WorksPhysics and Astronomy
Black Hole Love Numbers with Physics-Informed Neural Networks
Abstract
We develop and validate a physics-informed neural network (PINN) framework for computing the tidal Love numbers of black holes directly from their perturbation equations. We consider static perturbations of the Kerr black hole, whose exact hypergeometric solutions provide a controlled benchmark for both the numerical solver and the extraction of the physical response. The framework combines exact enforcement of the horizon boundary conditions with an -Sobolev loss that incorporates derivatives of the differential-equation residual. The Love number is then obtained by extracting the asymptotic coefficients of the learned solution, without using the analytic connection formula. Across the tested scalar, vector, and tensor perturbations, the method recovers the finite dissipative response with relative errors below . Controlled comparisons show that the higher-derivative loss terms improve solution accuracy by orders of magnitude, while independent tests of the extraction procedure distinguish solver errors from errors in recovering the response coefficients. The conservative response, which vanishes analytically, is constrained by the numerical error floor. These results demonstrate that PINNs can accurately compute black-hole tidal response and establish a validated framework for future extensions to frequency-dependent and non-separable perturbation problems.
Introduction
Physical observables are often determined indirectly from solutions of differential equations: the equation determines a function, and the quantity of interest is read off from its boundary or asymptotic behaviour. Examples include response coefficients, scattering amplitudes, eigenvalues, and quasinormal frequencies. The accuracy of the observable therefore depends not only on the accuracy of the numerical solution, but also on the procedure used to extract it.
The extraction step becomes important when the observable is small or vanishes exactly. Such a quantity is typically a subdominant coefficient in an asymptotic expansion, or the result of a cancellation between much larger contributions. Errors in the extracted quantity can then greatly exceed errors in the underlying solution.
Tidal Love numbers characterise the linear response of a compact object to an external gravitational field and are relevant to gravitational-wave observations and to tests of strong-field gravity. For black holes the Love number is a complex coefficient: its real, conservative part measures the deformability of the horizon under an external perturbation, and its imaginary, dissipative part the rate of energy dissipation. Both matter observationally in the inspiral of a binary, when each object is tidally deformed by its companion. For four-dimensional black holes in general relativity the conservative Love numbers vanish identically in the static limit [1, 2, 3, 4, 5], while the response of a spinning black hole also contains a finite dissipative part [6, 7, 8, 9]. The vanishing of the conservative part is explained by a set of hidden conformal symmetries [10, 11, 12, 13, 14, 15, 16, 17, 18, 19]. Both a vanishing and a finite response coefficient are therefore available within the same perturbation problem, which makes Love numbers a clean test of this effect.
In higher dimensions the pattern is richer and the static tidal Love numbers do not universally vanish [20, 21, 22, 23, 24, 25]. For a comprehensive review of tidal deformations of compact objects see [26, 27].
The calculation of tidal Love numbers has received considerable attention in recent years. Approaches include direct numerical methods, effective field theory [23, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37], master-equation analyses via perturbative schemes such as the MST formalism [38] or connection-formula implementations [39], and dualities such as the AGT correspondence [40, 41]. These methods are accurate within their perturbative domain of validity. The traditional route to the computation of the Love number involves the linear transformation formula relating the horizon and asymptotic boundary expansions, expressed in closed form via products of Gamma functions,
The coefficients appearing in the corresponding expression are defined and discussed in more detail later in the text. This approach is analytic, but at the integer angular numbers that arise in physical applications the two Gamma functions in the numerator and in the denominator both sit at poles, while every other factor stays finite provided (for the denominator factor is singular as well, and vanishes). The Love number is therefore defined only through the cancellation of these two divergences, and the limit must be approached rather than evaluated directly.
In this work we compute the Love numbers without the connection formula, with a fully numerical approach based on Physics-Informed Neural Networks (PINNs) [42]. The PINN solves the perturbation equation, and the Love number is then extracted from the asymptotic coefficients of the learned solution. The calculation therefore has exactly the two steps described above, and because the exact solution is known, the error of each step can be measured separately. PINNs have previously been applied to quasinormal modes [43, 44, 45] for the Schwarzschild and Kerr cases, and generalized solvers for the Regge–Wheeler and Teukolsky equations have been developed [46]. Extensions cover asymptotically dS and AdS solutions [47, 48], higher dimensions [49] and modified gravity [50, 51].
Neural-network solvers are also relevant in a setting where the standard methods do not apply at all. PINNs do not require separability. The Teukolsky construction, and with it the expansions and connection formulae, rests on the separation of the perturbation equation into radial and angular parts, a structure special to Kerr and to a handful of related backgrounds. For modified-gravity black holes, for numerically constructed stationary solutions, and for rotating solutions in higher dimensions, no exact separation is known in general, in particular for gravitational perturbations at finite spin, and the tidal response is known mostly in the non-rotating or slowly rotating limits [31, 52, 53]. A residual-minimising solver is not tied to separability: the residual loss carries over to a partial differential equation in several variables, although the boundary treatment and the extraction of the response would have to be adapted. This has recently been demonstrated for the quasinormal-mode problem, where the two-dimensional hyperboloidal Teukolsky equation was solved directly, without separation, as a prototype for beyond-Kerr backgrounds [54]. Spectral methods that avoid separation also exist and are substantially more accurate [55, 56]; the case for a PINN rests on flexibility and on requiring little analytical pre-processing, not on precision.
The PINN applications cited above are mostly quasinormal-mode calculations, which are eigenvalue problems. The boundary conditions select a discrete frequency, and the accuracy of that frequency is the accuracy of the result. A tidal response coefficient is a connection problem of a different character. The frequency is fixed (zero in the static case studied here), a solution exists for every parameter value, and the observable is a ratio of two coefficients in the expansion of that solution about the boundary, one of which multiplies a strongly suppressed local behaviour. The accuracy of the solution and the accuracy of the observable are then distinct questions, and the second does not follow from the first. That gap is the subject of this paper.
The Kerr black hole tidal Love number problem is well suited to a controlled numerical study. The full frequency-dependent perturbation equation is of confluent-Heun type [57], while in the static limit it reduces to a hypergeometric equation. In this paper, we use the static equation as a controlled test case, leaving the non-static case for a follow-up study. Because the static equation admits an exact hypergeometric solution, errors in solving the differential equation can be separated from errors introduced by extracting the physical response. Instead of evaluating the analytic connection formula, we train a neural network to represent the solution regular at the horizon directly and extract the asymptotic coefficients and by fitting the solution to the two independent solutions defined at the asymptotic boundary.
Our primary goal is to develop a highly accurate solver for the differential equation governing black-hole perturbations and, at the same time, a reliable procedure for extracting the associated physical observable, the Love number. The static problem provides an ideal controlled setting in which to develop and test both components: the exact solution is known, the relevant asymptotic data can be extracted explicitly, and the accuracy of the resulting Love number can be assessed independently. The object of study is accordingly the extraction step as much as the solver. The static Love number is used because its exact value is known, which is what makes it possible to separate solver error from extraction error. The conclusions drawn about that separation apply to any solver whose output is sampled and projected onto an asymptotic basis. For the solver, we use a weighted -Sobolev norm for the training loss. The loss penalizes the differential-equation residual and its first two derivatives, rather than the residual alone, providing greater control over the local behavior of the solution and its derivatives. The weight diverges at the two singular endpoints and so penalises the residual most strongly where the Frobenius structure of the solution is delicate.
A small differential-equation residual alone, however, does not guarantee an accurate physical observable. We therefore implement and analyze different methods for reliably extracting physical information encoded in the asymptotic behavior of the solution. This distinction is important here because the Love number is obtained from a delicate combination of terms in the near-boundary expansion, with multiplying a high-order asymptotic contribution.
Since the analytical static solution is known, we can disentangle errors in the neural-network solver from those introduced by the extraction procedure. We apply the same extraction method to both the Frobenius solution and the trained network. Errors common to the two calculations arise from the extraction procedure, whereas additional errors in the network result originate from the numerical solver.
The static problem is also the natural first step toward the more general, frequency-dependent problem. Our aim in future work is to retain the solver and observable-extraction framework developed here and subsequently extend it to nonzero , where the analytic structure is substantially more involved. Beyond the black-hole application, the same framework provides a general approach to second-order ODEs with singular points. In particular, from the perspective of singular-point classification, the present equation is the hypergeometric member of the broader class containing the general Heun equation and its confluent limits. Related approaches have recently been explored [58, 59], but, to the best of our knowledge, have been restricted to real parameters and real-valued solutions. Our solver instead treats the hypergeometric equation with complex parameters and complex-valued solutions.
The paper is organized as follows. Section 2 reviews the hypergeometric equation, introduces a regulator that lifts the degeneracy of the connection problem, and presents a field redefinition that regularizes the behavior at . It also establishes that the static Love number is identical for spin weights and . Section 2.2 recasts the resulting equation as a coupled system of two real ODEs. Section 3 introduces the PINN architecture, the weighted loss, and the training procedure, explaining how the horizon boundary condition is imposed and motivating our choice to solve the equation at spin weight . Section 4 analyzes how errors propagate from the loss function to the solution and the extracted coefficients, and presents the trained solution for our flagship configuration. Section 5 introduces eight methods for extracting and and presents the corresponding numerical results. Section 5.1 provides a closed-form explanation of why one family of extraction methods necessarily fails, while Section 5.2 examines the exponent convention that governs the accuracy of the entire computational pipeline. The subsequent subsections investigate the role of the Sobolev loss, the dependence of coefficient errors on the regulator , the resulting physical Love number, and the behavior across angular modes . Finally, Section 6 summarizes our conclusions and outlines directions for future work, while Appendix A derives the hypergeometric form of the static Teukolsky equation.
Set-up
Our starting point is the radial Teukolsky equation [60, 61, 62, 63] governing perturbations of the Kerr background, evaluated in the static limit . Throughout we work in units , with the black-hole mass, so that the horizon radii are
and spin parameter , with extremality occurring at . The details of the reduction are given in Appendix A; here we record the result. With the compactified radial variable
which maps the black hole horizon , the asymptotic boundary and Cauchy horizon , and with the substitution
the static radial equation becomes the hypergeometric equation [9]
with parameters
Here and are the angular and azimuthal numbers and is the spin weight of the perturbing field, constrained by and . The black-hole spin enters only through the single combination
Two features of (6) will be important in what follows. First, and are complex whenever , while is an integer. The complex parameters make complex-valued, which is why the PINN of Section 2.2 is trained on a coupled system for the real and imaginary parts. They also give rise to the dissipative (imaginary) part of the Love number: when , all parameters are real and is real. The integer makes the Gamma-ratio formula (13) indeterminate at physical parameter values, since vanishes exactly where has a pole. It has a second consequence. With a non-positive integer the hypergeometric series (18) for terminates, so at integer angular number is a polynomial of degree . Since implies , both local solutions at are then analytic and no logarithm appears. This is the degenerate case without a logarithmic branch, and it is the same fact that makes the conservative response vanish.
Second, the combination
is the difference of the Frobenius exponents at . Recall that the two Frobenius behaviors there are and , corresponding in the -coordinates to the response and the tidal field, respectively (Appendix § 9). For integer , this difference is the positive integer , placing the equation in the degenerate (resonant) case of the hypergeometric family.
Rather than working directly with the degenerate form of the connection formula, we regulate the exponent difference by analytically continuing the angular index,
where denotes the physical angular number. The corresponding hypergeometric parameters become
the last because does not depend on . The exponent difference at becomes
We reserve for this real valued, non-integer exponent difference and for its integer part. The two coincide as , but they differ at , which is the same order as the quantity ultimately being measured.
For nonzero , the exponent difference is noninteger and the standard Kummer connection formula may be applied without encountering the poles associated with the resonant case. We therefore first perform the analytic continuation and the expansion about at finite , and only subsequently take the limit . The tidal response coefficients can therefore be extracted before taking the resonant limit, after which the physical static Love number is obtained from the resulting finite ratio. The continuation is a formal device and carries no physical content. For non-integers the angular equation no longer admits solutions regular on the full domain, and we use (9) solely to define the limit of the radial connection problem, which is where the physical statement resides.
Of the two solutions to the second order differential equation in Eq. (5), the regularity condition at the horizon () selects the regular branch . Expanding this horizon-regular solution about using the standard linear transformation formula1 gives
where the connection coefficients and are determined by the hypergeometric transformation and are given by
Note that remains finite and non-zero as : its argument is positive, and neither nor has a pole there. The entire singular structure resides in , through the ratio of two divergent factors. The static Love number is then [9]
which agrees with Eq. (1). This analytic procedure motivates a complementary numerical approach with PINNs. Before Eq. (5) can be handed to a numerical solver, however, several considerations must be addressed, since the aim is a solver accurate enough to resolve the subleading coefficient and not merely the solution. First, the boundary conditions must be imposed on a function that stays bounded at both singular points, whereas itself grows as at . Second, a noise-free reference solution is needed against which the solver and the extraction of and can be tested independently. Third, the parameters and are complex, so the equation is complex-valued and the solver must handle a coupled system for its real and imaginary parts. We take these in turn.
The spin weight enters Eq. (1) only through the factor . Applying to both Gamma functions, the ratio of this factor at to its value at is , which equals one for integer and any , including . The static Love number is therefore the same whether it is computed from the equation with spin weight or , in line with the Teukolsky–Starobinsky identities relating the two [63]. The boundary conditions are also unchanged: with the prefactor of Eq. (4), the ingoing solution corresponds to the exponent $0$ of at for every , and the exponents at do not depend on $s$ ({xref}Appendix A`). We make use of this in later sections.
Boundary conditions
To impose boundary conditions appropriate for tidal deformations and to extract Love numbers from a smooth function, we adopt the field redefinition
with the exponent difference at , Eq. (11), after which is regular at both singular points. The exponent here is not a matter of choice. Under the exponents at shift from to , while those at are untouched. For the transformed equation to remain hypergeometric one exponent at must vanish, which forces or . (15) is thus the unique non-trivial rescaling within this family that preserves the hypergeometric class; any other exponent, the integer included. Under (15) Eq. (5) retains hypergeometric form,
with shifted parameters
so that . Equivalently, by Euler’s transformation,
which is the closed form against which every numerical statement below can be verified. The tilded solution is regular at both singular points, making it a convenient variable for the numerical treatment. The horizon boundary condition is simply
Near spatial infinity, , we expand in the two local solutions of the tilded hypergeometric equation,
where
The coefficients and therefore characterize the two asymptotic sectors and determine the corresponding response-to-tidal ratio. We note for later use that and need not be evaluated as hypergeometric functions at all. Substituting in Eq. (16) returns a hypergeometric equation in with parameters , so the exponents and the two series may be generated directly from the coefficients of the differential equation by Frobenius recursion,
Note that the Frobenius recursion only fixes the coefficient ratios so the leading coefficients and are free. We adopt the standard choice , The Love number identification therefore requires only the local structure at , and not the global connection formula.
Condition (19) alone does not fix the solution, because the horizon is a regular singular point. In Eq. (16) the coefficient of the highest derivative vanishes at , so the equation degenerates into an algebraic relation between and :
Combining Eq. (19) with Eq. (25) fixes the derivative,
We emphasize that Eq. (26) is not an additional physical input: it is forced by the ODE once one imposes Eq. (19) and demands that be analytic at the horizon.
Decomposition into real and imaginary parts
Because some of the coefficients carry the imaginary part , the tilded equation (16) has complex coefficients and is a complex-valued function of a real variable. A neural network with real weights and real activations cannot represent this directly, and automatic differentiation frameworks operate natively on real tensors. We therefore split the equation into its real and imaginary parts and solve the resulting coupled system for two real functions on . The two components are coupled precisely through the imaginary parts of the parameters, which is what carries the dissipative response.
We write
and define
Here and are the real and imaginary parts of the product , while and are those of the linear coefficient . Substituting into Eq. (16) and equating real and imaginary parts gives the coupled system
with boundary conditions , . The two equations share the same second-order operator and differ only in the sign with which and couple them; setting makes , the system decouples, and solves the imaginary part identically, recovering the real problem noted below Eq. (7). This is the system the PINN is trained to solve.
Physics-Informed Neural Network architecture
Having reduced the problem to the real system (30) with the boundary data (19) and (26), we now describe the neural-network approach to solving it. A PINN represents the unknown solution by a feed-forward network and trains it by minimising the residual of the differential equation itself, evaluated at collocation points drawn from the domain; no solution data are required. Two ingredients are specific to the present problem: how the boundary conditions at the regular singular point are imposed, and what norm of the residual is minimised.
A unique solution of a differential equation is singled out by its boundary conditions, and for PINNs these may be imposed in two ways [66, 67]. Under soft enforcement the boundary conditions enter as additional terms in the loss competing with the bulk residual during optimisation. Under hard enforcement the conditions are built into the architecture, so they hold identically for every value of the network parameters, either through a trial function that satisfies them by construction [68, 69, 70] or by constraining the network itself [71]. We use hard enforcement throughout, for reasons that are specific to this problem and are set out below.
Weighted -Sobolev loss
The standard PINN loss penalises the differential-equation residual, but does not explicitly penalise its derivatives. Define the real and imaginary parts of the ODE residual from (30) as
Standard PINN formulations minimize . Since we are solving a second-order equation and the Love number is obtained from the expansion coefficients, accurate recovery of the solution and its derivatives is important. We therefore also penalise the first and second derivatives of the residual. This gives an -Sobolev loss: a loss that combines the squared residual with its squared derivatives up to second order [72, 73]. For background on Sobolev spaces and weak derivatives, see [74, 75, 76, 77]. Independently of these derivative penalties, we introduce an endpoint weight to give greater emphasis to the regions near the two singular boundaries. The resulting weighted loss is
with weight
and Sobolev coefficients , . Here primes denote derivatives of the residuals with respect to . Their explicit expressions are given in Appendix § 10.
We comment below on several choices made in the construction of the loss.
In practice the integral is replaced by the empirical mean of the weighted integrand over 400 collocation points, drawn uniformly on and redrawn at each optimisation step.
We choose , the reciprocal of the coefficient of in Eq. (16), to increase the penalty near the horizon and the asymptotic boundary. These are the regions where the boundary conditions and expansion coefficients are determined. This is a numerical weighting choice, rather than a requirement of a Sobolev loss. Its effect is tested separately from that of the derivative penalties in Table 7. Because the weight diverges at both endpoints, we evaluate the loss on the truncated interval , where it remains finite.
| Sobolev | weight | loss | spread | ||
Table 7. Sobolev order crossed with endpoint weight, at , , 3000 epochs, two seeds, extraction by Method 1. Median over seeds. The loss column is the value of a different functional in every row and is not comparable between them.
We first derive the first and second derivatives of the residuals analytically by applying the product rule to Eqs. (31)–(32), using the explicit linear forms of and . The resulting expressions are given in Appendix § 10. We evaluate the derivatives of the network outputs and appearing in these expressions using automatic differentiation. The standard residual requires network derivatives up to second order; the first- and second-derivative Sobolev terms require network derivatives up to third and fourth order, respectively.
The first term alone in (33) without the Sobolev terms admits residuals that are small in mean but oscillatory. Including suppresses such oscillations, and further enforces smoothness of the residual itself. We use and as a conservative hierarchy in which each higher-derivative penalty is subdominant to the previous. These values were fixed before the sensitivity scan, by a short trial on the benchmark at as an order-of-magnitude choice making each penalty roughly ten times smaller than the previous one. The results depend only weakly on this choice (Section 5.3, Table 6).
| seed range in | ||||
| 0.1 | 0.01 | |||
| 0.01 | 0.01 | |||
| 1 | 0.01 | |||
| 0.1 | 0.001 | |||
| 0.1 | 0.1 |
Table 6. Sensitivity to the Sobolev coefficients at , , 5000 epochs, extraction by Method 1, three seeds per row; median over seeds, with the seed range for in brackets. The first row is the default. It was run separately from Table 5, and the difference from the corresponding row there reflects run-to-run variation.
Our training procedure uses Adam followed by an L-BFGS refinement stage. For the initial benchmark, Adam alone achieves the reported accuracy, and the L-BFGS stage is omitted (§ 4.1). Both stages are used for the angular-mode scan (§ 5.5). The Adam stage uses the optimizer Adam [78] at base learning rate , using 400 collocation points drawn uniformly at each epoch. During training we keep the network with the lowest validation loss, the same loss evaluated on a fixed grid of 4000 equally spaced points in that are not used for training, rather than the loss on the randomly drawn training points, which varies from epoch to epoch. For the angular-mode scan, the Adam stage (750 epochs) is followed by L-BFGS. L-BFGS is run in rounds of 20 iterations on a stratified grid of 2000 collocation points, which is redrawn every ten rounds so that the optimizer does not fit a fixed set of points. Training stops when the validation loss improves by less than 5% over three rounds. Each configuration is trained from scratch. Future work could reduce the cost of these scans using solution bundles that learn a family of solutions over the parameters [79], or transfer learning between nearby parameter values [80].
Hard enforcement of the horizon boundary condition
The physical boundary condition at the horizon is Eq. (19). The point is a regular singular point of the tilded equation, with the two Frobenius branches
The regular branch is the physical one. A naïve neural-network ansatz of the form enforces Eq. (19) exactly but leaves the derivative unconstrained. This freedom is problematic, since it allows the network to absorb a contribution from the singular branch , which is precisely the unphysical mode we seek to exclude. Since has , the singular branch and its first derivative both vanish at the horizon for , so neither nor detects it directly; what distinguishes it is that is non-analytic at whenever . A parametrisation that is smooth at the horizon by construction excludes it in principle, whereas a soft penalty on the boundary values cannot; we return to how well this works in practice below Eq. (36). We therefore impose both the boundary value and the first Frobenius recursion (26) directly through the architecture. Splitting with , we parametrise
where are the two real output channels of a single fully-connected feed-forward network whose weights and biases we denote collectively by . The factor guarantees that both Eq. (19) and Eq. (26) hold exactly at every training step, for any .
For the exclusion is not effective in practice. Near the horizon the singular branch behaves as . Only the slowly varying phase distinguishes it from a smooth function. We therefore solve the equation at spin weight and report the result for , which is allowed because . At the singular branch behaves as , diverges at the horizon, and cannot be represented by any bounded ansatz. The change has a second effect. At integer , is a polynomial of degree up to corrections, so lowering lowers the degree and the amplitude that the network must represent.
The ansatz also has a consequence for the weighted loss introduced above. Evaluating the residual of Eq. (30) at under the constraints (19) and (26) gives
identically, by Eq. (25). The loss weight diverges at , so for a generic ansatz the weighted norm would diverge there. The hard-enforced parametrisation is therefore required for the norm to be finite, and not only for the reasons given above.
In the implementation, is a fully-connected network with four hidden layers of input dimension 1, hidden width 128 and output dimension 2, with hyperbolic-tangent activations on the hidden layers and a linear output layer. The smooth activation is essential. The loss Eq. (33) requires up to fourth-order derivatives of the network output with respect to , obtained by repeated reverse-mode automatic differentiation through the tanh stack. Because tanh is , the resulting PINN solution is infinitely differentiable, which is what permits the higher-order Sobolev penalties to be defined at all. All arithmetic is performed in double precision (float64), including the network parameters, the collocation points and every automatic-differentiation pass; the exact reference solution is evaluated with mpmath at 50 significant digits.
From solution error to observable error
The quantity a PINN minimises is the equation residual, and the quantity reported in most PINN papers is its final value. Neither is the physical observable. The chain from one to the other has three links,
where is the trained loss, is the solution error, and are the relative errors in the extracted coefficients and in . This section quantifies the first link.
Training and solution profile
The network is trained with the weighted Sobolev loss, hard-enforced boundary conditions at , and 400 collocation points resampled per epoch. Unless stated otherwise, the results of this section and of Sections 5.1–5.4 are for the configuration , trained with Adam alone for 2500 epochs. The scan of Section 5.5 uses the two-stage Adam and L-BFGS training.
Figure 1 compares the trained PINN output and its first two -derivatives against the Frobenius series. The real and imaginary parts and are reproduced closely. The first derivatives match to comparable accuracy thanks to the term in the loss. The second derivatives also track the series, with a mild endpoint deterioration as where the collocation density vanishes. The hierarchy is quantitative: at the maximum pointwise errors are , and , each derivative costing roughly a factor of two. The first two maxima occur at the boundary, , and the third in the interior, at (Figure 2).

Figure 2. Pointwise error of the trained solution and of its first two derivatives against the exact series, on a logarithmic scale, for the configuration of Figure 1. Left: real part and right: imaginary part.

Figure 1. Trained PINN output (red dashed) versus the Frobenius series (black solid), together with the first and second -derivatives. Left column: and right column: . Configuration , , , and 2500 epochs.
Extraction of physical observables
Having established a high-accuracy solver for the differential equation, we can now focus on the extraction of the physical observable and quantify the errors in the latter link of the chain (38).
Once the network has been trained, we read off and from the expansion of at , Eq. (20), with the basis of Eqs. (23)(24) and the real exponent of Eq. (11). With this choice Eq. (20) is an identity with exactly the Gamma ratios (13), as follows from Euler’s transformation (18) together with the Gauss connection formula. Using the integer exponent in place of breaks that identity at ; § 5.2 quantifies the resulting error. However, for the third parameter of reaches the non-positive integer and the basis itself becomes ill-defined; for both series are well defined and we evaluate them in arbitrary-precision arithmetic, either from the closed forms (21)(22) or, equivalently, from the recursions (23)(24).
We implement eight extraction methods for , that can be grouped in three families.
Global projection (Methods 1, 2, 5): the solution is sampled on a dense subset of and projected onto by least squares. These average solver noise over many points and exploit the fact that and are well separated over the bulk of the domain even where their endpoint behaviours nearly coincide.
Matched (Methods 3, 4): the solution is combined with one or more of its derivatives at a small number of points, giving a square or mildly overdetermined system. These use less data and are correspondingly more sensitive to the conditioning of the local basis matrix.
Local-derivative (Methods 6, 7 and 8): is read directly off the -th derivative of the solution at the endpoint . This is the extraction method that Eq. (21) appears to suggest, and § 5.1 shows that the appearance is misleading.
These are
Method 1 — Integral residual. Minimize the weighted distance between and its decomposition,
with emphasizing both endpoints, discretized on 400 points in and solved by least squares.
Method 2 — Moment matching. For integer moments equate against the same moments of , giving an overdetermined system solved by least squares; the integrals are evaluated by the trapezoidal rule on the Method-1 grid. Explicitly,
The trapezoidal weights on the Method-1 grid are , , and for interior indices . Averaging over the domain makes this method robust to localised solver noise.
Method 3 — Singular-point asymptotics. Since and , evaluating at and and inverting the system gives directly. Two function evaluations, but maximally sensitive to error in the boundary layer.
Method 4 — Derivative matching. At , match and simultaneously in an overdetermined system solved by least squares. The linear system is
where the network derivative is obtained by automatic differentiation, and are computed from Eq. (21).
Method 5 — Weighted least squares near . As Method 4 but value-only, at with weights concentrating information near the endpoint.
Method 6 — Direct local extraction, pointwise. Evaluate and at by successive autograd passes, and read and from the local Frobenius expansion. Concretely,
This is the most direct implementation of Eq. (21) but amplifies any high-frequency noise in the network output by the factor , and evaluates at a point where the collocation density used in training vanishes.
Method 7 — Direct local extraction, Taylor fit. Fit a degree- polynomial in to on 15 points within half-width 0.02 of , identifying with the constant term and with . Writing the polynomial fit as , the identifications are
The 15 points are logarithmically spaced in from to 0.02; the polynomial degree gives a mildly overdetermined fit while retaining sensitivity to the term.
Method 8 — Endpoint-derivative extrapolation. Evaluate on a cluster inside the training domain, fit a low-degree polynomial in to both and , and extrapolate the fits to ; the constant term of each fit is the extrapolated value at the boundary. Explicitly,
where each is a degree-3 polynomial in fit by ordinary least squares. Extrapolating from inside the training domain avoids the point where the collocation density vanishes, but as shown in § 5.1 the method still suffers the exact-cancellation failure common to Methods 6, 7 and 8.
A direct comparison on the performance of these methods for the case is shown in Table 1.
| Method | Frobenius | PINN’s | ||
| Re | Im | Re | Im | |
| 1 | ||||
| 2 | ||||
| 3 | ||||
| 4 | ||||
| 5 | ||||
| 6 | ||||
| 7 | ||||
| 8 |
Table 1. Love number value for , with and using only the ADAM optimizer. The analytical reference is . The PINN values therefore differ from Table 8, which reports the median over three seeds at 5000 epochs.
The error for both and is summarized in Table 2.
| Method | Frobenius | PINN’s | ||
| 1 | ||||
| 2 | ||||
| 3 | ||||
| 4 | ||||
| 5 | ||||
| 6 | ||||
| 7 | ||||
| 8 |
Table 2. Error values for and for , with and using only the ADAM optimizer. The analytical reference is .
Every extraction method described above takes a function as input, so the same code path can be used with either the trained network or the exact series Eqs. (23)(24). The difference between the two solutions is the solver’s contribution; anything the Frobenius series already gets wrong is a property of the extraction method and cannot be repaired by a better solver. This decomposition is the main diagnostic of the paper.
The series is evaluated directly at each requested point, with no interpolation, and its truncation is contrasted against Eq. (18) in mpmath.
Why the local-derivative family cannot work
The local extraction methods are based on Eq. (42), which identifies the -th Taylor coefficient of about with . The contribution from the second independent solution modifies this coefficient. Expanding in powers of ,
the term occurs at the same order, , as the leading contribution from . The quantity extracted by Eq. (42) is therefore
The coefficient remains finite in the limit . As noted below Eq. (7), both and contain a factor of order . For the latter, this follows from and for . These factors cancel, giving rather than .
The limiting behavior can also be seen directly at . In this limit, reduces to a polynomial of degree , and hence its Taylor coefficient vanishes. (46) then implies
Thus, as , the local derivative extraction approaches zero rather than the desired coefficient . The resulting relative error therefore approaches unity. This behavior is observed for both the neural-network and Frobenius-series solutions.
| 0.216310 | 0.219416 | 0.029893 | |||
| 0.218198 | 0.218229 | 0.030508 | |||
| 0.218216 | 0.218219 | 0.030514 | |||
| 0.218218 | 0.218218 | 0.030514 |
Table 3. The quantity returned by the local-derivative extraction methods at (). is the coefficient of in , Eq. (45). The magnitude remains finite as , indicating that the contamination arises from cancellation rather than divergence. The last column shows linearly in , while for from to .
A separate effect enters Method 7. Since is not an integer for , the behavior introduces a branch point at . Consequently, is not analytic at this point, and a polynomial representation in cannot reproduce the local behavior exactly. The amplitude of this non-analytic contribution is , so this effect is subleading compared with the coefficient contamination described by Eq. (46).
The local extraction methods therefore provide a well-defined local coefficient, but that coefficient contains an contribution from the second independent solution. Increasing the accuracy of the underlying numerical solution or the arithmetic precision does not remove this systematic contribution. The global extraction methods avoid this contamination by using information from the bulk of the domain, where the two independent solutions remain separated.
The exponent convention
Before turning to the trained network, we quantify how the choice between the exact exponent and its integer part in the field redefinition propagates into the extracted connection coefficients, and assess whether this difference produces a bias comparable in magnitude to the physical observable. The uniqueness argument below Eq. (15) fixes the exponent in the field redefinition to , rather than to its integer part . Although the two differ only at , this distinction can have a non-negligible effect on the extracted coefficients, as we demonstrate below.
With the integer exponent, the shifted parameters become and , so the solver is handed an equation displaced from Eq. (16) at . The first basis function then acquires a spurious factor,
which at equals . Projecting the true solution onto this mismatched basis introduces coefficient errors at .
Two consequences follow. First, with the exact exponent Eq. (20) is an identity, so the extraction step introduces no bias, and any residual error is attributable to the solver and to the sensitivity analysed in Section 4. Second, and more insidiously, an extraction error scaling as shares its functional form with the physical vanishing of , so the two are easily confused.
Table 4 quantifies the difference on the Frobenius series, with no network involved. With the integer exponent, the error is linear in across four orders, ; with the exact exponent, it lies between and and shows no trend in .
| integer exponent | exact exponent | |||
Table 4. Relative error of the integral-residual extraction applied to the Frobenius series at (), computed in mpmath at 50 digits. Left: the superseded integer exponent . Right: the exact exponent . The apparent “conditioning floor” of the connection problem reported previously is an artifact of the exponent mismatch (48).
A related numerical point, easily missed, is that the hypergeometric parameters themselves must be constructed in extended precision. The combination is a near-integer, so in double precision its -part carries a relative error of order . The Frobenius recursion (23) divides by exactly that quantity at , and the resulting error propagates into , producing a spurious floor that grows as — at — independent of the working precision of the series itself. All parameters are therefore built in mpmath from exact decimal input.
Does the Sobolev loss help?
The terms in Eq. (33) are a design choice, and we test them directly. The same architecture is trained at the same configuration with three settings of the Sobolev coefficients — the default , the variant , and the plain weighted loss — holding the weight , the architecture, the optimiser and the collocation scheme fixed. Because the fourth derivative is required only when and the third only when , the implementation takes four, three or two autograd passes respectively, and the variants differ in cost per step as well as in objective.
Two quantities are comparable across the three rows and one is not. The trained loss is the value of a different functional in each case and is listed for reference only. The solution error and the coefficient error are the comparison.
The effect is large, and it is localised in the term. At equal epochs the loss reaches , the loss and the plain loss . Dropping the second-derivative penalty therefore costs a factor of 2300; dropping the first-derivative penalty as well costs only a further factor of 2.1. The seed spread (the range of values for the relative error) within each variant is a factor of 2.6, 1.4 and 1.2 respectively, so both statements are far outside the noise. The term is not a refinement of the loss but its active ingredient, and an penalty on its own is worth little more than none.
To check that the results do not hinge on the particular choice , we varied each coefficient by a factor of ten in both directions with the other held fixed, at the same configuration and with three seeds per setting (Table 6 and Figure 3). The solution error is insensitive to : for from 0.01 to 1 it stays between 1.1 and , within the seed spread. It depends more on , decreasing monotonically from at to at . Across all five settings varies by less than a factor of nine, whereas removing the term altogether costs more than three orders of magnitude (Table 5). The default weights are therefore not finely tuned; they are not even the best of the settings tried, and we keep them for consistency with the rest of the paper.
| variant | loss* | seed range in | ||
| (0.1, 0.01) | ||||
| (0.1, 0) | ||||
| (0, 0) |
Table 5. Effect of the Sobolev order at , , 5000 epochs, extraction by Method 1, three seeds per row. Median over seeds, with the seed range for in brackets. The loss column is not comparable between rows.

Figure 3. Solution error (circles; bars show the range over three seeds) and coefficient error (squares) when varying at (left) and at (right). Same runs as Table 6.
A second observation concerns the residual as a diagnostic. For the variant the trained loss and the solution error are of the same order, against . For the and variants the loss is smaller than by factors of 51 and 44. The two quantities carry different normalisations and the comparison is therefore indicative rather than exact.
The comparison above holds the weight fixed at , so it attributes to the Sobolev order whatever the two might jointly produce. To separate them we cross the two factors, training the same architecture with and with plain , each under the divergent weight and under no weight at all. Table 7 gives the result.
The two factors are of entirely different magnitude. Dropping the Sobolev terms costs a factor of 2600 under the divergent weight and 1350 under no weight. Changing the weight costs a factor of 2.1 when the Sobolev terms are present and 1.07 when they are not. The endpoint weight is therefore a second-order effect, and the small benefit it does confer appears only in combination with the higher-derivative penalties — which is consistent with its role, since it emphasises precisely the region where the higher derivatives of the solution are largest.
Physical observable: the Love number kell
The connection ratio is the tidal Love number. Its real and imaginary parts encode the conservative and dissipative tidal response respectively. Table 8 compares the trained network against the exact Gamma ratios as the regulator is taken toward zero.
| \multicolumn{2}{c}{ (PINN)} | \multicolumn{2}{c}{ (analytical)} | ||||
caption: tbl:8: Love number versus at , extraction by Method 1, median over three seeds at 5000 epochs. The analytical columns are exact and convention-independent. The last column is the absolute error in the conservative part; it is flat while itself falls by two orders of magnitude.
:::
The two parts of behave in opposite ways. The dissipative part is measured. agrees with the exact value to a relative error of order at every , and reproduces the approach to its limit to four significant figures. The conservative part is only bounded. The absolute error in is on average, between and across the four rows, and essentially independent of , while itself falls linearly as for . The relative accuracy therefore degrades from at to at and to about at .
For each run we measure the relative error of the extracted decaying coefficient,
where is the exact Gamma-ratio value. If this error dominates, the error it induces in the Love number is
evaluated with the median over the three seeds and the exact . At , for example, the three seeds give , and , and the median times gives . (50) thus predicts , , and for the four rows, against the observed , , and — agreement within a factor of 1.8 for from to . The prediction is low in every row, as expected since it neglects the error in . At the floor is already about half of itself, which is why the network no longer resolves it there. The vanishing of is therefore not established by the network; what the network establishes is that it is smaller than a noise floor of about , set by the accuracy of the extracted decaying coefficient.
It is worth emphasizing that the network does not use the Gamma-ratio formula, and by Eq. (23) the basis is generated from the coefficients of the differential equation itself. The procedure therefore requires the local Frobenius structure at , but not the global connection formula — a distinction that matters because the former survives the confluent limit and the latter does not. We have verified that the two constructions of , the recursions (23) and the closed forms (21), agree to machine precision at every sample point used. We stress also that the analytic values are recovered exactly by the Gamma-ratio formula; what the PINN result tests is whether a numerically obtained solution, put through the same extraction, reproduces them.
Scan over angular numbers (, , )
The results so far are for one configuration, . Here we check whether the method works across angular numbers and spin weights. Table 9 gives the Love number for scalar (), vector () and tensor () perturbations with from 1 to 5 and , and for several tensor rows with . All runs use the same network and loss as before, trained with Adam followed by L-BFGS (§ 3). For we solve the equation at spin weight , which gives the same Love number (§ 3.2).
| deg | PINN | analytical | analytical | rel. error in PINN | rel. error in | ||||
| 0 | 1 | 1 | 1 | ||||||
| 0 | 2 | 1 | 2 | 1.0 | |||||
| 0 | 3 | 1 | 3 | 6.4 | |||||
| 0 | 4 | 1 | 4 | 30 | |||||
| 1 | 1 | 1 | 0 | ||||||
| 1 | 2 | 1 | 1 | ||||||
| 1 | 3 | 1 | 2 | 0.59 | |||||
| 1 | 4 | 1 | 3 | 4.8 | |||||
| 1 | 5 | 1 | 4 | 25 | |||||
| 2 | 2 | 2 | 0 | ||||||
| 2 | 2 | 1 | 0 | ||||||
| 2 | 3 | 3 | 1 | ||||||
| 2 | 3 | 2 | 1 | ||||||
| 2 | 3 | 1 | 1 | ||||||
| 2 | 4 | 4 | 2 | ||||||
| 2 | 4 | 3 | 2 | ||||||
| 2 | 4 | 2 | 2 | ||||||
| 2 | 4 | 1 | 2 | 0.15 | |||||
| 2 | 5 | 1 | 3 | 1.5 |
Table 9. Love numbers at , , from Method 1. Rows with are computed at spin weight , which gives the same . “deg” is the polynomial degree of at for the equation that is solved. The last column is the error obtained with the network output set to zero, , i.e. from the hard boundary condition alone; rows where it is already small (deg ) are consistency checks. Single seed, 750 Adam epochs (1000 for ) followed by L-BFGS.
The last two columns of Table 9 compare two errors in : that of the trained network, and the one we would get if the network output in Eq. (36) were set to zero, so that comes only from the boundary condition at the horizon. The comparison is needed because, at integer , is close to a polynomial of degree (column “deg”). When this degree is 0 or 1, the boundary condition already fixes almost the whole solution and the network only adds a correction of order ; these rows check that the pipeline is consistent but do not test the solver. When the degree is 2 or more, the boundary condition alone gives errors between and 30, and the network has to learn the solution.
In the rows of degree 2 or more the network learns the solution accurately. The solution error is below and the error in is between and . Across the whole table, , which at is as small as , agrees with the exact value to better than 0.4%. An earlier version of this scan, trained with Adam alone at the physical spin weight, did not converge for at and for at ; with the L-BFGS stage and the change of spin weight these rows now converge. The equations at and give the same Love number, but they are not equally easy to solve. At the solution is a polynomial of lower degree and much smaller in size: for , for example, drops from at to 12 at . In addition, at the unphysical horizon solution diverges at , so the network cannot absorb it (§ 3.2).
Figure 4 shows the effect on training for , solved both ways. At , Adam lowers the validation loss only to about in 24 minutes. L-BFGS then brings it down to about , where it levels off; the solution error is and the error in is several percent. At , Adam alone reaches in about four minutes, and the final solution error is . The plateau at is caused by the unphysical horizon solution: subtracting the multiple of it that best fits the error of the network lowers to and the error in to .
Figure 4. Validation loss versus wall-clock time for , with the equation solved at and at , which give the same . The dotted lines mark the switch from Adam to L-BFGS (torch.optim.LBFGS), at about 24 min for (blue) and about 4 min for (red). The validation loss is the weighted residual on a fixed grid of 4000 points in that is not used for training; times are for one CPU. Final solution errors: () and ().
Discussion
We have developed and tested a PINNs framework for computing static black-hole tidal Love numbers. The method combines a decomposition of the complex hypergeometric equation into a coupled real system, an architecture that enforces the horizon boundary condition and first Frobenius recursion exactly, and a weighted -Sobolev loss. The response coefficients and are extracted from the trained solution using several independent methods, which are also applied to the exact Frobenius series. This provides a direct way to distinguish errors in the learned solution from limitations of the extraction procedure.
The two response coefficients present very different numerical challenges. The dominant coefficient, , is determined directly by the boundary value and is recovered accurately for all extraction methods considered. In contrast, multiplies the subdominant solution and is therefore much more sensitive to errors in the numerical solution. The comparison of extraction methods shows that global extraction methods provide the most reliable recovery of , while extraction based on points very close to the singular boundary becomes increasingly sensitive as increases. The local derivative methods considered here do not recover correctly. On the exact Frobenius solution they return the wrong quantity because the corresponding Taylor coefficient vanishes for the polynomial representation. This demonstrates that accurate extraction requires the extraction procedure itself to be tested independently of the neural-network solver.
The loss-function study shows that the higher-derivative terms in the Sobolev loss are essential for obtaining an accurate solution. In particular, removing the contribution produces a dramatic increase in the solution error, while changing the endpoint weight has a comparatively modest effect (Table 5 and Table 7). Thus, for the present problem, the principal improvement comes from controlling higher derivatives through the objective rather than from the endpoint weighting alone. This result also supersedes the corresponding conclusion of an earlier version of this work, which used a different exponent convention.
An important consequence is that the differential-equation residual should not be used by itself as a measure of solution accuracy. The loss-function comparisons show that a smaller residual does not necessarily correspond to a more accurate learned solution. The object being minimized therefore matters when interpreting PINN residuals, and solution accuracy should be assessed independently whenever a reference solution is available.
The angular-number scan extends these results to for scalar, vector and tensor perturbations. Two changes were needed: an L-BFGS stage after Adam, and, for , solving the equation at spin weight , which gives the same Love number and removes the singular horizon branch that the network otherwise absorbs at . With these, is recovered to a relative accuracy of or better in all rows.
The static problem provides a useful controlled setting for establishing these properties, but the framework is intended to extend to problems without closed-form solutions. Another important application is the frequency-dependent regime, in which the master equation takes the form of a confluent Heun equation. The architecture and Sobolev-loss construction can be retained, with the full -dependent operator replacing the present residual. In this setting, the ability to impose the near-horizon behavior directly through the network architecture may be particularly useful because an exact reference solution for validating the trained network is generally unavailable.
A broader application is to modified-gravity and numerically constructed black-hole backgrounds, where the perturbation equations are not separable. In such cases, the present approach provides a possible route toward computing response functions directly from the differential equations while retaining explicit control over boundary behavior and derivative regularity.
An important special case arises for and , where the parameter takes a non-positive integer value, and the construction of § 2 does not apply at the physical spin weight. At spin weight one has , so Schwarzschild and axisymmetric Kerr perturbations may be accessible with the same method; we have not tested this. The identity used here holds for the static problem, and the relation between the two spin weights has to be revisited for .
Overall, our results show that PINNs can achieve high solution accuracy for static black-hole tidal problems within a well-defined convergent regime. Achieving reliable response coefficients, however, requires the loss function, boundary conditions, neural-network solution, and extraction procedure to be considered together. The results provide a controlled demonstration of these requirements and establish a foundation for applying PINNs to frequency-dependent and non-separable black-hole perturbation problems where conventional analytic and numerical benchmarks are no longer available.
Code availability
The code used for all numerical results in this paper is publicly available at https://github.com/RodriguezMariaJ/pinn-love-numbers.
Acknowledgements
We thank Luca Santoni, Adam R. Solomon and Malcolm Perry for illuminating discussions on Love numbers and black hole perturbation theory. We thank Alessandro Tanzini, Giulio Bonelli, and Cristoforo Iossa for discussions on the connection problem in Heun-type equations. L.F.T. acknowledges support from the USU Howard L. Blood Fellowship and the USU Physics Department Presidential Doctoral Research Fellowship. M.J.R. acknowledges support from the U.S. National Science Foundation through grant PHY-2012036 during the initial stages of this work. M.J.R. also thanks the Mitchell Family Foundation for hospitality at the Cook’s Branch Workshop, USU Quantum and Gravitational Initiative and the Instituto de Física Teórica UAM/CSIC for hospitality during parts of this work. We acknowledge the use of the Anthropic Claude language model as an assistant for code and LaTeX manuscript review. All physics content, numerical results, and scientific claims are the responsibility of the authors.
From the Teukolsky equation to the hypergeometric form
This appendix presents the derivation of the equations used in § 2, and fixes the relation between the hypergeometric variable and the radial Teukolsky function.
Separated equations
We start from the linear perturbation equation derived by Teukolsky [60, 61, 62, 63]. With the ansatz
and with , separability gives
In the slowly rotating limit ,
Setting leaves
with .
Change of variable
Equation (55) has regular singular points at , and . The map
sends , and , placing the horizon at and spatial infinity at . Writing
Eq. (55) becomes the hypergeometric equation (5) with the parameters (6), where .
Exponents
At each singular point the two Frobenius exponents of the radial equation (55) and of the hypergeometric equation are related by the prefactor in Eq. (57):
| — | |||
Table 8.
The exponent differences agree in each column. At the two radial behaviours are , the external tidal field, and , the induced response. Since , the prefactor maps
(decaying: response),
(growing: source).
At the horizon the two exponents correspond to the ingoing and outgoing solutions; the prefactor in Eq. (57) selects the ingoing branch as the Frobenius root of at , which is the branch imposed by the network ansatz (36).
Explicit derivatives of the residuals
Applying the product rule to Eqs. (31)–(32) gives
Here and ; both and are linear in , so their second derivatives vanish.
References
- [1]B. Kol and M. Smolkin, “Black hole stereotyping: Induced gravito-static polarization,” JHEP 02 (2012) 010, arXiv:1110.3764 [hep-th].DOI
- [2]N. Gürlebeck, “Tidally Distorted Black Holes,” Springer Proc. Phys. 170 (2016) 87–93.
- [3]L. Hui, A. Joyce, R. Penco, L. Santoni, and A. R. Solomon, “Static response and Love numbers of Schwarzschild black holes,” JCAP 04 (2021) 052, arXiv:2010.00593 [hep-th].DOI
- [4]M. M. Riva, L. Santoni, N. Savić, and F. Vernizzi, “Vanishing of nonlinear tidal Love numbers of Schwarzschild black holes,” Phys. Lett. B 854 (2024) 138710, arXiv:2312.05065 [gr-qc].arxiv.org/abs/2312.05065
- [5]T. Hadad, B. Kol, and M. Smolkin, “Gravito-magnetic polarization of Schwarzschild black hole,” JHEP 06 (2024) 169, arXiv:2402.16172 [hep-th].arxiv.org/abs/2402.16172
- [6]A. Le Tiec, M. Casals, and E. Franzin, “Tidal Love Numbers of Kerr Black Holes,” Phys. Rev. D 103 no. 8, (2021) 084021, arXiv:2010.15795 [gr-qc].
- [7]H. S. Chia, “Tidal deformation and dissipation of rotating black holes,” Phys. Rev. D 104 no. 2, (2021) 024013, arXiv:2010.07300 [gr-qc].DOI
- [8]P. Charalambous, S. Dubovsky, and M. M. Ivanov, “On the Vanishing of Love Numbers for Kerr Black Holes,” JHEP 05 (2021) 038, arXiv:2102.08917 [hep-th].DOI
- [9]M. Perry and M. J. Rodriguez, “Dynamical Love Numbers for Kerr Black Holes,” arXiv:2310.03660 [gr-qc].arxiv.org/abs/2310.03660
- [10]L. Hui, A. Joyce, R. Penco, L. Santoni, and A. R. Solomon, “Ladder symmetries of black holes. Implications for love numbers and no-hair theorems,” JCAP 01 no. 01, (2022) 032, arXiv:2105.01069 [hep-th].DOI
- [11]P. Charalambous, S. Dubovsky, and M. M. Ivanov, “Hidden Symmetry of Vanishing Love Numbers,” Phys. Rev. Lett. 127 no. 10, (2021) 101101, arXiv:2103.01234 [hep-th].DOI
- [12]P. Charalambous, S. Dubovsky, and M. M. Ivanov, “Love symmetry,” JHEP 10 (2022) 175, arXiv:2209.02091 [hep-th].DOI
- [13]L. Hui, A. Joyce, R. Penco, L. Santoni, and A. R. Solomon, “Near-zone symmetries of Kerr black holes,” JHEP 09 (2022) 049, arXiv:2203.08832 [hep-th].DOI
- [14]J. Ben Achour, E. R. Levine, S. Mukohyama, and J.-P. Uzan, “Hidden symmetry of the static response of black holes: applications to Love numbers,” JHEP 07 (2022) 112, arXiv:2202.12828 [gr-qc].
- [15]T. Katagiri, M. Kimura, H. Nakano, and K. Omukai, “Vanishing Love numbers of black holes in general relativity: From spacetime conformal symmetry of a two-dimensional reduced geometry,” Phys. Rev. D 107 no. 12, (2023) 124030, arXiv:2209.10469 [gr-qc].DOI
- [16]A. Kehagias, D. Perrone, and A. Riotto, “Quasinormal modes and Love numbers of Kerr black holes from AdS₂ black holes,” JCAP 01 (2023) 035, arXiv:2211.02384 [hep-th].DOI
- [17]A. Guevara and U. Kol, “Self Dual Black Holes as the Hydrogen Atom,” arXiv:2311.07933 [hep-th].arxiv.org/abs/2311.07933
- [18]M. Cvetič, N. H. Rodríguez, M. J. Rodriguez, and O. Varela, “Kerr effective black hole geometries in supergravity,” Phys. Rev. D 112 no. 2, (2025) 026007, arXiv:2406.10458 [hep-th].arxiv.org/abs/2406.10458
- [19]F. Gray, C. Keeler, D. Kubiznak, and V. Martin, “Love symmetry in higher-dimensional rotating black hole spacetimes,” JHEP 03 (2025) 036, arXiv:2409.05964 [gr-qc].arxiv.org/abs/2409.05964
- [20]P. Charalambous and M. M. Ivanov, “Scalar Love numbers and Love symmetries of 5-dimensional Myers-Perry black holes,” JHEP 07 (2023) 222, arXiv:2303.16036 [hep-th].DOI
- [21]M. J. Rodriguez, L. Santoni, A. R. Solomon, and L. F. Temoche, “Love numbers for rotating black holes in higher dimensions,” Phys. Rev. D 108 no. 8, (2023) 084011, arXiv:2304.03743 [hep-th].DOI
- [22]P. Charalambous, “Love numbers and Love symmetries for p-form and gravitational perturbations of higher-dimensional spherically symmetric black holes,” JHEP 04 (2024) 122, arXiv:2402.07574 [hep-th].arxiv.org/abs/2402.07574
- [23]D. Glazer, A. Joyce, M. J. Rodriguez, L. Santoni, A. R. Solomon, and L. F. Temoche, “Higher-dimensional spinning black holes and effective field theory,” JHEP 03 (2026) 036, arXiv:2412.21090 [hep-th].arxiv.org/abs/2412.21090
- [24]P. Charalambous, S. Dubovsky, and M. M. Ivanov, “Love numbers of black p-branes: fine tuning, Love symmetries, and their geometrization,” JHEP 06 (2025) 180, arXiv:2502.02694 [hep-th].arxiv.org/abs/2502.02694
- [25]M. Cvetič, M. A. Liao, and M. M. Stetsko, “Tidal perturbations and Love symmetry for five-dimensional charged rotating black holes,” Phys. Rev. D 113 no. 8, (2026) 085008, arXiv:2601.20514 [hep-th].arxiv.org/abs/2601.20514
- [26]S. Chakraborty and P. Pani, “Tidal Response of Compact Objects,” arXiv:2604.08679 [gr-qc].arxiv.org/abs/2604.08679
- [27]M. J. Rodríguez, L. Santoni, and A. R. Solomon, “Love numbers of black holes and compact objects,” arXiv:2604.08653 [gr-qc].arxiv.org/abs/2604.08653
- [28]M. M. Ivanov and Z. Zhou, “Revisiting the matching of black hole tidal responses: A systematic study of relativistic and logarithmic corrections,” Phys. Rev. D 107 no. 8, (2023) 084030, arXiv:2208.08459 [hep-th].arxiv.org/abs/2208.08459
- [29]M. M. Ivanov and Z. Zhou, “Vanishing of Black Hole Tidal Love Numbers from Scattering Amplitudes,” Phys. Rev. Lett. 130 no. 9, (2023) 091403, arXiv:2209.14324 [hep-th].DOI
- [30]M. V. S. Saketh, Z. Zhou, and M. M. Ivanov, “Dynamical tidal response of Kerr black holes from scattering amplitudes,” Phys. Rev. D 109 no. 6, (2024) 064058, arXiv:2307.10391 [hep-th].arxiv.org/abs/2307.10391
- [31]C. G. A. Barura, H. Kobayashi, S. Mukohyama, N. Oshita, K. Takahashi, and V. Yingcharoenrat, “Tidal Love numbers from EFT of black hole perturbations with timelike scalar profile,” JCAP 09 (2024) 001, arXiv:2405.10813 [gr-qc].DOI
- [32]S. Caron-Huot, M. Correia, G. Isabella, and M. Solon, “Gravitational Wave Scattering via the Born Series: Scalar Tidal Matching to O(G7) and Beyond,” Phys. Rev. Lett. 135 no. 19, (2025) 191601, arXiv:2503.13593 [hep-th].
- [33]O. Combaluzier-Szteinsznaider, D. Glazer, A. Joyce, M. J. Rodriguez, and L. Santoni, “Dynamical tidal response of Schwarzschild Black Holes,” JHEP 06 (2026) 032, arXiv:2511.02372 [gr-qc].
- [34]S. Barbosa, P. Brax, S. Fichet, and L. de Souza, “Running Love numbers and the Effective Field Theory of gravity,” JCAP 07 (2025) 071, arXiv:2501.18684 [hep-th].DOI
- [35]M. P. Solon, “Universal Closed Form for Dynamical Love Numbers of Black Holes,” arXiv:2606.19281 [hep-th].DOI
- [36]S. Barbosa, S. Fichet, and L. de Souza, “Running Love numbers of charged black holes,” JCAP 09 (2026) 009, arXiv:2602.00349 [hep-th].DOI
- [37]A. Elkhidir, G. Martin, J. Parra-Martinez, and M. V. S. Saketh, “Quantum Love Numbers are Non-Zero,” arXiv:2609.25875 [hep-th].arxiv.org/abs/2609.25875
- [38]M. Perry and M. J. Rodriguez, “Love numbers for extremal Kerr black holes,” Phys. Rev. D 112 no. 12, (2025) 126004, arXiv:2412.19699 [hep-th].arxiv.org/abs/2412.19699
- [39]M. J. Rodriguez and L. F. Temoche, “Clines and the analytic structure of black hole perturbations,” JHEP 03 (2026) 150, arXiv:2511.08695 [gr-qc].arxiv.org/abs/2511.08695
- [40]G. Bonelli, C. Iossa, D. P. Lichtig, and A. Tanzini, “Exact solution of Kerr black hole perturbations via CFT2 and instanton counting: Greybody factor, quasinormal modes, and Love numbers,” Phys. Rev. D 105 no. 4, (2022) 044047, arXiv:2105.04483 [hep-th].DOI
- [41]S. N. Solodukhin and V. Tagiev, “Tidal Love numbers of wormholes as black-hole mimickers,” arXiv:2609.13129 [gr-qc].arxiv.org/abs/2609.13129
- [42]M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational physics 378 (2019) 686–707.DOI
- [43]R. Luna, J. Calderón Bustillo, J. J. S. Martínez, A. Torres-Forné, and J. A. Font, “Solving the Teukolsky equation with physics-informed neural networks,” Phys. Rev. D 107 no. 6, (2023) 064025, arXiv:2212.06103 [gr-qc].DOI
- [44]A. S. Cornell, A. Ncube, and G. Harmsen, “Using physics-informed neural networks to compute quasinormal modes,” Phys. Rev. D 106 no. 12, (2022) 124047, arXiv:2205.08284 [physics.comp-ph].DOI
- [45]N. Patel, A. Aykutalp, and P. Laguna, “Novel approach to solving Schwarzschild black hole perturbation equations via physics informed neural networks,” Gen. Rel. Grav. 56 no. 11, (2024) 137.DOI
- [46]A. S. Cornell, S. R. Herbst, A. M. Ncube, and H. Noshad, “Solving the Regge-Wheeler and Teukolsky equations: supervised versus unsupervised physics-informed neural networks,” arXiv:2402.11343 [gr-qc].arxiv.org/abs/2402.11343
- [47]A. Övgün, İ. Sakallı, and H. Mutuk, “Quasinormal modes of dS and AdS black holes: Feedforward neural network method,” Int. J. Geom. Meth. Mod. Phys. 18 no. 10, (2021) 2150154, arXiv:1904.09509 [gr-qc].DOI
- [48]A. M. Ncube, G. E. Harmsen, and A. S. Cornell, “Investigating a New Approach to Quasinormal Modes: Physics-Informed Neural Networks,” in 65th Annual Conference of the South African Institute of Physics. 8, 2021. arXiv:2108.05867 [gr-qc].arxiv.org/abs/2108.05867
- [49]N. J. L. S. Lobos, A. M. Ncube, R. C. Pantig, and A. S. Cornell, “Analyzing the effect of higher dimensions on the black hole silhouette, deflection angles, and PINN approximated quasinormal modes,” Annals Phys. 473 (2025) 169906, arXiv:2406.08078 [gr-qc].DOI
- [50]M. A. Haddou, “Quasi-normal modes of near-extremal black holes in dRGT massive gravity using Physics-Informed Neural Networks (PINNs),” arXiv:2303.02395 [gr-qc].arxiv.org/abs/2303.02395
- [51]R. Luna, D. D. Doneva, J. A. Font, J.-H. Lien, and S. S. Yazadjiev, “Quasinormal modes in modified gravity using physics-informed neural networks,” Phys. Rev. D 109 no. 12, (2024) 124064, arXiv:2404.11583 [gr-qc].DOI
- [52]V. Cardoso, E. Franzin, A. Maselli, P. Pani, and G. Raposo, “Testing strong-field gravity with tidal Love numbers,” Phys. Rev. D 95 no. 8, (2017) 084014, arXiv:1701.01116 [gr-qc]. [Addendum: Phys.Rev.D 95, 088901 (2017)].DOI
- [53]V. De Luca, J. Khoury, and S. S. C. Wong, “Implications of the weak gravity conjecture for tidal Love numbers of black holes,” Phys. Rev. D 108 no. 4, (2023) 044066, arXiv:2211.14325 [hep-th].DOI
- [54]A. Ferrer-Sánchez, D. D. Doneva, J. D. Martín-Guerrero, S. S. Yazadjiev, R. Ruiz de Austri-Bazan, Y. Vives-Gilabert, and J. A. Font, “Kerr Quasinormal Modes without Variable Separation: A Two-Dimensional Hyperboloidal Teukolsky Solver with Physics-Informed Neural Networks,” arXiv:2608.19774 [gr-qc].
- [55]A. K.-W. Chung, P. Wagle, and N. Yunes, “Spectral method for metric perturbations of black holes: Kerr background case in general relativity,” Phys. Rev. D 109 (2024) 044072, arXiv:2312.08435 [gr-qc].DOI
- [56]J. L. Blázquez-Salcedo, F. S. Khoo, J. Kunz, and L. M. González-Romero, “Quasinormal modes of Kerr black holes using a spectral decomposition of the metric perturbations,” Phys. Rev. D 109 (2024) 064028, arXiv:2312.10754 [gr-qc].
- [57]A. Ronveaux and F. M. Arscott, Heun’s differential equations. Clarendon Press, 1995.DOI
- [58]R. Cayuso, M. Herrero-Valea, and E. Barausse, “Deep learning solutions to singular ordinary differential equations: From special functions to spherical accretion,” Phys. Rev. D 111 no. 6, (2025) 064082, arXiv:2409.20150 [gr-qc].DOI
- [59]G. Boni, “Differential equations for feynman integrals and physics informed neural network,”.
- [60]S. A. Teukolsky, “Rotating black holes - separable wave equations for gravitational and electromagnetic perturbations,” Phys. Rev. Lett. 29 (1972) 1114–1118.DOI
- [61]S. A. Teukolsky, “Perturbations of a rotating black hole. 1. Fundamental equations for gravitational electromagnetic and neutrino field perturbations,” Astrophys. J. 185 (1973) 635–647.
- [62]W. H. Press and S. A. Teukolsky, “Perturbations of a Rotating Black Hole. II. Dynamical Stability of the Kerr Metric,” Astrophys. J. 185 (1973) 649–674.DOI
- [63]S. A. Teukolsky and W. H. Press, “Perturbations of a rotating black hole. III - Interaction of the hole with gravitational and electromagnetic radiation,” Astrophys. J. 193 (1974) 443–461.DOI
- [64]M. Abramowitz and I. A. Stegun, Handbook of mathematical functions with formulas, graphs, and mathematical tables, vol. 55. US Government printing office, 1948.
- [65]H. Bateman et al., “Higher transcendental functions [volumes i-iii],”.
- [66]S. Alkhadh r and M. Almekkawy, “Wave equation modeling via physics-informed neural networks: Models of soft and hard constraints for initial and boundary conditions,” Sensors 23 no. 5, (2023) 2792.
- [67]M. Sheikholeslami, S. Salehi, W. Mao, A. Esламdoost, and H. Nilsson, “Physics-informed neural networks with hard and soft boundary conditions for linear free surface waves,” Physics of Fluids 37 no. 8, (2025) 087158.
- [68]I. E. Lagaris, A. Likas, and D. I. Fotiadis, “Artificial Neural Networks for Solving Ordinary and Partial Differential Equations,” arXiv:physics/9706023.
- [69]F. Chen, D. Sondak, P. Protopapas, M. Mattheakis, S. Liu, D. Agarwal, and M. Di Giovanni, “Neurodiff eq: A python package for solving differential equations with neural networks,” Journal of Open Source Software 5 no. 46, (2020) 1931.DOI
- [70]S. Liu, P. Protopapas, D. Sondak, and F. Chen, “Recent advances of NeuroDiffEq – an open-source library for physics-informed neural networks,” 2025.arxiv.org/abs/2502.12177
- [71]M. Mattheakis, P. Protopapas, D. Sondak, M. Di Giovanni, and E. Kaxiras, “Physical symmetries embedded in neural networks,” 2019.arxiv.org/abs/1904.08991
- [72]W. M. Czarnecki, S. Osindero, M. Jaderberg, G. Swirszcz, and R. Pascanu, “Soboley training for neural networks,” Advances in neural information processing systems 30 (2017) 4281–4290.
- [73]H. Son, J. W. Jang, W. J. Han, and H. J. Hwang, “Sobolev training for physics-informed neural networks,” Communications in Mathematical Sciences 21 no. 6, (2023) 1679–1705.arxiv.org/abs/2101.08932
- [74]E. Kreyszig, Introductory functional analysis with applications. John Wiley & Sons, 1991.
- [75]R. A. Adams and J. J. Fournier, Sobolev spaces, vol. 140. Elsevier, 2003.
- [76]H. Brezis and H. Brézis, Functional analysis, Sobolev spaces and partial differential equations, vol. 2. Springer, 2011.DOI
- [77]L. C. Evans, Partial differential equations, vol. 19. American Mathematical Society, 2022.
- [78]D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” 2015. https://arxiv.org/abs/1412.6980.
- [79]C. Flamant, P. Protopapas, and D. Sondak, “Solving differential equations using neural network solution bundles,” 2020. https://arxiv.org/abs/2006.14372.
- [80]S. Desai, M. Mattheakis, H. Joy, P. Protopapas, and S. Roberts, “One-shot transfer learning of physics-informed neural networks,” 2022. https://arxiv.org/abs/2110.11286.
