Fiesta: A Multifaceted Research Acronym

• Fiesta is a multifaceted term used across disciplines to denote distinct computational frameworks and methodologies.
• Key applications include sector-decomposition for Feynman integrals, incremental mapping in robotics, adaptive model selection in NLP, and segmentation in medical imaging.
• Evaluations reveal that Fiesta variants enhance efficiency, speed, and accuracy in numerical integration, distance-field mapping, immersive analytics, and surrogate-based inference.

Searching arXiv for relevant "Fiesta/FIESTA" papers across domains. arXiv search query: Fiesta FIESTA titles across multiple research areas Searching for exact arXiv IDs: (Oh et al., 2024, Tentyukov et al., 2010, Moss et al., 2019, Han et al., 2019, Smirnov, 2015, Smirnov et al., 2021, Dumais et al., 2022, Honarmand et al., 29 Mar 2025, Pentland et al., 2024). Fiesta, more often stylized as FIESTA and occasionally as FiEstAS or fiesta, is not a single research construct but a recurrent label used across arXiv-indexed work for software systems, algorithms, protocols, prototype environments, and imaging sequences. In the literature surveyed here it denotes, among other things, a sector-decomposition framework for multiloop Feynman integrals, an incremental Euclidean distance-field mapper for aerial robots, adaptive model-selection and test-time-adaptation procedures in machine learning, an autoencoder-based tractography pipeline, a room-scale immersive analytics system, a tokamak equilibrium code, and a JAX-based transient-inference package (Tentyukov et al., 2010, Han et al., 2019, Moss et al., 2019, Honarmand et al., 29 Mar 2025, Dumais et al., 2022, Lee et al., 2020, Pentland et al., 2024, Koehn et al., 18 Jul 2025).

1. Nomenclature and scope

The term is best understood as a family of unrelated acronymic usages rather than a unified technical lineage. In practice, meaning is entirely domain-local.

Form	Expansion or use	Domain
FIESTA	Feynman Integral Evaluation by a Sector decomposiTion Approach	Multiloop quantum field theory
FIESTA	Fast Incremental Euclidean DiSTAnce Fields	Aerial-robot mapping and planning
FIESTA	Fast IdEntification of State-of-The-Art models using adaptive bandit algorithms	NLP model selection
FIESTA	Fisher Information-based Efficient Selective Test-time Adaptation	Video facial expression recognition
FIESTA	FIbEr Segmentation in Tractography using Autoencoders	White-matter bundle segmentation
FIESTA	Free-roaming Immersive Environment to Support Team-based Analysis	Collaborative immersive analytics
Fiesta	MATLAB-based free-boundary static equilibrium solver	Tokamak Grad–Shafranov equilibrium
FIESTA	MRI sequence name in fetal body segmentation study	Fetal MRI
fiesta	JAX-based Python package for surrogate-enabled Bayesian inference	Kilonovae and GRB afterglows

This suggests that “Fiesta” functions as a reusable acronymic attractor for efficiency-oriented or systems-oriented work, but not as a stable disciplinary term. A common misconception is therefore to treat FIESTA papers as variants of one framework; the arXiv record instead shows multiple independent constructs with distinct objectives, mathematical formalisms, and evaluation protocols (Tentyukov et al., 2010, Han et al., 2019, Moss et al., 2019, Honarmand et al., 29 Mar 2025, Dumais et al., 2022, Lee et al., 2020, Pentland et al., 2024, Fadida et al., 2022, Koehn et al., 18 Jul 2025).

2. Multiloop Feynman-integral evaluation

In high-energy theoretical physics, FIESTA denotes Feynman Integral Evaluation by a Sector decomposiTion Approach, a program for the automatic numerical evaluation of dimensionally regularized multiloop Feynman integrals. Its core workflow is to transform a loop integral into a Feynman-parameter representation with Symanzik polynomials $U$ and $F$, apply sector decomposition to factorize endpoint singularities, extract the Laurent expansion in $\epsilon$, and numerically integrate the finite coefficients. A schematic target form is

$$\int_0^1 dx_1\cdots dx_N\, \prod_{j=1}^N x_j^{b_j+c_j\varepsilon}\, G(x,\varepsilon),$$

where the singular structure has been isolated into the monomial prefactor and $G(x,\varepsilon)$ is finite at the boundaries (Tentyukov et al., 2010).

The software evolved through several major releases. FIESTA 3 added cluster parallelization, calculations in physical regions via contour deformation, new asymptotic expansion algorithms, and multiple sector-decomposition strategies (Smirnov, 2013). FIESTA 4 was explicitly performance-oriented, adding CUDA-based GPU support, CPU vectorization, MPI-oriented robustness, and speedups described as about $2$–$4\times$ on CPU plus another $2$–$4\times$ from GPU in suitable regimes (Smirnov, 2015). FIESTA5 rewrote the backend in C++17, introduced the QMC and Tensor Train integrators, made S2 the default sector-decomposition strategy, and emphasized improved contour transformation and complex-mode stability (Smirnov et al., 2021).

Its benchmark applications are canonical multiloop tasks. The 2010 application paper treated four-loop massless propagator master integrals, recalled a basis of 28 independent master integrals, and reported that with the Cuba library’s Vegas integrator and 1,500,000 sampling points, FIESTA reproduced analytically known coefficients to about 3–4 correct digits while also producing higher $\epsilon$-orders needed for future five-loop work (Tentyukov et al., 2010). More recent infrastructure work places Fiesta within a broader symbolic-to-numeric ecosystem: FeynCalc 10 and FeynHelpers expose interfaces connecting FeynCalc to external multiloop tools including Fiesta, so that GLI and FCTopology objects can serve as a common symbolic frontend before numerical master-integral evaluation (Shtabovenko, 2024).

3. Incremental distance fields and motion planning

In robotics, FIESTA means Fast Incremental Euclidean DiSTAnce Fields, a mapping framework for incrementally building a global Euclidean signed distance field for online motion planning of aerial robots. The system operates on an occupancy grid updated by raycasting, maintains separate insertion and deletion queues for obstacle changes, and uses indexing data structures plus doubly linked lists so that only affected voxels are revisited. Each voxel stores its position, occupancy, distance value, and closest-obstacle coordinate, with updates of the form

$F$0

and neighbor propagation based on inherited obstacle identities rather than quasi-Euclidean path-length accumulation (Han et al., 2019).

A central technical point is that the algorithm is near-optimal rather than exact. The paper argues that no finite-connectivity BFS-based ESDF update rule can be universally exact, yet FIESTA’s Euclidean update rule is still substantially more accurate than quasi-Euclidean propagation. Empirically, it is reported to outperform Voxblox by an order of magnitude in both performance and accuracy, and onboard experiments on a quadrotor with a Velodyne VLP-16 report ESDF updates finishing within 20 ms at 20 Hz (Han et al., 2019).

A later systems paper compared Voxblox, Fiesta, and GPU-Voxels as online distance-field backends for GPMP2-based whole-body planning. There Fiesta is characterized as a time-efficient incremental unsigned distance-field generator using doubly linked lists. The study adapted it to signed distance fields by computing transforms on the occupancy map and inverse occupancy map and subtracting them, but found signed mode much slower. At 0.05 m resolution the reported runtimes were $F$1 ms for Fiesta unsigned and $F$2 ms for Fiesta signed; GPU-Voxels remained fastest in all tested cases, and Fiesta had to be run single-threaded because it “did not operate successfully using multiple threads” (Finean et al., 2021).

4. Statistical learning, adaptation, and model comparison

In NLP, FIESTA stands for Fast IdEntification of State-of-The-Art models using adaptive bandit algorithms. The framework formalizes model selection as best-arm identification under both train-test split variability and random-seed variability, with the target

$F$3

and confidence objective

$F$4

It studies Sequential Halving for fixed budget, Top-Two Thompson Sampling for fixed confidence, and Batch Thompson Sampling for parallel fixed-confidence settings. On target-dependent sentiment analysis with eight candidate models, the paper reports that TTTS reduced mean evaluations from 281 to 130 at $F$5, from 206 to 96 at $F$6, and from 128 to 65 at $F$7, while Sequential Halving reached about 99% correct selection with 204 evaluations versus 85% for equal allocation (Moss et al., 2019).

In test-time adaptation, FIESTA has also been used for Fisher Information-based Efficient Selective Test-time Adaptation in video-based facial expression recognition. That method estimates diagonal empirical Fisher scores from only 1–3 sampled frames, ranks parameters by importance, freezes all non-selected weights, and adapts the selected subset using a temporal smoothing loss on median-filtered logits. On AffWild2, the best variant improved macro F1 from 0.325 to 0.350 while adapting only 22k parameters, compared with 456k for manual early-layer adaptation and about 91.8M for all-layer adaptation (Honarmand et al., 29 Mar 2025).

In medical image segmentation, a 2024 paper introduces FIESTA as Fourier-Based Semantic Augmentation with Uncertainty Guidance for Enhanced Domain Generalizability in Medical Image Segmentation. The abstract describes it as a Fourier-based semantic augmentation method for single-source domain generalization that manipulates amplitude and phase components in the frequency domain, uses a Fourier augmentative transformer for semantic amplitude modulation and structural coherence, and employs epistemic uncertainty to fine-tune augmentation for uncertain regions. The paper states that experiments across three cross-domain scenarios show superior performance over recent SDG approaches (Oh et al., 2024).

5. Biomedical imaging and tractography

In tractography, FIESTA means FIbEr Segmentation in Tractography using Autoencoders. The method learns a latent representation of streamlines with an undercomplete convolutional autoencoder and a contrastive loss based on QuickBundlesX clusters, then performs atlas-based latent nearest-neighbor bundle assignment and a generative bundle-population stage. The representation uses streamlines resampled to $F$8 vertices and a latent dimensionality of $F$9, with overall loss

$\epsilon$0

where $\epsilon$1. On the MyeloInferno test-retest benchmark over 17 common bundles, FIESTA reported voxel-wise Dice $\epsilon$2, weighted Dice $\epsilon$3, voxel-wise bundle adjacency $\epsilon$4 mm, and streamline density correlation $\epsilon$5, with strongest gains on small and hard-to-track bundles (Dumais et al., 2022).

The name also appears in MRI as a sequence designation rather than a method. A fetal body segmentation study compared TRUFI and FIESTA MRI sequences under low-annotation regimes. For the FIESTA sequence, the matched-effort comparison was 30 partially annotated training scans with 20% annotated slices each versus 6 fully annotated training scans. On out-of-distribution FIESTA data, the best partial-annotation configuration improved Dice from 0.836 to 0.899 and reduced 2D ASSD from 7.46 to 4.00 mm; on in-distribution data, mean performance was similar but standard deviations decreased, including a 27.5% reduction in Dice standard deviation and about 33% reduction in ASSD standard deviation (Fadida et al., 2022).

These two cases are terminologically distinct. In tractography, FIESTA is the algorithmic object; in fetal MRI, FIESTA is the acquisition sequence on which another segmentation method is evaluated. This distinction is methodologically important because identical naming does not imply shared mathematics or software.

6. Other scientific and engineering usages

Several additional usages reinforce the breadth of the term. In transient astrophysics, lowercase fiesta is a JAX-based Python package that trains surrogate models for afterglowpy, pyblastafterglow, and possis so that Bayesian inference on kilonovae and GRB afterglows becomes practical. The package reports full-lightcurve posterior evaluation “within minutes”; in one afterglowpy injection study, direct nmma inference took 19,700 s on 24 CPU cores, whereas fiesta inference took 96 s on an NVIDIA H100 GPU (Koehn et al., 18 Jul 2025). In tokamak equilibrium modeling, Fiesta is a MATLAB-based free-boundary static equilibrium solver used as a benchmark against FreeGSNKE and EFIT++ on MAST-U; the validation paper reports median Fiesta runtime of 6.7 s per forward solve and generally sub-centimetre agreement on shape targets such as X-points and strike points (Pentland et al., 2024).

In immersive analytics, FIESTA denotes the Free-roaming Immersive Environment to Support Team-based Analysis, a Unity/IATK-based VR prototype for co-located collaborative data analysis. A study with 10 groups of 3 participants showed that walls were frequently used to organize 2D visualizations, whereas 3D visualizations were typically suspended in open space; the virtual table was used far less than anticipated, and social ownership norms limited interaction with artifacts created by others even though the system imposed no hard ownership constraints (Lee et al., 2020). In quantum control, FIESTA is Floquet Interference Efficient Suppression of Transitions in the Adiabatic basis, a protocol for strongly driven qubits that suppresses nonadiabatic transitions between instantaneous Floquet states. For optimized pulses, the paper reports single-qubit gate fidelities of 0.99983 for $\epsilon$6 and 0.99986 for $\epsilon$7 (Deng et al., 2016). A related but orthographically distinct case is FiEstAS sampling, an adaptive importance-sampling Monte Carlo algorithm for multidimensional numerical integration with user parameters $\epsilon$8, $\epsilon$9, and $$\int_0^1 dx_1\cdots dx_N\, \prod_{j=1}^N x_j^{b_j+c_j\varepsilon}\, G(x,\varepsilon),$$0, designed especially for Bayesian evidence computation in multimodal and degenerate settings (0807.4479).

Taken together, these uses show that “Fiesta/FIESTA” has become a recurrent naming pattern for computational frameworks that emphasize automation, acceleration, or efficient structure exploitation. This suggests a semantic regularity at the level of branding, but not at the level of shared theory. Any technical interpretation of “Fiesta” therefore requires immediate disciplinary disambiguation by acronym expansion, problem class, and cited paper.