Rotating shallow water, the textbook model of oceans and atmospheres, has a topological property. Its wave bands carry an integer, the first Chern number, equal to $-2$ for the positive-frequency band. This integer predicts that waves travel one way only, without reflecting, along coasts and along the equator, where the Coriolis parameter $f$ changes sign. The prediction is derived for an idealized fluid: frictionless (inviscid), at rest, and rotationally symmetric. Real flows break every one of those assumptions. Viscous walls shed intense vortices, and a vortex array can reduce continuous rotational symmetry to a discrete $n$-fold one. This proposal asks, and settles numerically, which part of the topological prediction survives each of these breakdowns.
The organizing analogy comes from quantum field theory. There, an "anomaly" is a conserved topological obstruction. It must be reproduced at every energy scale, but the way it is carried can change when symmetry breaks. Global anomaly matching in the higher-dimensional CP^{N-1} model shows that when monopoles break a continuous $U(1)$ symmetry to $\mathbb{Z}_n$, a mixed anomaly survives if and only if $\gcd(n,N)>1$. That is an exact arithmetic rule for robustness under discrete symmetry breaking. We claim the fluid analogue is concrete and testable. The charge $-2$ is a Berry "monopole": a source of the geometric phase of the waves, sitting at the point where the two wave bands touch in the space of wavenumber and rotation, $(k_x,k_y,f)$. We predict the following. The total charge is always preserved, the analogue of anomaly matching. Whether it stays concentrated in a single double monopole, which fixes the dispersion and robustness of the edge waves, depends arithmetically on the residual $n$-fold symmetry. Beyond that, vortices strong enough to reverse the sign of the local absolute vorticity $f+\zeta$ create new interfaces that carry edge waves. That is exactly the regime where the inviscid, small-Rossby-number picture fails.
Core Claim
In rotating shallow water, the Coriolis term that controls band topology is the absolute vorticity $f+\zeta$, where $\zeta$ is the vorticity of the background flow. It is not $f$ alone. This gives us two claims.
The first is the matching claim. Under any deformation that keeps the bulk wave gap open away from the band-touching point, the charges summed over a sphere enclosing all degeneracies stay at $-2$. This holds for isotropy breaking, weak viscosity, and slowly varying vortex backgrounds. It plays the role that IR degrees of freedom play in (Mis-)Matching Type-B Anomalies on the Higgs Branch, which showed that apparent mismatches vanish once massless IR modes are counted.
The second is the splitting claim. When a deformation reduces continuous rotation in wavenumber space to an $n$-fold rotation $C_n$, the double monopole stays unsplit only for certain $n$ and otherwise breaks into two single monopoles. From standard rotation-eigenvalue arguments we expect this to happen for $n\in\{4,6\}$, and splitting for $n\in\{2,3\}$. This is a hypothesis we test, not a result we cite.
Splitting matters physically. It changes how many distinct edge branches exist at a given frequency, and how robustly they avoid backscattering from disorder. On the inviscid-flow side, as viscosity $\nu^*\to 0$ the boundary vortex gas of Gyre turbulence: Anomalous dissipation in a two-dimensional ocean model reaches peak vorticity $\omega_{\max}\sim 1/\sqrt{\nu^*}$. We predict that each anticyclone with $|\zeta|>f$ is enclosed by a sign-change contour of $f+\zeta$. Each such contour hosts $|\Delta Ch| = 2$ trapped one-way modes. These defects are injected by viscous detachment, and a strictly inviscid description never produces them.
Relation to Prior Work
Several papers supply the method. Counting topological interface modes using simplicial characteristic classes computes exact integer Chern numbers from sampled eigenvectors without taking derivatives, and recovers $Ch_1=-2$ for shallow water. We reuse its open-source Julia library, DiscreteVectorBundles.jl. We extend it in two ways: to spheres enclosing several candidate degeneracies, so we can see charge splitting, and to the weakly non-Hermitian operators produced by viscosity. That paper studies only linear, isotropic, inviscid operators, and this extension is exactly what we add.
Two papers motivate the inviscid-breakdown side. Quasi-geostrophy against the wall shows that coastal Kelvin waves steepen into shocks that produce cyclonic potential vorticity, while energy is conserved only as the Rossby number $Ro\to 0$. Gyre turbulence shows that no-slip walls produce viscosity-independent dissipation through a detached vortex gas. Neither paper examines the wave topology of the flows it produces. There is also a tension to state openly: in the quasi-geostrophic regime of both papers, $|\zeta|\ll f$, so no sign flips occur. Our vortex-defect mechanism therefore needs finite $Ro$, and it is precisely a statement about what the $Ro\to 0$ limit discards.
The field-theory papers constrain what the analogy can claim. The $\gcd$ criterion of Global anomaly matching in the higher-dimensional CP^{N-1} model concerns quantum 't Hooft anomalies, and our fluid rule is a classical band-topology statement. We borrow its logical structure (continuous symmetry broken to $\mathbb{Z}_n$, with an exact arithmetic survival condition), not its physics.
Neural geometric solvers are near neighbors but address a different problem. Neural Fluid Simulation on Geometric Surfaces and Topology-Preserving Neural Operator Learning via Hodge Decomposition preserve topology in physical space, meaning circulation around holes. We study topology in wavenumber space, and neither method models defect injection.
Research Program
The program has three parts.
The first is a bulk study. We linearize rotating shallow water around three families of background state. The first family is isotropic and gives the reference charge of $-2$. The second is $C_n$-symmetric, built from doubly periodic point-vortex lattices with $n=2,3,4,6$ that give Bloch wave bands, plus anisotropic topography as a vortex-free control. The third is viscous, using Laplacian friction of strength $\nu$. For each background, we sample eigenvectors on spheres of varying radius around the isotropic degeneracy. This yields both the total charge and the locations and charges of the individual degeneracies.
The second part is a direct edge test. We run time-domain simulations of an equatorial channel and of a coastal channel, seeded with vortex arrays of each symmetry. We count one-way modes by spectral flow, meaning the number of edge branches crossing the bulk gap. We also measure the transmission of wave packets past defects.
The third part is a defect-injection study. We run shallow-water versions of the wind-driven gyre at increasing Rossby number and decreasing $\nu^*$. We track the contours where $f+\zeta=0$ and test for trapped modes around them by local spectral analysis. The prediction is that the number of sign-flipping vortices, and so the number of trapped mode families, grows as $\omega_{\max}\sim 1/\sqrt{\nu^*}$ crosses $f$.
Baselines
The naive reference point is the inviscid, isotropic prediction applied everywhere. It says there is exactly one double monopole and two one-way branches at every equatorial or coastal interface, and no modes around vortices. Beating it means finding at least one controlled deformation where the measured edge spectrum departs from this prediction in the way our splitting or defect claims specify, while the measured total charge stays at $-2$.
The strongest existing method for the bulk invariant is the simplicial approach of Counting topological interface modes using simplicial characteristic classes. It reproduces $Ch_1=-2$ for shallow water exactly and recovers $Ch_1=-1$ for the Dirac model with 40 to 80 points on the phase sphere. It fails near 20 points, so the heuristic minimum is about 50. We run it unchanged on isotropic cases as a check. On deformed cases we give it the same sampling budget as our multi-sphere extension. Our extension beats it if it resolves individual charges after splitting, which a single enclosing sphere cannot report, at a sampling budget of 50 to 80 points per sphere.
The second bulk baseline is finite-difference Berry-curvature integration on identical eigenvector samples. The comparison is the fraction of runs rounding to the correct integer, measured as viscosity-induced non-Hermiticity grows.
The key ablation separates symmetry from vorticity. We break isotropy to $C_n$ with topography alone, which adds no vorticity, and separately with vortex lattices at matched anisotropy. If splitting follows $n$ in both cases, the arithmetic rule is the mechanism. If it appears only with vortices, the $f+\zeta$ shift is. For the edge and defect experiments, the reference is brute-force spectral counting in the same simulations. That count is treated as ground truth, not as a competitor.
Why not just…?
Why not just simulate the flow at high resolution and count the edge waves directly? This is a real threat to the value of the theory, since direct counting is our own ground truth. The answer is predictive reach. A bulk rule fixes the edge count for every coastline shape and every vortex arrangement in advance, whereas simulation answers one configuration at a time. The proposal succeeds only if the bulk prediction matches direct counts across all the symmetry and viscosity cases. If it fails on some, the note reports that the analogy breaks there.
Why not just integrate Berry curvature numerically? The simplicial-classes paper reports that derivative-based integration amplifies noise and yields non-integers that need ad hoc rounding. We include it as a baseline to measure how this failure worsens under viscous non-Hermiticity.
Why not just train a topology-aware neural operator such as Hodge Spectral Duality on the vortex flows? That method, which reports magnetostatics MSE $1.84\times10^{-4}$ against $2.88\times10^{-4}$ for DeepONet, preserves circulation around fixed holes in physical space. It assumes fixed topology and no defect injection, and it says nothing about wavenumber-space monopoles. Learning a surrogate would not answer which invariant survives. We therefore use no neural model.
Isn't the $C_n$ rule just known crystalline band theory? Partly. The protection of multi-charge degeneracies by rotation symmetry is expected from rotation-eigenvalue arguments. The open questions are three. Does the rule hold when the symmetry breaking comes from dynamical vortices rather than a fixed lattice? Does it survive viscosity? And does viscous detachment inject new charge?
Decisive Checks
Five outcomes decide the proposal. The total charge must read $-2$ in every gapped deformed case, or the matching claim fails. The $C_2$ and $C_3$ backgrounds must show two unit charges while the $C_4$ and $C_6$ backgrounds retain one double charge, with edge branch counts agreeing. If the topography ablation and the vortex runs agree, symmetry is the cause. In gyre runs, trapped-mode families must appear only once $\omega_{\max}>f$, and their number must track the count of sign-flip contours as $\nu^*$ decreases. Finally, the PDEBench shallow-water datasets (PDEBench) serve as an independent source of realistic backgrounds for checking the sign-flip criterion outside our own solver. No reported result was found on Lacuna for edge-mode counts in vortex-laden or viscous shallow water, so every number here is new.
References
- Global anomaly matching in the higher-dimensional <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mi mathvariant="double-struck">C</mml:mi><mml:msup><mml:mi mathvariant="double-struck">P</mml:mi><mml:mrow><mml:mi>N</mml:mi><mml:mo>−</mml:mo><mml:mn>1</mml:mn></mml:mrow></mml:msup></mml:mrow></mml:math> model
- (Mis-)Matching Type-B Anomalies on the Higgs Branch
- Gyre turbulence: Anomalous dissipation in a two-dimensional ocean model
- Counting topological interface modes using simplicial characteristic classes
- Quasi-geostrophy against the wall
- Neural Fluid Simulation on Geometric Surfaces
- Topology-Preserving Neural Operator Learning via Hodge Decomposition
- PDEBench: An Extensive Benchmark for Scientific Machine Learning