WorksComputer Science
Spin Spherical Fourier Neural Operators: Efficient Equivariant Learning for Vector and Spin Fields on the Sphere
Abstract
Many fields of physical interest on the sphere, such as wind, the polarisation of the cosmic microwave background and the shear of weak gravitational lensing, are spin fields, whose values change phase when the local frame rotates. We extend spherical Fourier neural operators (SFNOs) to fields of any spin, through lifted spherical convolutions that are exactly rotationally equivariant. Lifting spin fields to Mackey functions on the rotation group provides the conceptual framework, while computation is performed exactly and efficiently through their harmonic representation and by appealing to sampling theorems on the sphere and rotation group. We construct spin SFNOs (SFNOs) from these building blocks. We demonstrate empirically that rotational equivariance holds to machine precision for these layers. We achieve state-of-the-art performance on existing benchmarks, and introduce new benchmarks on which spin-aware models outperform their scalar counterparts.
Introduction
In many fields of science data are measured on the sphere, from the relic radiation of the Big Bang and the shapes of distant galaxies in cosmology, to the winds of the atmosphere and the currents of the ocean. Many of these data sets are not scalar. The horizontal wind and the ocean currents are vector, or spin-1, fields, while the polarisation of the cosmic microwave background (CMB) and the shear induced by weak gravitational lensing are spin-2 fields [58, 7, 48]. The value of a spin- field at a point is defined relative to a frame in the tangent plane and acquires the phase when that frame rotates by .
On the sphere, rotations play the role that translations play in the plane. Spherical CNNs consequently build on SO(3)-equivariant convolutions [9, 13, 24, 8, 40], and spherical Fourier neural operators (SFNOs) [5] bring the operator-learning perspective of Fourier neural operators [25] (Li et al. 2021) to the sphere. When applied to vector or spin fields, however, scalar spherical CNNs and SFNOs typically consider the components of spin fields (i.e. real and imaginary components) as unrelated scalar channels. This is the standard approach in data-driven weather models for wind [41, 26, 15] (Bi et al. 2023), which does not respect general rotations.
Equivariant networks for spin fields exist [14, 2] (Xu et al. 2022), but at a price. Those that convolve on the full rotation group are costly in computation and parameters. The group CNN of Ballerin et al. [2], for example, has about 2 million parameters on spherical MNIST (whereas we achieve similar performance with 50 thousand parameters). The equivariance of these networks, moreover, is asserted or illustrated rather than quantified, and benchmarks for spin fields are scarce. Our layers build on McEwen [33], who makes the lift–convolve–project construction of Cohen et al. (2019, Eq. (14)) and Xu et al. (2022) explicit on the sphere and shows that it unifies the spherical convolutions of the literature, so that our layers can clean filters of each of these forms.
Our contributions are fourfold. First, we extend spherical Fourier neural operators to fields of any spin (§ 4), building spin SFNOs (FNOs) from two blocks, the spin spherical convolution and the convolution on the rotation group, so that rotational equivariance is exact for every convolutional layer and holds in the continuum for every activation. Second, we describe the layers by lifting spin fields to Mackey functions on the rotation group, which provides their conceptual framework and proofs [33], but compute them in harmonic space to ensure FNOs are efficient in representation, computation and parameters (§ 5). Third, we show that the convolutional layers of FNOs are exactly equivariant on spin fields, with errors of order , that whole FNO blocks and UNets with the activations and the normalisation removed are equivariant to round-off, and we discuss the equivariance errors introduced by the other layers (§ 6). Fourth, we introduce three new public benchmarks for spin fields spanning classification, weather and cosmology (§ 7), on which our networks achieve state-of-the-art performance. Our layers, networks and benchmarks will be released with the s2ai code (https://github.com/astro-informatics/s2ai).
Background and related work
Spin fields and their harmonic analysis
A spin- field is a complex function on the sphere whose values are defined relative to a local tangent frame and transform as
under a right-handed rotation of the frame by ([39, 16]). We write , with , for its spin spherical harmonic coefficients and for the Wigner coefficients of , with the Wigner functions in Euler angles and the unnormalised Haar measure. The coefficients of every spin rotate by the same Wigner matrices under the spin rotation operator of McEwen et al. [36] (Appendix § 11.§ 1). Fields are bandlimited at if their coefficients vanish for , and functions on azimuthally bandlimited at if also for . By appealing to the sampling theorems of McEwen & Wiaux [34] and McEwen et al. [35], the s2fft code (Price & McEwen, 2024) transforms such signals exactly, at for spin spherical harmonic transforms and for Wigner transforms.
Neural operators and equivariant networks on the sphere
The spherical Fourier neural operator (SFNO) of Bonev et al. [5] learns one weight per spherical harmonic degree in the block structure of the Fourier neural operator (FNO) of Li et al. (2021). Like the FNO, it has a continuous representation through harmonic space, since the harmonic coefficients of bandlimited signals, computed exactly by appealing to sampling theorems, give access to the underlying continuous operator, as in the continuous approaches to spherical convolution described by Ocampo et al. [40]. Its layer is the non-localised zonal filter of Esteves et al. [13], the isotropic scalar member of the generalised convolutions of Cobb et al. [8]. Spherical CNNs that learn underlying continuous representations ([13, 8]) are thus SFNOs, neural operators in the sense of Kovachki et al. [25], although they did not carry the name (Appendix B). This point is also made for the discrete-continuous (DISCO) convolutions of Ocampo et al. [40], whose filters are continuous, in recent work that adopts them as local neural-operator layers ([32, 6]).
On the rotation group, spherical CNNs apply pointwise non-linearities (Cohen et al., [9]) or exact tensor products (Kondor et al., [24]; Cobb et al. [8] combine both and develop more efficient tensor-product activations. For spin fields, Esteves et al. [14] learn spin-changing harmonic filters, with non-linearities on each field, while Xu et al. (2022) lift to the rotation group only for a pointwise non-linearity before a fixed projection, and Ballerin et al. [2] apply full group convolutions and pointwise non-linearities. In § 4 to § 7 we combine these ingredients in a neural operator for spin fields with the block structure of the SFNO, form the lift only where it is needed, for directional filters and the non-linear coupling of spins, measure equivariance on spin fields, and introduce benchmarks that include spin 2, whereas the experiments of the networks above use spins of at most 1.
Equivariant spin convolutions by lifting
Our layers rest on the generalised lifted convolution of McEwen [33], which we recall concisely here, with the results that we use stated in Appendix § 11.
Lifting spin fields to the rotation group
Spin- fields are the sections of the line bundle associated by the character to the rotation group, the principal -bundle of frames over the sphere ([10, 52, 33]), and correspond to the Mackey functions of type on the rotation group, those with . A spin- field is lifted to its Mackey function by and returned by its adjoint, the projection , which integrates over the fibre against :
The lift is an isometry with , occupies the single Wigner column , and is equivariant, intertwining the regular representation with the spin rotation . Conversely, a function on is azimuthally bandlimited at if and only if it is a sum of at most lifted spin fields, one for each .
Lifted convolutions in harmonic space
We adopt the group convolution, which for reads:
The generalised lifted convolution lifts a spin field to the rotation group, convolves it there with a kernel and projects the result back to a spin field on the sphere. It resolves to four cases depending on whether a lift is applied to the input and a projection to the output:
All four are equivariant for every square-integrable kernel, and by the single-column lift only one kernel entry per degree reaches a spin- output [33],
so that the spin spherical convolution maps spin to spin by one multiplier per degree, without forming a signal on the rotation group. Specialising the equivariant kernel theorems [23, 24, 10] (Xu et al. 2022) to the sphere shows that these multipliers are complete, in the sense that every bounded equivariant complex-linear map from spin- to spin- fields acts as
for a unique bounded sequence , and that every such sequence defines one [33].
A directional layer whose kernel column is supported on is exactly a tuple of at most spin-changing layers, which relates the directional layers of Cobb et al. [8] to the multi-spin layers of Esteves et al. [14].
A key result of McEwen [33] is that directional kernels, a spherical output of one spin and rotational equivariance cannot be achieved simultaneously. Between two fixed spins the kernel of an equivariant map is radial apart from the frame phases, so that a layer whose output is a field of one spin cannot be both directional and equivariant. Directional information must therefore be kept on the rotation group or, equivalently, in a tuple of spin fields.
Since the generalised lifted convolution reduces to the spherical convolutions of the literature, each recovered by the support of its kernel, the choice of spins and whether the lift or the projection is omitted [33], layers built on it can learn filters of each of these forms rather than fix one by design. The spin spherical convolution contains every bounded equivariant complex-linear map between two spins, among them the axisymmetric convolutions of scalar spherical CNNs, the left convolution of Driscoll & Healy [12], the spin-weighted convolutions of Esteves et al. [14] and the spherical Kaiser–Squires relation of weak lensing [20, 48], and the convolution on the rotation group contains the directional, steerable and group convolutions [53, 36, 9, 8]. In our networks the filters are restricted by their parameterisation, through interpolated profiles in , and on the rotation group, a small internal spin set and a kernel diagonal on it.
Spin spherical Fourier neural operators
We build SFNOs from two blocks with exactly equivariant linear layers, joined by the lift and its adjoint .
Convolutional building blocks
The spin spherical convolution (Figure 1a) maps a tuple of spin fields to another by one multiplier (7) per degree for each pair of spins and channels, without explicitly forming the lift. It provides complete linear spin mixing (§ 3.2) but no non-linear coupling of spins, since its non-linearity acts on each field separately.
Figure 1. The spin spherical convolution (a), one multiplier per degree between spin spherical harmonic transforms (SHTs) without forming the lift, and the convolution on the rotation group (b), each followed by and computed in harmonic space below. Spheres denote fields on , spheres with a cone denote functions lifted to SO(3), triangles denote spin spherical harmonic coefficients and prisms Wigner coefficients, and indexes output spins in (a) and kernels in (b).
The convolution on the rotation group (Figure 1b) acts on the Wigner coefficients of a function on by (5), and its non-linearity acts pointwise on samples on the group, with samples in . In our networks its kernel is diagonal on an internal spin set , so that the linear part filters each spin separately and the spins are coupled by the non-linearity on the group, which is much less constrained than one on the sphere (§ 4.2). The projection then reads the Wigner column of each spin , and our implementation computes only these columns, at rather than the of full Wigner transforms (§ 5.2).
Equivariant non-linearities
Activations for spin fields may be applied in two settings, to the spin fields themselves on the sphere or pointwise to their lift on the rotation group; each setting brings its own constraint for equivariance. On the sphere a rotation turns the local frame and multiplies a spin- value by the phase of (1), so that for a pointwise activation is equivariant only if it commutes with that phase, , and may therefore act on the modulus alone, as magnitude, norm and gated non-linearities do [54, 51, 14]. On SO(3) rotations act on the argument alone, so that every pointwise map is equivariant [9, 8, 2], with a much weaker constraint. A tuple of spin fields may therefore be lifted and summed, passed through any pointwise map and returned (Appendix C).
On the sphere, for non-zero spin fields, we adopt the modulus non-linearity
with a learned real gain and offset , and , which keeps the phase and is therefore equivariant for every spin. For scalar fields on the sphere () we adopt standard activation functions. On SO(3) we apply a pointwise activation function to samples of the lifted tuple, such as the modulus non-linearity, and return to the sphere by the fixed projection followed by a learned filter for each spin.
SFNO blocks
The SFNO block (Figure 2), acting on a tuple of spin fields,
Figure 2. The SFNO block (10), with C the spin spherical convolution or the convolution on the rotation group, and harmonic convolutions and the activation (in detail in Figure 10).
with C the convolution of either building block, follows the SFNO block of Bonev et al. [5] in summing a harmonic-space convolution and a pointwise branch, but replaces the MLP that follows the sum by and has no identity residual connection. The operators , in parallel with the convolution, and , after the sum, are harmonic convolutions. Each transforms its input to harmonic space, truncates or pads the coefficients to the output bandlimit, mixes the channels with one matrix and the spins with one matrix , both constant in , and transforms back; on SO(3) it reads the columns , mixes them and lifts again. Between equal spins this is exactly the pointwise convolution of the bypass of the FNO. Between different spins it keeps the harmonic coefficients and changes their spin, which in pixel space is a normalised power of the spin raising or lowering operator [33]; since this is not local, it is computed in harmonic space. The role of the harmonic convolutions is to enrich the channel and spin mixing of the block.
Networks
We compare two constructions, each with Conv blocks, a convolution and its non-linearity, or with SFNO blocks, and name the four models by block and construction, Conv-S, Conv-SO(3), SFNO-S and SFNO-SO(3). The spherical construction never lifts explicitly, whereas the rotation-group construction lifts after a first spin spherical convolution, so that its filters act on the rotation group. We construct classifier and dense prediction networks based on these ararchitectural building blocks. The classifiers stack six blocks and pool rotation-invariant features, and we construct UNets with encoders and decoders built from the same blocks, with skip connections that concatenate features at equal bandlimit and spin set (Figure 11). The blocks include a normalisation, a group normalisation in the UNets and a batch normalisation in the classifiers, and the spin UNets for wind and weak lensing are Conv-SO(3) networks (Appendix § 14).
Figure 11. Composition of blocks into the UNets used for dense prediction. Every box is a block of a lifting convolution (LConv), the lift-and-sum (LSum), the activation (CAct) and group normalisation (GNorm). The first block maps the data spin onto the internal spin set and the last layer returns the output spin. Brackets give the five-level spin UNet of the mass-mapping benchmark (bandlimit; channels), whereas the MNIST UNets are shallower and the FNO UNets keep at every level.
Efficient representation, computation and parameterisation
Mackey functions and the lift give the conceptual description of the layers and their proofs (McEwen, [33]), whereas the harmonic representation provides exact and efficient computation. We form the lift explicitly only where it provides expressivity.
Representation and spin mixing
In harmonic space the lift need not be formed. The lift of a spin- field occupies a single Wigner column and so is held in the spin spherical harmonic coefficients of the field rather than the of a general function on SO(3), and a tuple of fields with spins in in coefficients. Only an explicit lift, its synthesis at samples in for a pointwise non-linearity, requires .
Linear spin mixing needs no lift. Every bounded equivariant complex-linear map between two spins is a spin spherical convolution, and a tuple of spin fields, one for each spin , is exactly a function on SO(3) of azimuthal bandlimit (McEwen et al., [35]). A group convolution mapping between such functions is therefore an array of spin spherical convolutions, one for each pair of spins, so that, as classes of linear maps, the spherical construction realises every convolution on SO(3) at that azimuthal bandlimit without forming the lift. The convolution on the rotation group in our networks uses only the diagonal of this array, so that its spins are coupled by the non-linearity alone, and the rotation-group construction mixes spins linearly in its first layer. A pointwise complex-linear map cannot mix spins equivariantly, since under a frame rotation (1) fields of different spins acquire different phases, so that an equivariant convolution on the samples mixes channels only within a spin and the pointwise bypass of the FNO, extended to spin fields, must mix spins in harmonic space.
General non-linear spin coupling is what the explicit lift achieves. On the sphere it needs maps whose outputs have a definite spin, such as the modulus , which takes every spin to spin and is the phase collapse of spherical scattering networks (McEwen et al., [37]; Mousset et al. [38]), or products of fields, whose spins add. The modulus non-linearity (9) keeps the phase, and hence the spin, of each field, so that the non-linearities of our spherical construction couple no spins. On SO(3) every pointwise map is equivariant, and one non-linearity after the lift couples all spins at once. In the rotation-group construction spins are thus coupled non-linearly through the lift.
Representing a vector or spin field as a spin field, rather than as a scalar channel, also matters under reflections of the sphere. Networks on scalar fields built from isotropic multipliers and pointwise maps, without positional inputs, are O(3)-equivariant (Weiler et al., [52]) and cannot represent maps that anticommute with the antipodal map , such as the advection of a vorticity treated as a scalar. Spin fields escape this restriction, since reverses the orientation of the tangent frame and so acts on them through conjugation as well as pull-back (Appendix § 13.§ 4). A multiplier between spins commutes with if and only if its coefficients are real, so that with real profiles our classifiers would be exactly invariant under reflections, and they therefore use complex profiles. Real profiles, whose multipliers commute with , could instead be useful for fields with antipodal symmetry, such as those of diffusion magnetic resonance imaging (Descoteaux et al., [11]).
Computation
We leverage sampling theorems and fast harmonic transforms on and SO(3) (McEwen & Wiaux, [34]; McEwen et al. [35]; Price & McEwen (2024)). Since the sampling theorems make the transforms of bandlimited spin fields and functions on SO(3) exact, every linear layer acting on bandlimited input is computed exactly, consistent with the machine-precision equivariance errors of § 6. A spin spherical convolution costs for each field, and so for a tuple, for a convolution on SO(3) at azimuthal bandlimit , whose output fills all spins, and for a full convolution on SO(3), with . These are upper bounds for the algorithms in harmonic space (McEwen, [33]). Against convolutions that fill every spin, the saving of the spherical construction lies in a spin set smaller than and in storing no samples on the rotation group, which the rotation-group construction samples only for its non-linearity. This non-linearity, which couples all spins, costs with full Wigner transforms and in our implementation, which computes only the Wigner columns that are read.
Parameterisation
We parameterise every linear layer in harmonic space. Our multipliers are interpolated in from ten knots placed at fixed fractions of the bandlimit, so that, as with the anchor points of Cobb et al. [8], filters at high resolution need few parameters. The knots reduce the multipliers of a spin spherical convolution for each pair of spins and channels to ten, against the values of a kernel column of azimuthal bandlimit and the of a full kernel on SO(3) [33]. In the UNets the profiles are real, with a channel matrix for each output spin, and in the classifiers complex (Appendix D). Each harmonic convolution has one channel matrix and one spin matrix, and the activations learn an offset for each channel and, in the UNets, a gain as well. Parameters therefore attach, through the knots, to fractions of the bandlimit rather than to samples, so that their number depends neither on the bandlimit nor on the sampling grid, and a trained layer applies on any grid that supports exact transforms at its bandlimit. Our layers are thus neural operators in the sense of Kovachki et al. [25].
Our layers are built on the s2fft code (Price & McEwen, 2024) and will be released with the s2ai code.
Equivariance
The convolutional layers of SFNOs are exactly equivariant to rotations of spin fields, which we confirm numerically, and we discuss the equivariance errors introduced by the other layers, such as the activations and the normalisation. Equivariance errors have been measured before for spherical CNNs on scalar signals [9, 13, 8, 40], but not for vector and spin fields on the sphere under general rotations.
Sources of equivariance error
As for spherical CNNs on scalar signals, convolutional layers are computed exactly in harmonic space. Any error therefore enters only through the non-linear pointwise activations and the normalisation. The output of an activation is not bandlimited, so that analysing its samples on the same grid aliases its content above the bandlimit, and on SO(3) beyond the azimuthal bandlimit. Oversampling reduces this aliasing [8]. The group normalisation of the UNets aliases for every spin, because it rescales the modulus by , up to constants for each group (Appendix D.1), and on the sphere it adds, for non-zero spin, a structural error, since the mean that it subtracts is not rotation invariant, of the kind that pointwise spin mixing would cause (Sec. 5.1). A normalisation with statistics computed in harmonic space and no mean subtracted for would be exactly equivariant, which we leave to future work.
Equivariance tests
We measure the mean relative equivariance error (MREE), a quadrature-weighted relative error, over 100 random bandlimited signals and 100 random rotations at in double precision (Appendix E). Linear layers attain errors
of order , the activations alias with errors of order 0.1 to 1, apart from the complex ReLU, measured at initialisation where it is the identity up to , and the group normalisation gives errors of order on the sphere and on the rotation group (Table 1). With the activations and the normalisation removed, whole SFNO blocks and UNets, with spin and channel mixing, are equivariant to the round-off floor of the transforms, of order at the larger bandlimit (Tables 1 and 7), which confirms that the errors of our blocks and networks arise only from their activations and normalisation.
Spin benchmarks and experiments
In this section we first introduce three new benchmarks for spin fields, spin spherical MNIST, global wind forecasting and weak gravitational lensing, and then compare our networks on each of them with scalar counterparts and published results (Appendix G).
New benchmarks
Public benchmarks for learning on spin fields are lacking. The benchmarks of spherical CNNs, such as spherical MNIST [9], are scalar, and the vector-field MNIST of Esteves et al. [14], built from the Sobel gradients of the images and also used by Xu et al. (2022), is limited to spin 1. We introduce three new benchmarks, a contribution of this article in their own right, whose inputs and targets have a definite spin, of any spin for spin spherical MNIST, and which span classification and regression on synthetic, atmospheric and cosmological fields. Their generation code will be released with s2ai, and Appendix § 16 describes them in full.
Spin spherical MNIST. MNIST digits ([28]) are projected onto the sphere [9] at . The Sobel-gradient construction of Esteves et al. [14], the only one available so far, gives a spin-1 field only. We instead synthesise spin fields of any spin from the same harmonic coefficients in the spin- basis, so that rotations act exactly on the bandlimited fields, and use the harmonic spin-1 and spin-2 fields alongside the Sobel spin-1 field. We consider the classification of the digit from one of these fields, and dense prediction, which maps the spin-0 image to its spin-1 field or back and so changes spin. Both follow the NR/R protocol of training on unrotated and testing on randomly rotated data ([9, 8, 40]), which tests the invariance of a classifier and the equivariance of a dense prediction to rotations not seen in training (Appendix § 16.§ 1).
Global wind forecasting. The task is to predict the horizontal wind at 50 hPa at 18:00 UTC on a given day from the wind at 00:00, 06:00 and 12:00 UTC of the same day, with the wind represented by the spin-1 field of its zonal and meridional components at . The data are the ERA5 reanalysis ([19]) of WeatherBench 2 ([44]) from 1980 to 2022, from which eight days drawn at random are held out in every year, so that the benchmark tests interpolation within the same years rather than generalisation to later years (Appendix § 16.§ 2).
Weak gravitational lensing. The task is to reconstruct the spin-0 convergence on the full sky from the noisy spin-2 shear . Matter along the line of sight, most of it dark matter, distorts the images of distant galaxies by the shear, which on the full sky is a spin multiplier of the convergence, a projection of the matter density ([3, 21]). The spherical Kaiser–Squires estimator inverts this multiplier as a spin spin spherical convolution ([20, 48]), but practical considerations make the recovery of ill-posed and call for priors learned from simulations, which on the full sky have so far taken the two shear components as scalar channels [49]. We simulate 10,000 pairs of convergence and noisy shear maps for training and validation and 600 for evaluation at with GLASS ([46]) and the shape noise of a Stage IV survey ([27]) (Appendix § 16.§ 3).
Spin spherical MNIST
We train four models as classifiers and as UNets for dense prediction (Appendix § 14). The classifiers are selected on a non-rotated validation split and the prediction UNets on the rotated held-out set, which is also their evaluation set. The Conv- classifier on the harmonic spin-1 field is trained with three seeds, whose mean accuracy we report, and the architecture of the classifiers was developed with reference to their test accuracies (Appendix § 17.§ 1).
The four classifiers generalise to rotations not seen in training (Table 2). On the Sobel field our best classifier comes within 0.15% of the published accuracy of the general spherical CNN (GsCNN) of Ballerin et al. [2], and on the harmonic spin-1 and spin-2 fields, for which no other results are published, all four exceed 99%. The classifiers use complex harmonic filters, since with real profiles they would be exactly invariant under reflections of the sphere (§ 5.1).
| Model | Params | Sobel spin 1 | Harmonic spin 1 | Harmonic spin 2 |
| Conv- | 50k | 99.15 | 99.19 | 99.17 |
| FNO- | 182k | 98.82 | 99.16 | 99.34 |
| Conv-SO(3) | 134k | 99.01 | 99.10 | 99.11 |
| FNO-SO(3) | 136k | 99.17 | 99.21 | 99.28 |
| Esteves et al. | 58k | 98.2 | – | – |
| Ballerin et al., SCNN | 2.2M | 99.20 | – | – |
| Ballerin et al., CsCNN | 2.0M | 99.32 | – | – |
Table 2. Spin spherical MNIST classification: NR/R accuracy (%) of our spherical () and rotation-group (SO(3)) networks (best accuracy and fewest parameters in bold).
In dense prediction, a spin-changing task, each rotation-group UNet has fewer parameters and a lower error than its spherical counterpart in all four mappings (Table 3 and Figure 3). Moreover, the Conv-SO(3) UNet gives the lowest errors of the models compared from images to fields, and the FNO-SO(3) UNet from fields to images. The published values, for another construction and given for context only, lie above those of the Conv-SO(3) UNet, which has about eight times fewer parameters than that of Esteves et al. [14].
| Model | Params | Sobel field 0 → 1 | Sobel field 1 → 0 | Harmonic field 0 → 1 | Harmonic field 1 → 0 |
| Conv-S | 23k | 3.681 | 3.212 | 5.580 | 3.229 |
| SFNO-S | 29k | 4.253 | 4.544 | 3.944 | 4.468 |
| Conv-SO(3) | 14k | 3.016 | 3.144 | 2.876 | 3.101 |
| SFNO-SO(3) | 20k | 3.866 | 2.948 | 3.063 | 3.005 |
| Esteves et al. | ≈112k | 4.3 | 4.0 | – | – |
| Xu et al. | – | 3.8 | – | – | – |
Table 3. Spin-changing prediction on spin spherical MNIST: NR/R MSE from the spin-0 image to its spin-1 field (0 → 1) and back (1 → 0), with the best model and the fewest parameters in bold.
Figure 3. Prediction of the spin-1 Sobel field from the spin-0 image of a rotated test digit, chosen in advance, by the Conv-SO(3) and Conv-S UNets, with their relative errors. Both predictions follow the rotated frame of the target (more in Figure 13).
Global wind forecasting
We compare a spin UNet, a Conv-SO(3) network that takes the spin field as input, with a scalar spherical UNet that takes its real and imaginary parts as scalar channels, and with persistence, which forecasts the 12:00 UTC field for 18:00 UTC (Appendix § 14.§ 3). Checkpoints are selected on the held-out days, which are also the evaluation set. The spin UNet reduces the MSE by 44.9% with 34.9% fewer parameters and improves every other metric (Table 4 and Figure 4. It is the only model that improves on persistence in every metric, whereas the scalar UNet is worse than persistence in every metric except the correlation. These results support the practical value of the spin representation.
| Model | Params | MSE ↓ | SNR ↑ (dB) | PSNR ↑ (dB) | PCC ↑ |
| Global wind forecasting (MSE in m s) | |||||
| Spin UNet | 48k | 9.52 | 5.25 | 25.06 | 0.979 |
| Scalar UNet | 74k | 17.27 | 11.43 | 22.64 | 0.969 |
| Persistence | – | 14.55 | 12.17 | 23.19 | 0.962 |
| Weak gravitational lensing (MSE ×10) | |||||
| Spin UNet | 81k | 1.28 | 7.53 | 23.47 | 0.907 |
| Scalar UNet | 93k | 2.06 | 5.45 | 21.40 | 0.846 |
Table 4. Wind forecasting at 50 hPa on the held-out days and mass mapping from the noisy spin-2 shear of weak lensing at , averaged over 600 maps (best results and fewest parameters in bold; all metrics in Table 8 and Table 9). All models are spherical, and the scalar UNet takes the spin field as two scalar channels. Persistence forecasts the 12:00 UTC field for 18:00 UTC.
Figure 4. Wind forecast for one held-out day: (a) the target speed, (b–d) the vector errors of the spin and scalar UNets and persistence, and (e–h) the same in the outlined region, with true (white) and predicted (coloured) vectors. The scalar UNet leaves larger, coherent patches of error (more in Figure 14).
Figure 14. Wind forecast for one held-out day at : the speed of the inputs, the target and the forecasts (a–f), the vector error of the spin UNet, the scalar UNet and persistence with the area-weighted MSE of this day (h–j), and a zoom into the region outlined in (g), with true (white) and predicted (coloured) vectors and the RMSE over the region (k–n).
Weak gravitational lensing
Table 4 and Figure 5 compare a spin UNet, which lifts the spin-2 shear and returns the spin-0 convergence, with a scalar UNet of DISCO convolutions (Ocampo et al., [40]) and residual blocks (He et al., 2016) that takes and as scalar channels, both five-level UNets (Ronneberger et al., [45]) (Appendix § 14.§ 3). The spin UNet reduces the MSE by 37.9% with 13.4% fewer parameters and improves every other metric.
Figure 5. Mass mapping for one evaluation map at : (a) PDFs of the true and reconstructed convergence, and residuals of (b) the spin and (c) the scalar UNet. The scalar UNet misses the tails and leaves large-scale structure in its residual, whereas the spin residual is noise-like (maps in Figure 15).
Conclusions
We have developed SFNOs (spin spherical Fourier neural operators), which extend spherical Fourier neural operators to fields of any spin, with rotational equivariance exact for every convolutional layer and holding in the continuum for every activation. Described by Mackey functions but computed in harmonic space, SFNOs are efficient in representation, computation and parameters, with far fewer parameters than networks that convolve spin fields on the full rotation group. On spin fields their convolutional layers
are equivariant to , and whole blocks and UNets without activations or normalisation to round-off. On three new benchmarks our networks achieve state-of-the-art results.
Impact Statement
This paper presents work whose goal is to advance the field of machine learning, and in particular the modelling of geophysical and astrophysical fields on the sphere. Improved treatment of spin fields such as wind and polarisation may benefit weather, climate and cosmological applications. There are many potential societal consequences of our work, none of which we feel must be specifically highlighted here.
Declaration on the use of large language models
LLMs were used in the writing of this manuscript, but the theory was worked out manually. We take full responsibility for the accuracy of the material presented.
Background: harmonic analysis, lifting and lifted convolutions
In this appendix we collect the conventions and results of McEwen [33] on which our layers rest, each cited in its heading by its number there, where it is proved. Definition 1, Lemma 2, Corollary 3, Theorem 4 and Proposition 6 are stated in the same words, and Theorem 5 and Corollary 7 in shortened form. Only the equivariance of the lifted non-linearity is proved in this article (Appendix C).
Notation and conventions
We adopt the conventions of McEwen [33]. Points of the sphere are written , with colatitude and longitude , and denotes the north pole. We adopt the measure and the inner product . Rotations are parameterised by Euler angles in the active convention, , with the unnormalised Haar measure of total mass . The map sends to the point , so that , with the stabiliser of the north pole.
The spherical harmonics carry the Condon–Shortley phase, , and the spin raising and lowering operators and (Newman & Penrose [39]; Goldberg et al. [16]) generate the spin spherical harmonics from them by
so that exists only for and, for fixed , these form an orthonormal basis of the space of square integrable spin- functions on the sphere.
The Wigner -functions, the matrix elements of the irreducible unitary representations of (Varshalovich et al. (1988)), may be written
where the Wigner -functions are real. They satisfy and and the orthogonality relation , and by the Peter–Weyl theorem every admits the Wigner expansion
The identity tying the sphere to the rotation group is
Since the lift of a spin- function thus occupies the Wigner column and intertwines the spin rotation with left translation (Lemma 2(c) and (e)), and left translation acts on the first Wigner index alone, the harmonic coefficients of a function of any spin rotate by the same matrices under the spin rotation operator of McEwen et al. [36],
Figure 6 and Figure 7 show spin at a point and over the sphere.
Figure 6. What spin means at a point. Each panel is the tangent plane at , with to the right. A spin- value is drawn as equal spokes, and rotating the frame by multiplies it by , as in (1). The ring, the fibre of SO(3) over (§ 3.1), is coloured by the phase of the lift of the field, which winds times around it.
Figure 7. Spin fields on the sphere: a scalar field and the fields , and obtained from it by the spin raising operator of (11), drawn with glyphs whose symmetry is their spin (Figure 6). Wind has spin 1, polarisation and shear spin 2, and the second flexion of weak lensing spin 3 [1].
Spin fields, Mackey functions and the lift
The rotation group is a principal -bundle over the sphere with projection , whose fibre over a point is the circle of unit tangent vectors there (Figure 8). We adopt the section
Figure 8. The rotation group as a principal SO(2)-bundle over the sphere. The fibre over , the rotations with , is drawn as the rim of a cone at , coloured by the fibre coordinate , with the section as a red dot and a general element in the inset.
for , with and at the poles. It selects at each point the frame aligned with the meridian direction and is measurable, which suffices since all constructions are in . The irreducible representations of
are the characters , , and the Mackey functions of type (McEwen, [33]) are the elements of
Changing the section to multiplies the coordinate expression of a Mackey function by , which is the frame law (1) of a spin- field (McEwen, [33]).
Definition 1 (Lift and projection; McEwen, [33]). For define the lift and its adjoint, the projection , by
Lemma 2 (Structure of the lift; McEwen, [33]). Let .
(a) is an isometry of onto ; and is the orthogonal projector onto . The subspaces are mutually orthogonal and .
(b) On the projection is evaluation on the section: .
(c) In harmonic space,
where the coefficients on the sphere are taken with respect to the spin- harmonics.
(d) : lifted spin harmonics are unit-norm Wigner functions.
(e) and for all , where is the left regular representation of on , , and is the representation of on spin- fields induced from ; both are developed in McEwen [33], where (e) is proved as Proposition IV.10.
By McEwen [33], the induced representation is the spin rotation operator of McEwen et al. [36].
Corollary 3. (Functions on SO(3) as tuples of spin fields; McEwen, [33]). Every decomposes uniquely as , its Fourier series in converging in , with ; the component is a spin- field whose harmonic coefficients are the Wigner column up to the constant in (20). In particular is azimuthally band-limited at , for , if and only if for spin fields .
Lifted convolutions
The statement summarised in § 3.2, for the convolutions (4) and (6) (McEwen, [33]), is the following.
Theorem 4. (Harmonic form; McEwen, [33]). Let , , and let . For , with whenever the input is a spin- field and whenever the output is a spin- field,
and the output of has no component of degree . All four operators are bounded and -equivariant for every : and correspondingly with in place of or . Their adjoints are , and .
In outline, the lift of occupies the Wigner column (Lemma 2(c)), and the convolution (5) multiplies each block of degree on the right by , so that only the kernel column is read, which gives (22). The projection (21) reads the column , so that only reaches the output and the -dependent constants cancel, which gives (23). Equivariance follows from Lemma 2(e) and the fact that the group convolution commutes with left translation, and the full proof is in McEwen [33].
By (15) the spaces and carry one copy of the irreducible representation for each , so that completeness follows from Schur’s lemma.
Theorem 5. (Completeness; McEwen, [33]). Let be a bounded complex-linear map with for all . Then there is a unique bounded sequence such that
with input degrees absent from the target annihilated. Conversely every bounded sequence defines such a , with .
Proposition 6. (Azimuthally band-limited and steerable kernels; McEwen, [33]). If the column of is supported on then
i.e. the output is represented by at most spin-changing components , , with admissible output slots at degree . Conversely, any finite tuple of bounded equivariant spin-changing maps from the same input spin assembles into a bounded map with a generally distributional kernel. A square integrable representative exists exactly when the sum of their Hilbert–Schmidt norms squared is finite. At finite band-limit the assembled kernel is always square integrable.
Corollary 7. (Incompatibility of directional kernels, spherical output and equivariance; McEwen, [33]). For a bounded complex-linear equivariant map between one fixed input spin and one fixed output spin,
the two-point kernel is
in the distributional sense in general. It is radial apart from the frame phases prescribed by the two spins. Directional kernels, output on the sphere and SO(3)-equivariance therefore cannot be achieved simultaneously beyond those prescribed phases for this input/output pair.
Spherical layers as neural operators
Table 5 summarises the spherical layers of § 2.2 as neural operators, with the year of the first public version of each work (the FNO and the generalised spherical CNN appeared in the same month), although resolution-independent parameters imply neither exact sampled non-linear equivariance nor learned weights at unseen degrees.
| Method | Year | Signals | Linear layer | Cost |
| Spherical CNN [13] | 2018 | , scalar | harmonic zonal filter, one weight per | |
| Spherical CNN [9] | 2018 | , scalar | grid filter via generalised FFT | |
| Clebsch–Gordan net [24] | 2018 | , harmonic | per- maps, CG products | |
| Spin-weighted CNN [14] | 2020 | , spins | harmonic filters interpolated over | |
| Generalised spherical CNN [8] | 2020 | , | per- maps, azimuthal bandlimit | |
| FNO (Li et al., 2021) | 2020 | plane | one multiplier per mode | |
| DISCO [40] | 2022 | , scalar | continuous filter, quadrature | |
| SFNO [5] | 2023 | , scalar | one weight per | |
| SFNO (this work) | 2026 | , , spins | real profile in per spin pair (10 knots) |
Table 5. Spherical convolutional layers as neural operators. A layer is a neural operator in the sense of Kovachki et al. [25] when its parameters do not depend on the sampling of the sphere. Representative costs at bandlimit (azimuthal bandlimit , planar grid points) assume fixed channel and retained spin counts and, for DISCO, bounded local neighbourhood size.
Equivariant non-linearities
We consider the two settings of § 4.2 in turn, the sphere and the rotation group.
Non-linearities on the sphere
A rotation moves a spin- field and multiplies its value at each point by a phase of the form (1), where is the angle by which the local frame turns there. A pointwise map commutes with the move and is therefore equivariant if and only if it commutes with every such phase, which for reads for all and , since the rotations about each point turn its frame by every angle. The modulus non-linearity (9) multiplies by a real function of , so that
and it is therefore equivariant for every spin. For scalar fields () every pointwise map is equivariant and standard activation functions apply.
Non-linearities on the rotation group
On a rotation only translates the argument of a function, so that every pointwise map commutes with it. The lift-and-sum places a tuple of spin fields, one for each distinct spin within a channel, in one function on , each in its own Wigner column by Lemma 2(c), a pointwise map acts on that function, and a return of projecting convolutions maps the result back to a tuple of spin fields. Definition 8 states this construction and Theorem 9 its equivariance.
Definition 8 (Lifted non-linearity). For finite input and output spin sets, indexed by and , kernels and a measurable pointwise map satisfying for finite nonnegative constants , let
Theorem 9 (Equivariance of the lifted non-linearity; Xu et al., 2022). Under the hypotheses of Definition 8, is a well-defined map between the corresponding finite sums of spin spaces and
Theorem 9 is the result of Xu et al. (2022, Sec. 3.4), who show that their lift-and-sum, Eq. (3), followed by an elementwise non-linearity and a projection satisfies their condition, Eq. (2), proved in the same way. The growth condition is ours.
Proof of Theorem 9. The finite sum is bounded into . The growth hypothesis gives , and each component of is bounded, so the composition is well defined. With the notation of Definition 8,
In our networks each component of the return is the fixed projection , read as the Fourier column of the sampled group signal, followed by a learned filter for each channel and spin (Appendix § 14). By (21) this projection is the projecting convolution (24) with the kernel row , a distributional kernel that lies in once truncated at the bandlimit. The filter multiplies this row by one value for each degree and channel, and the constants of the implemented lift-and-sum and Fourier read can be absorbed into and the filter, so that Theorem 9 covers the non-linearity of our networks in the continuum, and the group normalisation between and the return uses invariant statistics and so preserves equivariance (§ 6.1). Xu’s fixed fibre integral is with their unnormalised stabiliser measure (Xu et al., 2022), and the orthogonal group projection of Ballerin et al. [2] is .
Discretisation
Theorem 9 concerns continuum functions, and is generally not bandlimited even when is. The exact truncation to the Wigner coefficients with and commutes with rotations, which act only on , whereas computing these coefficients from samples aliases the content of beyond the bandlimits into them, which oversampling reduces [8]. Apart from the Conv classifiers, which oversample by a factor of two (Appendix § 14), our networks evaluate without oversampling.
Parity
The antipodal map commutes with rotations and multiplies scalar harmonic coefficients by . Networks on scalar fields built from isotropic multipliers and pointwise maps are therefore -equivariant [52] and cannot represent a map that anticommutes with , , such as the advection of a vorticity treated as a scalar. This is the counterpart, for such networks, of the selection rule of Gaunt tensor products [55]. Spin-weighted and vector spherical signals restore the couplings of odd total degree [31, 56], and spin fields escape the restriction, since acts on a spin- field by . The vorticity is carried by the imaginary part of , and one spin-1 stage represents its advection through with an appropriate pointwise map on that is not odd, unlike the modulus non-linearity. A multiplier between spins commutes with , equivalently with the mirror , if and only if its coefficients are real, the spin counterpart of the fixed phase of reflection-equivariant planar kernels [50]. The opposite parities of the and modes [58] are an instance of this criterion, which a mirror test of our classifiers with real profiles confirms. Completeness among real-linear maps requires the conjugate field , as the conjugate frequencies do for planar harmonic networks [50].
Network architectures
We describe the layers and networks of s2ai used for the reported results, with the notation and names of § 4.4.
Layers
Spin spherical convolution. The spherical layer computes the forward spin- transform of every input channel and spin, truncates or zero-pads the coefficients to the output bandlimit , and forms for each input channel and output spin
followed by a real channel-mixing matrix for each output spin and inverse spin- transforms, so that the multipliers of the layer form a real and parameter-efficient subfamily of those of Theorem 5, interpolated in . Each real profile , independent of the output channel, is the piecewise-linear interpolation of 10 learned values at knots evenly spaced on . The profile values are initialised Glorot-uniformly with fans and 10, and the channel matrices of shape , Glorot-uniformly with the default fan axes of JAX. Neither has a bias. The factor is that of evaluation on the section, and the factor and the conjugation of (7) are absorbed into the real weights. Figure 9 shows an block as implemented.
Figure 9. An block as implemented. The convolution C, here the spin spherical convolution (33), and a harmonic convolution () with one channel matrix and one spin matrix shared across degrees act in parallel, and their sum passes through a second such convolution (), the modulus non-linearity and group normalisation (GN). Only the rotation-group construction forms the lift-and-sum .
Lift-and-sum and projecting convolution. The lift-and-sum, which has no parameters, maps a tuple to the samples at , , without the factor of Definition 1. The projecting layer reads the columns , , of a group signal by the discrete Fourier sum , takes the forward spin- transform of each, truncates or pads, multiplies by with one real profile per channel and spin, mixes channels within each spin and applies the inverse spin- transform. Spin is thus mapped to spin only, by the fixed adjoint up to a constant followed by a learned spin-preserving filter.
Harmonic convolutions. The harmonic convolutions and of the blocks (Figure 10) mix channels by one real matrix of shape shared by all spins, and spins by one real matrix of shape shared by all degrees, orders and channels, both initialised Glorot-uniformly and without bias. The spin matrix gives spin-changing multipliers constant in , so that the stage is equivariant.
Figure 10. The generic S²FNO block (10) in detail. Stacked spheres with a dashed cone denote a tuple of spin fields or its lift to SO(3), and the normalisation is not shown.
Activation and normalisation. The activation (9) has a real gain and offset for each channel, initialised to 1 and 0, so that it is the identity up to at initialisation. Group normalisation splits the channels into groups and, for each sample and group, computes the quadrature-weighted mean of the complex field and the mean squared modulus of . It returns , with real and for each group initialised to 1 and 0, so that only the modulus is normalised and the phase of is kept. Its statistics are computed for each spin separately on the sphere and pooled over the samples on the group. The pooled mean on the group is the column, a spin-0 quantity, and hence an invariant constant, whereas the mean on the sphere of a field of non-zero spin is not rotation invariant, so that the mean subtraction on the sphere is not exactly equivariant for . A normalisation with statistics computed in harmonic space and no mean subtracted for would be exactly equivariant (Sec. 6.1), but the group normalisation already gives reasonable, if not exact, rotational equivariance, with errors of order on the sphere and on the rotation group (Table 1).
| Operation | Spin | SO(3) | |
| Conv- | 0 | – | |
| Conv- | 1 | – | |
| Conv-SO(3) | 2 | – | |
| Harmonic conv. | |||
| SFNO block | |||
| SFNO UNet | |||
| SFNO UNet | |||
| ReLU (complex data) | 0.453 | 0.592 | |
| Swish (complex data) | 0.581 | 1.047 | |
| Complex ReLU, init. | – | ||
| Complex Swish | – | 0.0885 | |
| Group norm. | 0.013 | 0.0033 |
Table 1. Equivariance errors in double precision on the sphere () and the rotation group (SO(3)), measured by the mean relative equivariance error (MREE) between rotating the input and rotating the output (Appendix E). Conv- and Conv-SO(3) are the spin spherical convolution and the directional convolution , ReLU and Swish are applied directly to complex data, whereas the complex activations act on the modulus with the phase kept (9), and the complex ReLU is measured at initialisation, where it is the identity up to . The SFNO blocks and the UNets of the spin spherical MNIST architecture have no activation or normalisation, are at , with on SO(3), and reach the round-off floor of the transforms. All other rows are at , with on SO(3).
Blocks and spin spherical MNIST networks
Blocks. Writing for the spin spherical convolution, for the projecting convolution, for the lift-and-sum, and for the harmonic convolutions and for the normalisation, the blocks of the four families are
and no block has an identity residual connection. The first layer of a Conv-SO(3) network is a spin spherical convolution followed by , and , and the first block of a FNO-SO(3) network is a FNO- block without activation and normalisation, followed by , and . In the UNets the input layer maps the data spin, 0 for images and 1 for vector fields, to , with in the rotation-group construction.
Classifiers. The classifiers of Table 2 follow Esteves et al. [14] in their normalisation, downsampling and learning-rate schedule. Each stacks six blocks of a convolution, a normalisation and an activation, with channels for Conv-, for Conv-SO(3) and FNO- and for FNO-SO(3), internal spins in the spherical construction and azimuthal bandlimit in the rotation-group construction. The bandlimit falls from 32 to 16 and 8 over successive pairs of blocks, and the Conv classifiers evaluate their activations on grids oversampled by a factor of two against aliasing. In the Conv classifiers each pair of channels has its own complex profile, and in the FNO classifiers each input channel has a complex profile for each pair of spins, followed by channel matrices. The batch normalisation takes statistics over the batch and the quadrature on the sphere (and on the group) and subtracts the mean only from spin-0 fields and on the group, so that it is equivariant. The activation is , with a bias for each channel, on spin-0 fields and on the samples on the group, and the modulus non-linearity , with initialised to , on fields of non-zero spin. The pooling returns only rotation-invariant features, the real and imaginary parts of the mean of the spin-0 field, or of the function on the group, and the mean modulus of every field. A linear layer to 128 units, a Swish activation ([43]) and a linear layer then give the 10 class scores.
UNets. The MNIST UNets (Figure 11) have an encoder of four levels with 8, 16, 32 and 32 channels and a decoder of two levels with 16 and 8 channels, each decoder level concatenating the previous output with the normalised encoder features of equal bandlimit and spin set. In the Conv UNets the bandlimit is 32 at the first two levels and 16 at the last two, and is restored to 32 in the decoder, by truncating and zero-padding the harmonic coefficients inside the convolutions, whereas the FNO UNets keep throughout. The final layer, which maps the concatenation of channels to one channel of the target spin, is a spin spherical convolution without activation in Conv-, a projecting convolution in the rotation-group construction and a FNO- block, with its activation and normalisation, in FNO-.
Networks for wind forecasting and weak lensing
For wind forecasting the spin UNet, built from Conv-SO(3) blocks like that of mass mapping, has a lifting input layer that maps the three complex input fields onto the spins , and the scalar UNet takes the real and imaginary parts of each input field as separate channels, six input channels in place of three.
The two mass-mapping UNets (Figure 12) have five levels that halve the bandlimit from to , with encoders of 4, 8, 16, 32 and 64 channels. In the spin UNet the first layer lifts the spin-2 shear to the internal spin set, every subsequent block acts on the lifted features, and the final layer returns the spin-0 convergence, whereas the scalar UNet combines DISCO convolutions [40] with residual layers (He et al., 2016), except in its last block, a residual layer alone, and its bottleneck adds a convolution from 128 to 64 channels.
Figure 12. The two mass-mapping UNets, with five levels from to . Each box gives the bandlimit and the number of channels of a block, and the dashed lines carry the encoder features, which are concatenated () with the decoder output at equal bandlimit. (a) The spin UNet of the rotation-group construction, whose first block lifts the spin-2 shear to the internal spin set and whose last layer, a projecting convolution , returns the spin-0 convergence. (b) The scalar UNet, of DISCO convolutions and residual layers, which takes the real and imaginary parts of the shear as two channels and returns the convergence as its single channel.
Parameter counts
Table 6 counts trainable parameters only, and its MNIST counts are equal for spin-0 and spin-1 inputs.
| Task | Architecture | Trainable parameters |
| Classification | Conv- | 49,674 |
| Classification | Conv-SO(3) | 133,562 |
| Classification | FNO- | 182,016 |
| Classification | FNO-SO(3) | 136,088 |
| Prediction | Conv- | 23,330 |
| Prediction | Conv-SO(3) | 13,890 |
| Prediction | FNO- | 29,097 |
| Prediction | FNO-SO(3) | 20,034 |
| Wind forecasting | Spin UNet | 48,446 |
| Wind forecasting | Scalar UNet | 74,398 |
| Mass mapping | Spin UNet | 80,702 |
| Mass mapping | Scalar UNet | 93,198 |
Table 6. Trainable parameter counts of the architectures evaluated in Sec. 7.
Equivariance tests
We measure equivariance by the round-trip test of Cobb et al. [8], also adopted by Ocampo et al. [40]. We consider test signals (spin- signals on the sphere or signals on the rotation group) with complex Gaussian random harmonic coefficients bandlimited at , and rotations drawn from the Haar distribution on . For each pair we rotate a copy of the coefficients exactly in harmonic space by (15), synthesise both copies on the sampling grid of McEwen & Wiaux [34], without oversampling, and pass them through the operator under test. We then transform both outputs to harmonic space, rotate the output from the unrotated input by the same rotation, transform both back to real space and compute the mean relative equivariance error
where denotes the rotated output obtained from the unrotated input, the output obtained from the rotated input, and is the norm, as in Cobb et al. [8], computed over samples, channels and spins or, for group signals, also the samples, with the colatitude quadrature weights of the sampling theorem. All errors are computed in double precision, at with on the rotation group, except for the linear models of Table 7, at .
| Model | Spins | Channels | Params | Equivariance error (MREE) | |
| Identity on | 1 | 1 | 0 | ||
| Identity on | 8 | 1 | 0 | ||
| Positive control | 1 | 1 | 0 | ||
| FNO- block | 1, 1, 1 | 15 | |||
| FNO- block | 1, 1, 1 | 15 | |||
| FNO- block | 1, 8, 8 | 138 | |||
| FNO- block | 8, 16, 16 | 1,506 | |||
| FNO- block | 4, 4, 4 | 1,162 | |||
| FNO- block 5 | 1, 1, 1 | 15 | |||
| FNO- block 8 | 8, 16, 16 | 1,026 | |||
| FNO- block 8 | 4, 4, 4 | 362 | |||
| FNO- UNet | 28,825 | ||||
| FNO- UNet | 28,825 | ||||
| FNO- UNet | 24,865 | ||||
| FNO- UNet 8 | 19,762 | ||||
| FNO- UNet 8 | 19,762 | ||||
| FNO- UNet 4 | 6,346 |
Table 7. Equivariance errors of linear models at , measured by the MREE (34) over the pairs of signal and rotation. Spins and channels are given as input, hidden and output, or as one set of spins used throughout, and is the base width of a UNet, with encoder widths , , and . The first block of the FNO- UNet.
The non-linear rows of Table 1 are tested on spin-0 signals on the sphere and on group signals, for which every pointwise map commutes with rotations in the continuum, so that the errors of the activations arise from aliasing. The rows ReLU and Swish (complex data) apply these functions directly to complex values, the ReLU keeping when and setting it to zero otherwise. The complex activations instead act on the modulus and keep the phase. The complex ReLU is the modulus non-linearity (9) at its initialisation, and , where it is the identity up to , so that its error of order is that of the round trip, and it would grow once and are learned. Complex Swish is the same non-linearity with the Swish (Ramachandran et al., [43]) in place of the ReLU, with the same initial gain and offset. Equivariance of the trained networks, with their learned activations and normalisations, is not measured.
We also test blocks and whole UNets with several spins and channels, with every activation and normalisation removed so that every model is linear (Table 7). The UNets with and have the architecture of our spin spherical MNIST UNets and are tested in both directions, and . The identity maps give the round-off floor of the transforms, of order on the sphere and on the rotation group. The positive control, a layer that adds one to the samples in the northern hemisphere, gives an error of order , which shows that the test detects broken equivariance. Every linear block and UNet gives errors of order , the floor on the sphere, as expected.
Datasets
We describe the three benchmarks of § 7.1 in turn.
Spin spherical MNIST
We project planar digits onto the sphere following Cohen et al. [9] and retain their scalar harmonic coefficients up to bandlimit . We also construct a Sobel variant, motivated by [14] (Xu et al. 2022), in which the Sobel derivatives of the projected image along the colatitude and longitude axes of the sampling grid, with their means subtracted, are combined into , whose modulus is rescaled to for each sample with the phase kept, and the result is transformed as a spin-1 field before any rotation.
The harmonic variants are computed with s2fft (Price & McEwen, 2024), by a forward spherical harmonic transform of the scalar image at bandlimit , a rotation in harmonic space by (15), exact at that bandlimit, and an inverse spin spherical harmonic transform at the desired spin, which synthesises . The synthesis keeps the scalar coefficients of the degrees , without the degree-dependent factors of differential spin raising, and commutes with rotations by (15). A spin-1 field so constructed has no monopole, whereas the scalar fields keep a small one, since the mean subtracted from each image is the unweighted mean over the grid rather than the quadrature-weighted spherical mean.
For the prediction experiments we generate 60,000 samples at on the sampling grid of McEwen & Wiaux [34] with samples, from the first 51,000 images of the MNIST training set and the first 9,000 of the MNIST test set, which form an 85/15 split into training and held-out sets. Rotations of the held-out samples use Euler angles drawn uniformly from , which is not the Haar distribution on (that would draw uniformly), so that the rotated test set over-represents some rotation axes. The generation code and the exact split indices will be released with s2ai.
Splits and protocol. The classification experiments use a regenerated split of the same construction, in the image, Sobel spin-1, harmonic spin-1 and harmonic spin-2 variants. The first 55,000 images of the MNIST training set are used for training and the remaining 5,000, unrotated, for validation, and the 10,000 images of the MNIST test set for testing, with Haar-distributed rotations drawn once with fixed seeds. Following Cohen et al. [9], models are trained on unrotated samples and evaluated on unrotated (NR/NR) and randomly rotated (NR/R) test samples, and all results reported in § 7.2 are NR/R.
Global wind forecasting
The training set is a subset of the WeatherBench 2 1.5° archive of ERA5 reanalysis data ([44, 19]) at the 50 hPa pressure level, covering four six-hourly observations per day between 1980 and 2022. The observations at 00:00, 06:00 and 12:00 UTC form the input and the observation at 18:00 UTC the target, with the zonal and meridional components combined into the spin-1 field (Sec. 7.1), which is the field of the meridian frame up to the constant factor . Data are converted from the equiangular latitude–longitude grid to the sampling of McEwen & Wiaux [34] at , with samples, and standardised by a real scale, after subtracting the mean of the training set. For evaluation, eight indices are drawn at random in the range , ignoring leap years, and the corresponding days, which cover all seasons, are held out from every year. The preprocessing code and the held-out indices will be released with s2ai.
Weak gravitational lensing
Weak lensing is described by two fields derived from the lensing potential , namely the spin-0 convergence , which traces the projected matter density, and the spin-2 shear . Since both derive from the same potential, they are related in harmonic space by the spherical Kaiser–Squires relation ([20, 48]), which in our conventions reads
The forward map (35) is therefore the spin layer of Theorem 5 with , and spherical Kaiser–Squires is the spin layer with , where , on the degrees that the shear determines. Practical considerations make the recovery of ill-posed and call for priors learned from simulations.
The convergence maps are simulated with GLASS [46] and CAMB [29] for a Planck 2018 cosmology [42], with lognormal matter shells of 200 Mpc from redshift to 1, one Gaussian source bin at , HEALPix resolution [17] and . The shear maps are forward modelled with DarkMax¹ with the noise of a Stage IV survey [27], a galaxy number density and an ellipticity dispersion , with a different noise seed for each map. Maps are used at on the sampling grid of McEwen & Wiaux [34], with samples. The training set holds 10,000 shear–convergence pairs, split 80/20 into training and validation sets, and the evaluation set 600 further pairs from the same pipeline.
Training details and additional results
Each model is trained once, except where stated, so that our results carry no estimate of their variation between runs.
Spin spherical MNIST
Training. The classifiers are trained in single precision on a shared NVIDIA GB10 with Adam [22] () at a learning rate of , multiplied by 0.2 at 50% and again at 83% of the run, in batches of 32 for 12 epochs (5 or 6 for Conv-SO(3), Table 2), minimising the categorical cross-entropy, and the reported checkpoint is the epoch of best accuracy on the non-rotated validation split. The prediction UNets are trained in double precision with Adam () at a learning rate of for up to 2000 epochs in batches of 32, minimising the squared modulus of the difference between prediction and target, integrated with the quadrature weights of the sampling theorem, on NVIDIA GH200 and A100 GPUs with identical code. The reported prediction model is the checkpoint with the lowest loss on the rotated held-out set, on which the NR/R results are also evaluated, so that, with no separate validation set, the reported prediction errors are those of the best checkpoint on the evaluation data.
Prediction examples. The samples of Figure 13 are fixed in advance, not chosen by their errors.
Figure 13. Prediction by the Conv-S and Conv-SO(3) UNets at from the spin-0 image to the spin-1 Sobel field (a–c) and back (d–f), for the first three test digits under one rotation. Errors (prediction minus target) are on the scales of the target, spin-1 fields are drawn as arrows over their modulus, and percentages give the relative error .
Wind forecasting
Training. Both UNets of Appendix D.3 are trained for 200 epochs with Adam [22] at its default settings and a constant learning rate, minimising the mean squared error computed on the sphere, with the learning rates and batch sizes of Table 8, chosen with ablation studies. The reported model of each kind is the checkpoint with the lowest loss on the held-out days.
| Model | Params | LR | Input shape | MSE | RMSE | MAE | SNR | PSNR | PCC |
| Spin UNet | 48,446 | 9.52 | 3.09 | 2.67 | 14.02 | 25.06 | 0.979 | ||
| Scalar UNet | 74,398 | 17.27 | 4.16 | 3.56 | 11.43 | 22.64 | 0.969 | ||
| Persistence | – | – | – | 14.55 | 3.82 | 3.31 | 12.17 | 23.19 | 0.962 |
Table 8. Wind forecasting at 50 hPa and : configurations and all metrics on the held-out days, in physical units (MSE in , RMSE and MAE in , SNR and PSNR in dB). Input shapes are batch channels, and best results are in bold.
Weak gravitational lensing
Training. Both UNets of Appendix D.3 are trained for 200 epochs on the 10,000 shear–convergence pairs of Appendix F.3, split 80/20 into training and validation sets, minimising a spherically weighted mean squared error with Adam [22], with the inputs, batch sizes and learning rates of Table 9. Evaluation is on the 600 held-out maps.
| Model | Params | LR | Input shape | MSE | RMSE | MAE | SNR (dB) | PSNR (dB) | PCC |
| Spin UNet | 80,702 | 1.28 | 1.13 | 0.90 | 7.53 | 23.47 | 0.907 | ||
| Scalar UNet | 93,198 | 2.06 | 1.43 | 1.14 | 5.45 | 21.40 | 0.846 |
Table 9. Weak lensing mass mapping at with shape noise: configurations and all metrics, averaged over the 600 evaluation maps. Input shapes are batch channels, and best results are in bold.
Results. The residual of the scalar UNet retains high-amplitude, large-scale structure (Figure 15), and the reconstruction of the spin UNet follows the tails of the one-point distribution of the target more closely (Figure 15d).
Figure 15. Mass mapping for one evaluation map at , with the true convergence (a), its reconstructions by the spin and scalar UNets (b, c) and their one-point distributions with that of the target (d), the noisy input shear (e, f) and the residuals of the two UNets with their MSE on this map (g, h). The scalar residual retains large-scale structure that the spin residual does not.
References
- [1]Bacon, D. J., Goldberg, D. M., Rowe, B. T. P., and Taylor, A. N. Weak gravitational flexion. Monthly Notices of the Royal Astronomical Society, 365(2):414–428, 2006. doi: 10.1111/j.1365-2966.2005.09624.x. arXiv:astro-ph/0504478.DOI
- [2]Ballerin, F., Blaser, N., and Grong, E. SO(3)-equivariant neural networks for learning from scalar and vector fields on spheres. arXiv:2503.09456, 2025.DOI
- [3]Bartelmann, M. and Schneider, P. Weak gravitational lensing. Physics Reports, 340(4–5):291–472, 2001. doi: 10.1016/S0370-1573(00)00082-X.DOI
- [5]Bonev, B., Kurth, T., Hundt, C., Pathak, J., Baust, M., Kashinath, K., and Anandkumar, A. Spherical Fourier neural operators: Learning stable dynamics on the sphere. In International Conference on Machine Learning (ICML), 2023. arXiv:2306.03838.arxiv.org/abs/2306.03838
- [6]Bonev, B., Kurth, T., Mahesh, A., Bisson, M., Kossaifi, J., Kashinath, K., Anandkumar, A., Collins, W. D., Pritchard, M. S., and Keller, A. FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale. arXiv:2507.12144, 2025.
- [7]Castro, P. G., Heavens, A. F., and Kitching, T. D. Weak lensing analysis in three dimensions. Physical Review D, 72(2):023516, 2005. doi: 10.1103/PhysRevD.72.023516. arXiv:astro-ph/0503479.DOI
- [8]Cobb, O. J., Wallis, C. G. R., Mavor-Parker, A. N., Marignier, A., Price, M. A., d’Avezac, M., and McEwen, J. D. Efficient generalized spherical CNNs. In International Conference on Learning Representations (ICLR), 2021. arXiv:2010.11661.arxiv.org/abs/2010.11661
- [9]Cohen, T. S., Geiger, M., Köhler, J., and Welling, M. Spherical CNNs. In International Conference on Learning Representations (ICLR), 2018. arXiv:1801.10130.arxiv.org/abs/1801.10130
- [10]Cohen, T. S., Geiger, M., and Weiler, M. A general theory of equivariant CNNs on homogeneous spaces. In Advances in Neural Information Processing Systems (NeurIPS), pp. 9142–9153, 2019. arXiv:1811.02017.arxiv.org/abs/1811.02017
- [11]Descoteaux, M., Angelino, E., Fitzgibbons, S., and Deriche, R. Regularized, fast, and robust analytical Q-ball imaging. Magnetic Resonance in Medicine, 58(3):497–510, 2007. doi: 10.1002/mrm.21277.
- [12]Driscoll, J. R. and Healy, Jr., D. M. Computing Fourier transforms and convolutions on the 2-sphere. Advances in Applied Mathematics, 15(2):202–250, 1994. doi: 10.1006/aama.1994.1008.
- [13]Esteves, C., Allen-Blanchette, C., Makadia, A., and Daniilidis, K. Learning SO(3) equivariant representations with spherical CNNs. In European Conference on Computer Vision (ECCV), pp. 54–70, 2018. arXiv:1711.06721.
- [14]Esteves, C., Makadia, A., and Daniilidis, K. Spin-weighted spherical CNNs. In Advances in Neural Information Processing Systems (NeurIPS), 2020. arXiv:2006.10731.arxiv.org/abs/2006.10731
- [15]Esteves, C., Slotine, J.-J., and Makadia, A. Scaling spherical CNNs. In International Conference on Machine Learning (ICML), 2023. arXiv:2306.05420.arxiv.org/abs/2306.05420
- [16]Goldberg, J. N., Macfarlane, A. J., Newman, E. T., Rohrlich, F., and Sudarshan, E. C. G. Spin-s spherical harmonics and ð. Journal of Mathematical Physics, 8(11):2155–2161, 1967. doi: 10.1063/1.1705135.DOI
- [17]Górski, K. M., Hivon, E., Banday, A. J., Wandelt, B. D., Hansen, F. K., Reinecke, M., and Bartelmann, M. HEALPix: A framework for high-resolution discretization and fast analysis of data distributed on the sphere. The Astrophysical Journal, 622(2):759–771, 2005. doi: 10.1086/427976.
- [19]Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., et al. The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730):1999–2049, 2020.
- [20]Kaiser, N. and Squires, G. Mapping the dark matter with weak gravitational lensing. The Astrophysical Journal, 404:441–450, 1993. doi: 10.1086/172297.DOI
- [21]Kilbinger, M. Cosmology with cosmic shear observations: a review. Reports on Progress in Physics, 78(8):086901, 2015. doi: 10.1088/0034-4885/78/8/086901.DOI
- [22]Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015.arxiv.org/abs/1412.6980
- [23]Kondor, R. and Trivedi, S. On the generalization of equivariance and convolution in neural networks to the action of compact groups. In International Conference on Machine Learning (ICML), pp. 2747–2755, 2018. arXiv:1802.03690.
- [24]Kondor, R., Lin, Z., and Trivedi, S. Clebsch–Gordan nets: A fully Fourier space spherical convolutional neural network. In Advances in Neural Information Processing Systems (NeurIPS), 2018. arXiv:1806.09231.arxiv.org/abs/1806.09231
- [25]Kovachki, N., Li, Z., Liu, B., Azizadenesheli, K., Bhattacharya, K., Stuart, A., and Anandkumar, A. Neural operator: Learning maps between function spaces with applications to PDEs. Journal of Machine Learning Research, 24(89):1–97, 2023. arXiv:2108.08481.DOI
- [26]Lam, R., Sanchez-Gonzalez, A., Wilson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., and Battaglia, P. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416–1421, 2023. doi: 10.1126/science.adi2336. arXiv:2212.12794.DOI
- [27]Laureijs, R. et al. Euclid definition study report, 2011. arXiv:1110.3193.arxiv.org/abs/1110.3193
- [28]LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
- [29]Lewis, A., Challinor, A., and Lasenby, A. Efficient computation of cosmic microwave background anisotropies in closed Friedmann-Robertson-Walker models. The Astrophysical Journal, 538(2):473–476, 2000. doi: 10.1086/309179.
- [31]Liang, C., Lin, Y., Kryvenko, A., Yu, W., Li, C., Xie, J., Qian, X., and Ji, S. Spin-weighted spherical harmonics enable complete and scalable E(3)-equivariant networks. arXiv:2607.01408, 2026.
- [32]Liu-Schiaffini, M., Berner, J., Bonev, B., Kurth, T., Azizadenesheli, K., and Anandkumar, A. Neural operators with localized integral and differential kernels. In International Conference on Machine Learning (ICML), 2024. arXiv:2402.16845.
- [33]McEwen, J. D. Unifying spherical convolutions through lifting. arXiv:2610.13892, 2026.arxiv.org/abs/2610.13892
- [34]McEwen, J. D. and Wiaux, Y. A novel sampling theorem on the sphere. IEEE Transactions on Signal Processing, 59(12):5876–5887, 2011.DOI
- [35]McEwen, J. D., Büttner, M., Leistedt, B., Peiris, H. V., and Wiaux, Y. A novel sampling theorem on the rotation group. IEEE Signal Processing Letters, 22(12):2425–2429, 2015a. doi: 10.1109/LSP.2015.2490676. arXiv:1508.03101.DOI
- [36]McEwen, J. D., Leistedt, B., Büttner, M., Peiris, H. V., and Wiaux, Y. Directional spin wavelets on the sphere. arXiv:1509.06749, 2015b.arxiv.org/abs/1509.06749
- [37]McEwen, J. D., Wallis, C. G. R., and Mavor-Parker, A. N. Scattering networks on the sphere for scalable and rotationally equivariant spherical CNNs. In International Conference on Learning Representations (ICLR), 2022. arXiv:2102.02828.arxiv.org/abs/2102.02828
- [38]Mousset, L., Allys, E., Price, M. A., Aumont, J., Delouis, J.-M., Montier, L., and McEwen, J. D. Generative models of astrophysical fields with scattering transforms on the sphere. Astronomy & Astrophysics, 691:A269, 2024. doi: 10.1051/0004-6361/202451396. arXiv:2407.07007.DOI
- [39]Newman, E. T. and Penrose, R. Note on the Bondi–Metzner–Sachs group. Journal of Mathematical Physics, 7(5):863–870, 1966. doi: 10.1063/1.1931221.DOI
- [40]Ocampo, J., Price, M. A., and McEwen, J. D. Scalable and equivariant spherical CNNs by discrete-continuous (DISCO) convolutions. In International Conference on Learning Representations (ICLR), 2023. arXiv:2209.13603.arxiv.org/abs/2209.13603
- [41]Pathak, J., Subramanian, S., Harrington, P., Raja, S., Chatopadhyay, A., Mardani, M., Kurth, T., Hall, D., Li, Z., Azizadenesheli, K., Hassanzadeh, P., Kashinath, K., and Anandkumar, A. FourCastNet: A global data-driven high-resolution weather model using adaptive Fourier neural operators. arXiv:2202.11214, 2022.arxiv.org/abs/2202.11214
- [42]Planck Collaboration, Aghanim, N., et al. Planck 2018 results. VI. cosmological parameters. Astronomy & Astrophysics, 641:A6, 2020. doi: 10.1051/0004-6361/201833910. Price, M. A. and McEwen, J. D. Differentiable and accelerated spherical harmonic and Wigner transforms. Journal of Computational Physics, 510:113109, 2024. arXiv:2311.14670.
- [43]Ramachandran, P., Zoph, B., and Le, Q. V. Searching for activation functions, 2017. arXiv:1710.05941.arxiv.org/abs/1710.05941
- [44]Rasp, S., Hoyer, S., Merose, A., Langmore, I., Battaglia, P., Russell, T., Sanchez-Gonzalez, A., Yang, V., Carver, R., Agrawal, S., et al. WeatherBench 2: A benchmark for the next generation of data-driven global weather models. Journal of Advances in Modeling Earth Systems, 16(6):e2023MS004019, 2024.arxiv.org/abs/2308.15560
- [45]Ronneberger, O., Fischer, P., and Brox, T. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), volume 9351 of Lecture Notes in Computer Science, pp. 234–241, 2015. doi: 10.1007/978-3-319-24574-4_28.DOI
- [46]Tessore, N., Loureiro, A., Joachimi, B., von Wietersheim-Kramsta, M., and Jeffrey, N. GLASS: Generator for large scale structure. The Open Journal of Astrophysics, 6:11, 2023. doi: 10.21105/astro.2302.01942.DOI
- [48]Wallis, C. G. R., Price, M. A., McEwen, J. D., Kitching, T. D., Leistadt, B., and Plouviez, A. Mapping dark matter on the celestial sphere with weak gravitational lensing. Monthly Notices of the Royal Astronomical Society, 509(3):4480–4497, 2022. doi: 10.1093/mnras/stab3235. arXiv:1703.09233.DOI
- [49]Wang, Y. and Yu, Y. Advancing weak lensing mass mapping with a mask-aware HEALPix transformer. Physical Review D, 113(4):043553, 2026. doi: 10.1103/kc9z-jllp. arXiv:2603.25471.DOI
- [50]Weiler, M. and Cesa, G. General E(2)-equivariant steerable CNNs. In Advances in Neural Information Processing Systems (NeurIPS), 2019. arXiv:1911.08251.arxiv.org/abs/1911.08251
- [51]Weiler, M., Geiger, M., Welling, M., Boomsma, W., and Cohen, T. S. 3D steerable CNNs: Learning rotationally equivariant features in volumetric data. In Advances in Neural Information Processing Systems (NeurIPS), 2018. arXiv:1807.02547.arxiv.org/abs/1807.02547
- [52]Weiler, M., Forré, P., Verlinde, E., and Welling, M. Coordinate independent convolutional networks – isometry and gauge equivariant convolutions on Riemannian manifolds. arXiv:2106.06020, 2021.arxiv.org/abs/2106.06020
- [53]Wiaux, Y., Jacques, L., Vielva, P., and Vandergheynst, P. Fast directional correlation on the sphere with steerable filters. The Astrophysical Journal, 652(1):820–832, 2006. doi: 10.1086/507692. arXiv:astro-ph/0508516.DOI
- [54]Worrall, D. E., Garbin, S. J., Turmukhambetov, D., and Brostow, G. J. Harmonic networks: Deep translation and rotation equivariance. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7168–7177, 2017. doi: 10.1109/CVPR.2017.758. arXiv:1612.04642.DOI
- [55]Xie, Y., Daigavane, A., Kotak, M., and Smidt, T. The price of freedom: Exploring expressivity and runtime tradeoffs in equivariant tensor products. In International Conference on Machine Learning (ICML), 2025. arXiv:2506.13523.arxiv.org/abs/2506.13523
- [56]Xie, Y., Daigavane, A., Kotak, M., and Smidt, T. Asymptotically fast Clebsch–Gordan tensor products with vector spherical harmonics. arXiv:2602.21466, 2026.arxiv.org/abs/2602.21466
- [58]Zaldarriaga, M. and Seljak, U. All-sky analysis of polarization in the microwave background. Physical Review D, 55 (4):1830–1840, 1997. doi: 10.1103/PhysRevD.55.1830. arXiv:astro-ph/9609170.DOI














