Pollux: Star, Spectropolarimeter & Beyond

• Pollux is a multifaceted term referring primarily to a benchmark K0 III giant star, noted for its long-period radial-velocity variations, weak surface magnetism, and potential planetary companion.
• In space instrumentation, Pollux designates a proposed high-resolution UV-to-NIR spectropolarimeter with full-Stokes capability, aimed at capturing detailed spectral and polarimetric data for diverse astrophysical phenomena.
• Beyond astronomy, the name Pollux is adopted in fields like deep learning scheduling, LLM benchmarking, and political science, underscoring its cross-disciplinary impact and varied applications.

Searching arXiv for papers on “Pollux” to ground the article in the current literature. Pollux denotes several distinct research referents. In stellar astrophysics it is Pollux, or $\beta$ Gem and HD 62509, a K0 III giant whose weak surface magnetism, long-period radial-velocity variability, and possible planetary companion have made it a benchmark object for spectropolarimetry, Zeeman Doppler imaging, magnetohydrodynamic simulation, and radio diagnostics (Aurière et al., 2021). In astronomical instrumentation, Pollux is a high-resolution spectrograph and spectropolarimeter proposed by a European consortium for the Habitable Worlds Observatory, with nearly continuous coverage from the far ultraviolet to the near infrared and an emphasis on full-Stokes capability (Neiner et al., 10 Feb 2026). The name has also been adopted for technically unrelated systems, including a deep-learning cluster scheduler, a Russian-language LLM benchmark, and the Pollux information service in political science (Qiao et al., 2020); (Martynov et al., 30 May 2025); (Holtdirk et al., 2024).

1. Principal referents of the name

The modern literature uses “Pollux” across several domains. The most prominent referents are summarized below.

Referent	Domain	Defining description
Pollux ($\beta$ Gem, HD 62509)	Stellar astrophysics	K0 III giant (Aurière et al., 2021)
Pollux spectropolarimeter	Space instrumentation	high-resolution spectrograph and spectropolarimeter working from 100 nm to 1.8 microns proposed for HWO (Neiner et al., 10 Feb 2026)
POLLUX benchmark	LLM evaluation	open-source, criteria-driven benchmark for assessing the generative capabilities of LLMs in Russian (Martynov et al., 30 May 2025)
Pollux scheduler	Distributed systems	co-adaptive cluster scheduling for goodput-optimized deep learning (Qiao et al., 2020)
Pollux information service	Political-science infrastructure	Specialised Information Service (FID) for Political Science in Germany (Holtdirk et al., 2024)

In astronomy, the name therefore spans both a physical object and a major instrument concept. Outside astronomy, later uses preserve the name but not the astrophysical content. This division is important because the underlying literatures are technically independent.

2. Pollux as a giant star

Pollux is described in the spectropolarimetric literature as a K0 III giant. Adopted stellar parameters vary with the analysis and model context. One observational synthesis gives $M_\star \simeq 1.9\,M_\odot$, $R_\star \simeq 8.8\,R_\odot$, $L_\star \simeq 42\,L_\odot$, $T_{\rm eff} \simeq 4865\,{\rm K}$, and $\log g \simeq 2.75$ (Auriere et al., 2013). A later study gives $M_\star \simeq 1.7\,M_\odot$, $R_\star = 8.8\,R_\odot$, $T_{\rm eff} \simeq 5000\,{\rm K}$, $\beta$0, and identifies the star as a red-clump giant on the first ascent of the giant branch (Aurière et al., 2021). The 3-D MHD simulation of the deep convective envelope instead adopts a $\beta$1, $\beta$2, $\beta$3 model extending from the base of the envelope to the stellar surface (Palacios et al., 2013). The coexistence of these values reflects differences between observational inference and simulation setup rather than a settled disagreement over the object’s classification.

The star has long been associated with a stable long-period radial-velocity signal. One summary reports strictly sinusoidal RV variations with period $\beta$4 days and semi-amplitude $\beta$5, coherent for more than 25 years and historically interpreted as evidence for a $\beta$6–$\beta$7 planet at $\beta$8 (Auriere et al., 2013). A later reanalysis gives $\beta$9 d and $M_\star \simeq 1.9\,M_\odot$0, yielding $M_\star \simeq 1.9\,M_\odot$1 and $M_\star \simeq 1.9\,M_\odot$2 under the planet hypothesis (Aurière et al., 2021). The persistence of the RV signal is not in dispute; the controversy concerns its physical origin.

The magnetic field was established through repeated spectropolarimetry with ESPaDOnS and Narval. The 2021 study reports 265 circular-polarization sequences on 41 epochs between 2007 September and 2012 February, with $M_\star \simeq 1.9\,M_\odot$3 spectra from 370 nm to 1000 nm reduced with Libre-ESpRIT and analyzed with Least-Squares Deconvolution using a solar-abundance ATLAS9 mask centered at $M_\star \simeq 1.9\,M_\odot$4 and effective Landé factor $M_\star \simeq 1.9\,M_\odot$5 (Aurière et al., 2021). The earlier 2013 synthesis reports 41 observing nights and 266 Stokes $M_\star \simeq 1.9\,M_\odot$6 sequences over 4.25 years, with LSD combining $M_\star \simeq 1.9\,M_\odot$7 photospheric lines (Auriere et al., 2013). Both analyses identify a very weak but stable surface field.

3. Magnetic topology, radial-velocity interpretation, and outer-atmosphere diagnostics

The line-of-sight field is measured through the standard first-moment relation applied to LSD profiles,

$M_\star \simeq 1.9\,M_\odot$8

with the 2021 study finding $M_\star \simeq 1.9\,M_\odot$9 between about $R_\star \simeq 8.8\,R_\odot$0 and $R_\star \simeq 8.8\,R_\odot$1 and a semi-amplitude $R_\star \simeq 8.8\,R_\odot$2 (Aurière et al., 2021). Zeeman Doppler imaging produces a largely dipolar topology in both major analyses, but the inferred geometry differs. One ZDI reconstruction, using a period consistent with the RV signal, finds a predominantly poloidal field with 71% of the magnetic energy, 99% of the poloidal energy in the dipolar component, a mean surface field strength $R_\star \simeq 8.8\,R_\odot$3, and a magnetic obliquity $R_\star \simeq 8.8\,R_\odot$4 (Auriere et al., 2013). The later study instead reports $R_\star \simeq 8.8\,R_\odot$5 d, 99.95% of the magnetic energy in the poloidal field, 97% of the poloidal energy in the dipole mode, $R_\star \simeq 8.8\,R_\odot$6, and $R_\star \simeq 8.8\,R_\odot$7 (Aurière et al., 2021). Both results support the same qualitative conclusion: Pollux hosts an exceptionally weak, low-order, long-lived large-scale field.

The relation between magnetic variability and RV variability remains the main interpretive issue. The 2013 study found the longitudinal field to vary sinusoidally with a period close to that of the RV variations and a phase lag of $R_\star \simeq 8.8\,R_\odot$8 d, and suggested that large-scale spots associated with the magnetic dipole could reproduce the RV signal, potentially rendering the close-in exoplanet hypothesis unnecessary (Auriere et al., 2013). The 2021 reanalysis favors contemporaneous dynamo action as the field origin and presents two scenarios: if $R_\star \simeq 8.8\,R_\odot$9, the RV signal is planetary; if $L_\star \simeq 42\,L_\odot$0, magnetic activity could account for all or part of the RV modulation (Aurière et al., 2021). A common misconception is that Pollux is either unambiguously a planet host or unambiguously a false positive. The literature instead treats the system as an unresolved but sharply constrained case in which a weak giant-star dynamo and a long-period RV signal may overlap.

The dynamo interpretation has been explored numerically with the ASH code in a rotating anelastic spherical shell. Palacios and Brun’s first 3-D MHD simulations use $L_\star \simeq 42\,L_\odot$1 d, a seed multipole with $L_\star \simeq 42\,L_\odot$2, $L_\star \simeq 42\,L_\odot$3, and $L_\star \simeq 42\,L_\odot$4, and operate in a regime with $L_\star \simeq 42\,L_\odot$5 of order $L_\star \simeq 42\,L_\odot$6–$L_\star \simeq 42\,L_\odot$7 and $L_\star \simeq 42\,L_\odot$8 (Palacios et al., 2013). After $L_\star \simeq 42\,L_\odot$9 d of simulated time, the dynamo remains in a linear growth phase with $T_{\rm eff} \simeq 4865\,{\rm K}$0, local rms field strengths up to $T_{\rm eff} \simeq 4865\,{\rm K}$1, a global mean around $T_{\rm eff} \simeq 4865\,{\rm K}$2, and an emerging low-order dipolar topology. The study identifies the two hallmarks of an $T_{\rm eff} \simeq 4865\,{\rm K}$3–$T_{\rm eff} \simeq 4865\,{\rm K}$4 dynamo and argues that an $T_{\rm eff} \simeq 4865\,{\rm K}$5–$T_{\rm eff} \simeq 4865\,{\rm K}$6 dynamo is expected in the strongly stratified, rotating convective envelope (Palacios et al., 2013).

Radio observations extend the magnetic and atmospheric picture beyond optical spectropolarimetry. VLA detections at 21.2 and 9.0 GHz yield flux densities of $T_{\rm eff} \simeq 4865\,{\rm K}$7 and $T_{\rm eff} \simeq 4865\,{\rm K}$8, with a $T_{\rm eff} \simeq 4865\,{\rm K}$9 upper limit of $\log g \simeq 2.75$0 at 3.0 GHz (O'Gorman et al., 2016). The corresponding disk-averaged brightness temperatures are approximately $\log g \simeq 2.75$1 at 21.2 GHz, $\log g \simeq 2.75$2 at 9.0 GHz, and $\log g \simeq 2.75$3 at 3.0 GHz, consistent with a quiet-Sun-like outer atmosphere dominated by optically thick thermal chromospheric emission at 9–21 GHz (O'Gorman et al., 2016). The same study constrains the mass-loss rate to $\log g \simeq 2.75$4 for $\log g \simeq 2.75$5.

4. Pollux as a space spectropolarimeter

As an instrument concept, Pollux originated in LUVOIR studies as a high-resolution UV spectropolarimeter and later expanded within the HWO framework into a UV-to-NIR spectrograph and spectropolarimeter. The LUVOIR Phase 0 science case defined POLLUX as a three-channel instrument covering 90–390 nm or 90–400 nm with $\log g \simeq 2.75$6 and full-Stokes capability (Bouret et al., 2018); (Muslimov et al., 2018). The current HWO concept defines Pollux as a high-resolution spectrograph and spectropolarimeter working from 100 nm to 1.8 microns, with five contiguous channels: FUV 100–123 nm, FMUV 101–236 nm, NUV 234–472 nm, OPT 472–944 nm, and NIR 944–1888 nm (Neiner et al., 10 Feb 2026). In both generations of the design, resolving power is expressed conventionally as

$\log g \simeq 2.75$7

The recent HWO baseline describes Pollux as a focal-plane instrument fed after the primary and secondary mirrors, with a dedicated pick-off mirror feeding the FUV arm and a second pick-off plus three dichroics splitting the remainder of the beam to the FMUV, NUV, VIS, and NIR channels (Hue et al., 11 Mar 2026). Each channel is a classical high-dispersion echelle spectrograph. The instrument is designed for $\log g \simeq 2.75$8 over its entire range; in the current baseline the FUV, NUV, VIS, and NIR arms reach $\log g \simeq 2.75$9, while the FMUV arm reaches $M_\star \simeq 1.7\,M_\odot$0 (Hue et al., 11 Mar 2026). A related HWO overview gives typical $M_\star \simeq 1.7\,M_\odot$1 values from $M_\star \simeq 1.7\,M_\odot$2 to $M_\star \simeq 1.7\,M_\odot$3, with $M_\star \simeq 1.7\,M_\odot$4, $M_\star \simeq 1.7\,M_\odot$5, $M_\star \simeq 1.7\,M_\odot$6, and $M_\star \simeq 1.7\,M_\odot$7 in OPT and NIR (Neiner et al., 10 Feb 2026).

The optical and detector architecture is strongly wavelength dependent. The FUV arm is permanently polarimetric; retractable dual-beam polarimetric modules are installed in FMUV, NUV, VIS, and NIR (Hue et al., 11 Mar 2026). Modulation is achieved via rotating retarders, including MgF$M_\star \simeq 1.7\,M_\odot$8 waveplates in the UV and achromatic FLCs in VIS/NIR, followed by a calcite or Wollaston beam-splitting analyzer (Hue et al., 11 Mar 2026). Large-format back-illuminated CMOS detectors are specified from FUV through VIS, with $M_\star \simeq 1.7\,M_\odot$9–25%, $R_\star = 8.8\,R_\odot$0–VIS$R_\star = 8.8\,R_\odot$1–80%, read noise $R_\star = 8.8\,R_\odot$2 rms, and dark current $R_\star = 8.8\,R_\odot$3; the NIR arm uses a Teledyne H2RG HgCdTe array with $R_\star = 8.8\,R_\odot$4, read noise $R_\star = 8.8\,R_\odot$5, and dark current $R_\star = 8.8\,R_\odot$6 (Hue et al., 11 Mar 2026).

Optical design studies show that the exact channel partition and achievable $R_\star = 8.8\,R_\odot$7 depend on telescope size, detector choice, and disperser design. A 2024 design trade study gives four channels—FUV, MUV, NUV, and VIS–NIR—with the VIS–NIR channel extending either to 1050 nm with silicon detectors or to 1800 nm with HgCdTe detectors (Muslimov et al., 2024). For an 8 m telescope, the same study reports realized $R_\star = 8.8\,R_\odot$8 values of $R_\star = 8.8\,R_\odot$9–$T_{\rm eff} \simeq 5000\,{\rm K}$0 in MUV, $T_{\rm eff} \simeq 5000\,{\rm K}$1–$T_{\rm eff} \simeq 5000\,{\rm K}$2 in NUV, $T_{\rm eff} \simeq 5000\,{\rm K}$3 in VIS–NIR with Si-CCD, $T_{\rm eff} \simeq 5000\,{\rm K}$4 for a single-grating FUV design, and $T_{\rm eff} \simeq 5000\,{\rm K}$5 for an FUV echelle design (Muslimov et al., 2024). Earlier LUVOIR modeling based on freeform holographic gratings likewise reported MUV resolving power of $T_{\rm eff} \simeq 5000\,{\rm K}$6–$T_{\rm eff} \simeq 5000\,{\rm K}$7 over 118.5–195 nm (Muslimov et al., 2018).

5. Polarimetric architecture, calibration, and technology maturation

Pollux’s polarimetric design is split between transmissive birefringent solutions above about 120 nm and mirror-only solutions below that threshold. For wavelengths $T_{\rm eff} \simeq 5000\,{\rm K}$8 nm, MgF$T_{\rm eff} \simeq 5000\,{\rm K}$9 is the baseline material for the MUV and NUV channels, using a rotating-modulator/analyzer scheme built entirely of MgF$\beta$00 (Neiner et al., 7 Mar 2025). Pollux employs six equally spaced modulator positions, $\beta$01, rather than the minimum four-step cycle, with a demodulation matrix chosen to achieve the theoretical extraction efficiency $\beta$02 for simultaneous recovery of $\beta$03, $\beta$04, and $\beta$05 (Neiner et al., 7 Mar 2025). Below $\beta$06 nm, MgF$\beta$07 is opaque and no transmissive optics exist, so the FUV polarimeter is entirely reflective: a rotating stack of mirrors provides modulation and a final analyzer mirror selects one linear polarization component (Neiner et al., 7 Mar 2025).

The FUV reflection polarimeter was first studied in detail for the LUVOIR-era design. The prototype uses a three-mirror “K-mirror” modulator and a Brewster-like reflective analyzer (Gal et al., 2019). A nominal incidence of $\beta$08 was examined for the modulator mirrors, while the analyzer angle was optimized near $\beta$09, where the mean polarization contrast is $\beta$10 over 90–135 nm for the gold prototype (Gal et al., 2019). For a gold K-mirror at $\beta$11 and four modulation angles, the calculated $\beta$12 is only $\beta$13–15%, which motivated the shift toward alternative coatings such as SiC and ta-C (Gal et al., 2019). More recent HWO-era work describes the design evolution from initial B$\beta$14C+SiC and taC choices toward an all-SiC modulator with a SiO$\beta$15-substrate analyzer over-coated with B$\beta$16C/MgF$\beta$17 (Neiner et al., 7 Mar 2025).

Calibration and stability requirements are correspondingly stringent. A recent technical overview gives a dual-beam polarimetric uncertainty

$\beta$18

so that for $\beta$19 one obtains $\beta$20, with design goals of a few $\beta$21 after binning or multiple exposures (Hue et al., 11 Mar 2026). The same overview specifies a dedicated Polarimetric Calibration Unit upstream of the polarimeter, hollow-cathode lamps yielding wavelength solutions better than 0.005 Å in the UV, a stabilized Fabry–Pérot etalon tracking drift to $\beta$22 equivalent error over hours, on-sky validation with polarized standards known to $\beta$23, and instrumental cross-talk confirmed below $\beta$24 (Hue et al., 11 Mar 2026). Predicted performance for an 8 m HWO telescope includes $\beta$25 per resolution element in 3600 s on a $\beta$26 mag point source in the VIS arm at $\beta$27, $\beta$28–100 in the FUV in 1 hr for an adjusted $\beta$29 target, and $\beta$30 on a $\beta$31 star in 1 hr in NIR (Hue et al., 11 Mar 2026).

Technology maturation is centered on dedicated ground test benches. The 2026 Pollux test-bench description gives two reconfigurable optical layouts in a clean-room vacuum chamber: MgF$\beta$32-based NUV/MUV benches covering 115–300 nm and a mirror-only FUV bench covering 100–123 nm (Girardot et al., 11 Feb 2026). The chamber reaches $\beta$33 mbar in 12 min, uses a 1 $\beta$34m pinhole and two-mirror collimator with wavefront error $\beta$35, and targets $\beta$36 through repeated generation of known Stokes states and demodulation with the combined Mueller matrix (Girardot et al., 11 Feb 2026). No full polarimetric data set is yet published, but the reported goal is an increase from TRL 3 to TRL 5 for the NUV/MUV polarimeters and from TRL 4 to TRL 5 for the FUV mirror-based design (Girardot et al., 11 Feb 2026). Parallel CNES-funded R&D on space UV polarimeters reports optical contacting tests between $\beta$37 and $\beta$38, with no failure in that range but slippage below $\beta$39, underscoring the need for controlled in-flight thermal conditions (Neiner et al., 7 Mar 2025).

6. Scientific scope of the instrument

The science portfolio of Pollux is unusually broad because the instrument combines high dispersion, wide spectral coverage, and UV spectropolarimetry. The recent HWO concept groups the science drivers into stars, planets and exoplanets, and cosmic ecosystems, including magnetospheric accretion in protostars, massive-star magnetism across metallicity, exoplanet atmospheres and escape, solar-system aurorae and cometary charge exchange, the baryon cycle in the ISM, CGM, and IGM, AGN and starburst feedback, and cosmological probes (Neiner et al., 10 Feb 2026). A more explicit performance-driven summary lists cosmology and large-scale structure through the Lyman-$\beta$40 forest and metal-line absorbers at $\beta$41, stellar magnetism through Zeeman Doppler imaging, exoplanet atmospheres through transmission spectroscopy of H$\beta$42O, CO, CO$\beta$43, and CH$\beta$44, and solar-system ocean worlds through spectropolarimetry of icy surfaces and tenuous atmospheres (Hue et al., 11 Mar 2026).

The ocean-worlds case is the most detailed recent application. Pollux is proposed to measure the linear polarization $\beta$45 across 100 nm–1.9 $\beta$46m at phase angles $\beta$47 from $\beta$48 to $\beta$49 in order to map the Negative Polarization Branch, inversion angle, and Polarization Opposition Effect (Hue et al., 11 Mar 2026). The same program uses absorption bands of H$\beta$50O ice at 1.04, 1.25, 1.5, and 2.0 $\beta$51m, CO$\beta$52 at 1.57 $\beta$53m, and salts in the 1–1.2 $\beta$54m region to infer composition, while UV spectroscopy of O I 130.4 and 135.6 nm targets airglow brightness and Doppler shifts in giant-planet moon environments (Hue et al., 11 Mar 2026). In Titan’s haze, repeated observations of limb polarization across 200–600 nm are intended to constrain aerosol size distributions, shape, refractive index, and seasonal change (Hue et al., 11 Mar 2026).

AGN science was a core LUVOIR-era justification for the instrument. The AGN study emphasizes full UV coverage from the Lyman limit to 4000 Å, $\beta$55, and full-Stokes retrieval for diagnosing accretion-disk physics, jet magnetic fields, BAL-quasar winds, and dust composition (Marin et al., 2018). It cites continuum linear polarization up to $\beta$56 from electron scattering in ionized accretion-disk atmospheres, 40–60% linear polarization in blazars and radio galaxies from synchrotron emission, several-percent UV polarization in BAL QSOs, and polarimetric sensitivity $\beta$57 with absolute accuracy $\beta$58 after calibration (Marin et al., 2018).

Massive-star studies provide a complementary stellar-use case. The LUVOIR massive-star paper defines a three-arm UV Pollux covering 90–400 nm at constant $\beta$59 and emphasizes UV resonance diagnostics such as N V 124 nm and C IV 155 nm for mass loss and terminal wind speed, polarized line profiles for 3D magnetospheres, and UV signatures of very massive stars, including He II 164 nm emission, O V 137 nm absorption, and the absence of P Cygni in Si IV 140 nm (Neiner et al., 2018). The broader POLLUX science case also includes exoplanet reflected-light polarimetry, where a hot Jupiter at full phase can reach a net polarized signal of $\beta$60 when unresolved from its host star (Bouret et al., 2018). Taken together, these papers position Pollux less as a single-purpose instrument than as a mission-scale platform for high-resolution spectropolarimetric astrophysics.

7. Later computational and information-science uses of the name

Outside astronomy, “Pollux” has been reused for technically unrelated systems. In distributed deep learning, Pollux is a scheduler that defines

$\beta$61

and co-optimizes GPU allocation, per-GPU batch size, and gradient-accumulation steps by combining per-job modeling with cluster-wide scheduling (Qiao et al., 2020). In experiments with real DL jobs and trace-driven simulations, it reduces average job completion times by 37–50% relative to state-of-the-art DL schedulers, even when those baselines are provided with ideal resource and training configurations (Qiao et al., 2020). Here the term has no astronomical content; it identifies a scheduling framework centered on “goodput.”

In LLM evaluation, POLLUX is an open-source benchmark for Russian generative models. It defines a taxonomy of 35 task groups, 2,100 manually crafted prompts, 66 evaluation criteria, and a transparent scoring protocol in which models rate outputs against explicit rubrics and provide justifications (Martynov et al., 30 May 2025). The release includes 7B and 32B LLM-as-a-Judge evaluators. On the zero-shot test, the 32B judge reaches $\beta$62 points against expert means, Spearman $\beta$63, and verdict confidence $\beta$64 (Martynov et al., 30 May 2025). In this usage, “POLLUX” denotes an interpretability-oriented benchmarking framework rather than a scientific instrument.

In political-science information infrastructure, Pollux is the Specialised Information Service for Political Science in Germany. A 2024 technical report describes the addition of political blogs by introducing the new record types blog_feed and blog_entry, two Python micro-services rss_downloader.py and rss_converter.py, and BERTopic-based analysis of translated summaries (Holtdirk et al., 2024). The weekly pipeline covers $\beta$65 feeds, processes $\beta$66 GB of XML in $\beta$67 minutes, and added 22,739 posts to the index during the reported integration period (Holtdirk et al., 2024). The shared name therefore spans a star, an instrument, a scheduler, a benchmark, and an academic information service, but only the astronomical usages are historically connected by subject matter.