edge computing

PlanarBench: Evaluating LLM Spatial Reasoning via Planar Graph Drawing

Existing LLM graph benchmarks typically ask models to answer graph-theoreticquestions or compute symbolic solutions rather than construct spatial layouts.Within-task difficulty is also primarily stratified by vertex count. However, existingresearch also suggests that task difficulty is more closely related to the number ofconstraints imposed by the edges than to the number of vertices being arranged.We introduce PlanarBench, a benchmark that asks models to produce crossing-free ASCII drawings of planar graphs given only an edge list. Across 91 modelconfigurations and 199 non-isomorphic connected planar graphs with 2–7 vertices,edge count is more strongly associated with mean task score than vertex count(r = −0.85 versus r = −0.47) and remains strongly associated after controllingfor vertex count (rpartial = −0.80). PlanarBench provides a controlled settingfor separating these two difficulty axes. In addition, neither drawing area nortotal response length demonstrated a meaningful correlation with score, which isevidence against a simple output-size explanation. Performance varies widely: thebest model scores 159.5 out of 199, most models below 30B parameters scoreunder 25, and substantial failures remain among frontier systems.

Artificial Intelligence and Edge Computing for Sustainable Smart Water-Safety Monitoring in Low-Resource Communities: A Critical Review

Poor water quality monitoring and delayed responses to pollution remain major challenges in low-resource areas. Traditional methods of monitoring water and wastewater resources are ineffective because they take a long time to report contamination. Therefore, this critical review aims to examine how a combination of artificial intelligence (AI) and edge computing can comprehend decentralised, real-time water quality monitoring, even in areas with limited infrastructure and internet, constrained maintenance capacity, and shortages of skilled personnel. To our knowledge, this study is a first-of-its-kind integrated framework that showcases edge AI architectures and refers to specific operational, societal, and infrastructural limitations in the water, sanitation, and hygiene (WASH) sector. The synthesis clearly shows that the implementation of edge AI techniques has the capability to improve global water quality through immediate pollution detection, disaster forecasting, and automatic filter or alarm response without the need for cloud infrastructure. The examples given from developing countries support this statement by demonstrating that the technologies are low-cost and implementable in the long term. The problem of sensor calibration, data quality, and energy efficiency was identified as the most important implementation challenge. There is enough evidence of pilot-scale tests, but long-term validation of the technology in field trials is needed. The authors also present future research directions, such as the integration of AI, edge computing, machine learning, and IoT, and open-source edge frameworks. Edge AI systems provide promising avenues for decentralised water safety surveillance in low-resource communities in real time and can support the sustainable development goals of the United Nations through the realisation of clean water and sanitation for all.

What must a theory of perturbational complexity explain? Nine preregistration-hardened constraints, six dead hypotheses, and a minimal visibility account

Across five preprints (doi:10.5281/zenodo.22100546, .22100826, .22101059, .22120069, .22133403), we measured how perturbational complexity behaves once its estimator is debiased, what quantity it tracks — the reproducible dimensionality of the evoked response (R-dim) — and how that quantity depends on dynamics, trial count, species, and arousal state, always under preregistrations sealed before data contact and with failures published. This paper consolidates the empirical residue into nine constraints that any theory of perturbational complexity must satisfy, reports the autopsies of six hypotheses of ours that died against them (including our own formal theory, killed in two sealed prediction cycles), and offers the minimal account we know of that survives: reproducible components with rapidly decaying amplitudes become visible one by one as trial averaging lowers an effective noise floor. With an approximately exponential amplitude spectrum, this account reproduces, with three lines of algebra, five constraints at once: logarithmic growth of R-dim with trial count, absence of a detectable human ceiling at protocol-scale trial counts, finite system-specific ceilings in simulated networks, level scaling with electrode coverage without growth scaling, and — via uniform attenuation of component amplitudes — the downward displacement of the growth curve under anesthesia. (v2) v1 described that displacement as parallel (level, not slope); preprint 6 v2 withdraws the parallel-shift shape, so C9 now constrains the existence of the contrast and the recovery lag rather than the shape. The account is explicitly not a theory of consciousness: it is silent on why rich spectra require edge-of-chaos dynamics, why networks self-organize toward that regime, and why recovery from anesthesia lags the state change. It does, however, make a quantitative, falsifiable prediction — the vertical displacement between two states' curves should equal twice the within-state slope times the log of the evoked-amplitude gain ratio — and we tested it: the test protocol was sealed publicly before the one unpublished quantity it requires (the gain ratio) was computed for any animal, and run once. The prediction held on both sealed criteria at the field level (ordering ρ = +0.41, one-sided p = 0.038, n = 20; median observed/predicted ratio 1.00). It then failed its cross-level replication: sealed identically and run once on spiking populations of the same brains (commit 76f2fea), ordering inverted (ρ = −0.29, p = 0.84) and observed displacements exceeded predictions by a median factor of 3.0 — anesthesia does more to neuronal dimensionality than uniform amplitude gain allows. The visibility account therefore survives as an economical summary of the trial-scaling constraints and of C9's shape, and dies as a mechanistic claim at the neuronal level; we report the refutation of our own model here rather than elsewhere. We then chased the structural change itself with a further sealed protocol: anesthesia rotates the population response subspace to near-orthogonality (median between-state overlap 0.082 against a within-state control of 0.532; lower in 15 of 15 animals, p = 3.1 × 10⁻⁵) — anesthesia does not attenuate a fixed response so much as replace it — yet the extent of rotation does not scale the excess displacement either (ρ = −0.09, p = 0.62), killing the rotation-as-explanation hypothesis in the same run that established the rotation. The last pre-declared candidate, per-component decoherence, is reported as untested for a reason we document in full: our estimator of trial-level coherence at matched amplitude failed its pre-registered known-truth validation bench (three versions, criteria fixed before each run, all archived), and under our rules an uncertified instrument runs no confirmatory test. The theory this field needs must pass through all nine constraints; we mark where every account we tried has failed so others can start further ahead.

Frustration and holonomy in sheaves of qualia structures: quantitative local-to-global principles for phenomenal unity

In [1] we introduced the frustration f(S)f(S) of a sheaf of pseudometric spaces — the least consistency radius (in the sense of Robinson [3, 4]) achievable by any assignment — and proved metric local-to-global principles: holonomy displacement bounds, exact values on cycles, and rectification on acyclic covers. Prompted by correspondence on that paper, we here restrict the stalks to finite-dimensional normed vector spaces and show that the entire theory becomes cohomological. For an affine sheaf — a linear cellular sheaf SS on a multigraph twisted by a 11-cochain zz — we prove the cochain-level reformulation fp(S,z)=21/p−1dist⁡p(z,im⁡δ0)=21/p−1∥[z]∥H1,f_p(S,z) = 2^{1/p-1}\operatorname{dist}_p(z,\operatorname{im}\delta^0) = 2^{1/p-1}\|[z]\|_{H^1}, one half (for p=∞p=\infty) the quotient norm of the obstruction class in sheaf cohomology: frustration is a norm on H1H^1, its optimal assignments are minimal-norm cocycle representatives (harmonic for p=2p=2, minimax for p=∞p=\infty), and in finite dimensions the obstruction is complete: f=0f=0 iff [z]=0[z]=0 iff a global section exists. Duality identifies the frustration with a maximization over the dual cycle space ker⁡δT\ker\delta^T; by Rockafellar’s theory of elementary vectors, for rank-one (gain-graph) sheaves at p=∞p=\infty the optimal dual certificates are supported on the circuits of Zaslavsky’s frame matroid — balanced cycles and unbalanced theta/handcuff pairs — yielding a closed combinatorial formula that strictly extends the mean-cycle theorem of [1], and explaining exactly when cycle holonomy fails to determine frustration: the maximizing certificate can be a handcuff. For translation systems we identify the extremal frustration-to-cycle-bound ratio on two-vertex (banana) graphs with the Jung constant J(V)J(V) of the stalk norm — the equilateral-theta gap 2/32/\sqrt3 of [1] is exactly J(ℓ22)J(\ell_2^2) — settling the two-vertex case and the necessity direction of the Helly boundary question posed there, with sufficiency conjectured (and numerically supported): bananas appear to be the worst case in general. For homogeneous linear sheaves the natural invariant is the unit-normalized frustration, which equals σmin⁡(δ)/2\sigma_{\min}(\delta)/\sqrt2 — equivalently, its square is half the spectral gap λmin⁡\lambda_{\min} of the Hansen–Ghrist sheaf Laplacian; on cycles with orthogonal restriction maps λmin⁡=2−2cos⁡(θ∗/n)\lambda_{\min}=2-2\cos(\theta^\ast/n), and we give the exact dictionary between this spectral theory and the metric cycle formula of [1]. Finally we connect all of this to Robinson’s transmission-line sheaves on quantum graphs [J. Differential Equations 260 (2016) 872–896]: in the worked loop-with-tail example his resonance conditions and cohomology-dimension jumps are precisely the zero locus of our quantitative obstruction σmin⁡\sigma_{\min} of the secular map, which we compute in closed form; loss imposes a uniform positive lower bound on it; and his gauge freedom under edge collapse leaves cohomology invariant while transforming frustration by condition-number factors — frustration sees the geometry that cohomology forgets. The main identities and examples are machine-verified (18 further checks, all passing).

A Confound-Annotated Curriculum Dataset with Parsed Prerequisite Logic for the Universities of the United Arab Emirates

A machine-readable, course-level corpus of the curricula of the universities of the United Arab Emirates, assembled from their published course catalogs. The corpus comprises twenty-two institutions across 56 catalog editions, 52,802 course records, and 33,945 parsed prerequisite relations, of which 7,030 (20.7%) are disjunctive alternatives rather than mandatory obligations. Its distinguishing property is that prerequisites are parsed into conjunctive-normal form, so that the alternatives a catalog states with the word "or" are preserved as boolean structure rather than flattened into a list of mandatory courses; the group index and alternative flag in the edge file recover the full conjunctive-normal structure. Every record is annotated with the measurement confounds that make document-derived curriculum data misleading if they are ignored, namely notation drift, selective disclosure, subject-code renumbering, and prerequisite-operator ambiguity, each exposed as a filterable field. One institution, the United Arab Emirates University, is covered by an eleven-edition panel spanning the decade from 2015-2016 to 2025-2026. All twenty-two institutions are represented at the course level. The verbatim course-description prose is not redistributed; its availability, language, and length are recorded in the course table, and its semantic content is provided as non-reproducing sentence embeddings. A pre-registered sampled correctness audit, included with the deposit, places course-code agreement at 100%, credit and prerequisite agreement in the mid-to-high nineties, and substantive title accuracy near 99%. The deposit includes the analysis code that computes curricular complexity under both the standard all-conjunctive reading and the alternative-aware reading, the integrity-verification script, and the full audit bundle.

Cumulative Intraoperative Hypothermic Burden, Transfusion, and Estimated Blood Loss: A Retrospective Cohort Study

Superseded. This record is retired and should not be used or cited. It is replaced by 10.5281/zenodo.22234051, which is the archive accompanying the manuscript. Version v2.0.0 of this record (10.5281/zenodo.22232689) additionally contained files that should not have been distributed and its removal has been requested. Analysis pipeline for a retrospective cohort study of cumulative intraoperative hypothermic burden, intraoperative red cell transfusion, and estimated blood loss in 2,567 adults undergoing non-cardiac surgery, using the open VitalDB perioperative database. Scripts run in numeric order: 01-14 reproduce the originally submitted analysis, and 15-23 produce the first revision, including the edge-trimmed thermal exposure metric, threshold and dose analyses, and the table and figure builders. The repository carries code only; the manuscript, cover letters and peer-review correspondence are not distributed. VitalDB source data are openly available at https://vitaldb.net and are not redistributed here. v2.2.0 is the clean release accompanying the first revision submitted to BMC Anesthesiology. The repository was rebuilt from a single commit containing code only. Key methodological change from v1.1.0: core temperature had been integrated across the whole anaesthesia window, which counts probe equilibration at insertion and probe withdrawal at emergence as hypothermia, implying an implausible cohort nadir of 33.30 °C. Exposures are now computed after excluding the first and last 5 minutes of each case's monitored window. Cumulative burden is essentially unchanged (Pearson r=1.00, Spearman rho=0.99); the nadir corrects to 35.04 °C. Earlier v2.x versions are superseded. v2.0.0 in particular should not be used.

TinyHLS, a Python-based Hardware Compiler for 1D and 2D Convolutional Neural Networks

TinyHLS is a Python-based hardware compiler that automatically generates hardware accelerators in the form of hardware description language (HDL) code for Convolutional Neural Networks (CNNs). The description of the CNN architecture as well as the training is done in advance using Python TensorFlow Keras. TinyHLS reduces the development effort to implement inference calculations in digital hardware regarding cost and time. Furthermore, tinyHLS offers a platform independent alternative to commercial high-level synthesis tools like AMD Vivado HLS TM [ 1 ]. Each hardware accelerator generated by tinyHLS is a full hardware implementation of its CNN, allowing low-latency and low-power inference. In this work, the concept of this hardware compiler is presented. The workflow of tinyHLS is demonstrated based on a smart farming use case. For this use case a CNN to detect oranges in images is developed in TensorFlow Keras, translated using tinyHLS and implemented on a field programmable gate array (FPGA). The results in terms of accuracy, latency, energy consumption and hardware requirements are then discussed based on the implementation of the use case 144 CNN. Finally, a brief outlook on the improvement of tinyHLS is given to meet requirements of edge artificial intelligence (AI) computing in the future.

Reproducibility package — Source-Code Analysis of iFogSim for Simulating Distributed IoT Architectures

Simulation is an indispensable tool for validating distributed IoT architectures before physical deployment, and iFogSim has emerged as one of the most widely adopted platform in the fog and edge computing research community. Yet the experience of using iFogSim for non-canonical, application-specific architectures remains incompletely documented, leaving practitioners without guidance on when the tool is appropriate, which scientific objectives it can address, and how to manage the modelling approximations it imposes. This article helps in providing that guidance through two complementary contributions. First, we present a structured state of the art covering iFogSim and iFogSim2, a taxonomy of ten scientific objectives that motivate IoT architecture simulation, and a comparative survey of eight simulation tools assessed against those objectives. Second, we report our experience of simulating a four-tier smart emergency response system for resource-constrained urban environments, covering a 25-node synthetic road topology, four experimental configurations, and quantitative results including end-to-end alert latency (near 205 ms), FPGA-accelerated Dijkstra path computation (x10 CPU speedup), concurrent incident conflict rates (75% under dual load), and path-cache acceleration (x197). The analysis is organised around five practitioner questions: whether iFogSim fits the target architecture, which objectives it covers natively versus partially, what modelling challenges arise and how their workarounds bias reported results, what changes to the iFogSim source code would close the identified gaps, and whether tool co-simulation can provide comprehensive coverage. Seven modelling challenges are documented with source-code-grounded root causes and explicit bias assessments; finally, seven developer recommendations are proposed as an actionable improvement roadmap for the iFogSim community.

Improved Conservative Scheme for Lenard‐Bernstein Collisions in Gyrokinetic Turbulence Simulations

ABSTRACT Turbulence in the edge and scrape‐off layer regions plays a critical role for the performance of future magnetic confinement fusion power plants. Gyrokinetic simulations allow studying this regime with high fidelity. A key aspect in these regions is the high concentration of impurities, which can radiate energy, leading to significant losses. Due to large mass and high charge state, impurities are highly collisional, making them difficult to model accurately. This work presents discretization and algorithmic improvements for Lenard‐Bernstein collisions in gyrokinetic simulations based on previous conservative finite‐volume scheme. The new discretization improves numerical consistency by eliminating conservation errors, which were previously circumvented through the use of free parameters. While small boundary corrections remain necessary, we show that numerical conservation can be improved through careful stencil design, reducing reliance on free parameters. Its implementation is verified through conservation and relaxation tests. The algorithmic improvements focus on computational performance, achieving compute and communication performance gains in a scaled‐down TCV‐X21 benchmark. It also scales as with the number of species , significantly improving upon the previous naive implementation.

Bursty workload-adaptive scheduling with data protections for industrial cloud–edge collaboration

Gray Langurs Optimizer-Optimized Feature Mode Decomposition for Adaptive Denoising of Multi-Source Monitoring Data from Floating Offshore Wind Turbines

Feature Mode Decomposition (FMD) adaptively decomposes signals into band-limited modes through an adaptive finite impulse response (FIR) filter bank optimized via correlated kurtosis (CK) maximization, yet its denoising performance is highly sensitive to four hyperparameters—the number of decomposition modes nm, the filter length L, the CK shift order M, and the characteristic-period scaling Tscale—whose manual tuning is impractical for multi-channel floating offshore wind turbine monitoring deployments. We propose GLO-FMD, an adaptive denoising framework coupling the Gray Langurs Optimizer (GLO) with FMD. GLO autonomously optimizes the FMD parameters, thereby aligning the CK objective with structural modal periods rather than impulsive fault periods. Although the search space spans (nm,L,M,Tscale), the CK shift order M is fixed at 2 and Tscale is estimated automatically from the dominant autocorrelation peak; consequently, only (nm,L) are actively optimized. The optimized FMD decomposes multi-axis tower-base signals into band-limited modes through iterative CK-maximizing FIR filter optimization; each mode identifies a dominant periodic component, and the original signal is zero-phase band-pass filtered around the identified frequencies to preserve physical phase during reconstruction. Validation employs (i) semi-synthetic signals reproducing the measured tower-base structure (a smooth 0.15 Hz structural mode plus an impulse-excited 3.77 Hz resonance) with exactly known ground truth—a best-case benchmark by construction that isolates denoising capability from reference uncertainty—and (ii) real strapdown inertial sensor data acquired at 8 Hz from the tower-base interface of a floating offshore wind turbine at an operational site in Chinese coastal waters, over a six-day measurement campaign (18–23 April 2023). Six kinematic channels spanning triaxial acceleration (north, up, east) and triaxial velocity (north, up, east) are analyzed, with 200-s (1600-sample) continuous windows extracted for algorithmic evaluation. On the semi-synthetic data, GLO-FMD achieves a 9.6–10.2 dB SNR improvement over default wavelet thresholding against the known ground truth, and the GLO optimization is essential for reliability—the default FMD configuration is unstable across noise realizations, whereas the optimized parameters recover the clean components consistently. GLO-FMD also achieves pseudo-reference-relative SNR gains of 5.3–7.8 dB over default wavelet thresholding across all six real-data channels. Bootstrap resampling over 12 independent segments confirms statistical significance (p<0.001, Cohen’s d>8), and a no-reference smoothness index provides complementary evaluation independent of the pseudo-reference assumption. Multi-day consistency analysis yields coefficients of variation below 5%, demonstrating short-term consistency across the environmental conditions represented in the six-day dataset. The online denoising stage requires approximately 1.5 s per channel, supporting potential deployment on edge-computing hardware at the turbine controller level.

Cooperative Computation for Multiuser Task Offloading in Wireless-Powered MEC Systems

This paper investigates joint computing and relaying for multiuser task offloading in a wireless-powered mobile edge computing (MEC) system comprising an energy node (EN), an edge server (ES), and multiple energy-harvesting users. One user is selected as the helper for the remaining task users. Each task user partitions its workload among local computing, cooperative computing at the helper, and remote execution at the ES. During a parallel cooperation stage, the helper computes one portion of the uploaded tasks locally while forwarding the remaining portion to the ES and also processes its own task through local computing or edge offloading. The weighted sum computation rate (WSCR) is maximized by jointly optimizing helper selection, task partitioning, time allocation, transmission-energy allocation, and CPU-resource allocation under frame-duration, energy-neutrality, communication, and computation constraints. For each candidate helper, transmission-energy variables are introduced to decouple transmission time and power, and the perspective structure of the achievable-rate functions is exploited to reformulate the continuous resource-allocation problem as an equivalent convex problem. By solving the convex problem for all the candidate helpers, the globally optimal helper selection and resource allocation are obtained. The numerical results show that the proposed joint computing-and-relaying scheme consistently outperforms computing-only, relaying-only, and dedicated-helper cooperation. The performance gain stems from adaptively balancing helper computing and ES processing according to the prevailing communication, computation, and energy bottlenecks.