betaFeedback

WorksPhysics and Astronomy

Star-forming disc galaxies align tangentially with the cosmic tidal field

Abstract

The alignments of galaxy shapes, with one another and with the large-scale structure of the Universe, have been a topic of considerable investigation within cosmology and extragalactic astrophysics for the last 60 years. Galaxy alignments are a major systematic for weak gravitational lensing and help us learn about the processes of galaxy formation. Red early-type galaxies are known to show strong radial alignments around other galaxies; however, blue star-forming galaxies have shown little evidence of alignment. We combine SPHEREx spectrophotometry with DESI Legacy Imaging Surveys to build a full-sky catalogue of 71 million galaxies with precise redshifts. With this catalogue we are able to measure alignment signals by splitting galaxies into colour and morphological samples. We detect tangential alignment of blue disc galaxies at 22.5σ22.5σ and radial alignment of red spheroidal galaxies at 65σ65σ. By measuring the alignment signal separately around overdensities and underdensities, we show that each of these alignments has the scale and environmental dependence of a linear response to the cosmic tidal field; a response purely quadratic in the Gaussian tidal field, as in classic tidal torque theory is ruled out.

The large-scale alignments of galaxy shapes that we observe in the Universe are thought to be set during the process of galaxy formation. As the process of galaxy formation varies significantly with galaxy type, from spirals and discs to irregulars and spheroids, we should also expect the alignment of galaxy shapes to depend strongly upon galaxy type. Observations of galaxy alignments will help us understand the process of galaxy formation in more detail, and may hold a fossil record of the tidal field at the time of a galaxy’s formation.

There are two physical galaxy formation mechanisms which are expected to create large-scale alignments between galaxy shapes. Both are related to the tidal field of the Universe. The tidal field is the differential pull of gravity, and therefore describes the stretching and squeezing produced by the gravitational potential at a given point. A forming galaxy can be stretched directly by the tidal field, or it can be spun up by it through tidal torquing. These lead to the two principal alignment mechanisms, a linear and a quadratic response of the galaxy shape to the tidal field. However, as we explain below, the formation mechanisms and the alignment mechanisms do not map one to one: tidal torquing can also produce a linear response.

For early-type galaxies the acquisition of an alignment is expected to be relatively simple. As a proto-galaxy forms from gravitational collapse, the cosmic tidal field sets a preferential collapse axis, such that the galaxy’s form contains a small additional shear which is proportional to the tidal field [8] (Hirata and Seljak 2004). As a result, the projected image of the galaxy has a small linear contribution from the tidal shear. For disc galaxies ⓘ the situation is significantly more complex. Their alignments are instead expected to be largely determined by the direction of the galaxy’s angular momentum [8] (Lee and Pen 2000). As the proto-galaxy forms, its own inertia tensor may be misaligned with the large-scale tidal field. The tidal field will then torque the proto-galaxy, causing its angular momentum to grow [21] (Peebles 1969; White 1984). As the proto-galaxy inertia also depends upon the tidal field, the resulting alignment of the disc galaxy will respond quadratically to the tidal field.

Considering galaxy shapes projected on the sky, instead of the full three-dimensional bodies, we can approximate the shape of a galaxy as an ellipse. We can write the observed galaxy ellipticity in terms of a contribution from a linear response to the tidal field, C1C_1, a contribution from a quadratic response to the tidal field, C2C_2, and a term from the weighting of the tidal field by the local density field, C1δC_{1\delta}, which arises because we can only measure galaxy shapes where there are galaxies. We can sketch the observed galaxy shape field, the average ellipticity of galaxies, γ\gamma, at a position x\mathbf{x}, as

γ(x)=C1slin(x)+C1δδ(x)slin(x)+C2squad(x),(1)\gamma(\mathbf{x}) = C_1 s_{\rm lin}(\mathbf{x}) + C_{1\delta}\delta(\mathbf{x})s_{\rm lin}(\mathbf{x}) + C_2 s_{\rm quad}(\mathbf{x}), \tag*{(1)}

where δ\delta is the density field, slins_{\rm lin} is the tidal field, and squads_{\mathrm{quad}} is quadratic in the tidal field. Therefore the amplitudes of C1C_1, C1δC_{1\delta} and C2C_2 will determine the shape of the galaxy alignment field. The full form of this equation is given in Methods, Modelling linear and quadratic responses to the tidal field, Eq. (6).

C1C_1 is often referred to as the tidal alignment model, and terms proportional to C2C_2 are often referred to as the tidal torquing model. The term C1δC_{1\delta} only appears when considering the inhomogeneous sampling of the tidal field from galaxies. Together, these terms form the tidal alignment and tidal torquing (TATT) model. Physically, however, this is a misnomer: torquing models can also produce a linear response to the tidal field [4] (Hui and Zhang 2008; Larsen and Challinor 2016). Once the density field has evolved non-linearly, a large-scale mode modulates the growth of the small-scale tidal field that torques the galaxy, and the resulting alignment contains a term that responds linearly to the large-scale tidal field. By classic tidal torque theory we mean the case without this effect, in which the alignment is purely quadratic in the initial, Gaussian tidal field (Methods, Modelling linear and quadratic responses to the tidal field). Therefore we refer to C1C_1 as the linear response to the tidal field, C1δC_{1\delta} as the density-weighting response, and C2C_2 as the quadratic response. The galaxy ellipticity is defined as e=∣e∣e2iψ=e1+ie2e = \lvert e\rvert e^{2i\psi} = e_1 + ie_2, where the factor of two comes from the fact that a rotation of an ellipse by π\pi gives an identical ellipse. The galaxy ellipticity is a spin-2 quantity. The possible alignment patterns of galaxies are radial (e+>0e_+ > 0), tangential (e+<0e_+ < 0) and cross (e×≠0e_\times\ne0). In Figure 2 we illustrate the alignment of galaxy shapes from linear and quadratic response. The linear response switches sign between a void and a cluster, whereas the quadratic response does not.

Figure 2. The alignment patterns from linear and quadratic alignment models. Shapes of galaxies (purple) around a void (open circle) and a cluster (filled circle). The top row shows the linear response to the tidal field where the shape alignment switches from tangential to radial between voids and clusters and the bottom row the quadratic response where the signal must keep the same sign for both voids and clusters.

Galaxy alignments (or intrinsic alignments) are a crucial systematic for weak gravitational lensing surveys (Hirata and Seljak, 2004; Joachimi et al., 2015; Troxel and Ishak, 2015). Weak gravitational lensing describes the deformation of galaxy images on the sky through the deflection of light by massive objects. These deformations lead to correlations between galaxy shapes, and between galaxy shapes and positions. Correlations from intrinsic alignments can therefore be mistaken for correlations from lensing. It is for this reason that it is essential to improve our understanding of intrinsic alignments of galaxies for precise weak gravitational lensing cosmology.

There have been many efforts to measure galaxy alignments for different galaxy types. For red galaxies, significant radial alignments of galaxies have been found consistently (Mandelbaum et al., 2006; Hirata et al., 2007; Joachimi et al., 2011; Singh et al., 2015; Johnston et al., 2019; Yao et al., 2020; Georgiou et al., 2025; Siegel et al., 2026). These observations have been well described by the linear response model. For blue galaxies and late-type galaxies, the observational situation is more complex. Within the field of galaxy formation there have been detections of blue galaxy spin alignments with filaments of the cosmic web (Tempel and Libeskind, 2013; Welker et al., 2020), however, using integral-field kinematics, other authors found inconsistent results (Krolewski et al., 2019). These results use the reconstructed cosmic web, instead of the alignment of galaxies around other galaxies, therefore their importance for weak gravitational lensing is difficult to assess. With an approach more relevant to weak lensing surveys, blue galaxies have given null detections (Mandelbaum et al., 2011; Johnston et al., 2019; Yao et al., 2020; Georgiou et al., 2025; Siegel et al., 2026), until the recent 3.3σ hint of tangential alignment for massive blue galaxies in Paviot et al. (2026).

While the linear alignment model for early-type galaxies is well established, both in theory and in observations, to produce radial alignments around overdensities, the predictions for disc galaxies are far less clear. Many hydrodynamical simulations studies find that disc or blue galaxies have either no detectable large-scale alignment (Tenneti et al., 2016; Hilbert et al., 2017) or a radial alignment [17] (Samuroff et al. 2021; Zjupa et al. 2022; van Heukelum and Chisari 2026). In the Horizon-AGN simulation, however, disc galaxies have large-scale tangential alignments in 3D [11], but the projected alignment of theses discs is consistent with zero. Furthermore the alignment of simulated galaxies within the cosmic web appears to depend on the particular hydrodynamical code used for the simulation (Tenneti et al., 2016). These simulations are currently limited by the simulation volume and the difficulty of creating realistic disc galaxies within large volume simulations [11, 12].

One reason for the strong detections for luminous red galaxies (LRGs) and the lack of detections for blue galaxies is the selection inherent to large spectroscopic galaxy redshift surveys, which have generally targeted red early-type galaxies, such as LRGs, which reside in massive halos. These galaxies are excellent tracers of the large-scale structure for cosmology, as they are bright, allowing deeper selections in redshift, and strongly clustered, giving a large signal in their galaxy auto-correlation functions. This has biased the literature of direct alignment measurements away from the blue discs which we study in this work. Due to the relatively weak constraints on the alignments of blue galaxies, they have been assumed to have small or no intrinsic alignment signals (Joachimi et al., 2011). The majority of galaxies used as weak gravitational lensing sources for the upcoming Euclid (Mellier et al., 2025), LSST (Ivezić et al., 2019) and Roman (Spergel et al., 2015) surveys will be blue star-forming galaxies. Therefore it is crucial that we understand the amplitudes of their alignments, so that cosmological inferences are not biased.

In order to measure galaxy alignments, we measure the cross-correlation between galaxy shapes and galaxy positions and the auto-correlation of galaxy positions. The cross-correlation between galaxy shapes and positions, wg+(rp)w_{\mathrm{g}+}(r_{\mathrm{p}}), is closely related to the mean shape of galaxies, ⟨e+⟩\langle e_{+}\rangle, at projected separation rpr_{\mathrm{p}} from another galaxy, integrated along the line of sight.

wg+(rp)=bg[A1L(rp)+A1δDw(rp)+A2Q(rp)],(2)w_{\mathrm{g}+}(r_{\mathrm{p}}) = b_{\mathrm{g}}\left[A_{1}L(r_{\mathrm{p}})+A_{1\delta}D_{\mathrm{w}}(r_{\mathrm{p}})+A_{2}Q(r_{\mathrm{p}})\right], \tag*{(2)}

where A1A_{1}, A1δA_{1\delta} and A2A_{2} are dimensionless versions of C1C_{1}, C1δC_{1\delta} and C2C_{2}, which make it possible to compare the amplitudes across different redshifts (the relations between AA and CC are given in Methods, Modelling linear and quadratic responses to the tidal field, Eq. (7)). Positive A1A_{1} leads to radial alignments and negative A1A_{1} leads to tangential alignments. LL, DwD_{\mathrm{w}} and QQ are functions that depend upon the assumed cosmology and the redshift distributions of the shape and position samples; they are defined in Methods, Modelling linear and quadratic responses to the tidal field (Eqs. (13)–(15)).

The auto-correlation of galaxy positions measures the galaxy bias,

wgg(rp)=bg2G(rp),(3)w_{\mathrm{gg}}(r_{\mathrm{p}}) = b_{\mathrm{g}}^{2}G(r_{\mathrm{p}}), \tag*{(3)}

where GG is the projected matter correlation function; the full expression is given in Methods, Modelling linear and quadratic responses to the tidal field (Eq. (12)). By combining measurements of wggw_{\mathrm{gg}} and wg+w_{\mathrm{g+}} at fixed cosmology we can directly constrain the amplitude of intrinsic alignment. Unless stated otherwise, we fit our measurements with the linear response alone, setting A1δ=A2=0A_{1\delta}=A_{2}=0.

Blue discs show tangential alignments

In order to study the alignments of galaxies within the cosmic web, we require precise distances to each of the galaxies. We selected resolved sources with an rr-band magnitude r<20.5r < 20.5 from the DESI Legacy Imaging Surveys and obtained their spectrophotometry using data from the SPHEREx satellite [20, 5]. From this we obtained a redshift catalogue of 71.2 million galaxies with a σNMAD\sigma_{\mathrm{NMAD}} uncertainty of 0.0065 and 0.46% catastrophic outliers across the full sky (Methods, The Legacy Surveys–SPHEREx redshift catalogue). This sample has no colour or morphological selection, so we can measure intrinsic alignment signals for any type of galaxy. A full-sky projection of the cosmic web from this catalogue is shown in Figure 1. From this redshift catalogue, we selected a high-quality redshift subsample by keeping the 90% of galaxies with the narrowest redshift probability distributions. This leaves us with a sample of 63.0 million galaxies, with σNMAD=0.0055\sigma_{\mathrm{NMAD}} = 0.0055–0.00770.0077 and 0.13–0.52% catastrophic outliers in the three redshift bins of our analysis.

The cosmic web in an equatorial Mollweide projection

Figure 1. The cosmic web. The black dots are the positions of galaxies from the Legacy-SPHEREx catalogue (LSx) within the redshift range 0.115<z<0.1450.115 < z < 0.145, in an equatorial Mollweide projection. The filaments, voids and clusters of the cosmic web are directly visible. An animated version of this figure can be found at https://calumhrmurray.github.io/spherex-sky.html.

We divide this sample into disc-type galaxies and spheroidal galaxies, using the Sérsic index measurements from the DESI Legacy Imaging Surveys. These two groups are further subdivided into a red sample, composed of dusty galaxies filled with old stars, and a blue sample, composed of star-forming galaxies filled with young stars, using a colour cut in thin redshift slices (Methods, Position and shape samples; Extended Data Figure 2). For our shape samples we use shape measurements from the Tractor; we repeat our measurements with DES Y6 shapes in Extended Data Figure 6 (Methods, Consistent results using DES Y6 shapes), where we show that the two shape measurement methods give consistent results. Example galaxies of each sample are shown in Figure 3. Table 1 shows the number of galaxies in each of these shape samples for each redshift bin; the position samples contain 11.0, 16.1 and 12.1 million galaxies in the three bins. We then measured the cross-correlation of each of these galaxy shape samples with the full galaxy position sample.

Figure 3. An example galaxy from each of the shape samples. Galaxy images using grzgrz bands from the Legacy surveys. Each image is 38′′38^{\prime\prime} on a side, and all galaxies are within the redshift range 0.05<z<0.080.05 < z < 0.08.

Figure 6. The alignment switches sign between voids and clusters, which cannot be explained by the quadratic response. rpwq+r_pw_{q+} around random points in the densest 20% of the density field smoothed on 5h−1 Mpc5h^{-1}\,\mathrm{Mpc}, at 0.20<z<0.350.20 < z < 0.35 (qhq_h, δ>0.25\delta> 0.25; filled points, solid lines) and around the emptiest 20% of the field (qlq_l, δ<−0.28\delta< -0.28; open points, dashed lines). The coloured lines show the best fitting linear response. Black lines show the quadratic response with its amplitude set by the densest 20% at 10h−1 Mpc10h^{-1}\,\mathrm{Mpc}. The models are described in Methods, Alignments around underdense and overdense regions. The quadratic response must have the same alignment pattern around underdense and overdense regions, which is in clear contradiction with our blue disc and red spheroid observations.

0.05<z<0.200.05 < z < 0.200.20<z<0.350.20 < z < 0.350.35<z<0.550.35 < z < 0.55
Blue discs2,797,1431,844,256862,574
Red discs529,710326,499121,670
Blue spheroids270,776332,238426,514
Red spheroids1,185,1481,598,3841,428,209

Table 1. Galaxies in each shape sample. Number of galaxies in each redshift bin.

In Figure 4 we see clear tangential alignment for blue disc galaxies and the opposite, radial alignment for red spheroidal galaxies. Both samples are well described by a linear response to the tidal field (probabilities to exceed, PTE, of 0.58 and 0.59) and poorly described by a quadratic response alone (PTE of 3×10−63 \times10^{-6} and 7×10−257 \times10^{-25}). The quadratic response model has the wrong shape: it peaks near 10h−1 Mpc10h^{-1}\,\mathrm{Mpc} and then decreases too rapidly to explain the shape of the measured correlation function. For each model we fit between 10 and 100 h−1 Mpch^{-1}\,\mathrm{Mpc}. The cut at 10h−1 Mpc10h^{-1}\,\mathrm{Mpc} avoids the regime where structure formation becomes non-linear. We fit wg+w_{g+} and wggw_{gg} at the same time. The models are described in Methods, Modelling linear and quadratic responses to the tidal field, the galaxy auto-correlation measurements in Methods, Galaxy clustering correlation measurements (Extended Data Figure 3), and the fitting in Methods, Model fitting and covariance matrices. The predicted gravitational lensing signal, given the stellar masses of these galaxies and their redshift distribution, is only 3% of the observed blue-disc correlation function. Therefore lensing cannot explain this observation. The inconsistency of lensing with this observation is further explored in Methods, Quantifying contamination from weak lensing.

Figure 4. Blue discs align tangentially and red spheroids align radially. The cross-correlation of galaxy shapes with galaxy positions, rpwg+r_p w_{g+}, for the redshift bin 0.20<z<0.350.20 < z < 0.35. Both samples are well fitted by the linear response (solid line) and poorly fitted by the quadratic response (dashed line). The predicted gravitational lensing signal (dotted line) is very small compared to the observed correlations.

In Figure 5 we show the alignment correlation functions for each colour, morphological type and redshift bin. We use two measures of the detection significance. The first is the χ2\chi^2 of the measured wg+w_{g+} against zero over all scales, 0.4<rp<335h−1 Mpc0.4 < r_p < 335h^{-1}\,\mathrm{Mpc}, which tests the consistency of the observation with no alignment and does not depend on a model. The second is the amplitude of the linear response divided by its uncertainty, ∣A1∣/σA|A_1|/\sigma_A (Table 2).1 For the intermediate and high redshift bins, blue discs show strong tangential alignment signals. For the intermediate redshift bin this is inconsistent with no alignment at 15.7σ15.7\sigma, and for the high redshift bin at 16.2σ16.2\sigma (22.5σ22.5\sigma combined; ∣A1∣/σA=12.4|A_1|/\sigma_A = 12.4 and 11.611.6). In the lowest redshift bin blue disc galaxies show a much smaller alignment signal; as shown in Figure 7, this is possibly because the lowest redshift bin contains the blue disc sample with the lowest mass on average. The intermediate redshift bin for the red discs shows an 8.4σ8.4\sigma detection of tangential alignment, the same sign as the blue discs. However, no signal is seen in the other two redshift bins. Red spheroidal galaxies show strong radial alignments in every redshift bin, at 16.2σ16.2\sigma, 42.4σ42.4\sigma and 46.8σ46.8\sigma (65.0σ65.0\sigma

Figure 5. Alignment signal for each galaxy type in redshift bins. Each of the alignment signals except for the red spheroids at 0.35<z<0.550.35 < z < 0.55 are well modelled by the linear response. The cross component is consistent with a null detection throughout, as expected for both intrinsic alignment and weak lensing measurements at this precision. rpwg+r_{\mathrm{p}}w_{\mathrm{g}+} (coloured points) and the cross component rpwg×r_{\mathrm{p}}w_{\mathrm{g}\times} (grey open points) of blue discs, red discs, blue spheroids and red spheroids (rows) in the three redshift bins (columns). The grey band shows the range of scales used for the model fit, 10≤rp≤100h−1 Mpc10 \le r_{\mathrm{p}} \le100h^{-1}\,\mathrm{Mpc}. Line and band: the linear-response fit and its 68% interval, including the predicted lensing. The errors are estimated using a 100-region jackknife.

Figure 7. The alignment signal strengthens with stellar mass. Here we show the amplitude of the linear response A1A_1 for samples in bins of stellar mass. a, blue discs and b, red spheroids. The marker and the shade of the colour show the redshift bin with the darker shades showing higher redshift. Horizontal bars span 16–84% of each cell’s masses.

Table 2. The alignment amplitudes for every sample with different models. We fit each measurement with six different models which are different combinations of the three parameters A1A_1, A1δA_{1\delta} and A2A_2, with linear galaxy-biasing used throughout. Unless otherwise stated all other parameters are set to zero. Model (1) A1A_1 free; model (2) A1A_1 free with the density weighting fixed at A1δ=bSA1A_{1\delta}=b_SA_1; model (3) A1A_1 and A1δA_{1\delta} free; model (4) A2A_2 free; model (5) A1A_1 and A2A_2 free with A1δ=bSA1A_{1\delta}=b_SA_1; model (6) A1A_1, A1δA_{1\delta} and A2A_2 all free. The density-sample bias bDb_D is fitted jointly, constrained by wggw_{gg}, and the lensing term wgGw_{gG} is fixed at its prediction. Amplitudes are posterior means and standard deviations. bSb_S is the measured galaxy bias for each shape sample. ∣A1∣/σA|A_1|/\sigma_A is the detection significance from model (1). χν2\chi^2_\nu is evaluated at the best fit over the five bins with 10≤rp≤100h−1 Mpc10 \le r_p \le100h^{-1}\,\mathrm{Mpc}, using the jackknife covariance with the Hartlap correction, with ν\nu the number of bins minus the number of free amplitudes; the PTE is the probability of a χ2\chi^2 at least this large for ν\nu degrees of freedom. Rows with ∣A1∣/σA≥5|A_1|/\sigma_A \ge5 are shaded: green where the model is acceptable (PTE ≥0.05\ge0.05) and red where it is not.

SamplezzbSb_S∣A1∣/σA|A_1|/\sigma_A(1) A1A_1χ42\chi^2_4 (PTE)(2) A1A_1, A1δ=bSA1A_{1\delta}=b_SA_1χ42\chi^2_4 (PTE)
Blue discs0.05–0.200.960.7−0.15±0.21-0.15\pm0.210.3 (0.99)−0.11±0.13-0.11\pm0.130.2 (0.99)
0.20–0.350.9712.4−2.81±0.23-2.81\pm0.232.9 (0.58)−2.00±0.16-2.00\pm0.164.5 (0.34)
0.35–0.551.1211.6−3.79±0.33-3.79\pm0.332.3 (0.68)−2.75±0.24-2.75\pm0.245.6 (0.23)
Red discs0.05–0.201.260.6+0.40±0.63+0.40\pm0.630.8 (0.94)+0.23±0.37+0.23\pm0.370.8 (0.93)
0.20–0.351.255.9−4.32±0.74-4.32\pm0.742.8 (0.59)−2.62±0.48-2.62\pm0.487.8 (0.10)
0.35–0.551.531.4−1.34±0.98-1.34\pm0.980.9 (0.93)−0.72±0.63-0.72\pm0.631.4 (0.84)
Blue spheroids0.05–0.200.961.1−0.70±0.61-0.70\pm0.612.9 (0.58)−0.56±0.39-0.56\pm0.392.1 (0.71)
0.20–0.351.111.3+0.52±0.40+0.52\pm0.400.4 (0.99)+0.34±0.28+0.34\pm0.280.6 (0.97)
0.35–0.551.270.2+0.07±0.43+0.07\pm0.433.0 (0.57)−0.02±0.30-0.02\pm0.303.0 (0.56)
Red spheroids0.05–0.201.3812.2+2.69±0.22+2.69\pm0.221.1 (0.90)+1.45±0.12+1.45\pm0.128.1 (0.09)
0.20–0.351.5122.9+3.72±0.16+3.72\pm0.162.8 (0.59)+2.13±0.10+2.13\pm0.1033.4 (0.00)
0.35–0.551.8227.1+5.25±0.19+5.25\pm0.1917.9 (0.00)+3.26±0.12+3.26\pm0.1240.6 (0.00)
(3) A1A_1, A1δA_{1\delta}(4) A2A_2
SamplezzA1A_1A1δA_{1\delta}χ32\chi^2_3 (PTE)A2A_2χ42\chi^2_4 (PTE)
Blue discs0.05–0.20−0.02±0.38-0.02\pm0.38−0.3±0.6-0.3\pm0.60.2 (0.98)−0.3±0.3-0.3\pm0.30.1 (1.00)
0.20–0.35−2.54±0.38-2.54\pm0.38−0.7±0.8-0.7\pm0.82.1 (0.56)−4.3±0.4-4.3\pm0.431.0 (0.00)
0.35–0.55−3.52±0.46-3.52\pm0.46−0.9±1.1-0.9\pm1.11.7 (0.63)−6.0±0.6-6.0\pm0.639.1 (0.00)
Red discs0.05–0.20+0.36±0.91+0.36\pm0.91+0.1±1.4+0.1\pm1.40.8 (0.85)+0.4±0.9+0.4\pm0.91.0 (0.91)
0.20–0.35−5.48±1.22-5.48\pm1.22+3.1±2.5+3.1\pm2.51.4 (0.72)−5.6±1.2-5.6\pm1.216.4 (0.00)
0.35–0.55−2.14±1.57-2.14\pm1.57+2.5±3.8+2.5\pm3.80.4 (0.93)−1.4±1.7-1.4\pm1.72.1 (0.72)
Blue spheroids0.05–0.20+0.32±1.00+0.32\pm1.00−1.9±1.4-1.9\pm1.41.2 (0.74)−1.3±0.8-1.3\pm0.81.5 (0.83)
0.20–0.35+0.61±0.60+0.61\pm0.60−0.2±1.2-0.2\pm1.20.3 (0.96)+0.6±0.7+0.6\pm0.71.1 (0.89)
0.35–0.55+0.40±0.64+0.40\pm0.64−1.2±1.5-1.2\pm1.52.4 (0.50)−0.3±0.8-0.3\pm0.82.9 (0.58)
Red spheroids0.05–0.20+2.43±0.37+2.43\pm0.37+0.5±0.6+0.5\pm0.60.3 (0.97)+3.3±0.3+3.3\pm0.330.9 (0.00)
0.20–0.35+3.53±0.27+3.53\pm0.27+0.4±0.5+0.4\pm0.52.1 (0.55)+5.1±0.2+5.1\pm0.2119.4 (0.00)
0.35–0.55+4.65±0.28+4.65\pm0.28+2.1±0.7+2.1\pm0.78.9 (0.03)+8.8±0.4+8.8\pm0.4216.6 (0.00)
(5) A1A_1, A1δ=bSA1A_{1\delta}=b_SA_1, A2A_2(6) A1A_1, A1δA_{1\delta}, A2A_2
SamplezzA1A_1A2A_2χ32\chi^2_3 (PTE)A1A_1A1δA_{1\delta}A2A_2χ22\chi^2_2 (PTE)
Blue discs0.05–0.20+0.07±0.56+0.07\pm0.56−0.4±1.2-0.4\pm1.20.1 (0.99)−0.20±0.85-0.20\pm0.85+0.9±4.2+0.9\pm4.2−1.3±4.6-1.3\pm4.60.1 (0.95)
0.20–0.35−2.90±0.53-2.90\pm0.53+2.2±1.2+2.2\pm1.21.4 (0.71)−3.24±0.80-3.24\pm0.80−5.7±4.9-5.7\pm4.9+5.0±4.9+5.0\pm4.91.0 (0.61)
0.35–0.55−4.06±0.66-4.06\pm0.66+3.5±1.7+3.5\pm1.71.1 (0.77)−4.39±1.11-4.39\pm1.11−7.5±7.8-7.5\pm7.8+6.2±7.2+6.2\pm7.21.0 (0.61)
Red discs0.05–0.20+0.67±1.35+0.67\pm1.35−1.1±3.1-1.1\pm3.10.7 (0.87)+2.19±2.34+2.19\pm2.34+10.9±13.0+10.9\pm13.0−11.7±13.9-11.7\pm13.90.1 (0.95)
0.20–0.35−7.05±1.83-7.05\pm1.83+11.6±4.6+11.6\pm4.61.6 (0.66)−6.02±2.60-6.02\pm2.60−0.4±15.7-0.4\pm15.7+3.5±15.6+3.5\pm15.61.3 (0.52)
0.35–0.55−3.12±2.34-3.12\pm2.34+6.9±6.4+6.9\pm6.40.3 (0.96)−3.28±3.65-3.28\pm3.65+6.5±26.3+6.5\pm26.3+8.8±24.0+8.8\pm24.00.3 (0.86)
Blue spheroids0.05–0.20+0.70±1.35+0.70\pm1.35−2.7±2.8-2.7\pm2.81.2 (0.76)+0.70±1.97+0.70\pm1.97+0.7±10.9+0.7\pm10.9−2.8±11.6-2.8\pm11.61.2 (0.55)
0.20–0.35+0.73±0.81+0.73\pm0.81−1.0±2.0-1.0\pm2.00.3 (0.96)+0.65±1.23+0.65\pm1.23+0.1±8.8+0.1\pm8.8−0.3±8.5-0.3\pm8.50.3 (0.85)
0.35–0.55+0.71±0.86+0.71\pm0.86−2.0±2.3-2.0\pm2.32.2 (0.53)+1.00±1.25+1.00\pm1.25+4.2±10.0+4.2\pm10.0−4.9±9.0-4.9\pm9.02.1 (0.35)
Red spheroids0.05–0.20+2.77±0.51+2.77\pm0.51−3.3±1.2-3.3\pm1.20.9 (0.82)+2.44±0.65+2.44\pm0.65+0.6±4.1+0.6\pm4.1−0.1±4.2-0.1\pm4.20.3 (0.87)
0.20–0.35+4.18±0.39+4.18\pm0.39−5.5±1.0-5.5\pm1.04.1 (0.25)+3.55±0.60+3.55\pm0.60+0.6±4.1+0.6\pm4.1−0.1±4.0-0.1\pm4.02.1 (0.35)
0.35–0.55+5.51±0.38+5.51\pm0.38−7.2±1.2-7.2\pm1.20.8 (0.84)+5.74±0.46+5.74\pm0.46+13.3±3.8+13.3\pm3.8−10.0±3.4-10.0\pm3.40.1 (0.97)

combined; ∣A1∣/σA=12.2|A_1|/\sigma_A = 12.2, 22.922.9 and 27.127.1), while blue spheroids show no alignment in any bin.

The cross component wg×w_{\mathrm{g}\times}, which is expected to be very small for intrinsic alignments, is consistent with zero for every measurement. A non-zero ×\times component would indicate systematic issues with the measurement, therefore this is a reassuring quality check. There are many systematic issues that would produce only a ++ signal; these are tested in Methods (Correlation functions, Quantifying contamination from weak lensing and Consistent results using DES Y6 shapes), where we find no evidence of significant systematic effects that can change the conclusions of this work.

The linear response alone provides a good fit to every measurement except the red spheroids at 0.35<z<0.550.35 < z < 0.55, with probabilities to exceed between 0.57 and 0.99 elsewhere. We discuss the red spheroids at 0.35<z<0.550.35 < z < 0.55 in the Discussion and in Methods, Alignment amplitudes for discs and spheroids. The quadratic response alone is rejected for every measurement with a significant detection. For the blue spheroids we do not detect an alignment signal for any of the redshift bins. The red spheroids in our highest redshift bin are poorly fit by the linear model (PTE 0.001), but they are fitted well when the linear matter power spectrum is replaced with the non-linear matter power spectrum (the Non-linear Linear Alignment (NLA) model (Bridle and King [7]); PTE 0.12). Therefore we expect that the linear model that we use breaks down, within this redshift bin, which is the redshift bin containing the most massive galaxies within our sample (given the flux-limit of the galaxy sample).

The tangential alignment of blue discs cannot be explained by gravitational lensing as discussed in Methods, Quantifying contamination from weak lensing.

The alignment switches from tangential to radial between clusters and voids

We can provide further evidence that the response is linear in the tidal field by considering how galaxies align around overdense and underdense regions. A quadratic response must produce alignments of the same sign around overdense and underdense regions, whereas for a linear response tangential alignments around clusters would imply radial alignments around voids. We test this directly with our observations. We estimate the galaxy density field, δg=ng/nˉ−1\delta_g = n_g/\bar{n} - 1, from all galaxies of the position sample at 0.20<z<0.350.20 < z < 0.35, smoothed with a Gaussian of 5h−1 Mpc5h^{-1}\,\mathrm{Mpc} across the line of sight and 30h−1 Mpc30h^{-1}\,\mathrm{Mpc} along the line of sight. We then place 2 million random positions within the field, and select those within the top 20% of density values, qhq_h, and those within the lowest 20%, qlq_l. We then measure the alignment signals wqh+w_{q_h+} and wql+w_{q_l+} around each of these position samples in the same manner as for galaxy positions (Methods, Alignments around underdense and overdense regions). These correlation functions are presented in Figure 6.

The sign of the alignment signal switches between underdensities and overdensities for each of the samples with a detected alignment signal. Red spheroids align tangentially around voids and radially around clusters. Blue discs align radially around voids and tangentially around clusters. It is not possible to explain these observations with a quadratic response or with density weighting; therefore we can rule out that the alignments of disc galaxies are dominated by a physical mechanism whose alignment keeps the same sign when the density field changes sign. Fitting both regions together, the linear response describes the blue discs with A1=−2.43±0.22A_1 = -2.43 \pm0.22 (PTE 0.80), whereas the quadratic response alone does not (PTE 7×10−237 \times10^{-23}). The model used here is explained in Methods, Alignments around underdense and overdense regions.

The alignment is stronger at higher stellar mass

We can now extend the study further to look at the dependence of the alignment amplitudes on galaxy stellar mass. In order to estimate stellar masses we ran CIGALE ([6]) on the LSx catalogue, using the DESI Legacy Imaging photometric bands and the SPHEREx spectrophotometry (Methods, Stellar masses).

In Figure 7 we see that both the blue disc and red spheroid alignments are strongly correlated with galaxy stellar mass. The red spheroids show a plateau at low stellar masses and a steep rise at the high stellar mass end, in agreement with multiple other studies (Hirata et al., 2007; Joachimi et al., 2011; Singh et al., 2015; Hervas Peters et al., 2025; Paviot et al., 2026). The blue disc tangential alignment increases significantly at high stellar mass. The blue disc sample does not reach the high stellar masses of the red spheroid sample. We find that for both discs and spheroids the most massive galaxies have the strongest alignment signals.

Discussion

The significant detection of the tangential alignments of blue disc galaxies that we observe in this work does not have an adequate explanation from theoretical models or hydrodynamical simulations. The amplitude of alignment is large, A1=−2.81±0.23A_1 = -2.81 \pm0.23 and −3.79±0.33-3.79 \pm0.33 at 0.20<z<0.350.20 < z < 0.35 and 0.35<z<0.550.35 < z < 0.55, and comparable in size to the amplitude of radial alignments of red spheroids. The measurements cannot be explained by redshift outliers, gravitational lensing contamination or the shape measurement method (Methods: Quantifying contamination from weak lensing and Consistent results using DES Y6 shapes). The blue alignment signal is well described by a linear response to the tidal field. The alignment switches sign between underdensities and overdensities, which provides further evidence that the mechanism creating these alignments produces a linear response to the tidal field, and not a quadratic response, which has previously been the theoretically preferred model for the alignment of disc galaxies ([8] (Lee and Pen 2000; Hirata and Seljak 2004)).

Past measurements of the alignments of blue galaxies found no alignment (Mandelbaum et al., 2011; Johnston et al., 2019; Yao et al., 2020; Georgiou et al., 2025; Siegel et al., 2026), until Paviot et al. (2026) measured a 3.3σ3.3\sigma hint of tangential alignment for massive blue galaxies. There are two reasons for this. First, the large spectroscopic surveys used for direct alignment measurements target red galaxies, so samples of blue galaxies with precise redshifts have been small. Second, blue samples have been selected by colour alone, without a morphological selection, which dilutes the alignment signal of the discs with that of other blue galaxies. As the LSx catalogue has no colour selection we are able to probe the alignment of any type of galaxy, provided that they are bright enough to enter the sample. We see this dilution in Extended Data Figure 7: when the shape sample is selected on galaxy colour alone, the alignment amplitude is significantly smaller. For the redshift bin 0.20<z<0.350.20 < z < 0.35 all resolved blue galaxies have an alignment amplitude of A1=−2.06±0.19A_1 = -2.06 \pm0.19, compared with −2.81±0.23-2.81 \pm0.23 for the blue disc galaxies. Therefore it is specifically blue disc galaxies which show the strongest alignment signals. Our blue galaxy sample is large enough, and our redshifts are precise enough that even when measuring the ensemble of blue galaxies we have a significant detection of tangential alignments.

Extended Data Figure 7. The blue galaxy alignment is smaller than blue disc alignment. rpwg+r_pw_{g+} for blue discs (dark blue circles), all well-resolved blue galaxies (light blue diamonds) and all of the blue galaxies with res >> 0.5 (grey triangles), in each redshift bin. We see that in each case the intrinsic alignment signal, when detected, is strongest for blue discs followed by all blue galaxies and then the smallest least resolved galaxies (grey triangles).

The detection of the linear tangential alignments of disc galaxies contradicts classic tidal torque theory ([8] (Lee and Pen 2000; Hirata and Seljak 2004)), in which the alignment is quadratic in the Gaussian tidal field and its cross-correlation with galaxy positions should only give a weak quadratic alignment (Hirata and Seljak, 2004). This is also in contradiction with hydrodynamical simulations, in which disc galaxies show either no large-scale alignment or radial alignments ([17] (Tenneti et al. 2016; Hilbert et al. 2017; Samuroff et al. 2021; Zijupa et al. 2022; van Heukelum and Chisari 2026)) with the exception of the Horizon-AGN simulations ([11, 12]).

Significant spin alignments of blue discs with their host filament have been found in both observations (Tempel and Libeskind, 2013; Welker et al., 2020) and simulations ([14, 13] (Kraljic et al. 2020)). Low-mass disc galaxies have spins aligned to their filament, and more massive galaxies have perpendicular spins ([13] (Welker et al. 2020)). The disc major axis is perpendicular to the disc spin. As filaments are connected to clusters, this would naively lead to tangential alignments for low-mass discs and radial alignments for high-mass discs, the converse of what we observe (Figure 7).

Therefore the physical mechanism responsible for such strong tangential alignments in disc galaxies is unclear. The measurements do not exclude more complicated tidal-torquing mechanisms. For example, non-linear evolution can produce an effective linear response from tidal-torquing ([4] (Hui and Zhang 2008)). On large-scales this cannot be distinguished from the linear response that we measure. Therefore torquing may be the cause of the linear alignment that we observe (Hui and Zhang, 2008; Larsen and Challinor, 2016; Blazek et al., [4]. Blazek et al. [4] does not make predictions for the sign of this physical mechanism and Larsen and Challinor (2016) found that it should be radial, in contradiction with our observations. Furthermore this effect has not be found in hydrodynamical simulations, so this remains a tentative explanation. More theoretical research is needed to understand the sign, amplitude and mass dependence of these results.

These results are important for future weak gravitational lensing measurements. The majority of source galaxies for weak lensing in Euclid (Mellier et al., 2025), LSST (Ivez̆ić et al., 2019) and Roman (Spergel et al., 2015) will be blue star-forming galaxies. We have shown that large blue disc galaxies align with an amplitude comparable to that of red spheroids. In Extended Data Figure 6 (Methods, Consistent results using DES Y6 shapes) we also see that small blue galaxies align radially. Therefore the intrinsic alignment signal within a lensing survey will depend sensitively upon the source selection of galaxies. It is therefore essential that intrinsic alignment models are calibrated for the specific sample used for the lensing sources, in terms of survey depth, galaxy size and galaxy morphology. For surveys as deep as Euclid (Mellier et al., 2025), LSST (Ivez̆ić et al., 2019) and Roman (Spergel et al., 2015) this will be observationally challenging. Alternatively we need to develop a much deeper physical understanding of how disc galaxies acquire the large tangential alignments that we have found in this work.

Extended Data Figure 6. An independent shape measurement gives the same alignments. rpwg+r_p w_{g+} measured with DES Y6 galaxy shapes. Top: blue discs matched between DES Y6 and the Tractor, with Tractor (filled) and DES Y6 (open) shapes of the same galaxies; the amplitudes are consistent in each bin. Middle: blue galaxies passing the DES Y6 shape cuts, without the morphological or size cuts of the Tractor sample. The alignment amplitude is much smaller, but still significant in the highest redshift bin. Bottom: the galaxies of the middle row that are not in the Tractor shape sample. These galaxies show significant radial alignments in the low and intermediate redshift bins.

Methods

The Legacy Surveys–SPHEREx redshift catalogue

To create a high quality redshift catalogue (Legacy x SPHEREx, LSx catalogue) we select sources using the DESI Legacy Imaging Surveys and perform forced photometry in SPHEREx images for each source. The DESI Legacy Imaging Surveys (Dey et al., [19]) (DR9 north, DR10 south and DR11 south) combine imaging from several telescopes, processed in a consistent manner using the Tractor (Lang et al., 2016). As noted in the main text, our selection is deliberately simple: we select sources with dereddened r<20.5r < 20.5, at least one observation in each of the gg, rr and zz bands, and we exclude sources that are identified by the Tractor as unresolved (PSF morphological type). This last cut primarily removes stars, and more than halves the number of sources to be measured. SPHEREx (Doré et al., [20]; Bock et al., [5]) is a near-infrared satellite that surveys the entire sky in 102 bands between 0.750.75–5.0 μm5.0\,\mu\mathrm{m}. By combining the Legacy imaging source selection with SPHEREx we can measure precise redshifts. For this catalogue we used SPHEREx observations between 24 April 2025 and 20 July 2026.

The LSx catalogue was validated using spectroscopic redshifts from DESI DR1 (DESI Collaboration et al., [18]), SDSS (Abdurro’uf et al., [1]), GAMA (Driver et al., [22]), 6dFGS (Jones et al., 2004, 2009) and 2dF-GRS (Colless et al., [15]). Before redshift quality cuts the sample has a redshift uncertainty of σNMAD=0.0065\sigma_{\mathrm{NMAD}} = 0.0065 and an outlier rate of 0.46%0.46\%, where outliers are defined as galaxies with ∣Δz∣>0.15|\Delta z| > 0.15 and Δz=(zest−zspec)/(1+zspec)\Delta z = (z_{\mathrm{est}} - z_{\mathrm{spec}})/(1 + z_{\mathrm{spec}}).

Position and shape samples

Position sample. To define a high-quality position sample for galaxy clustering, we first apply the redshift quality cut, keeping the 90% of galaxies with the narrowest redshift probability distributions (63.0 million galaxies). We then remove regions of high Galactic extinction (E(B−V)>0.10E(B-V) > 0.10 (Chiang, [9])), high stellar density (above 2000 deg−2^{-2} for bright Gaia sources G<16.5G < 16.5), and regions within 8∘8^\circ and 5∘5^\circ of the Large and Small Magellanic Clouds. We do not include the DR11 south region of the Legacy Imaging Surveys as we found higher contamination of the catalogue within this area. After removing these regions, the clustering sample contains 39.2 million galaxies with 0.05≤zest<0.550.05 \le z_{\mathrm{est}} < 0.55, over 19,312 deg2^2. We create systematic weights for each galaxy using the iterative systematics decontamination method (Rodríguez-Monroy et al., 2022). This removes the linear dependence of the galaxy density on E(B−V)E(B-V), stellar density, the gg, rr and zz depths and the rr-band seeing. The probability that a given galaxy passes the quality cut of the clustering catalogue selection depends upon the SPHEREx depth, which varies significantly across the sky, due to the scanning strategy of SPHEREx. To correct for this we also weight the randoms with a linear correction based upon the SPHEREx depth which removes this dependence.

Shape samples. For the shape sample we select galaxies identified as Tractor SER, EXP or DEV sources, with a finite ellipticity ∣e∣<1|e| < 1 and a resolved shape, res≡rhalf/FWHMPSF>1\mathrm{res} \equiv r_{\mathrm{half}}/\mathrm{FWHM}_{\mathrm{PSF}} > 1, where rhalfr_{\mathrm{half}} is the half-light radius of the Tractor model (SHAPE_R) and FWHMPSF_{\mathrm{PSF}} is the full width at half maximum of the rr-band point-spread function at the position of the source (PSFSIZE_R). This leaves us with 5.61, 4.78 and 3.21 million galaxies in each redshift bin. We select a disc sample, as galaxies with either EXP morphological type or SER models with Sérsic index n<1.5n < 1.5. We select a spheroid sample, as galaxies with either DEV morphological type or SER models with Sérsic index n≥2.5n \ge2.5. SER models with 1.5≤n<2.51.5 \le n < 2.5, 15, 14 and 12% of the shape sample in the three redshift bins, are in neither sample, so the four samples of Table 1 contain 4.78, 4.10 and 2.84 million galaxies. Each sample is then split into red and blue using the dereddened g−rg-r colour (Extended Data Figure 2).

Extended Data Figure 2. The red-blue galaxy separation is clear for LSx. The dereddened g−rg-r colour against rr magnitude for every galaxy in each Δz=0.02\Delta z = 0.02 slice of redshift (grey, logarithmic counts). The red sequence is shown with the dotted orange line. The cut we use between the blue and red galaxies is parallel to the red-sequence.

The number of galaxies in each sample is given in Table 1. We show the n(z)n(z) for each of the shape samples in Extended Data Figure 1.

Extended Data Figure 1. The LSx redshifts of the shape samples are precise, with few catastrophic outliers. Each panel shows the redshift distribution for each of the shape samples with a spectroscopic match, where the redshift distribution is shown using the matched spectroscopic redshift.

Ellipticities. We use the Tractor ellipticities SHAPE_E1 and SHAPE_E2. The Tractor forward models each source with different light profiles; a point source, a round exponential, an exponential, a de Vaucouleurs or a Sérsic profile, each convolved with the point-spread function. If a source is well-fit by a round exponential then no shape is measured, therefore these sources are excluded from our catalogue. With this forward-modelling method the shapes are corrected for the point-spread function, but they are not calibrated for weak lensing shear. We show in Consistent results using DES Y6 shapes that our results are robust against the shape measurement method used. An additional issue is that the Tractor chooses the round exponential when an ellipticity does not lead to a significantly better fit, which is primarily the case for small and faint sources. These objects make up 5–7% of our shape samples. Their exclusion means that our selection has a small dependence upon ∣e∣|e|. This could lead to a small change in the measured amplitude, but it will not affect the detection. For example, with DES Y6 shapes, which do not exclude round galaxies, we still obtain significant tangential alignments for blue galaxies (Extended Data Figure 6).

Stellar masses

We estimate stellar masses for the LSx catalogue using CIGALE ([6]). CIGALE estimates the galaxy stellar mass using spectral energy distribution models based upon the galaxy’s star-formation history, stellar populations, nebular emission, dust attenuation and dust emission. We use the same models that were used for the CIGALE value-added catalogue of the DESI Early Data Release ((Siudek et al., 2024)). The stellar mass fit succeeds for 97% of the quality sample. Compared with the DESI DR1 CIGALE catalogue ([18]), our stellar masses have a median offset of +0.04 dex and a scatter of σNMAD=0.16\sigma_{\mathrm{NMAD}} = 0.16 dex.

Correlation functions

In this work we measure galaxy position-shape correlations and galaxy position–position correlations. For the position-shape correlations we use the same estimator as Singh et al. (2015),

ξg+(rp,Π)=S+D−S+RSR,(4)\xi_{g+}(r_p,\Pi) = \frac{S_{+D}-S_{+R}}{S_R}, \tag*{(4)}

where S+DS_{+D} is the sum of e+e_+ over shape–position pairs at projected separation rpr_p and line-of-sight separation Π\Pi, S+RS_{+R} is the same sum around random positions, and SRS_R is the shape random counts. The subtraction of S+RS_{+R} removes potential additive shape measurement systematics; we find that our S+RS_{+R} measurements are consistent with zero on the scales that we fit.

For the galaxy autocorrelation, wggw_{gg}, we use the Landy–Szalay estimator ((Landy and Szalay, 1993))

ξgg(rp,Π)=DD−2DR+RRRR,(5)\xi_{gg}(r_p,\Pi) = \frac{DD-2DR+RR}{RR}, \tag*{(5)}

where DDDD, DRDR and RRRR are the position–position, position–random and random–random pair counts at a separation (rp,Π)(r_p,\Pi), each normalised by the weighted total number of pairs. We then integrate each of the correlation functions over the line-of-sight between −Πmax⁡-\Pi_{\max} and Πmax⁡\Pi_{\max} with Πmax⁡=100h−1 Mpc\Pi_{\max}=100h^{-1}\,\mathrm{Mpc}, to measure the projected correlation function wg+(rp)w_{g+}(r_p) and wgg(rp)w_{gg}(r_p).

We measure each correlation function with 15 logarithmic bins in physical separation rpr_p, from 0.3 to 427 h−1 Mpc427\,h^{-1}\,\mathrm{Mpc}. For each of the correlation functions we use TreeCorr ((Jarvis et al., 2004)). The position sample is weighted using the systematic weights described in Position and shape samples. The shape sample is unweighted. Using inverse-variance weights for the shape sample changes the inferred amplitude A1A_1 of the blue discs at 0.20<z<0.350.20<z<0.35 by 6×10−46\times10^{-4}. This is because the measurement uncertainty, ∼0.004\sim0.004, is much smaller than the intrinsic scatter, ≈0.30\approx0.30.

For each correlation function measurement we estimate the covariance using a jackknife with 100 angular regions. The same angular regions are used for every correlation function, so that we can correctly calculate the covariance between different measurements. We correct the inverse covariances using the Hartlap factor ((Hartlap et al., 2007)).

Modelling linear and quadratic responses to the tidal field

In this work we use the tidal alignment and tidal torquing (TATT) model ([4]) to model our observations. As noted in the main body of the text, this model has a somewhat confusing name, but correctly models the physics that we are interested in.

In Eq. (1), we sketched out each of the terms of our model. Following Blazek et al. [4], the full expression is,

γij=C1sij+C1δ δsij+C2(sikskj−13δijsklskl),(6)\gamma_{ij}=C_1s_{ij}+C_{1\delta}\,\delta s_{ij}+C_2\left(s_{ik}s_{kj}-\frac{1}{3}\delta_{ij}s_{kl}s_{kl}\right), \tag*{(6)}

where δ\delta is the matter density, δij\delta_{ij} the Kronecker delta, repeated indices are summed, and sijs_{ij} is the tidal field. In Fourier space the tidal field is sij(k)=(kikj/k2−δij/3)δ(k)s_{ij}(\mathbf{k})=(k_i k_j/k^2-\delta_{ij}/3)\delta(\mathbf{k}). The first term, C1sijC_1s_{ij}, is the linear response of observed galaxy shapes to the tidal field. The second term, C1δ δsijC_{1\delta}\,\delta s_{ij}, arises because we necessarily sample the shape field at the positions of galaxies, so the tidal field is weighted by the local density. The third term, C2sikskjC_2s_{ik}s_{kj}, is the quadratic response. Following Blazek et al. [4], we can write the three coefficients in terms of dimensionless amplitudes,

C1=−A1C~1ρcritΩmD(z),C_1=-\frac{A_1\widetilde{C}_1\rho_{\mathrm{crit}}\Omega_m}{D(z)},
C1δ=−A1δC~1ρcritΩmD(z),C_{1\delta}=-\frac{A_{1\delta}\widetilde{C}_1\rho_{\mathrm{crit}}\Omega_m}{D(z)},
C2=5A2C~1ρcritΩm2D2(z),(7)C_2=\frac{5A_2\widetilde{C}_1\rho_{\mathrm{crit}}\Omega_m}{2D^2(z)}, \tag*{(7)}

where the parameters that we choose to measure in our analyses are A1A_1, A1δA_{1\delta} and A2A_2. D(z)D(z) is the linear growth factor, normalised to unity today, and C~1ρcrit=0.0134\widetilde{C}_1\rho_{\mathrm{crit}}=0.0134. Throughout our analysis we use different combinations of A1A_1, A1δA_{1\delta} and A2A_2 as free parameters. If the A1δA_{1\delta} term is generated purely by density weighting then we can set A1δ=bSA1A_{1\delta}=b_SA_1, where bSb_S is the linear galaxy bias for the shape sample. However, like A1A_1 and A2A_2, A1δA_{1\delta} may be induced by other physical mechanisms, so we explore both A1δ=bSA1A_{1\delta}=b_SA_1 and A1δA_{1\delta} as a free parameter.

The wg+w_{g+} correlations that we measure can be modelled by projecting the cross-power spectrum of the matter density and the E-mode of the shape field. Following Blazek et al. [4], we have

PδE(k,μ)=(1−μ2)[C1Pδ(k)⏟linear+C1δ(A0∣0E+C0∣0E)(k)⏟density weighting+C2(A0∣E2+B0∣E2)(k)⏟quadratic],(8)P_{\delta E}(k,\mu)=(1-\mu^2)\left[\underbrace{C_1P_\delta(k)}_{\text{linear}}+\underbrace{C_{1\delta}\left(A_{0|0E}+C_{0|0E}\right)(k)}_{\text{density weighting}}+\underbrace{C_2\left(A_{0|E2}+B_{0|E2}\right)(k)}_{\text{quadratic}}\right], \tag*{(8)}

where for Pδ(k)P_\delta(k) we use the linear power spectrum, μ=k∥/k\mu=k_\parallel/k is the cosine of the angle between kk and the line of sight, and the factor 1−μ21-\mu^2 arises from the projection of the tidal field into the plane of the sky. The full expressions for the density-weighting and quadratic terms can be found directly in Blazek et al. [4].

For a position sample with linear bias bgb_g, each of the three terms of Eq. (8) gives a contribution to the three-dimensional correlation between galaxy positions and the ++ component of galaxy shapes.

ξg+(rp,Π,z)=bgCˉ1ρcritΩm[A1D(z)ξlin+(rp,Π)+A1δD3(z)ξden+(rp,Π)−5A2D2(z)ξquad+(rp,Π)],(9)\xi_{g+}(r_p,\Pi,z)=b_g\bar{C}_1\rho_{\mathrm{crit}}\Omega_m\left[A_1D(z)\xi_{\mathrm{lin}}^+(r_p,\Pi)+A_{1\delta}D^3(z)\xi_{\mathrm{den}}^+(r_p,\Pi)-5A_2D^2(z)\xi_{\mathrm{quad}}^+(r_p,\Pi)\right], \tag*{(9)}

where ξlin+\xi_{\mathrm{lin}}^+, ξden+\xi_{\mathrm{den}}^+ and ξquad+\xi_{\mathrm{quad}}^+ can be calculated from their respective power spectra with Singh et al. (2015),

ξX+(rp,Π)=∫0∞k⊥3 dk⊥2πJ2(k⊥rp)∫−∞∞dk∥2πcos⁡(k∥Π)PX(k)k2,(10)\xi_X^+(r_p,\Pi)=\int_0^\infty\frac{k_\perp^3\,\mathrm{d}k_\perp}{2\pi}J_2(k_\perp r_p)\int_{-\infty}^{\infty}\frac{\mathrm{d}k_\parallel}{2\pi}\cos(k_\parallel\Pi)\frac{P_X(k)}{k^2}, \tag*{(10)}

J2J_2 is the Bessel function of order two, and k⊥2/k2=1−μ2k_\perp^2/k^2=1-\mu^2 is the projection factor of Eq. (8).

Then we project these to obtain the projected correlation functions which we use to constrain our intrinsic alignment models. Following Mandelbaum et al. (2011),

wg+(rp)=∫dz W(z)∫dΠ p(Π∣z)ξg+(rp,Π,z)+wgG(rp),(11)w_{g+}(r_p)=\int\mathrm{d}z\,W(z)\int\mathrm{d}\Pi\,p(\Pi|z)\xi_{g+}(r_p,\Pi,z)+w_{gG}(r_p), \tag*{(11)}
wgg(rp)=∫dz Wgg(z)∫dΠ pgg(Π∣z)ξgg(rp,Π,z).(12)w_{gg}(r_p)=\int\mathrm{d}z\,W_{gg}(z)\int\mathrm{d}\Pi\,p_{gg}(\Pi|z)\xi_{gg}(r_p,\Pi,z). \tag*{(12)}

where W(z)∝nS(z)nD(z)/(dVc/dz)W(z)\propto n_S(z)n_D(z)/(\mathrm{d}V_c/\mathrm{d}z), which is normalised to 1, nSn_S and nDn_D are the redshift distributions of the shape and position samples and VcV_c is the comoving volume, Wgg(z)∝nD2(z)/(dVc/dz)W_{gg}(z)\propto n_D^2(z)/(\mathrm{d}V_c/\mathrm{d}z). p(Π∣z)p(\Pi|z) is the probability that a shape–position pair at redshift zz and true line-of-sight separation Π\Pi has an observed separation ∣Πobs∣<Πmax⁡|\Pi_{\mathrm{obs}}|<\Pi_{\max}, and pggp_{gg} is the same for position–position pairs. We estimate the product W(z) p(Π∣z)W(z)\,p(\Pi|z) directly using Monte Carlo sampling. We draw random shape–position pairs from the two samples, with the position galaxies weighted by their systematic weights, and give each galaxy an independently matched redshift error ztrue−zestz_{\mathrm{true}}-z_{\mathrm{est}}, drawn from spectroscopically matched galaxies of similar estimated redshift, magnitude and colour. We keep the pairs with ∣Πobs∣≤Πmax⁡|\Pi_{\mathrm{obs}}|\leq\Pi_{\max}, weighted by 1/χˉ21/\bar{\chi}^2, where χˉ\bar{\chi} is the comoving distance at the observed redshifts. From this we can measure the histogram in bins of true redshift and true line-of-sight separation. As the redshift errors are drawn from the measured distribution, catastrophic outliers and non-Gaussian tails are included within this method. wgGw_{gG} is the galaxy–galaxy lensing term, for which we follow Georgiou et al. (2025). From Eq. (11) we can define the LL, DwD_w and QQ terms from the main text,

L(rp)=Cˉ1ρcritΩm∫dz W(z)D(z)∫dΠ p(Π∣z)ξlin+(rp,Π),(13)L(r_p)=\bar{C}_1\rho_{\mathrm{crit}}\Omega_m\int\mathrm{d}z\,W(z)D(z)\int\mathrm{d}\Pi\,p(\Pi|z)\xi_{\mathrm{lin}}^+(r_p,\Pi), \tag*{(13)}
Dw(rp)=Cˉ1ρcritΩm∫dz W(z)D3(z)∫dΠ p(Π∣z)ξden+(rp,Π),(14)D_w(r_p)=\bar{C}_1\rho_{\mathrm{crit}}\Omega_m\int\mathrm{d}z\,W(z)D^3(z)\int\mathrm{d}\Pi\,p(\Pi|z)\xi_{\mathrm{den}}^+(r_p,\Pi), \tag*{(14)}
Q(rp)=−5Cˉ1ρcritΩm∫dz W(z)D2(z)∫dΠ p(Π∣z)ξquad+(rp,Π).(15)Q(r_p)=-5\bar{C}_1\rho_{\mathrm{crit}}\Omega_m\int\mathrm{d}z\,W(z)D^2(z)\int\mathrm{d}\Pi\,p(\Pi|z)\xi_{\mathrm{quad}}^+(r_p,\Pi). \tag*{(15)}

The linear power spectrum is computed with CAMB (Lewis et al., 2000) for the Planck 2018 cosmology (Aghanim et al., [2]). We calculate A0∣0EA_{0|0E}, C0∣0EC_{0|0E}, A0∣E2A_{0|E2} and B0∣E2B_{0|E2} using FAST-PT (Fang et al., [23]).

Galaxy clustering correlation measurements

As explained in the main body of the text, we must measure the galaxy clustering bias in combination with the wg+w_{g+} correlation functions in order to constrain the amplitudes for intrinsic alignments. We measure the galaxy bias from the auto-correlation wggw_{gg}, for each of the redshift bins. We also require the galaxy bias for the shape samples for one of the alignment models that we explore (where we fix A1δ=bSA1A_{1\delta}=b_SA_1). We measured the galaxy bias for each shape sample by using the cross-correlation between the shape and position samples. These correlation functions are presented in Extended Data Figure 3. We see that for the position sample the galaxy bias increases with redshift, bg=1.08b_g=1.08, 1.211.21 and 1.371.37. This is expected for a flux-limited galaxy sample: at larger redshifts we can only see the most luminous galaxies which in turn will be more massive and have a larger galaxy bias. The red spheroids have much larger galaxy bias values (1.38±0.081.38\pm0.08, 1.51±0.051.51\pm0.05 and 1.82±0.041.82\pm0.04) compared to the blue discs (bS=0.96±0.08b_S=0.96\pm0.08, 0.97±0.040.97\pm0.04 and 1.12±0.031.12\pm0.03) which again is expected as red spheroids are expected to live in more massive halos than the blue discs. The PTE for all nine galaxy correlation functions are greater than 0.3, indicating that linear theory describes these observations well on our scales of interest. For the two higher-redshift bins the measurements appear to lie consistently above the prediction; this is a result of the strong correlations between the rpr_p bins. As the PTE are acceptable, our model provides a good fit to the observations.

Extended Data Figure 3. The galaxy clustering of the position and shape samples follow linear theory. Black diamonds: auto-correlations rpwggr_p w_{gg} of the galaxy position sample for each redshift bin. Red squares: the cross-correlation between the galaxy position sample and the red spheroid shape sample positions. Blue circles: the cross-correlation between the galaxy position sample and the blue disc sample positions. The solid lines show the best fit for the linear galaxy-bias model. Lower panels: residuals between the observation and the model, divided by the jackknife errors. The triangles mark disagreements beyond 3σ3\sigma. The small-scale excess is due to non-linear galaxy clustering.

Alignment amplitudes for discs and spheroids

In Table 2 we show the fits of six different models to every measurement presented in Figure 5, in combination with the galaxy clustering measurements presented in Extended Data Figure 3. The linear response model (1) fits every measurement with a PTE between 0.57 and 0.99, except the red spheroids at 0.35<z<0.550.35 < z < 0.55 (χ2=17.9\chi^2 = 17.9 for 4 degrees of freedom, PTE 0.001). The quadratic response alone, model (4), is rejected for every significant detection. For red spheroids in the highest redshift bin, we can obtain a reasonable fit with models (5) and (6), however this requires a large A2A_2 which is counterbalanced by a large A1A_1 of the opposite sign. We interpret this as non-linear evolution on small scales instead of evidence for a non-zero A2A_2 term. As mentioned previously, this observation obtains an acceptable fit (PTE 0.12) when we use the NLA model [7].

Alignments around underdense and overdense regions

As described in the main text, we smooth the galaxy density field with a Gaussian of width σ⊥=5h−1 Mpc\sigma_\perp= 5h^{-1}\,\mathrm{Mpc} in the plane of the sky and σ∥=30h−1 Mpc\sigma_\parallel= 30h^{-1}\,\mathrm{Mpc} along the line of sight. We then generate 2 million random positions within the survey footprint. We compute the smoothed density contrast δc\delta_c at the position of each random, and the normalised peak height ν=δc/σc\nu= \delta_c/\sigma_c, by dividing the smoothed density contrast by its standard deviation over all random positions. We select the 20% of positions with the lowest density, the quintile qlq_l, and the 20% with the highest, qhq_h, following the approach used in density-split statistics in galaxy clustering [16] (Paillas et al. 2023; Paillas et al. 2024). We then measure the alignment correlations around the positions of each quintile with the estimator of Eq. (4), except with the quintile positions instead of galaxy positions.

To model the signal, we first start from the correlation function definition. The mean alignment of shapes around the quintile positions qq, at a separation r′−r\mathbf{r}'-\mathbf{r}, is

ξq+(r′−r)≡E[(1+bSδ(r′))γ+(r′)∣ν(r)∈q],(16)\xi_{q+}(\mathbf{r}'-\mathbf{r}) \equiv\mathop{\mathrm{E}}\left[(1+b_S\delta(\mathbf{r}'))\gamma_+(\mathbf{r}') \mid\nu(\mathbf{r}) \in q\right], \tag*{(16)}

where the factor 1+bSδ1+b_S\delta appears because we only measure shapes at positions where galaxies are observed. This equation is then projected along the line of sight to calculate the model predictions shown in Figure 6.

wq+(rp)=∫−Πmax⁡Πmax⁡dΠobs∫−∞∞dΠe−(Πobs−Π)2/2σΔz,s22πσΔz,s ξq+(rp,Π),(17)w_{q+}(r_p) = \int_{-\Pi_{\max}}^{\Pi_{\max}} \mathrm{d}\Pi_{\mathrm{obs}} \int_{-\infty}^{\infty} \mathrm{d}\Pi\frac{e^{-(\Pi_{\mathrm{obs}}-\Pi)^2/2\sigma_{\Delta z,s}^2}}{\sqrt{2\pi}\sigma_{\Delta z,s}}\,\xi_{q+}(r_p,\Pi), \tag*{(17)}

where Π\Pi and Πobs\Pi_{\mathrm{obs}} are the true and observed line-of-sight separations and σΔz,s\sigma_{\Delta z,s} is the redshift error of the shape sample in comoving distance.

Throughout we take the x-axis to be the projected separation between the quintile position and the galaxy shape position and the z-axis as the line of sight. Therefore r′−r=(rp,0,Π)\mathbf{r}'-\mathbf{r}=(r_p,0,\Pi) and γ+=γxx−γyy\gamma_+ = \gamma_{xx}-\gamma_{yy}. Writing out explicitly Eq. (13) of Blazek et al. [4] (the radial component of Eq. (6)), we have

(1+bSδ)γ+=C1s++bSC1δs++C2(−szzs++sxz2−syz2),(18)(1+b_S\delta)\gamma_+ = C_1s_+ + b_SC_1\delta s_+ + C_2(-s_{zz}s_+ + s_{xz}^2 - s_{yz}^2), \tag*{(18)}

where every field is at the position r′\mathbf{r}' and s+≡sxx−syys_+ \equiv s_{xx}-s_{yy}. There is a small difference with Blazek in that here we have written explicitly that C1δs=bSC1C_1\delta_s = b_SC_1, instead of leaving this as a free parameter, which can absorb any response of the galaxy shape field to the product of the density and tidal fields.

Now we consider the expectation of Eq. (18) for a given value of ν\nu. Within linear theory, the density, tidal and peak height are jointly Gaussian random fields with zero mean [3] (van de Weygaert and Bertschinger 1996). For jointly Gaussian random fields we use the following relations for their conditional expectations: E[X∣ν]=⟨Xν⟩ν\mathrm{E}[X\mid\nu]=\langle X\nu\rangle\nu and E[XY∣ν]=⟨XY⟩+⟨Xν⟩⟨Yν⟩(ν2−1)\mathrm{E}[XY\mid\nu]=\langle XY\rangle+\langle X\nu\rangle\langle Y\nu\rangle(\nu^2-1) [3]. To simplify the following equations, we define cδ≡⟨δν⟩c_\delta\equiv\langle\delta\nu\rangle, c+≡⟨s+ν⟩c_+ \equiv\langle s_+\nu\rangle, czz≡⟨szzν⟩c_{zz}\equiv\langle s_{zz}\nu\rangle, and cxz≡⟨sxzν⟩c_{xz}\equiv\langle s_{xz}\nu\rangle. Therefore we arrive at,

E[(1+bSδ)γ+∣ν]=C1c+ν+bSC1cδc+(ν2−1)+C2(cxz2−czzc+)(ν2−1).(19)\mathrm{E}\left[(1+b_S\delta)\gamma_+\mid\nu\right] = C_1c_+\nu+b_SC_1c_\delta c_+(\nu^2-1)+C_2(c_{xz}^2-c_{zz}c_+)(\nu^2-1). \tag*{(19)}

The linear response is proportional to an odd term of ν\nu, whereas the quadratic response is proportional to ν2−1\nu^2-1, meaning that as the peak height goes from negative to positive the sign of the linear response changes whereas the quadratic response keeps the same sign. We then average Eq. (19) over all positions within the quintile, using ⟨ν⟩qh=+1.4\langle\nu\rangle_{q_h}=+1.4 and −1.4-1.4 for qhq_h and qlq_l, and ⟨ν2−1⟩q=1.178\langle\nu^2-1\rangle_q=1.178 for both quintiles. Therefore,

ξq+=±1.4C1c++1.178bSC1cδc++1.178C2(cxz2−czzc+),(20)\xi_{q+} = \pm1.4C_1c_+ + 1.178b_SC_1c_\delta c_+ + 1.178C_2(c_{xz}^2-c_{zz}c_+), \tag*{(20)}

where c+c_+, cδc_\delta, cxzc_{xz} and czzc_{zz} are calculated as

cX(r′−r)=bgσc∫d3k(2π)3W(k)Pδ(k)KX(k)eik⋅(r′−r),(21)c_X(\mathbf{r}'-\mathbf{r}) = \frac{b_g}{\sigma_c}\int\frac{\mathrm{d}^3k}{(2\pi)^3}W(\mathbf{k})P_\delta(k)K_X(\mathbf{k})e^{i\mathbf{k}\cdot(\mathbf{r}'-\mathbf{r})}, \tag*{(21)}

where we need to account for the smoothing of the galaxy density field with the window function, W(k)=exp⁡[−12k⊥2σ⊥2−12kz2(σ∥2+σΔz2)]W(k) = \exp[-\frac{1}{2}k_{\perp}^{2}\sigma_{\perp}^{2}-\frac{1}{2}k_{z}^{2}(\sigma_{\parallel}^{2}+\sigma_{\Delta z}^{2})], which relates ν\nu to the true matter field. This takes into account the smoothing that we apply to the Gaussian field in order to estimate the peak-height, σ⊥=5h−1 Mpc\sigma_{\perp}=5h^{-1}\,\mathrm{Mpc} and σ∥=30h−1 Mpc\sigma_{\parallel}=30h^{-1}\,\mathrm{Mpc}, as well as the scatter of the photometric redshifts σΔz=21.3h−1 Mpc\sigma_{\Delta z}=21.3h^{-1}\,\mathrm{Mpc}. σc2=bg2∫d3kW2(k)Pδ(k)/(2π)3\sigma_{c}^{2}=b_{g}^{2}\int d^{3}kW^{2}(k)P_{\delta}(k)/(2\pi)^{3} is the variance of the smoothed galaxy density. The kernel KX(k)K_{X}(k) takes the values Kδ=1K_{\delta}=1, K+=(kx2−ky2)/k2K_{+}=(k_{x}^{2}-k_{y}^{2})/k^{2}, Kzz=kz2/k2−1/3K_{zz}=k_{z}^{2}/k^{2}-1/3 and Kxz=kxkz/k2K_{xz}=k_{x}k_{z}/k^{2} for each of cδc_{\delta}, c+c_{+}, czzc_{zz} and cxzc_{xz}.

Alignment in bins of stellar mass

In Figure 7 we presented the alignment amplitudes as a function of stellar mass. The shape samples in each redshift bin are divided into five stellar mass bins of width 0.5 log⁡10(M⋆/M⊙)\log_{10}(M_{\star}/M_{\odot}) from 9 to 11.5, and one bin for all galaxies above 11.5.

It is important to note that our shape samples are not complete in stellar mass. As our sample is flux limited we will miss many low mass galaxies. Additionally our resolution cut, res>1\mathrm{res}>1 (a half-light radius larger than the full width at half maximum of the point-spread function), means that at fixed mass we preferentially select large galaxies. This second point is very important. For different sized galaxies the Tractor shape information will come from different physical scales within the galaxy. Therefore we cannot tell from these results alone whether the mass trends we see arise from the stellar mass of the shape galaxies, or from measuring their shapes at different radii.

Model fitting and covariance matrices

Throughout the paper, each model is fitted between 10≤rp≤100h−1 Mpc10\leq r_{p}\leq100h^{-1}\,\mathrm{Mpc} with five logarithmic bins. The lower limit is a conservative choice. Our models assume linear galaxy bias and perturbative alignment models, so they are only valid when the density field is close to linear. We can test where this breaks down with the clustering of our own samples (Extended Data Figure 3). The position sample and the blue disc and red spheroid shape samples agree with linear bias to within 1.2σ1.2\sigma at rp≈7h−1 Mpcr_{p}\approx7h^{-1}\,\mathrm{Mpc}, but lie 1–6% above it (up to 4σ4\sigma) at 4h−1 Mpc4h^{-1}\,\mathrm{Mpc} and 5–15% above it (up to 11σ11\sigma) at 3h−1 Mpc3h^{-1}\,\mathrm{Mpc}. Past direct measurements of galaxy alignments have fitted the non-linear linear alignment model above 6h−1 Mpc6h^{-1}\,\mathrm{Mpc} ((Joachimi et al., 2011; Samuroff et al., 2023)).

We limit the fits to rp≤100h−1 Mpcr_{p}\leq100h^{-1}\,\mathrm{Mpc} because on larger scales it is more difficult to estimate the errors when the size of the jackknife regions approaches the scales that we measure. The jackknife regions are approximately 14∘14^{\circ} across, which is 37–140, 140–240 and 240–350 h−1 Mpch^{-1}\,\mathrm{Mpc} over the three redshift bins. In the lowest redshift bin the largest fitted scales are already comparable to the size of a region, and its errors on those scales are the least reliable; the jackknife errors there are larger than the analytic Gaussian errors (Extended Data Figure 4).

Extended Data Figure 4. The jackknife errors are larger than the analytic Gaussian errors on every scale. σ[wg+]\sigma[w_{g+}] for each sample (rows) and redshift bin (columns), as in Figure 5. Points: the 100-region jackknife used in every fit. Black: the analytic Gaussian covariance (total, solid) and its terms, clustering ×\times shape noise (dotted), shot noise (dashed) and cosmic variance (dash-dotted). Grey line: the shot noise computed with the measured pair counts, which includes the clustering of tracers around the shape galaxies. Grey band: fitted scales.

The measurements are then fit using a Gaussian likelihood, using the covariance estimated from the 100 jackknife regions (Correlation functions). We compute posteriors using the code emcee ([24]). We use large flat priors for each parameter: −20<A1<20-20<A_{1}<20, −100<A1δ<100-100<A_{1\delta}<100, −200<A2<200-200<A_{2}<200 and 0.5<bg<40.5<b_{g}<4. In Extended Data Figure 4 we compare the diagonal of our jackknife covariance against the analytic Gaussian covariance for projected correlation functions, derived by Chisari and Dvorkin [10] and applied to simulated intrinsic alignment measurements by Samuroff et al. (2021). We find that our jackknife errors are consistently larger than the theoretical prediction. We therefore adopt the jackknife covariance errors throughout as we consider this to be the more conservative choice.

Quantifying contamination from weak lensing

Weak gravitational lensing of the shape sample by the position sample (galaxy–galaxy lensing) would also produce a tangential shear signal for the blue discs, as we see in the observations. Figure 4 includes the predicted lensing signal (the lensing term wgGw_{gG}, following Georgiou et al. (2025)). As the shape and position samples are at the same redshift, the predicted contamination of our intrinsic alignment signal by weak lensing is very small: 3% of the blue disc correlation function at 0.20<z<0.350.20<z<0.35. Given the redshift distributions of the shape and position samples, to explain the observed signal with gravitational lensing, every galaxy of the position sample would need a halo mass of ∼1015.7h−1M⊙\sim10^{15.7}h^{-1}M_{\odot}. In other words every position galaxy would have to be within the most massive clusters in the Universe, which is clearly an unreasonable proposition.

However we could underestimate the lensing prediction if our redshift distributions are incorrect. This is unlikely given the low outlier rates, 0.13–0.52% (Extended Data Figure 1), and the small redshift errors, σNMAD=0.0055\sigma_{\mathrm{NMAD}}=0.0055–0.0077, of the quality sample in our three redshift bins. We test this independently of the spectroscopic validation by cross-correlating the position and shape samples of within redshift bins. In Extended Data Figure 5 we show these wg+w_{g+} correlations without the condition ∣Πobs∣<Πmax⁡\lvert\Pi_{\mathrm{obs}}\rvert<\Pi_{\max}, so that all position–shape pairs are used. From lensing and incorrect redshift distributions, we would see a significant signal in every combination of bins. If the redshift distributions are correct, we expect a small lensing signal where the position sample is in front of the shape sample, no signal where the shapes are in front of the positions, and a strong intrinsic alignment signal where both are in the same redshift bin. We observe the second scenario, therefore our observations are inconsistent with gravitational lensing.

Extended Data Figure 5. There is no evidence of significant lensing contamination. rpwg+r_p w_{g+} measurements for every pair of position and shape tracers across different redshift bins (there are no line-of-sight distance limits for these correlation functions). The green panels show the shape-position correlations for galaxies at the same redshift, where we see significant non-zero correlation function measurements, which are consistent with the previous alignment amplitude measurements. The panels with shapes in front of position tracers are consistent with null-detections, as anticipated from precise redshift measurements with low-outlier rates. The panels with shapes behind the position tracers show small non-zero correlations, consistent with the prediction from gravitational lensing.

Comparison with published alignment measurements

The large alignment amplitude that we measure for blue discs may appear surprising given past constraints on the alignments of blue galaxies. In Extended Data Figure 7 we show that the alignment amplitude is much smaller when we select all blue galaxies, without the disc selection: at 0.20<z<0.350.20 < z < 0.35, A1=−2.06±0.19A_1 = -2.06 \pm0.19 for all blue galaxies with res >1> 1 and −0.62±0.12-0.62 \pm0.12 with res >0.5> 0.5, compared with −2.81±0.23-2.81 \pm0.23 for the blue discs. This is similar to the result of Paviot et al. (2026), where low-mass blue galaxies show no alignment and higher-mass blue galaxies show a 3.3σ3.3\sigma tangential alignment. Our blue galaxy sample is large enough, and its redshifts precise enough, that entire blue galaxy sample without stellar mass or morphological cuts give a significant tangential alignment.

In Extended Data Figure 8 we compare our results with published measurements. As past work has generally used the NLA model, we refit our measurements with the NLA model for this comparison. Our blue discs show a significant and large amplitude tangential alignment, in contrast to previous measurements. When we instead select all blue galaxies with res >0.5> 0.5, we still detect an alignment in every redshift bin, but the amplitudes are comparable with those of past measurements. This detection is due to the improved statistical power of this sample compared to past observations.

Extended Data Figure 8. Our results are consistent with past results for similar galaxy types. Alignment amplitudes of the NLA model for red galaxies (top) and blue galaxies (bottom): the published measurements (open squares) (Joachimi et al., 2011; Mandelbaum et al., 2011; Singh et al., 2015; Johnston et al., 2019; Yao et al., 2020; Fortuna et al., 2021; Kurita and Takada, 2023; Samuroff et al., 2023; Tonegawa et al., 2025; Hervas Peters et al., 2025; Georgiou et al., 2025; Siegel et al., 2026; Navarro-Gironés et al., 2026; Paviot et al., 2026)} and this work (filled circles). The right-hand column gives the significance of each measurement, $|A_{\mathrm{IA}}|/\sigma_A$. The 95% limits of Mandelbaum et al. (2011); {citet}Tonegawa et al. (2025) are divided by 1.96 to approximate the $1-\sigma$ constraints of the other analyses. The two redMaGiC samples from {citet}Samuroff et al. (2023) (†) are shown as published, however, {citet}Hervas Peters et al. (2025)` found that the covariance used in that work omits the term from clustering power and shape noise. We therefore expect their significances to be overestimated by a factor of around 1.5.

For red spheroids our results are consistent with the literature. Red spheroids show significant radial alignments, which increase with galaxy stellar mass. Our measurement at 0.35<z<0.550.35 < z < 0.55 (27.4σ27.4\sigma by ∣AIA∣/σA|A_{\mathrm{IA}}|/\sigma_{A}) is the most significant detection of radial alignment of red galaxies to date.

We also measured shape-shape correlations w++w_{++} and w××w_{\times\times} for each of our samples. None of these correlations is significantly different from zero, but they are consistent with the amplitudes predicted from the observed wg+w_{g+} with the linear response model.

Consistent results using DES Y6 shapes

The Tractor shapes are corrected for the point-spread function but they are not calibrated for weak lensing shear, therefore the multiplicative bias is unknown. A multiplicative bias rescales the measured A1A_1, however it cannot produce a correlation that does not otherwise exist or even change the significance of the detected correlation. This is therefore not a major concern for this work, the aim of which is primarily to improve our understanding of galaxy intrinsic alignments rather than to provide alignment priors for weak lensing experiments. The Tractor shapes of the Legacy Imaging Surveys have not been extensively tested. The Dark Energy Survey Year 6 (DES Y6) shape catalogue, was built specifically for weak lensing and has been thoroughly validated (Yamamoto et al., 2025). We therefore repeat our measurements with DES Y6 shapes.

We match LSx positions to DES Y6 shapes within 1′′1''. Of the 3.92 million galaxies of the LSx shape sample within the 4750 deg2^2 DES footprint, 3.02 million have a DES Y6 match. The majority of the unmatched galaxies are large and bright; DES excludes large bright galaxies as a large fraction of their large bright galaxies are imaging systematics. In each sample and redshift bin we divide the DES shapes by the mean metacalibration shear response. This response is for calibrating weak lensing shear, rather than intrinsic galaxy shapes so it is unclear that this correction is correct. The effect is simply to rescale the alignment amplitude.

The results are shown in Extended Data Figure 6. For the same blue disc galaxies, the DES Y6 and Tractor shapes give consistent amplitudes ΔA1=−0.07±0.04\Delta A_1 = -0.07 \pm0.04 and −0.14±0.07-0.14 \pm0.07 at 0.20<z<0.350.20 < z < 0.35 and 0.35<z<0.550.35 < z < 0.55. With the DES Y6 shapes the tangential alignment of the blue discs is detected at 4.7σ4.7\sigma and 6.9σ6.9\sigma in these bins; the detection significance is lower than in the main analysis as the survey footprint of DES is smaller 4750 deg2^2 rather than 19,312 deg2^2 of the LSx catalogue. For all red galaxies (not shown in the Figure for brevity) we also find significant radial alignments with DES Y6 shapes, however the amplitudes are 16–29% lower than with the Tractor shapes of the same galaxies (A1=+1.39±0.40A_1 = +1.39 \pm0.40 against +1.94±0.36+1.94 \pm0.36 at 0.20<z<0.350.20 < z < 0.35, and +3.75±0.47+3.75 \pm0.47 against +4.44±0.40+4.44 \pm0.40 at 0.35<z<0.550.35 < z < 0.55). It has been shown previously that the alignment amplitude is sensitive to the shape measurement method (Singh and Mandelbaum, 2016; Georgiou et al., 2019).

We also consider a broader DES Y6 shape sample, where the only selection are the DES Y6 shape selection and the LSx colour cut. Therefore no morphological or resolution cuts. This sample has a weaker alignment than the blue discs, A1=+0.43±0.21A_1 = +0.43 \pm0.21, −0.07±0.19-0.07 \pm0.19 and −0.84±0.19-0.84 \pm0.19 in the three redshift bins, against −0.42±0.40-0.42 \pm0.40, −2.07±0.44-2.07 \pm0.44 and −4.06±0.59-4.06 \pm0.59 for the matched blue discs. The main difference is that this sample contains galaxies of a smaller angular size. Further we consider the galaxies that our in this sample, which are not within the Tractor sample. These galaxies align radially, with A1=+1.57±0.32A_1 = +1.57 \pm0.32 and +0.78±0.19+0.78 \pm0.19 in the two lower redshift bins, and are consistent with no alignment +0.15±0.17+0.15 \pm0.17 in the highest redshift bin. The alignment of blue galaxies is therefore strongly dependent upon not only morphology but size and resolution. Small blue galaxies align radially, and large blue discs align tangentially.

Use of a large language model

Claude (Anthropic), was used to write code for this research project and language editing of the manuscript. The authors take full responsibility for the contents of this paper.

Data availability

The DESI Legacy Imaging Surveys, SPHEREx, DES, DESI DR1, SDSS, GAMA, 6dFGS and 2dFGRS data used in this work are publicly available from the respective surveys. The measured correlation functions and the data underlying each figure are available on request.

Code availability

The correlation functions in this work were made with TreeCorr (Jarvis et al., 2004), the model predictions were computed with CAMB (Lewis et al., 2000) and FAST-PT (Fang et al., [23]), the posteriors were sampled using emcee (Foreman-Mackey et al., [24]) and the stellar masses estimated with CIGALE [6], all of which are publicly available. The analysis and plotting code is available on request.

Acknowledgements

Calum Murray would like to thank Sandrine Codis, Olivier Doré, Jamie Bock, Richard M. Feder, and Zhaoyu Huai for useful discussions about this work. This publication makes use of data products from SPHEREx, a joint project of the Jet Propulsion Laboratory and the California Institute of Technology, funded by the National Aeronautics and Space Administration. The DESI Legacy Imaging Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS), the Beijing-Arizona Sky Survey (BASS), and the Mayall z-band Legacy Survey (MzLS); the complete acknowledgments can be found at https://www.legacysurvey.org/acknowledgment/. This project used data from the Dark Energy Survey (DES) and the Dark Energy Spectroscopic Instrument (DESI), whose full acknowledgments are at https://www.darkenergysurvey.org/the-des-project/data-access/ and https://data.desi.lbl.gov/doc/acknowledgments/. Funding for SDSS-IV has been provided by the Alfred P. Sloan Foundation, the U.S. Department of Energy Office of Science, and the Participating Institutions. GAMA is a joint European-Australasian project based around a spectroscopic campaign using the Anglo-Australian Telescope (https://www.gama-survey.org/). We also use redshifts from the 6dF Galaxy Survey and the 2dF Galaxy Redshift Survey.

References

  1. [1]Abdurro’uf et al. The seventeenth Data Release of the Sloan Digital Sky Surveys: Complete release of the MaNGA, MaStar, and APOGEE-2 data. Astrophys. J. Suppl. Ser., 259:35, 2022. doi: 10.3847/1538-4365/ac4414.
  2. [2]N. Aghanim et al. Planck 2018 results. VI. Cosmological parameters. Astron. Astrophys., 641:A6, 2020. doi: 10.1051/0004-6361/201833910. [Erratum: Astron.Astrophys. 652, C4 (2021)].DOI
  3. [3]J. M. Bardeen, J. R. Bond, N. Kaiser, and A. S. Szalay. The Statistics of Peaks of Gaussian Random Fields. Astrophys. J., 304:15–61, 1986. doi: 10.1086/164143.
  4. [4]Jonathan Blazek, Niall MacCrann, M. A. Troxel, and Xiao Fang. Beyond linear galaxy alignments. Phys. Rev. D, 100:103506, 2019. doi: 10.1103/PhysRevD.100.103506.DOI
  5. [5]James J. Bock et al. The SPHEREx Satellite Mission. arXiv e-prints, 2025.DOI
  6. [6]M. Boquien, D. Burgarella, Y. Roehlly, V. Buat, L. Ciesla, D. Corre, A. K. Inoue, and H. Salas. CIGALE: a python Code Investigating GALaxy Emission. Astron. Astrophys., 622:A103, 2019. doi: 10.1051/0004-6361/201834156.DOI
  7. [7]Sarah Bridle and Lindsay King. Dark energy constraints from cosmic shear power spectra: impact of intrinsic alignments on photometric redshift requirements. New J. Phys., 9:444, 2007. doi: 10.1088/1367-2630/9/12/444.DOI
  8. [8]P. Catelan, M. Kamionkowski, and R. D. Blandford. Intrinsic and extrinsic galaxy alignment. Mon. Not. R. Astron. Soc., 320:L7–L13, 2001. doi: 10.1046/j.1365-8711.2001.04105.x.DOI
  9. [9]Yi-Kuan Chiang. Corrected SFD: A More Accurate Galactic Dust Map with Minimal Extragalactic Contamination. Astrophys. J., 958:118, 2023. doi: 10.3847/1538-4357/acf4a1.DOI
  10. [10]Nora Elisa Chisari and Cora Dvorkin. Cosmological information in the intrinsic alignments of luminous red galaxies. J. Cosmol. Astropart. Phys., 12:029, 2013. doi: 10.1088/1475-7516/2013/12/029.DOI
  11. [11]Nora Elisa Chisari, Sandrine Codis, Clotilde Laigle, Yohan Dubois, Christophe Pichon, Julien Devriendt, Adrianne Slyz, Lance Miller, Raphael Gavazzi, and Karim Benabed. Intrinsic alignments of galaxies in the Horizon-AGN cosmological hydrodynamical simulation. Mon. Not. Roy. Astron. Soc., 454(3):2736–2753, 2015. doi: 10.1093/mnras/stv2154.DOI
  12. [12]Nora Elisa Chisari, Clotilde Laigle, Sandrine Codis, Yohan Dubois, Julien Devriendt, Lance Miller, Karim Benabed, Adrianne Slyz, Raphael Gavazzi, and Christophe Pichon. Redshift and luminosity evolution of the intrinsic alignments of galaxies in Horizon-AGN. Mon. Not. Roy. Astron. Soc., 461(3):2702–2721, 2016. doi: 10.1093/mnras/stw1409.DOI
  13. [13]S. Codis, A. Jindal, N. E. Chisari, D. Vibert, Y. Dubois, C. Pichon, and J. Devriendt. Galaxy orientation with the cosmic web across cosmic time. Mon. Not. Roy. Astron. Soc., 481(4):4753–4774, 2018. doi: 10.1093/mnras/sty2567.DOI
  14. [14]Sandrine Codis, Christophe Pichon, and Dmitry Pogosyan. Spin alignments within the cosmic web: a theory of constrained tidal torques near filaments. Mon. Not. Roy. Astron. Soc., 452(4):3369–3393, 2015. doi: 10.1093/mnras/stv1570.DOI
  15. [15]Matthew Colless et al. The 2dF Galaxy Redshift Survey: spectra and redshifts. Mon. Not. R. Astron. Soc., 328:1039, 2001. doi: 10.1046/j.1365-8711.2001.04902.x.DOI
  16. [16]Carolina Cuesta-Lazaro et al. SUNBIRD: a simulation-based model for full-shape density-split clustering. Mon. Not. Roy. Astron. Soc., 531(3):3336–3356, 2024. doi: 10.1093/mnras/stae1234.
  17. [17]Ana Maria Delgado et al. The MillenniumTNG project: intrinsic alignments of galaxies and haloes. Mon. Not. Roy. Astron. Soc., 523(4):5899–5914, 2023. doi: 10.1093/mnras/stad1781.DOI
  18. [18]DESI Collaboration, M. Abdul Karim, et al. Data Release 1 of the Dark Energy Spectroscopic Instrument. Astron. J., 171:285, 2026. doi: 10.3847/1538-3881/ae4c43.
  19. [19]Arjun Dey et al. Overview of the DESI Legacy Imaging Surveys. Astron. J., 157:168, 2019. doi: 10.3847/1538-3881/ab089d.DOI
  20. [20]Olivier Doré et al. Cosmology with the SPHEREx All-Sky Spectral Survey. arXiv e-prints, 2014.DOI
  21. [21]A. G. Doroshkevich. Spatial structure of perturbations and origin of galactic rotation in fluctuation theory. Astrophysics, 6:320–330, 1970. doi: 10.1007/BF01001625.
  22. [22]Simon P. Driver et al. Galaxy And Mass Assembly (GAMA): Data Release 4 and the z < 0.1 total and z < 0.08 morphological galaxy stellar mass functions. Mon. Not. R. Astron. Soc., 513:439, 2022. doi: 10.1093/mnras/stac472.DOI
  23. [23]Xiao Fang, Jonathan A. Blazek, Joseph E. McEwen, and Christopher M. Hirata. FAST-PT II: an algorithm to calculate convolution integrals of general tensor quantities in cosmological perturbation theory. J. Cosmol. Astropart. Phys., 2017(02):030, 2017. doi: 10.1088/1475-7516/2017/02/030.DOI
  24. [24]Daniel Foreman-Mackey, David W. Hogg, Dustin Lang, and Jonathan Goodman. emcee: The MCMC Hammer. Publ. Astron. Soc. Pac., 125:306–312, 2013. doi: 10.1086/670067.

Paper details

vlnc:W_86k2yc68
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents