Machine-learning interatomic potentials are widely used as computationally efficient surrogates for density functional theory in atomistic simulations, enabling large-scale, long-time modeling of materials systems. We investigate how different fine-tuning strategies influence the prediction of harmonic phonon band structures, thermal properties, and the potential energy surface along imaginary phonon modes. We achieve substantial accuracy improvements with minimal additional data, with as few as 10 additional training structures already yielding significant gains. In addition to existing approaches, we introduce Equitrain, a finetuning framework that implements LoRA-based adaptation. Across 53 materials systems, we show that fine-tuned models consistently outperform both the underlying pretrained model and models trained from scratch. Equitrain achieves the best overall performance, and our results demonstrate that fine-tuning enables accurate phonon predictions.

Truncated Grassmannians are defined as closures of orbits of abelian unipotent groups acting on the degree truncations of projectivized wedge powers. We show that such truncations in a more general setup show up in the description of the blow-ups of general flag varieties along Schubert subvarieties. We work out the case of Grassmannians in detail. In particular, we show that our blow-ups are members of a larger family of varieties projecting onto Grassmannians, and describe the fibers of these projections via the spaces of collineations.

Attrition in survey and field experiments presents a challenge for social science research. Common approaches to deal with this problem -- such as complete case analysis, multiple imputation, and weighting methods -- rely on strong assumptions that may not hold in practice. This paper introduces a new method that combines recent advances in statistical inference with established tools for handling missing data. The approach produces prediction intervals for treatment effects that are both robust and precise. Evidence from simulation studies shows that the method achieves better coverage and produces narrower intervals than common alternatives. The reanalysis of two recently published experiment studies illustrates how this framework allows researchers to compare treatment effects across participants who remain in the study, those who drop out, and the full sample. Taken together, these results highlight how the proposed approach provides a stronger foundation for causal inference in the presence of attrition.

This paper investigates an initial-Neumann boundary value problem for a Keller--Segel system with parabolic-parabolic-ODE coupling. The model incorporates a signal-dependent, non-increasing motility function that, through indirect signal production, captures a self-trapping effect suppressing cellular movement at high densities. We establish the global existence of classical solutions in arbitrary spatial dimensions for a broad class of non-increasing motility functions, both with and without external source terms. Furthermore, we demonstrate that any external damping source exhibiting superlinear growth ensures uniform-in-time boundedness. Conversely, in the absence of such damping, solutions may become unbounded as time tends to infinity. More precisely, in the two-dimensional homogeneous case with the exponentially decaying motility function $γ(v) = e^{-v}$, a critical mass phenomenon emerges: classical solutions remain uniformly bounded for subcritical initial mass, while supercritical initial masses can lead to infinite-time blow-up. Our analysis relies on the construction of carefully designed auxiliary functions along with refined comparison methods and iteration arguments.

We calculate the $K$-theory of the reduced $C^*$-algebra $C^*_r(G)$ of a reductive $p$-adic group $G$. To do so, we show that each direct summand in Plymen's Plancherel decomposition of $C^*_r(G)$ is Morita equivalent to a twisted crossed product for an action of a finite group on the blow-up of a compact torus along the zero-locus of a certain Plancherel density. It follows that the $K$-theory of $C^*_r(G)$ is the direct sum of the twisted equivariant $K$-theory groups of these blow-ups, which can be computed using an Atiyah-Hirzebruch spectral sequence. As an illustration, the case of $\operatorname{Sp}_4$ is treated in some detail. Our main result is obtained from a more general study of $C^*$-algebras of compact operators on twisted equivariant Hilbert modules, from which we also recover results due to Wassermann for real groups, and to Afgoustidis and Aubert in the $p$-adic case.

We study the K-moduli stack of prime Fano threefolds of genus twelve, known as $V_{22}$. We prove that its boundary, which parametrizes singular members, is purely divisorial and consists of four irreducible components corresponding to the four families of Prokhorov's one-nodal $V_{22}$. A key ingredient is a modular relation between Fano threefolds $X$ and their anticanonical K3 surfaces $S$. We prove that the forgetful morphism from the moduli of Fano--K3 pairs $(X,S)$ where $X$ is a K-semistable degeneration of $V_{22}$ to the moduli space of genus $12$ polarized K3 surfaces $(S,{-K_X}|_S)$ is an open immersion. In particular, the K-moduli of $V_{22}$ is governed by the moduli of their anticanonical K3 surfaces, providing a modular realization of Mukai's philosophy. Along the way, we develop a general deformation framework for Fano threefolds of large volume, which may be useful beyond the study of K-moduli.

Recent advances in data-driven turbulence modeling have established tensor basis neural networks (TBNN) as a physically grounded framework for Reynolds-stress closure in Reynolds-averaged Navier-Stokes (RANS) simulations. However, their robustness in wall-bounded turbulent flows remains limited across Reynolds numbers and geometries due to the lack of an intrinsic scaling mechanism. In this work, we propose a self-scaling tensor basis neural network (STBNN) for Reynolds-stress modeling of wall-bounded turbulence. The model incorporates an invariant velocity-gradient normalization derived from the first two invariants of the velocity-gradient tensor, providing an intrinsic and geometry-independent scale that balances strain and rotation effects without relying on empirical coefficients or wall-distance inputs. Owing to its frame-indifferent formulation, the approach preserves Galilean and rotational invariance while maintaining a physically interpretable representation of Reynolds-stress anisotropy. STBNN is evaluated through a priori and a posteriori studies using direct numerical simulation (DNS) data of canonical wall-bounded flows, including plane channel and periodic hill flows. In a priori tests, the model accurately reproduces Reynolds-stress anisotropy, with correlation coefficients exceeding 99% and relative errors below 10%, while capturing near-wall scaling and logarithmic-layer behavior. In a posteriori RANS simulations, STBNN predicts mean velocity profiles in close agreement with DNS and improves prediction of separation and reattachment compared with linear and quadratic eddy-viscosity models and the baseline TBNN. Notably, a model trained at low Reynolds numbers generalizes to higher Reynolds numbers and unseen geometries. These results demonstrate the effectiveness of the proposed framework for data-driven Reynolds-stress modeling in wall-bounded turbulent flows.

We consider the homogenization of random integral functionals which are possibly unbounded, that is, the domain of the integrand is not the whole space and may depend on the space-variable. In the vectorial case, we develop a complete stochastic homogenization theory for nonconvex unbounded functionals with convex growth of generalized Orlicz-type, under a standard set of assumptions in the field, in particular a coercivity condition of order $p^->1$, and an upper bound of order $p^+<\infty$. The limit energy is defined in a possibly anisotropic Musielak-Orlicz space, for which approximation results with smooth functions are provided. The proof is based on the localization method of $Γ$-convergence and a careful use of truncation arguments.

This paper considers for the first time pursuit-evasion (PE) differential games with irrational perceptions of both pursuer and evader on probabilistic characteristics of environmental uncertainty. Firstly, the irrational perceptions of risk aversion and probability sensitivity are modeled and incorporated within a Bayesian PE differential game framework by using Cumulative Prospect Theory (CPT) approach; Secondly, several sufficient conditions of capturability are established in terms of system dynamics and irrational parameters; Finally, the existence of CPT-Nash equilibria is rigorously analyzed by invoking Brouwer's fixed-point theorem. The new results reveal that irrational behaviors benefit the pursuer in some cases and the evader in others. Certain captures that are unachievable under rational behaviors can be achieved under irrational ones. By bridging irrational behavioral theory with game-theoretic control, this framework establishes a rigorous theoretical foundation for practical control engineering within complex human-machine systems.

For any compact set $K$ lying on a closed surface $\mathcal{S}$ we introduce a closed equivalence relation $\sim$, called the {\em Schönflies equivalence} on $K$. We show that every class $[x]_\sim$ of $\sim$ is a continuum and that the resulting quotient space $K\!/\!\sim$ is a {\em Peano compactum}. By definition, all components of a Peano compactum are locally connected and for any $\varepsilon>0$ only finitely many of them have diameter greater than $\varepsilon$. The decomposition $\mathcal{D}_K=\{[x]_\sim: x\in K\}$ refines every other upper semicontinuous decomposition of $K$ into subcontinua that has a Peano compactum as its quotient space. In other words, $\mathcal{D}_K$ is the {\em core decomposition of $K$} with Peano quotient. The elements of $\mathcal{D}_K$ are called {\em atoms} of $K$. We also show that for any branched covering $f: \mathcal{S}^*\rightarrow \mathcal{S}$ from a closed surface $\mathcal{S}^*$ to $\mathcal{S}$, every atom of $f^{-1}(K)$ is sent into an atom of $K$. If $f$ is even a covering, it sends every atom of $f^{-1}(K)$ onto an atom of $K$. We illustrate our theory with examples and show that it cannot be generalized to $n$-manifolds with $n\ge 3$ by providing a detailed counterexample in~$\mathbb{R}^3$.

The Bollobás--Nikiforov conjecture asserts that for any graph $G \neq K_n$ with $m$ edges and clique number $ω(G)$, \[ λ_1^2(G) + λ_2^2(G) \;\leq\; 2\!\left(1 - \frac{1}{ω(G)}\right)m, \] where $λ_1(G) \geq λ_2(G) \geq \cdots \geq λ_n(G)$ are the adjacency eigenvalues of $G$. We prove the conjecture for all complete multipartite graphs $K_{n_1,\ldots,n_r}$ with $n_1 + \cdots + n_r > r$. The proof computes the full spectrum via a secular equation, establishes that $λ_2 = 0$ whenever the graph has more vertices than parts, and then applies Nikiforov's spectral Turán theorem; equality holds if and only if all parts have equal size. We also prove a stability result for $K_4$-free graphs whose spectral radius is near the Turán maximum: such graphs are structurally close to the balanced complete tripartite graph, and as a consequence the conjecture holds for all $K_4$-free graphs with $m = Ω(n^2)$ when $n$ is sufficiently large. Finally, we identify the precise obstruction preventing a Hoffman-bound approach from settling the conjecture for $K_4$-free graphs with independence number $α(G) \geq n/3$.

Differential privacy (DP) has established itself as one of the standards for ensuring privacy of individual data. However, reasoning about DP is a challenging and error-prone task, hence methods for formal verification and refutation of DP properties have received significant interest in recent years. In this work, we present a novel method for automated formal refutation of $ε$-DP. Our method refutes $ε$-DP by searching for a pair of inputs together with a non-negative function over outputs whose expected value on these two inputs differs by a significant amount. The two inputs and the non-negative function over outputs are computed simultaneously, by utilizing upper expectation supermartingales and lower expectation submartingales from probabilistic program analysis, which we leverage to introduce a sound and complete proof rule for $ε$-DP refutation. To the best of our knowledge, our method is the first method for $ε$-DP refutation to offer the following four desirable features: (1)~it is fully automated, (2)~it is applicable to stochastic mechanisms with sampling instructions from both discrete and continuous distributions, (3)~it provides soundness guarantees, and (4)~it provides semi-completeness guarantees. Our experiments show that our prototype tool SuperDP achieves superior performance compared to the state of the art and manages to refute $ε$-DP for a number of challenging examples collected from the literature, including ones that were out of the reach of prior methods.

Starting from the free surface Euler equations, we derive a leading-order system in terms of surface variables, depending on the surface current and on the bathymetry through the depth-dependent Dirichlet-to-Neumann (DN) operator. The resulting system is shown to be well-posed using the theory of hyperbolic systems of pseudo-differential operators. We then consider wave propagation in slowly varying environments. As an explicit approximation to the DN operator, the semiclassical Weyl quantization of the symbol $g_b(X,ξ)=|ξ|\tanh(b(X)|ξ|)$ is shown to be both asymptotically accurate and consistent with the self-adjoint structure of the true operator, and to provide the natural framework for asymptotic analysis of the wave system. A central consequence of the resulting framework is that classical asymptotic models - including the wave action equation, the mild-slope equation, the Schrödinger equation, and the action balance equation - emerge systematically from a single formulation. By deriving these equations, we show how the simple leading order system with the Weyl quantization of the DN operator provides a unified and mathematically consistent framework for the asymptotic linear theory of wave-current-bathymetry interaction, hence providing a transparent, rigorous and accessible route from the primitive Euler equations to the mentioned asymptotic models. Throughout, numerical experiments are included to illustrate the analysis.

This paper studies the degree to which a bivariate copula fails to be symmetric under coordinate permutation, a property known as non-exchangeability. Working within an axiomatic framework that quantifies this asymmetry through a family of $L^p$-based measures, we establish sharp bounds linking non-exchangeability to classical dependence and concordance measures, prove exact scaling laws under convex mixing that enable explicit construction of copulas with any prescribed degree of asymmetry, and characterise the class of maximally non-exchangeable copulas together with the feasible range of asymmetry--concordance pairs. On the inferential side, we propose a nonparametric permutation test for exchangeability with exact finite-sample error control and consistency against all asymmetric alternatives, validated by Monte Carlo simulation and illustrated on a real data set.

We study the estimation of optimal transport (OT) maps between an arbitrary source probability measure and a log-concave target probability measure. Our contributions are twofold. First, we propose a new evolution equation in the set of transport maps. It can be seen as the gradient flow of a lift of some user-chosen divergence (e.g., the KL divergence, or relative entropy) to the space of transport maps, constrained to the convex set of optimal transport maps. We prove the existence of long-time solutions to this flow as well as its convergence toward the OT map as time goes to infinity, under standard convexity conditions on the divergence. Second, we study the practical implementation of this constrained gradient flow. We propose two time-discrete computational schemes-one explicit, one implicit-, and we prove the convergence of the latter to the OT map as time goes to infinity. We then parameterize the OT maps with convexity-constrained neural networks and train them with these discretizations of the constrained gradient flow. We show that this is equivalent to performing a natural gradient descent of the lift of the chosen divergence in the neural networks' parameter space, similarly to drifting generative models. Empirically, our scheme outperforms the standard Euclidean gradient descent methods used to train convexity-constrained neural networks in terms of approximation results for the OT map and convergence stability, and it still yields better results than the same approach combined with the widely used Adam optimizer.

We introduce a canonical operator-theoretic construction associated to a finite geometric lattice, in which a simple nonassociative ``diamond product'' on the lattice basis gives rise to a family of creation operators indexed by atoms and a corresponding self-adjoint Hamiltonian on $\mathbb R[L]$. A key structural feature is that the Hamiltonian changes rank by at most one, so that its compression to the rank-radial subspace is a Jacobi matrix. In this way, geometric lattices give rise in a direct and uniform manner to finite orthogonal polynomial systems. The Jacobi coefficients admit explicit combinatorial formulas. For Boolean lattices one obtains the centered Krawtchouk Jacobi matrix, while for projective geometries one obtains natural $q$-deformations consistent with the $q$-Hahn family. The construction applies to arbitrary geometric lattices and requires no symmetry assumptions.

We simulate the thermalization dynamics for minimally truncated SU(2) pure gauge theory on linear plaquette chains with up to 151 plaquettes using IBM quantum computers. We study the time dependence of the entanglement spectrum, Rényi-2 entropy and anti-flatness on small subsystems. The quantum hardware results obtained after error mitigation agree with extrapolated classical simulator results for chains consisting of up to 101 plaquettes. Our results demonstrate the feasibility of local thermalization studies for chaotic quantum systems, such as nonabelian lattice gauge theories, on current noisy quantum computing platforms.

In a remarkable paper, Tatsuyuki Hikita settled a longstanding e-positivity conjecture of Stanley and Stembridge. Among many other things, he wrote down a certain formula $φ_k$, and proved that the $φ_k$ sum to one, thereby defining a probability distribution. Though Hikita's proof was simple, it remains surprising that the $φ_k$ sum to one. In this note, we give a combinatorial interpretation of Hikita's probability distribution. The main tool is a certain permutation statistic that we call the watershed. After seeing an early version of our work, Darij Grinberg noticed that the permutation statistic was implicit in a so-called "bulldozer problem" that was on the short list for the 2015 International Mathematics Olympiad. However, our description of the statistic, which makes use of the Renyi-Foata bijection, appears to be new.

We investigate the relation between non-Hermitian Hamiltonian and Lindblad dynamics in nonequilibrium open quantum systems. Non-Hermitian models can extend phase diagrams and enable sensing advantages, but such effects often rely on postselection, raising questions about their relevance for unconditional dynamics. Using a minimal two-qubit setup mediating a heat current, we compare local and global Markovian master equations with their non-Hermitian counterparts. We observe that exceptional points emerge only in the local master equation and in the corresponding non-Hermitian Hamiltonian at sufficiently strong nonequilibrium. We further consider hybrid configurations, where one bath is treated with a Lindblad description and the other with a non-Hermitian approach, interpolating between the two extremes. Our results contribute understanding the role of quantum jumps and exceptional points in nonequilibrium open quantum systems and identify a simple, experimentally accessible architecture, realizable, for instance, in circuit-QED platforms, for their exploration.

Governments routinely adjust capacity in rationed programs such as university fields, medical training and public housing, where admitting one individual displaces others and triggers chains of reallocation. We show that in such settings, the standard multi-treatment two-stage least squares (2SLS) coefficient identifies exactly the total societal effect of a marginal expansion, including all downstream reallocations. The result is an algebraic identity: under instrument relevance and a single alignment condition, satisfied in centralized admissions systems, the 2SLS coefficient equals the general-equilibrium shadow value of relaxing a capacity constraint, while the single-instrument Wald ratio captures only the direct effect. Their difference recovers the full equilibrium adjustment without additional structure. Monotonicity is not required. The identity extends beyond queue-based allocation to any fixed-supply setting, including competitive markets with price instruments. We apply the framework to two policy questions in Swedish university admissions, where marginal students are allocated across fields through a centralized lottery mechanism. First, revisiting the debate on whether economics and business education erodes prosocial values, we find that the direct effect of expanding business on charitable giving is precisely zero, but expanding the less competitive fields that business students are displaced from has large prosocial effects. Second, analyzing gender-targeted STEM policies, we find that admitting four women to competitive STEM generates one additional male STEM degree through downstream vacancies. Both are general-equilibrium effects invisible to single-instrument methods.

Photonic computing using chalcogenide phase-change materials (PCMs) is under active development for energy-efficient artificial intelligence (AI) applications. A key requirement is to enable as many optically programmable levels per device as possible, while maintaining relatively low optical loss. In this work, we carry out multiscale simulations using density functional theory and finite-difference time-domain methods, proposing a "the shorter the better" strategy to optimize the performance of Sb2Te photonic waveguide devices. Our subsequent experimental characterizations of Sb2Te thin films and optical device measurements fully verify our theoretical predictions. In particular, we reveal the unconventional optical properties of metastable crystalline Sb2Te, and utilize these features for device design, yielding a simultaneous improvement in both the programming window and the optical loss. Overall, an optical programming precision exceeding 7-bit is achieved using a single waveguide cell, setting a new record for all-optical phase-change memory devices. Our work serves as a compelling example of computational material design, which demonstrates the predictive power of multiscale simulations in guiding the design of phase-change photonic devices for enhanced performance.

Bubble nucleation at catalyst surfaces plays a critical role in the operation of electrolyzers. However, achieving controlled bubble nucleation remains challenging due to limited understanding of the underlying mechanisms. Here, we present a free-energy model that quantitatively predicts both the activation energy and critical nucleus size of bubbles at given supersaturation, temperature, pressure, and surface wettability. We find that the activation energy $ΔG_{max}$ decreases with increasing supersaturation $ζ$, following a power-law scaling of $ΔG_{max} \sim ζ^{-2}$, while the critical nucleus radius $R_c$ scales as $R_c\sim ζ^{-1}$. Our theoretical predictions for the critical nucleus radius of hydrogen, oxygen and nitrogen bubbles are in quantitative agreement with experimental measurements. Finally, we present a simple model that couples gas diffusion and electrochemical reaction kinetics to determine the maximum gas supersaturation at a given current density. Our results advance the fundamental understanding of bubble nucleation at catalyst surfaces and provide practical guidelines for catalyst layer design to improve the performance of electrolyzers.

In a recent paper [arXiv:2205.06279], Danielson et al. demonstrated that the mere presence of a black hole causes universal decoherence of quantum superpositions (dubbed the DSW decoherence). We analyze decoherence in a superconducting analogue [arXiv:1709.06154] of the event horizon of a black hole, where Andreev reflection plays the role of Hawking radiation. We consider a normal metal interferometer threaded by an Aharonov-Bohm flux, where one of the arms of the interferometer is coupled to a superconductor by a tunnel coupling of varying strength. At absolute zero temperature and for weak coupling, we find that the scattering states of the interferometer are decohered by Andreev reflection, a nontrivial manifestation of the proximity effect analogous to DSW decoherence from the event horizon of a black hole. However, for increasing coupling strength to the superconductor, we find a reemergence of coherence via resonant tunneling through Andreev bound states. This suggests the existence of an analogue gravitational phenomenon wherein transmission mediated by virtual Hawking radiation leads to a reemergence of coherence in an interferometer placed within a few Compton wavelengths of a black hole's event horizon.

We investigate the one-loop renormalization group evolution in four dimensions of the leading operators in the effective field theories of shift-symmetric scalars, photons, and gravitons. We show that certain non-minimal three-point interactions induce a negative running of the corresponding Wilson coefficients, with beta-functions suppressed by the Planck scale. The decrease of the coefficients toward the infrared prompts us to revisit their dispersive bounds, in particular accounting for graviton loops. Gravitational interactions generate positive infrared contributions which, after smearing over the momentum transfer, are argued to dominate over the negative running, provided the number of non-minimally coupled particles is bounded from above according to the species bound.

The Variational Quantum Eigensolver (VQE) is widely regarded as a promising algorithm for calculating ground states of quantum systems that are intractable for classical computers. This promise is typically motivated by the hope of mitigating the exponential growth of Hilbert space with system size. Here we scrutinize how the computational cost of adaptive VQE scales with the size of the target system. We demonstrate that the Rényi entropy derived from classical simulations predicts the required number of adaptive iterations of VQE with high accuracy ($R^2 \approx 0.99$). We validate this on a benchmarking set of more than 20 different molecules with active spaces ranging from four to ten orbitals. For these molecules, we find an exponential scaling of the number of adaptive iterations, and in turn, of the circuit depth with the system size. We therefore conclude that it is unlikely that VQE in its current form is able to simulate large molecular systems with high fidelity without exponential resource requirements.

The optimization and interpretation of microlensing surveys depends on having an accurate model of the Milky Way. However, existing population synthesis Galactic modeling tools often perform poorly in replicating the stellar contents of the inner Galactic bulge region and reproducing microlensing survey results. We present an updated Galactic model implementation within the \synthpop framework that has been tuned for simulating the upcoming {\it Nancy Grace Roman Space Telescope}'s Galactic Bulge Time Domain Survey (RGBTDS). We evaluate the model against stellar catalogs and kinematics from optical and infrared surveys toward the Galactic bulge, finding good agreement in much of the bulge, including the RGBTDS' contiguous lower bulge fields. However, within Galactic latitudes of $b\lesssim0.5^\circ$ of the Galactic plane, some inconsistencies arise which may impact projections for the RGBTDS' Galactic center field. The model over-predicts optical microlensing event rate per star measurements by a $\sim20$\%, but detailed comparisons to near-infrared measurements are hampered by their lack of detection efficiencies. {\it Roman}'s GBTDS and Galactic Plane Survey will be instrumental in resolving the remaining model inconsistencies and improving our understanding of the structure of the central few degrees of our Galaxy.

We study penalization coupled with time discretization for decoupled Markovian doubly reflected BSDEs with obstacles \(p_b(t,X_t)\le Y_t\le p_w(t,X_t)\). The DRBSDE is approximated by a penalized BSDE with parameter \(λ\) and discretized by an implicit Euler scheme with step \(Δt\). A key difficulty is that the forward approximation used to evaluate the obstacles generates an error term that is amplified by \(λ\). In the single-obstacle case this amplification can be removed by the shift \(Y-p_b(t,X)\), but no analogous transformation eliminates both obstacles simultaneously; this motivates simulating the forward SDE on a finer grid \(\tilde{Δt}\) and projecting onto the backward grid (two-grid scheme). Under structural assumptions motivated by financial barriers we sharpen penalization rates and obtain a uniform \(O(λ^{-1})\) bound for the value process. We derive an explicit error bound in \((Δt,\tilde{Δt},λ)\) and tuning rules; for \(Z\)-independent drivers, \(λ\asymp Δt^{-1/2}\) with \(\tilde{Δt}=O(Δt/λ^2)\) yields the target \(O(Δt^{1/2})\) rate. Nonsmooth barriers/payoffs are handled via a multivariate Itô--Tanaka and local-time-on-surfaces argument. We also provide numerical experiments for a one-dimensional game put under the Black--Scholes model. The observed grid-refinement errors are consistent with the predicted \(O(n^{-1/2})\) behavior, while the penalty sweep indicates that the tested regime remains pre-asymptotic with respect to the penalty parameter.

$k$-means has historically been positioned primarily as an offline processing primitive, typically used for dataset organization or embedding preprocessing rather than as a first-class component in online systems. In this work, we revisit this classical algorithm under the lens of modern AI system design and enable $k$-means as an online primitive. We point out that existing GPU implementations of $k$-means remain fundamentally bottlenecked by low-level system constraints rather than theoretical algorithmic complexity. Specifically, the assignment stage suffers from a severe IO bottleneck due to the massive explicit materialization of the $N \times K$ distance matrix in High Bandwidth Memory (HBM). Simultaneously, the centroid update stage is heavily penalized by hardware-level atomic write contention caused by irregular, scatter-style token aggregations. To bridge this performance gap, we propose flash-kmeans, an IO-aware and contention-free $k$-means implementation for modern GPU workloads. Flash-kmeans introduces two core kernel-level innovations: (1) FlashAssign, which fuses distance computation with an online argmin to completely bypass intermediate memory materialization; (2) sort-inverse update, which explicitly constructs an inverse mapping to transform high-contention atomic scatters into high-bandwidth, segment-level localized reductions. Furthermore, we integrate algorithm-system co-designs, including chunked-stream overlap and cache-aware compile heuristics, to ensure practical deployability. Extensive evaluations on NVIDIA H200 GPUs demonstrate that flash-kmeans achieves up to 17.9$\times$ end-to-end speedup over best baselines, while outperforming industry-standard libraries like cuML and FAISS by 33$\times$ and over 200$\times$, respectively. Our code is open-sourced at https://github.com/svg-project/flash-kmeans.

The presence of a quintessence-like field can influence the black hole shadow through three primary mechanisms: the dynamics of accretion flows, the trajectories of photons, and the motion of observers. Unlike standard shadow analyses that assume a static observer at spatial infinity, the non-asymptotically flat nature of quintessence-corrected spacetimes motivates the consideration of freely falling (geodesic) observers. Using a perturbative approach, we derive analytical expressions for the event-horizon location, photon-sphere radius, innermost stable circular orbit, and critical impact parameter. We compute the observed intensity profiles for both static and infalling spherical accretion flows. We find that, although the photon-sphere radius and the critical impact parameter are invariant properties of the spacetime, the apparent angular size of the shadow depends sensitively on the observer's motion and location. Freely infalling observers systematically measure smaller angular radii than static observers at the same radius, whereas freely outgoing observers measure larger ones, in agreement with relativistic aberration. In contrast to the Schwarzschild case, the impact parameter alone is insufficient to characterize the observed angular structure in non-asymptotically flat spacetimes. Applying our results to the Event Horizon Telescope observation of M87$^\ast$, we show that more negative equations of state lead to stronger constraints on the quintessence parameter, largely independent of the observer prescription. Our analysis highlights the importance of carefully specifying the observer in shadow studies of non-asymptotically flat black-hole spacetimes.

Finite element simulations play a critical role in a wide range of applications, from automotive design to tsunami modeling and computational electromagnetics. Performing these simulations efficiently at the high resolutions needed for practical applications and scientific insights necessitates the use of high-order methods and large-scale supercomputing. While much progress has been made in porting finite element codes to GPU systems in recent years, additional improvements in the efficiency and computational speed of GPU-accelerated high-order finite element simulations are in constant demand. In this paper, we demonstrate that the FP64 tensor cores on NVIDIA GPUs can be used to further accelerate such simulations, achieving significant speedups in key kernels of MFEM, a scalable open-source finite element library widely used in HPC applications. By integrating FP64 tensor cores with kernel fusion optimizations, we were able to achieve up to 2$\times$ performance gains and up to 83% energy efficiency gains on NVIDIA's Grace Hopper GH200 and Grace Blackwell GB200 architectures. To the best of our knowledge, this is the first time that FP64 tensor cores have been directly programmed to accelerate large-scale finite element scientific computing applications. We demonstrate the performance of the optimized kernels at exascale by showing near-perfect weak scaling efficiency and 90% strong scaling efficiency across nearly 10,000 GPUs on the Alps system. The new algorithms and MFEM enhancements directly benefit complex production codes, including the 2025 Gordon Bell Prize-winning application for real-time tsunami forecasting.