Constraining Axion Dark Matter with Galactic-Centre Resonant Dynamics

This paper investigates the dynamical influence of fuzzy dark matter (FDM) — composed of axions or axion-like particles — on the orbits of stars at the Galactic centre, and derives new constraints on the axion particle mass m a from the stability of the observed young stellar disc.

Physical setup. FDM haloes form Bose–Einstein condensates described by a single field psi solving the Schrödinger–Poisson system. Such haloes develop dense solitonic cores at their centres, with core mass M c and radius R c related to the halo virial mass M vir by:

[

R c = 100, pc times (10-22, eV over m a) squared (10 9 M over M c),

]

[

M c = 6.7 times10 7 M (10-22, eV over m a) (M vir over 10 10 M) 1/3.

]

Around the supermassive black hole (SMBH) of mass M about 4 times10 6 M, this core forms a gravitational atom whose eigenstates are hydrogen-atom-like, with Bohr radius

[

r B = squared over G M m a squared.

]

The paper focuses on m a about 10-20 – 10-19 eV, for which r B about parsec, comparable to the stellar disc radius.

Orbit-averaged potential. Assuming the hierarchy m s M a M (star mass, atom mass, SMBH mass), the atom's gravitational potential on a star is

[

(x) = -integral d 3y G rho a(y) over x-y, rho a(y) = M a psi(y) squared.

]

The authors double-average this over the star's mean anomaly and argument of pericentre (fast time-scales), obtaining

[

da = -sum l,m J lm Y* lm,

]

with coefficients J lm given by equation (37), involving 3-j symbols, radial hydrogen wave-functions, and the functions s l(a,y,e,0) from resonant-relaxation theory. For a superposition of non-degenerate eigenstates, only m=0 terms survive, giving

[

da = -sum l 0 sum l=0 2l 0 J l,l 0 P l(times a),

]

where a is the atom's angular-momentum direction. This induces precession of stellar angular momenta about a with frequency

[

= -G M a m s over L r B sum l 0,l I l,l 0 P'l(times a) a,

]

where I l,l 0 are dimensionless coefficients (equation 12). The authors conclude: Equation (11) implies that an axion core with some angular momentum would participate in the VRR dynamics of the Galactic centre, and thus must be accounted for in the latter's modelling.

Disc stability criterion. The paper derives a stability condition for a stellar disc under the external torque from the gravitational atom, comparing it with the disc's internal torque. The criterion (equation 15–16) is:

[

sum l,l 0 2 J l,l 0 over G M d P''l(times a) 2 theta < xi d over a d,

]

where M d is the disc mass, theta is the angle between the disc angular momentum d and a, xi d is an order-unity parameter, and a d is the disc semi-major axis. This is stated as the main theoretical result of this paper.

Specific model for the atom's state. The authors argue, based on energy minimisation and angular-momentum constraints, that the gravitational atom's wave-function is plausibly

[

psi about alpha 100100 + alpha 211211,

]

where 211 is the n=2, l=1, m=1 hydrogen-like state. The rotating-mass fraction is parameterised by

[

alpha squared l not equal to 0 psi squared M a over M a, exp,

]

with M a, exp the expected atom mass from equations (1–2) and (17):

[

M a about M a, exp M c ((GM/sigma squared, R c) over R c) cubed.

]

For the Milky Way halo spin parameter lambda c about 0.05 – 0.09, equation (21) gives

[

alpha squared about 8.4 times10-2 (lambda c over 0.06),

]

with the angle between the disc and halo angular momenta theta about 79.

Constraints. For this model, the stability inequality (26) becomes

[

alpha squared M a, exp < M d r B xi d over a d, 6 I 2,1 2 theta.

]

Using parameters from Table I (M vir = (1.1 plus or minus0.1) times10 12 M, M d = (6500 plus or minus3500) M, iota = 14 plus or minus4 or 16 plus or minus4.8, lambda = 0.067 plus or minus0.018), the authors find the disc would be disrupted for

[

m a in [4.2, 5.48] times10-20, eV (2 sigma, conservative iota),

]

or

[

m a in [3.82, 5.98] times10-20, eV (2 sigma, iota = 16 plus or minus4.8).

]

The disruption time-scale is estimated as d-1 < 10 Myr for m a in [3,7] times10-20 eV.

Generalisations. The paper also treats the case where only a fraction f FDM of dark matter is fuzzy, showing the constraint can be recast as a joint constraint on (m a, alpha) via equation (26), with alpha alpha f FDM 2/3. Figure 4 shows the resulting 1σ and 2σ exclusion contours in the (m a, alpha) plane.

Conclusions. The authors state: "we showed that the dynamics of the Galactic centre are strongly influenced by ultra-light axions, which cannot be neglected in the dynamical modelling of the stellar distribution there—if these particles indeed constitute a sizeable fraction of dark matter. They emphasise that the constraints will tighten significantly with future, improved data" and could also apply to nuclear discs in other galaxies.

Based on this paper, I can identify several specific improvements for AI systems, particularly in the domains of astrophysical modeling, scientific computation, and data-driven inference.

1. AI for Orbit-Averaged Potential Computation

• Improvement: Train a neural network or symbolic regression model to directly compute the double-averaged gravitational potential coefficients J lm (equation 37) for arbitrary axion field configurations psi, without needing to evaluate the computationally expensive 3-j symbols, radial integrals, and s l functions each time.

• What the improved AI can do: Given a set of quantum numbers (n,l,m), the star’s orbital parameters (a,e), and the black hole mass M, the AI can output the full set of J lm coefficients in milliseconds, enabling rapid parameter-space scans (e.g., for m a and alpha) that would otherwise take hours or days.

2. AI for Stability Criterion Optimization

• Improvement: Implement a gradient-based or reinforcement-learning optimizer to solve the inequality (16) for the disc stability boundary, rather than the current brute-force minimization over semi-major axes and disc thickness.

• What the improved AI can do: Given a disc mass M d, opening angle iota, and axion parameters (m a, alpha), the AI can directly output the critical stability curve (like the blue line in Figure 2) and its uncertainty band, accounting for correlated parameter errors (e.g., M vir and lambda) via a probabilistic surrogate model.

3. AI for Bayesian Parameter Inference with Correlated Uncertainties

• Improvement: Build a hierarchical Bayesian model that treats alpha, lambda c, M vir, M d, and iota as latent variables with priors from Table I, and uses a normalizing flow or variational autoencoder to sample the joint posterior of (m a, alpha) given the disc’s existence.

• What the improved AI can do: Instead of the current 1σ/2σ shaded regions, the AI can produce a full posterior distribution for m a and alpha, automatically propagating all uncertainties (including the 4.8° error on iota) and providing a quantitative probability that the disc is stable for any given m a. This would directly yield the 2σ exclusion interval 4.2 times 10-20 eV m a 5.4 times 10-20 eV with a rigorous statistical interpretation.

4. AI for General Wave-Function State Decomposition

• Improvement: Develop an AI-based spectral analyzer that, given a simulated or observed FDM core density profile rho a(y), automatically decomposes it into the eigenstate basis nlm (equation 6) using a learned mapping from density snapshots to quantum numbers.

• What the improved AI can do: For any arbitrary core configuration (not just the simplified 100 + 211 state), the AI can extract the dominant alpha nlm coefficients, allowing the stability analysis to be applied to realistic, non-idealized cores from simulations (e.g., those with turbulence or partial accretion), without manual quantum-number guessing.

5. AI for Time-Scale and Disruption-Time Prediction

• Improvement: Train a regression model on N-body simulations of VRR (like those in refs. [19, 25]) to predict the disc disruption time d-1 as a function of (m a, alpha, M d, iota, theta), rather than using the analytic estimate.

• What the improved AI can do: Given a set of parameters, the AI can output a probability distribution for the time until the disc breaks, enabling a direct comparison with the observed disc age (10 Myr). This would allow the constraint to be tightened by requiring not just stability but also survival for the disc’s actual lifetime.

6. AI for Multi-Galaxy Generalization

• Improvement: Create a transfer-learning model that takes the Milky Way constraints and generalizes them to other galactic nuclei (e.g., M31, NGC 4248) by scaling with M, M vir, and lambda, using a physics-informed neural network that respects the scaling relations in equations (1–3, 17).

• What the improved AI can do: Given a target galaxy’s basic properties (black hole mass, halo mass, spin parameter), the AI can predict whether its nuclear stellar disc would be stable against FDM torques, and output the corresponding excluded m a range, enabling multi-wavelength observational tests without re-deriving the full formalism.

7. AI for Real-Time Data Assimilation

• Improvement: Implement an online learning system that updates the (m a, alpha) constraints in real-time as new Galactic center stellar orbit data (e.g., from GRAVITY or future ELT observations) become available, using a Kalman-filter-like approach on the angular-momentum vectors of disc stars.

• What the improved AI can do: As individual stellar orbits are refined, the AI can continuously re-evaluate the stability criterion (16) and update the exclusion contours, providing live alerts if the data begin to favor or disfavor a particular m a value, without requiring a full re-analysis of the entire dataset.

8. AI for Handling Degeneracies with Nuclear Cluster

• Improvement: Build a multi-component dynamical model that includes the nuclear cluster’s angular momentum (which is misaligned with the disc) and use a differentiable simulator to jointly fit the disc stability, the cluster’s torque, and the FDM atom’s torque.

• What the improved AI can do: The AI can disentangle the FDM contribution from the cluster’s contribution to the disc’s warp, providing a cleaner constraint on m a even in the presence of the cluster, and potentially resolving the degeneracy mentioned in the paper’s discussion.