quick-Look Pipeline ([17]), restricted to Tmag<16T_{\rm mag} < 16 and to targets observed in more than one sector. Each light curve is resampled to a 30-minute grid and truncated to 1000 points in four channels: standardized flux, median-filtered flux, ACF, and LS. These are the same channels as in Kepler, but without the activity envelopes, as the sectors are too short for it. Compared to Kepler, TESS has a wider coverage (1,503,260 encoded stars) but has a shorter per-sector baseline, bigger pixel size (21″ vs. 4″), and is generally noisier (Van Cleve et al. (2016)).

LAMOST. LAMOST (Zhao et al., 2012) is a 4-m telescope with 4000 fibers, conducting the largest optical spectroscopic survey of the northern sky. We use low-resolution (R∼1800R \sim1800) spectra from [11], continuum-normalized and resampled onto a common 4096-point wavelength grid. Its combination of scale (5.1M spectra) and northern footprint makes LAMOST the dominant spectroscopic partner for both Kepler and TESS.

APOGEE. APOGEE ([24]) is a high-resolution (R∼22,500R \sim22{,}500) HH-band spectrograph. Working in the near-infrared lets it observe through dust into the Galactic plane and bulge, and it targets predominantly giants—a population complementary to the dwarf-dominated Kepler and GALAH samples. We use the continuum-normalized combined spectra from [1], resampled to 8576 points.

GALAH. GALAH ([10]) is a high-resolution (R∼28,000R \sim28{,}000) southern optical survey using the HERMES spectrograph on the Anglo-Australian Telescope, designed for Galactic archaeology and delivering up to ∼\sim30 elemental abundances per star. We use the continuum-normalized spectra of [7] over 4700–7900 Å, concatenated across the four HERMES arms onto a 14,420-point grid, and learning, in addition to basic stellar parameters and vsin⁡iv \sin i, also 11 abundance elements ([Mg/Fe], [Si/Fe], [Ca/Fe], [Ti/Fe], [Na/Fe], [Al/Fe], [Mn/Fe], [Ni/Fe], [Cr/Fe], [Ba/Fe], [Y/Fe]). GALAH’s southern footprint is disjoint from Kepler’s.

Gaia. Gaia ([14]) provides all-sky astrometry and low-resolution spectrophotometry, and serves here as the universal hub. We use two products from DR3 ([15]): (i) The BP/RP (XP) sampled spectra ([9]): 340 flux points spanning 330–1050 nm, scaled per-spectrum with a robust (median/IQR) scaler and deliberately not continuum-normalized, so that the absolute spectral-energy-distribution shape, which carries temperature and extinction information, is preserved. (ii) The five-parameter astrometric solution plus radial velocities, from which we

KeplerTESSLAMOSTAPOGEEGALAH
Kepler91,21369,70166,61520,7120
TESS1,247,706935,262226,347188,513
LAMOST1,053,269135,42960,388
APOGEE310,34755,623
GALAH247,444

caption: tbl:1: Survey census and pairwise overlaps in the cross-match master (unique stars; Gaia is the anchor and covers essentially all 1,386,210 stars). The diagonal gives each survey’s total; off-diagonal entries give the number of stars observed by both surveys.

:::

ModalityArchitectureParamsInput Lengthdid_i
Gaia XPSpectral ViT38.5 M340512
LAMOSTSpectral ViT615.3 M40962048
APOGEESpectral ViT154.2 M85761024
GALAHSpectral ViT154.4 M14,4201024
KeplerDual-view LC129.6 M6×21,6006 \times21{,}600768
TESSDual-branch LC139.5 M4×10004 \times1000512

Table 2. The six frozen unimodal encoders. Input Length is the number of flux bins (spectra) or channels and time samples (light curves). did_i is the feature dimensionality.

derive UVWUVW space velocities via the equatorial-to-Galactic transformation of Johnson and Soderblom [18].

Cross-match

All surveys are cross-matched onto a common star identity keyed by the Gaia DR3 source_id. The master cross-match retains every star observed by at least two of the five non-Gaia surveys, yielding 1,613,142 observations of 1,386,210 unique stars. Table 1 lists the per-survey census and the pairwise overlaps. The pair frequencies span roughly two orders of magnitude—from TESS×LAMOST (935,262 stars) down to Kepler×APOGEE (20,712). Notably, Kepler and GALAH have zero common stars (northern vs. southern footprints): these two modalities are never observed together and can only be related transitively, through the shared latent space. Counting the number of non-Gaia surveys per star, 1,225,913 stars have exactly two, 143,035 have three, and 17,262 have four; no star is covered by all five.

Training

Pre-training was done on a mixture of H100 and B200 GPUs. Each encoder was trained until convergence, using early stopping criteria. Training times vary between several hundred and several thousand GPU-hours. Training the alignment model is fast: of the order of 10 B200 GPU-hours. The alignment model is trained on a fixed stratified subset that deliberately reserves half its budget for single-modality stars, so the shared space is shaped by the long unimodal tail and not only by the well-covered multimodal core. Multimodal stars are weighted by their rarest constituent modality pair. The invariance and variance terms in VICReg were given a weight of 25. All other loss terms had a weight of 1.

Why a continuity penalty produces a Fourier geometry

We sketch why minimising the smoothness penalty organises the shared coordinates into sinusoidal modes. The argument motivates the observed geometry rather than proving a theorem about the full objective; idealisations are stated as we go.

The unnormalised penalty carries the density squared. Write s(c)s(c) for the shared embedding at parameter cc. An unnormalised graph penalty 1n2∑a≠bwab∥sa−sb∥2\frac{1}{n^2}\sum_{a\ne b}w_{ab}\lVert s_a-s_b\rVert^2 is a sum over pairs of stars drawn i.i.d. from the sampling density pp, so its expectation integrates against p(x)p(y)p(x)p(y). Expanding at small bandwidth,

1n2∑a≠bwab∥sa−sb∥2⟶ε2∫∥s′(c)∥2p(c)2 dc,(2)\frac{1}{n^2}\sum_{a\ne b} w_{ab}\lVert s_a-s_b\rVert^2 \longrightarrow\varepsilon^2 \int\lVert s'(c)\rVert^2p(c)^2\,dc, \tag*{(2)}

one factor of pp per summation index. More generally, dividing the kernel by qaαqbαq_a^\alpha q_b^\alpha (with qa∝p(ca)q_a \propto p(c_a) the degree) gives a continuum limit weighted by p2−2αp^{2-2\alpha} [8]. A global normaliser such as ∑w\sum w cannot change this exponent; only the per-node division can.

Normalizing the density Taking α=1\alpha=1, i.e. w~ab=wab/(qaqb)\widetilde{w}_{ab}=w_{ab}/(q_aq_b), cancels the density to leading order and leaves the unweighted Dirichlet energy E[s]=∫∥s′(c)∥2 dcE[s]=\int\lVert s'(c)\rVert^2\,dc. Its Euler–Lagrange operator, under Neumann boundary conditions, is the ordinary Laplacian −d2/dc2-d^2/dc^2.

The scale constraint carries the density too. The energy alone is minimised by a constant; a scale must be fixed, and this is done by the VICReg variance and covariance terms. But these are batch averages, hence estimates of integrals against pp: the variance constraint is “unit variance under pp”, which reintroduces the density through the measure and warps the minimisers by a factor ∝p\propto\sqrt{p}. We therefore apply the same α=1\alpha=1 correction a second time, importance-weighting every variance and covariance moment by 1/qa∝1/p(ca)1/q_a\propto1/p(c_a) before the (unchanged) hinge and off-diagonal penalties. Under a quadratic-constraint idealisation the effective constraint becomes “unit variance under Lebesgue measure”, and the Rayleigh–Ritz problem is the density-free

−sn′′=λnsn.(3)-s_n''=\lambda_n s_n. \tag*{(3)}

Normalising the kernel alone leaves the p\sqrt{p} warp; normalising both is what makes the geometry density-free. The invariance term is left unweighted: it compares two modalities of the same star, a per-pair quantity with no density to cancel.

Fourier modes and their robustness. On [0,L][0,L] with Neumann boundaries the eigenfunctions are sn(c)=cos⁡(nπc/L)s_n(c)=\cos(n\pi c/L), λn=(nπ/L)2\lambda_n=(n\pi/L)^2: the observed ladder of increasingly oscillatory coordinates. Even where the correction is imperfect, the operator remains of Sturm–Liouville type, whose nn-th eigenfunction has exactly nn nodes, so the node ladder is preserved regardless of residual density weighting. This robustness is why the ladder survives approximate stationarity.

Why two coordinates degrade. The one-dimensional argument does not extend directly to two dimensions. The distribution in (Teff,log⁡g)(T_{\mathrm{eff}}, \log g) is neither translation invariant nor defined on a rectangular domain, so higher eigenfunctions distort and mix through near-degenerate eigenvalues. The lowest-frequency mode is far more robust, explaining why only the fundamental remains clean in the two-dimensional analysis.

Bandwidth. ε\varepsilon is fixed (chosen for ∼10\sim10 graph neighbours per batch); edges are retained within a 3ε3\varepsilon ball. Optimising the bandwidth, including per-coordinate, is left to future work.

Additional Results

Marginal shape

A Fourier geometry constrains the feature distribution: a direction close to cos⁡(kπc)\cos(k\pi c) should be bimodal across stars (the arcsine law). We test this per principal direction on frozen embeddings, calibrated against a synthetic Fourier manifold (Table 5); per-block coverage is in Table 4. Past the third direction every model is heavy-tailed rather than Gaussian, so “arcsine head over a Gaussian bulk” is not the geometry these encoders produce.

KeplerLAMOSTAPOGEETESSGaia XPGALAH
nn1,81914,16120,00020,00020,00020,000
Raw ziz_i+449.380−1.162−1.561+22.440+4.210−1.691
Vanilla VICReg−0.950−0.190−0.240−0.050−0.310+0.340
Shared, no Lsmooth\mathcal{L}_{\mathrm{smooth}}−1.261−1.132−1.002+0.040−1.072+0.820
Shared, with Lsmooth\mathcal{L}_{\mathrm{smooth}}−1.402−1.372−1.502−1.332−1.382−1.541
Residual, with Lsmooth\mathcal{L}_{\mathrm{smooth}}+446.250−1.102−0.890+5.730+3.350−1.641

Table 4. Marginal shape by survey block: excess kurtosis of PC1 (γ2\gamma_2; pure 1-D arcsine limit −1.5, Gaussian 0), with strong depth as a subscript: the number of leading contiguous PCs with γ2<−1\gamma_2 < -1. Bold marks γ2(PC1)<−1\gamma_2(\mathrm{PC1}) < -1. Star sets differ per block (nn in the header row), so this table reports coverage, not cross-arm magnitude; for the latter see Table 5. Kepler is the minimum stratum (n=1,819n = 1,819) and its verdict is the least secure. Lsmooth\mathcal{L}_{\mathrm{smooth}} is the density-normalised (α=1\alpha=1) form throughout: a fixed-ϵ\epsilon kernel divided by qaqbq_a q_b, with the VICReg variance and covariance statistics importance-weighted by 1/q1/q. The prior takes the signature from three blocks to six and tightens it onto the arcsine value; without the prior the same architecture retains it in four blocks, marginally, and loses TESS and GALAH.

ArmPanelγ2(PC1)\gamma_2(\mathrm{PC1})γ2(PC2)\gamma_2(\mathrm{PC2})Strong depthMed. PC1–8Bulk
Raw ziz_iLAMOST−1.38−1.052+0.100.00
APOGEE−1.53−0.451+1.040.00
Vanilla VICRegLAMOST−0.29−0.320−0.020.00
APOGEE−0.42−0.520−0.080.00
pooled−0.48−0.600−0.080.00
Shared, no Lsmooth\mathcal{L}_{\mathrm{smooth}}LAMOST−0.86−1.100+1.520.00
APOGEE−0.84−1.110+1.520.00
pooled−0.22−1.210+1.300.00
Shared, with Lsmooth\mathcal{L}_{\mathrm{smooth}}LAMOST−1.54−1.132+2.730.00
APOGEE−1.54−1.132+2.740.00
pooled−1.58−1.092+2.050.00
Residual, with Lsmooth\mathcal{L}_{\mathrm{smooth}}LAMOST−1.37−0.781+1.730.00
APOGEE−0.88−0.650+0.520.00
Control: Fourier, uniform coordinate—−1.49−1.434−0.730.90
Control: Fourier, measured density—−1.51−1.322−0.080.88
Control: isotropic Gaussian—−0.06−0.010+0.020.81

Table 5. Common star set and calibration. Upper block: tier cross (clean main-sequence stars with both LAMOST and APOGEE present, n=10,214n = 10{,}214; pooled is the modality-averaged token on the largest exact availability cell, n=4,584n = 4{,}584). Identical stars in every row, so this is the only cross-arm comparison we make. Lower block: synthetic manifolds pushed through the identical pipeline at the measured signal fraction (0.899). The density-matched control is the honest reference: a perfect Fourier manifold at this sample’s own TeffT_{\mathrm{eff}} density reaches strong depth 2, so the achievable depth here is two to three directions. Lsmooth\mathcal{L}_{\mathrm{smooth}} is the density-normalised (α=1\alpha= 1) form throughout. Single seed, ε=0.0139\varepsilon= 0.0139.

Shared latent space with and without $L_{\mathrm{smooth}}$

Figure 9. Effect of LsmoothL_{\mathrm{smooth}} on the latent space. Left: UMAP of the shared latent space with LsmoothL_{\mathrm{smooth}}. Right: the same model without it. The prior yields a more continuous latent space with a clean log⁡g\log g gradient.

References

  1. [1]Abdurro’uf, Katherine Accetta, Conny Aerts, Víctor Silva Aguirre, Romina Ahumada, Nikhil Ajgaonkar, N. Filiz Ak, Shadab Alam, Carlos Allende Prieto, Andrés Almeida, Friedrich Anders, Scott F. Anderson, Brett H. Andrews, Borja Anguiano, Erik Aquino-Ortíz, Alfonso Aragón-Salamanca, Maria Argudo-Fernández, Metin Ata, Marie Aubert, Vladimir Avila-Reese, Carles Badenes, Rodolfo H. Barbá, Kat Barger, Jorge K. Barrera-Ballesteros, Rachael L. Beaton, Timothy C. Beers, Francesco Belfiore, Chad F. Bender, Mariangela Bernardi, Matthew A. Bershady, Florian Beutler, Christian Moni Bidin, Jonathan C. Bird, Dmitry Bizyaev, Guillermo A. Blanc, Michael R. Blanton, Nicholas Fraser Boardman, Adam S. Bolton, Médéric Boquien, Jura Borissova, Jo Bovy, W. N. Brandt, Jordan Brown, Joel R. Brownstein, Marcella Brusa, Johannes Buchner, Kevin Bundy, Joseph N. Burchett, Martin Bureau, Adam Burgasser, Tuesday K. Cabang, Stephanie Campbell, Michele Cappellari, Joel K. Carlberg, Fábio Carneiro Wanderley, Ricardo Carrera, Jennifer Cash, Yan-Ping Chen, Wei-Huai Chen, Brian Cherinka, Cristina Chiappini, Peter Doohyun Choi, S. Drew Chojnowski, Haeun Chung, Nicolas Clerc, Roger E. Cohen, Julia M. Comerford, Johan Comparat, Luiz da Costa, Kevin Covey, Jeffrey D. Crane, Irene Cruz-Gonzalez, Connor Culhane, Katia Cunha, Y. Sophia Dai, Guillermo Damke, Jeremy Darling, James W. Davidson, Jr., Roger Davies, Kyle Dawson, Nathan De Lee, Aleksandar M. Diamond-Stanic, Mariana Cano-Díaz, Helena Domínguez Sánchez, John Donor, Chris Duckworth, Tom Dwelly, Daniel J. Eisenstein, Yvonne P. Elsworth, Eric Emsellem, Mike Eracleous, Stephanie Escoffier, Xiaohui Fan, Emily Farr, Shuai Feng, José G. Fernández-Trincado, Diane Feuillet, Andreas Filipp, Sean P. Fillingham, Peter M. Frinchaboy, Sebastien Fromenteau, Lluís Galbany, Rafael A. García, D. A. García-Hernández, Junqiang Ge, Doug Geisler, Joseph Gelfand, Tobias Géron, Benjamin J. Gibson, Julian Goddy, Diego Godoy-Rivera, Kathleen Grabowski, Paul J. Green, Michael Greener, Catherine J. Grier, Emily Griffith, Hong Guo, Julien Guy, Massinissa Hadjara, Paul Harding, Sten Hasselquist, Christian R. Hayes, Fred Hearty, Jesús Hernández, Lesley Hill, David W. Hogg, Jon A. Holtzman, Danny Horta, Bau-Ching Hsieh, Chin-Hao Hsu, Yun-Hsin Hsu, Daniel Huber, Marc Huertas-Company, Brian Hutchinson, Ho Seong Hwang, Héctor J. Ibarra-Medel, Jacob Ider Chitham, Gabriele S. Ilha, Julie Imig, Will Jaekle, Tharindu Jayasinghe, Xihan Ji, Jennifer A. Johnson, Amy Jones, Henrik Jönsson, Ivan Katkov, Arman Khalatyan, Dr., Karen Kinemuchi, Shobhit Kisku, Johan H. Knapen, Jean-Paul Kneib, Juna A. Kollmeier, Miranda Kong, Marina Kounkel, Kathryn Kreckel, Dhanesh Krishnarao, Ivan Lacerna, Richard R. Lane, Rachel Langgin, Ramon Lavender, David R. Law, Daniel Lazarz, Henry W. Leung, Ho-Hin Leung, Hannah M. Lewis, Cheng Li, Ran Li, Jianhui Lian, Fu-Heng Liang, Lihwai Lin, Yen-Ting Lin, Sicheng Lin, Chris Lintott, Dan Long, Penélope Longa-Peña, Carlos López-Cobá, Shengdong Lu, Britt F. Lundgren, Yuanze Luo, J. Ted Mackereth, Axel de la Macorra, Suvrath Mahadevan, Steven R. Majewski, Arturo Manchado, Travis Mandeville, Claudia Maraston, Berta Margalef-Bentabol, Thomas Masseron, Karen L. Masters, Savita Mathur, Richard M. McDermid, Myles Mckay, Andrea Merloni, Michael Merrifield, Szabolcs Meszaros, Andrea Miglio, Francesco Di Mille, Dante Minniti, Rebecca Minsley, and Antonela Monachesi. The Seventeenth Data Release of the Sloan Digital Sky Surveys: Complete Release of MaNGA, MaStar, and APOGEE-2 Data. ApJS, 259(2):35, April 2022. doi: 10.3847/1538-4365/ac4414.
  2. [2]Randall Balestriero and Yann LeCun. LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics. arXiv e-prints, art. arXiv:2511.08544, November 2025. doi: 10.48550/arXiv.2511.08544.DOI
  3. [3]Adrien Bardes, Jean Ponce, and Yann LeCun. VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning. arXiv e-prints, art. arXiv:2105.04906, May 2021. doi: 10.48550/arXiv.2105.04906.DOI
  4. [4]Mikhail Belkin, Partha Niyogi, and Vikas Sindhwani. Manifold regularization: A geometric framework for learning from labeled and unlabeled examples. Journal of Machine Learning Research, 7(85):2399–2434, 2006.DOI
  5. [5]William J. Borucki, David Koch, Gibor Basri, Natalie Batalha, Timothy Brown, Douglas Caldwell, John Caldwell, Jørgen Christensen-Dalsgaard, William D. Cochran, Edna DeVore, Edward W. Dunham, Andrea K. Dupree, Thomas N. Gautier, John C. Geary, Ronald Gilliland, Alan Gould, Steve B. Howell, Jon M. Jenkins, Yoji Kondo, David W. Latham, Geoffrey W. Marcy, Søren Meibom, Hans Kjeldsen, Jack J. Lissauer, David G. Monet, David Morrison, Dimitar Sasselov, Jill Tarter, Alan Boss, Don Brownlee, Toby Owen, Derek Buzasi, David Charbonneau, Laurance Doyle, Jonathan Fortney, Eric B. Ford, Matthew J. Holman, Sara Seager, Jason H. Steffen, William F. Welsh, Jason Rowe, Howard Anderson, Lars Buchhave, David Ciardi, Lucianne Walkowicz, William Sherry, Elliott Horch, Howard Isaacson, Mark E. Everett, Debra Fischer, Guillermo Torres, John Asher Johnson, Michael Endl, Phillip MacQueen, Stephen T. Bryson, Jessie Dotson, Michael Haas, Jeffrey Kolodziejczak, Jeffrey Van Cleve, Hema Chandrasekaran, Joseph D. Twicken, Elisa V. Quintana, Bruce D. Clarke, Christopher Allen, Jie Li, Haley Wu, Peter Tenenbaum, Ekaterina Verner, Frederick Bruhweiler, Jason Barnes, and Andrej Prsa. Kepler Planet-Detection Mission: Introduction and First Results. Science, 327 (5968):977, February 2010. doi: 10.1126/science.1185402.DOI
  6. [6]Amit Boyarski, Sanketh Vedula, and Alexander M. Bronstein. Spectral geometric matrix completion. In Mathematical and Scientific Machine Learning, 2019. URL https://api.semanticscholar.org/CorpusID:235363958.arxiv.org/abs/1911.07255
  7. [7]Sven Buder, Janez Kos, Xi Ella Wang, Madeleine McKenzie, Madeleine Howell, Sarah Martell, Michael R. Hayden, Daniel B. Zucker, Thomas Nordlander, Benjamin Montet, Gregor Traven, Joss Bland-Hawthorn, Gayandhi M. de Silva, Kenneth Freeman, Geraint Lewis, Karin Lind, Sanjib Sharma, Jeffrey D. Simpson, Dennis Stello, Tomaz Zwitter, Anish M. Amarsi, Joseph J. Armstrong, Kirsten Banks, Mark Beavis, Kevin-Luke Beeson, Boquan Chen, Ioana Ciucă, Gary S. da Costa, Richard de Grijs, Bailey Martin, David Moise Nataf, Melissa Ness, Adam D. Rains, Tim Scarr, Rok Vogrinčić, Zixian Purmort al Wang, Rob A. Wittenmyer, Yi Anne Xie, and The Galah Collaboration. The GALAH survey: Data release 4. PASA, 42:e051, May 2025. doi: 10.1017/pasa.2025.26.DOI
  8. [8]Ronald R. Coifman and Stéphane Lafon. Diffusion maps. Applied and Computational Harmonic Analysis, 21(1):5–30, 2006. ISSN 1063-5203. doi: https://doi.org/10.1016/j.acha.2006.04.006. URL https://www.sciencedirect.com/science/article/pii/S1063520306000546. Special Issue: Diffusion Maps and Wavelets.
  9. [9]F. De Angeli, M. Weiler, P. Montegriffo, D. W. Evans, M. Riello, R. Andrae, J. M. Carrasco, G. Busso, P. W. Burgess, C. Cacciari, M. Davidson, D. L. Harrison, S. T. Hodgkin, C. Jordi, P. J. Osborne, E. Pancino, G. Altavilla, M. A. Barstow, C. A. L. Bailer-Jones, M. Bellazzini, A. G. A. Brown, M. Castellani, S. Cowell, L. Delchambre, F. De Luise, C. Diener, C. Fabricius, M. Fouesneau, Y. Frémat, G. Gilmore, G. Giuffrida, N. C. Hambly, S. Hidalgo, G. Holland, Z. Kostrzewa-Rutkowska, F. van Leeuwen, A. Lobel, S. Marinoni, N. Miller, C. Pagani, L. Palaversa, A. M. Piersimoni, L. Pulone, S. Ragaini, M. Rainer, P. J. Richards, G. T. Rixon, D. Ruz-Mieres, N. Sanna, L. M. Sarro, N. Rowell, R. Sordo, N. A. Walton, and A. Yoldas. Gaia Data Release 3. Processing and validation of BP/RP low-resolution spectral data. A&A, 674:A2, June 2023. doi: 10.1051/0004-6361/202243680.DOI
  10. [10]G. M. De Silva, K. C. Freeman, J. Bland-Hawthorn, S. Martell, E. Wylie de Boer, M. Asplund, S. Keller, S. Sharma, D. B. Zucker, T. Zwitter, B. Anguiano, C. Bacigalupo, D. Bayliss, M. A. Beavis, M. Bergemann, S. Campbell, R. Cannon, D. Carollo, L. Casagrande, A. R. Casey, G. Da Costa, V. D’Orazi, A. Dotter, L. Duong, A. Heger, M. J. Ireland, P. R. Kafle, J. Kos, J. Lattanzio, G. F. Lewis, J. Lin, K. Lind, U. Munari, D. M. Nataf, S. O’Toole, Q. Parker, W. Reid, K. J. Schlesinger, A. Sheinis, J. D. Simpson, D. Stello, Y.-S. Ting, G. Traven, F. Watson, R. Wittenmyer, D. Yong, and M. Žerjal. The GALAH survey: scientific motivation. MNRAS, 449(3):2604–2617, May 2015. doi: 10.1093/mnras/stv327.
  11. [11]Ming-Yi Ding, Jian-Rong Shi, Yue Wu, Hugh R. A. Jones, Hong-Liang Yan, Chun-Qian Li, Qi Gao, Tian-Yi Chen, Jing-Hua Zhang, Shuai Liu, Tai-Sheng Yan, and Xiao-Jin Xie. Stellar Atmospheric Parameters of M-type Stars from LAMOST DR8. ApJS, 260(2):45, June 2022. doi: 10.3847/1538-4365/ac6754.DOI
  12. [12]Davide Eynard, Artiom Kovnatsky, Michael M. Bronstein, Klaus Glashoff, and Alexander M. Bronstein. Multimodal manifold analysis by simultaneous diagonalization of laplacians. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(12):2505–2517, 2015. doi: 10.1109/TPAMI.2015.2408348.DOI
  13. [13]Bernhard N. Flury. Common principal components in k groups. Journal of the American Statistical Association, 79(388):892–898, 1984. doi: 10.1080/01621459.1984.10477108. URL https://doi.org/10.1080/01621459.1984.10477108.
  14. [14]Gaia Collaboration, T. Prusti, J. H. J. de Bruijne, A. G. A. Brown, A. Vallenari, C. Babusiaux, C. A. L. Bailer-Jones, U. Bastian, M. Biermann, D. W. Evans, L. Eyer, F. Jansen, C. Jordi, S. A. Klioner, U. Lammers, L. Lindegren, X. Luri, F. Mignard, D. J. Milligan, C. Panem, V. Poinsignon, D. Pourbaix, S. Randich, G. Sarri, P. Sartoretti, H. I. Siddiqui, C. Soubiran, V. Valette, F. van Leeuwen, N. A. Walton, C. Aerts, F. Arenou, M. Cropper, R. Drimmel, E. Høg, D. Katz, M. G. Lattanzi, W. O’Mullane, E. K. Grebel, A. D. Holland, C. Huc, X. Passot, L. Bramante, C. Cacciari, J. Castañeda, L. Chaoul, N. Cheek, F. De Angeli, C. Fabricius, R. Guerra, J. Hernández, A. Jean-Antoine-Piccolo, E. Masana, R. Messineo, N. Mowlavi, K. Nienartowicz, D. Ordoñez-Blanco, P. Panuzzo, J. Portell, P. J. Richards, M. Riello, G. M. Seabroke, P. Tanga, F. Thévenin, J. Torra, S. G. Els, G. Gracia-Abril, G. Comoretto, M. Garcia-Reinaldos, T. Lock, E. Mercier, M. Altmann, R. Andrae, T. L. Astaatmadja, J. Bellas-Velidis, K. Benson, J. Berthier, R. Blomme, G. Busso, B. Carry, A. Cellino, G. Clementini, S. Cowell, O. Creevey, J. Cuypers, M. Davidson, J. De Ridder, A. de Torres, L. Delchambre, A. Dell’Oro, C. Ducourant, Y. Frémat, M. García-Torres, E. Gosset, J.-L. Halbwachs, N. C. Hambly, D. L. Harrison, M. Hauser, D. Hestroffer, S. T. Hodgkin, H. E. Huckle, A. Hutton, G. Jasniewicz, S. Jordan, M. Kontizas, A. J. Korn, A. C. Lanzafame, M. Manteiga, A. Moinho, K. Muinonen, J. Osinde, E. Pancino, T. Pauwels, J.-M. Petit, A. Recio-Blanco, A. C. Robin, L. M. Sarro, C. Siopis, M. Smith, K. W. Smith, A. Sozzetti, W. Thuillot, W. van Reeven, Y. Viala, U. Abbas, A. Abreu Aramburu, S. Accart, J. J. Aguado, P. M. Allan, W. Allasia, G. Altavilla, M. Álvarez, J. Alves, R. I. Anderson, A. H. Andrei, E. Anglada Varela, E. Antiche, T. Antoja, S. Antón, B. Arçay, A. Atzei, L. Ayache, N. Bach, S. G. Baker, L. Balaguer-Núñez, C. Barache, C. Barata, A. Barbier, F. Barblan, M. Baroni, D. Barrado y Navascués, M. Barros, M. A. Barstow, U. Becciani, M. Bellazzini, G. Bellei, A. Bello García, V. Belokurov, P. Bendjoya, A. Berihuete, L. Bianchi, O. Bienaymé, F. Billebaud, N. Blagorodnova, S. Blanco-Cuaresma, T. Boch, A. Bombrun, R. Borrachero, S. Bouquillon, G. Bourda, H. Bouy, A. Bragaglia, M. A. Bredells, N. Brouillet, T. Brüsemeister, B. Bucciarelli, F. Budnik, P. Burgess, R. Burgon, A. Burlacu, D. Busonero, R. Buzzi, E. Caffau, J. Cambras, H. Campbell, R. Cancelliere, T. Cantat-Gaudin, T. Carlucci, J. M. Carrasco, M. Castellani, P. Charlot, J. Charnas, P. Charvet, F. Chassat, A. Chiavassa, M. Clotet, G. Cocozza, R. S. Collins, P. Collins, and G. Costigan. The Gaia mission. A&A, 595:A1, November 2016. doi: 10.1051/0004-6361/201629272.
  15. [15]Gaia Collaboration, A. Vallenari, A. G. A. Brown, T. Prusti, J. H. J. de Bruijne, F. Arenou, C. Babusiaux, M. Biermann, O. L. Creevey, C. Ducourant, D. W. Evans, L. Eyer, R. Guerra, A. Hutton, C. Jordi, S. A. Klioner, U. L. Lammers, L. Lindegren, X. Luri, F. Mignard, C. Panem, D. Pourbaix, S. Randich, P. Sartoretti, C. Soubiran, P. Tanga, N. A. Walton, C. A. L. Bailer-Jones, U. Bastian, R. Drimmel, F. Jansen, D. Katz, M. G. Lattanzi, F. van Leeuwen, J. Bakker, C. Cacciari, J. Castañeda, F. De Angeli, C. Fabricius, M. Fouesneau, Y. Frémat, L. Galluccio, A. Guerrier, U. Heiter, E. Masana, R. Messineo, N. Mowlavi, C. Nicolas, K. Nienartowicz, F. Pailler, P. Panuzzo, F. Riclet, W. Roux, G. M. Seabroke, R. Sordo, F. Thévenin, G. Gracia-Abril, J. Portell, D. Teyssier, M. Altmann, R. Andrae, M. Audard, I. Bellas-Velidis, K. Benson, J. Berthier, R. Blomme, P. W. Burgess, D. Busonero, G. Busso, H. Cánovas, B. Carry, A. Cellino, N. Cheek, G. Clementini, Y. Damerasji, M. Davidson, P. de Teodoro, M. Nuñez Campos, L. Delchambre, A. Dell’Orо, P. Esquej, J. Fernández-Hernández, E. Fraile, D. Garabato, P. García-Lario, E. Gosset, R. Haighron, J.-L. Halbwachs, N. C. Hambly, D. L. Harrison, J. Hernández, D. Hestroffer, S. T. Hodgkin, B. Holl, K. Janßen, G. Jevardat de Fombelle, S. Jordan, A. Krone-Martins, A. C. Lanzafame, W. Löffler, O. Marchal, P. M. Marrese, A. Moitinho, K. Muinonen, P. Osborne, E. Pancino, T. Pauwels, A. Recio-Blanco, C. Reylé, M. Riello, L. Rimoldini, T. Roegiers, J. Rybizki, L. M. Sarro, C. Siopis, M. Smith, A. Sozetti, E. Utrilla, M. van Leeuwen, U. Abbas, P. Ábrahám, A. Abreu Aramburu, C. Aerts, J. J. Aguado, M. Ajaj, F. Aldea-Montero, G. Altavilla, M. A. Álvarez, J. Alves, F. Anders, R. I. Anderson, E. Anglada Varela, T. Antoja, D. Baines, S. G. Baker, L. Balaguer-Núñez, E. Balbinot, Z. Balog, C. Barache, D. Barbato, M. Barros, M. A. Barstow, S. Bartolomé, J.-L. Bassilana, N. Bauchet, U. Becciani, M. Bellazzini, A. Berihuete, M. Bernet, S. Bertone, L. Bianchi, A. Binnenfeld, S. Blanco-Cuaresma, A. Blazere, T. Boch, A. Bombrun, D. Bossini, S. Bouquillon, A. Bragaglia, L. Bramante, E. Breedt, A. Bressan, N. Brouillet, E. Brugaletta, B. Bucciarelli, A. Burlacu, A. G. Butkevich, R. Buzzi, E. Caffau, R. Cancelliere, T. Cantat-Gaudin, R. Carballo, T. Carlucci, M. I. Carnerero, J. M. Carrasco, L. Casamiquela, M. Castellani, A. Castro-Ginard, L. Chaoul, P. Charlot, L. Chemin, V. Chiaramida, A. Chiavassa, N. Chornay, G. Comoretto, G. Contursi, W. J. Cooper, T. Cornez, S. Cowell, F. Crifo, M. Cropper, M. Crosta, C. Crowley, C. Dafonte, A. Daperoglas, M. David, P. David, P. de Laverny, F. De Luise, and R. De March. Gaia Data Release 3. Summary of the content and survey properties. A&A, 674:A1, June 2023. doi: 10.1051/0004-6361/202243940.
  16. [16]Jihun Ham, Daniel D. Lee, and Lawrence K. Saul. Semisupervised alignment of manifolds. In International Conference on Artificial Intelligence and Statistics, 2005. URL https://api.semanticscholar.org/CorpusID:2913820.
  17. [17]Chelsea X. Huang, Andrew Vanderburg, Andras Pál, Lızhou Sha, Liang Yu, Willie Fong, Michael Fausnaugh, Avi Shporer, Natalia Guerrero, Roland Vanderspek, and George Ricker. Photometry of 10 million Stars from the First Two Years of TESS Full Frame Im- ages: Part II. Research Notes of the American Astronomical Society, 4(11):206, November 2020. doi: 10.3847/2515-5172/abca2d.
  18. [18]Dean R. H. Johnson and David R. Soderblom. Calculating Galactic Space Velocities and Their Uncertainties, with an Application to the Ursa Major Group. AJ, 93:864, April 1987. doi: 10.1086/114370.
  19. [19]Vassilis Kalofolias, Xavier Bresson, Michael M. Bronstein, and Pierre Vandergheynst. Matrix completion on graphs. ArXiv, abs/1408.1717, 2014. URL https://api.semanticscholar.org/CorpusID:14746688.arxiv.org/abs/1408.1717
  20. [20]Dhruva Karkada, Daniel J. Korchinski, Andres Nava, Matthieu Wyart, and Yasaman Bahri. Symmetry in language statistics shapes the geometry of model representations. arXiv e-prints, arXiv:2602.15029, February 2026. doi: 10.48550/arXiv.2602.15029.DOI
  21. [21]Arto Klami, Seppo Virtanen, and Samuel Kaski. Bayesian canonical correlation analysis. J. Mach. Learn. Res., 14(1):965–1003, April 2013. ISSN 1532-4435.DOI
  22. [22]S. Lafon, Y. Keller, and R.R. Coifman. Data fusion and multicue data matching by diffusion maps. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(11):1784–1797, 2006. doi: 10.1109/TPAMI.2006.223.DOI
  23. [23]N. R. Lomb. Least-Squares Frequency Analysis of Unequally Spaced Data. Ap&SS, 39(2):447–462, February 1976. doi: 10.1007/BF00648343.DOI
  24. [24]Steven R. Majewski, Ricardo P. Schiavon, Peter M. Frinchaboy, Carlos Allende Prieto, Robert Barkhouser, Dmitry Bizyaev, Basil Blank, Sophia Brunner, Adam Burton, Ricardo Carrera, S. Drew Chojnowski, Kátia Cunha, Courtney Epstein, Greg Fitzgerald, Ana E. García Pérez, Fred R. Hearty, Chuck Henderson, Jon A. Holtzman, Jennifer A. Johnson, Charles R. Lam, James E. Lawler, Paul M. Masana, Szabolcs Mészáros, Matthew Nelson, Duy Cuong Nguyen, David L. Nidever, Marc Pinsonneault, Matthew Shetrone, Stephen Smee, Verne V. Smith, Todd Stolberg, Michael F. Skrutskie, Eric Walker, John C. Wilson, Gail Zasowski, Friedrich Anders, Sarbani Basu, Stephane Beland, Michael R. Blanton, Jo Bovy, Joel R. Brownstein, Joleen Carlberg, William Chaplin, Cristina Chiappini, Daniel J. Eisenstein, Yvonne Elsworth, Diane Feuillet, Scott W. Fleming, Jessica Galbraith-Frew, Rafael A. García, D. Aníbal García-Hernández, Bruce A. Gillespie, Léo Girardi, James E. Gunn, Sten Hasselquist, Michael R. Hayden, Saskia Hekker, Inese Ivans, Karen Kinemuchi, Mark Klaene, Suvrath Mahadevan, Savita Mathur, Benoît Mosser, Demitri Muna, Jeffrey A. Munn, Robert C. Nichol, Robert W. O’Connell, John K. Parejko, A. C. Robin, Helio Rocha-Pinto, Matthias Schultheis, Aldo M. Serenelli, Neville Shane, Victor Silva Aguirre, Jennifer S. Sobeck, Benjamin Thompson, Nicholas W. Troup, David H. Weinberg, and Olga Zamora. The Apache Point Observatory Galactic Evolution Experiment (APOGEE). AJ, 154(3):94, September 2017. doi: 10.3847/1538-3881/aa784d.DOI
  25. [25]A. McQuillan, S. Aigrain, and T. Mazeh. Measuring the rotation period distribution of field M dwarfs with Kepler. MNRAS, 432(2):1203–1216, June 2013. doi: 10.1093/mnras/stt536.DOI

Paper details

vlnc:W_4syy846x
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents