Many calculations in physics and machine learning save effort the same way. They hold one part of a coupled system fixed and let only the other part move. In holographic models of superfluids, which describe a strongly interacting fluid through a dual problem of fields living on a black-hole spacetime, the usual shortcut is the probe limit. The spacetime geometry is frozen, and the fluid's order parameter evolves on top of it without bending it. The most detailed recent studies of how multiply wound vortices break apart were done entirely in this limit. Those studies find that a doubly wound vortex splits fastest at a "turning temperature" $T/T_c \approx 0.745$, and that a four-times-wound vortex changes its splitting symmetry from 2-fold to 3-fold at $T/T_c \approx 0.35$ and from 3-fold to 4-fold at $T/T_c \approx 0.55$ (Lan et al., N=2; Lan et al., N=4). The authors themselves say the approximation fails at low temperature, and that is exactly where the 0.35 transition sits. No one knows whether these numbers survive once the geometry is allowed to respond.
We propose to answer this without the full cost of solving the coupled Einstein–Maxwell–scalar equations in 3+1 dimensions. The method is first-order backreaction. We compute the leading correction to the geometry sourced by the probe-limit vortex. We then compute how that correction shifts the complex frequency of each splitting mode, using a biorthogonal (left/right eigenvector) perturbation formula suited to dissipative problems. This gives a predicted shift for every transition temperature at the cost of a few linear solves. We check the predictions against a small number of full nonlinear backreacted runs. As a secondary and deliberately falsifiable test, we carry the same diagnostic into vision-language model adaptation. Text-only prompt tuning with a frozen image encoder has the same structure as the probe limit, and we ask whether the first-order response of the frozen side predicts when unfreezing it pays off.
Core Claim
The probe-limit splitting thresholds are not equally robust. Backreaction enters with the local energy density of the condensate, which grows as temperature falls. So the first-order correction should move the $p=2\to3$ transition near 0.35 substantially more than the $p=3\to4$ transition near 0.55 or the $N=2$ turning point near 0.745. It should also preferentially destabilize the near-degenerate transient $p=5$ and $p=6$ channels. These claims would be new, and each is checkable against a direct simulation.
The claim is narrow on purpose. Annala and Möttönen prove that multiply charged defects are non-generic critical points that split under any generic perturbation. That result constrains us: backreaction cannot stabilize the $N=4$ vortex. It can only reorder which splitting channel wins and shift where the reordering happens.
Research Program
The setting is the holographic superfluid of Lan et al.: a charged scalar and a Maxwell field on planar Schwarzschild–AdS. It is solved in ingoing Eddington–Finkelstein coordinates with Chebyshev pseudo-spectral grids.
We introduce a backreaction parameter $\epsilon$ that measures the matter's gravitational pull. It plays the role of the couplings $\alpha$ and $q$ in Chang et al.. The work proceeds in three steps.
- Metric correction. We solve the linearized Einstein equations for the metric correction $\delta g$. This correction is sourced by the stress tensor of the stationary probe-limit $N$-vortex, and it is axisymmetric by construction.
-
Frequency shift. Each splitting mode is a quasinormal mode, meaning a damped or growing oscillation of the black-hole-plus-vortex system. Its frequency $\omega_p$ with angular index $p$ satisfies $L\,\psi_p = \omega_p\,\psi_p$ on the frozen background. At first order its shift is $$\delta\omega_p = \epsilon\,\frac{\langle \tilde\psi_p,\ \delta L\,\psi_p\rangle}{\langle \tilde\psi_p,\ \psi_p\rangle}.$$ Here $\delta L$ is the change in the linearized operator. That change comes from both $\delta g$ and the correction to the vortex profile it induces. $\tilde\psi_p$ is the corresponding left (adjoint) mode. The left mode is needed because the horizon makes the problem non-Hermitian, so the textbook expectation-value formula is wrong here.
-
Transition shifts. A transition between channels $p$ and $p'$ sits at the temperature $T_\ast$ where $\mathrm{Im}\,\omega_p = \mathrm{Im}\,\omega_{p'}$. To first order it moves by $$\delta T_\ast = -\,\frac{\mathrm{Im}(\delta\omega_p-\delta\omega_{p'})}{\partial_T\,\mathrm{Im}(\omega_p-\omega_{p'})}.$$ This follows from Taylor-expanding the crossing condition. The same logic, applied to the maximum of $\mathrm{Im}\,\omega_{p=2}$, gives the shift in the $N=2$ turning temperature.
All temperatures are reported in units of the $T_c$ of the backreacted theory. Chang et al. find that $T_c$ follows scalings in the couplings that are insensitive to backreaction, but we do not rely on this.
The machine-learning arm reuses only the logic, not the equations. CoOp-style text prompt tuning on frozen CLIP is the "probe solution." The "stress tensor" is the gradient of the task loss with respect to vision-side prompt parameters, evaluated at that solution. The first-order prediction of the benefit of unfreezing is the loss decrease of a single Gauss–Newton step on the vision prompts, taken with the text prompts held at their tuned values. The question is whether this per-dataset quantity ranks where coupled methods help.
Baselines
Physics, naive reference. The probe-limit values themselves: turning temperature $T/T_c \approx 0.745$ (Lan et al., N=2) and transitions near 0.35 and 0.55 (Lan et al., N=4). We reproduce them with our own code before any correction is added, so the do-nothing answer is measured on identical grids.
Physics, strongest competitor and ground truth. Full nonlinear evolution of the coupled Einstein–matter system at a small grid of $(\epsilon, T)$ points around each transition. The splitting channel is read off from the vortex-core symmetry, and the growth rate from the early exponential phase. Lan et al. found $\tau \propto 1/\mathrm{Im}\,\omega$, which justifies reading the rate this way. The comparison is fair because both computations share discretization, boundary conditions and the definition of the backreacted $T_c$. The perturbative method beats this baseline if two conditions hold together. First, its predicted $\delta T_\ast$ agrees with the nonlinear value within the nonlinear run's own temperature resolution, for $\epsilon$ up to the value where the $0.35$ shift exceeds that resolution. Second, it costs at least an order of magnitude less compute.
Physics, ablations. Two ablations isolate the key ingredients.
- Hermitian formula. We replace the adjoint mode $\tilde\psi_p$ with $\psi_p^\ast$ in the shift formula. This tests whether the biorthogonal structure matters.
- Homogeneous backreaction. We place the vortex on the fully backreacted but spatially uniform hairy black brane of Chang et al.'s type, dropping the vortex's localized stress tensor. This tests whether the vortex core's own gravitational pull matters, or only the renormalized background.
Machine learning. The naive reference is CoOp, which tunes text prompts only. The strongest coupled methods are MaPLe, which derives vision prompts from language prompts through a linear map, and BMIP, which uses bidirectional projections with learned aggregation. BMIP reports an open-world harmonic mean of 79.04% against MaPLe's 78.22%, and EuroSAT accuracy of 86.10% against MaPLe's 81.43%. DPLQ reports a remote-sensing average harmonic mean of 83.5% against MaPLe's 80.42%. The file gives no per-dataset CoOp numbers apart from the FGVC-Aircraft figure quoted by SpecPL, so we rerun CoOp, MaPLe and BMIP ourselves on the same seeds, shots, backbone and tuning budget.
The target quantity is the per-dataset gain of MaPLe or BMIP over CoOp. The diagnostic competes against two cheap predictors: the zero-shot CLIP accuracy, and the raw gradient norm on vision prompts. It "wins" only if its Spearman rank correlation with the realized gain exceeds both predictors across the pooled datasets, with confidence intervals from resampling seeds.
Why not just…?
Why not just run full backreacted simulations everywhere? That is the ground truth, and we do run it, but only at a handful of points. Chang et al. note that multi-dimensional backreacted configurations require coupled nonlinear PDEs that remain computationally challenging. Mapping transition curves in the full $(\epsilon, T, N)$ space this way is what the perturbative method is meant to avoid. If first-order predictions already fail at small $\epsilon$, that is itself a finding: it would show that splitting channels are nonperturbatively sensitive to geometry, and we would report it.
Why not just use a backreacted homogeneous black hole as a new fixed background? This is a real threat, and it is the second physics ablation. If this cheap replacement reproduces the nonlinear shifts, the local vortex stress tensor is irrelevant and our method is unnecessary. We expect it to underpredict the shift at $T/T_c \approx 0.35$, because the core of a four-times-wound vortex concentrates energy in a way a uniform condensate does not. The experiment settles this directly.
Why not just model splitting with a damped Gross–Pitaevskii equation? The Gross–Pitaevskii equation is the standard mean-field wave equation for weakly interacting condensates. As Lan et al. and Yao et al. note, it describes weakly coupled condensates near zero temperature. Finite-temperature dissipation then enters through a phenomenological damping constant. The thresholds we test are temperature-dependent features of a strongly coupled theory, and there dissipation comes from the horizon. A damped Gross–Pitaevskii model has no analogue of gravitational backreaction to test.
Why not just always tune both modalities (use BMIP) and skip any diagnostic? If coupled tuning never loses to text-only tuning at equal budget, then a predictor of when it helps has little practical value. The machine-learning arm would then reduce to an interpretability result. We report the fraction of datasets where CoOp matches or beats the coupled methods. If that fraction is near zero, we demote this arm to an analysis section rather than claim a method.
Why not just predict the gain from zero-shot accuracy? Low zero-shot accuracy plausibly signals a frozen image encoder that is badly suited to the domain, as with EuroSAT or the medical sets in DPLQ. It may predict the gain as well as our diagnostic. That is why it is a required baseline. We drop the transfer claim if the first-order step does not beat it.
Decisive Checks
The physics claim stands or falls on four observations at $\epsilon$ values where the nonlinear runs resolve a shift:
- The first-order prediction of how far the $p=2\to3$ transition moves matches the nonlinear result.
- That transition moves more than the $p=3\to4$ transition and the $N=2$ turning point.
- The adjoint-weighted formula outperforms the Hermitian one.
- The local-source calculation outperforms the homogeneous-background ablation.
A secondary check tests whether the transient $p=5$ and $p=6$ modes lose or gain growth-rate margin under backreaction. They are the channels closest to degeneracy, so first-order theory predicts they are the most sensitive.
The machine-learning check covers the 11 base-to-novel datasets, the 8 remote-sensing and 3 medical datasets of DPLQ, and the ImageNet shift sets. It asks one question: whether the first-order response of the frozen image encoder ranks per-dataset unfreezing gains better than zero-shot accuracy does. If the answer is no, we say so and keep the physics result as the contribution.
References
- Splitting of doubly quantized vortices in holographic superfluid of finite temperature
- Heating Up Quadruply Quantized Vortices: Splitting Patterns and Dynamical Transitions
- Charge ambiguity and splitting of monopoles
- Holographic superconductors: An analytic method revisit
- MaPLe: Multi-modal Prompt Learning
- BMIP: Bi-directional Modality Interaction Prompt Learning for VLM
- Domain Prompt Learning with Quaternion Networks (Extended Abstract)
- SpecPL: Disentangling Spectral Granularity for Prompt Learning
- Baby skyrmion in two-component holographic superfluids