REVOLVER Method: Cross-Domain Techniques

• REVOLVER Method is a cross-domain label that encompasses diverse technical procedures—from continuous policy transfer in robotics to watershed reconstruction in cosmology and optimized data layouts in homomorphic computations.
• In robotics, REvolveR employs continuous interpolation with local randomized evolution and reward shaping to bridge morphology gaps, notably improving performance metrics on simulated tasks.
• Applications in cosmology and machine learning use Voronoi-watershed techniques and ciphertext scheduling to reconstruct cosmic structures and enable secure neural network operations, highlighting the method’s versatile implementations.

Searching arXiv for recent and foundational papers using the term “REVOLVER” or closely related variants to ground the article in the literature. REVOLVER Method designates several unrelated technical procedures in the arXiv literature rather than a single standardized formalism. In the cited corpus, the label appears as an acronym, a software-library name, a matrix-encoding strategy, a watershed-pruning rule, and a descriptor for physical configurations in optics, photonics, and inertial-confinement fusion. The best-known uses include REvolveR for robot-to-robot policy transfer (Liu et al., 2022), REVOLVER for Voronoi-watershed reconstruction of cosmic voids and clusters (Arsenov et al., 22 Sep 2025), Volley Revolver for homomorphic neural-network inference and training (Chiang, 2022), and REvolver for renormalization-group running and matching in QCD (Hoang et al., 2021). Across these usages, the shared name does not imply a shared mathematical core.

1. Nomenclature and recurrent meanings

The term appears in both uppercase and mixed-case forms. In cosmology, REVOLVER stands for REal-space VOid Locations from surVEy Reconstruction and denotes a ZOBOV-like watershed procedure operating on a Voronoi density field (Arsenov et al., 22 Sep 2025). In robotics, REvolveR denotes “continuous evolutionary models for robotic policy transfer,” with a curriculum over interpolated robot morphologies and dynamics (Liu et al., 2022). In privacy-preserving machine learning, Volley Revolver and Double Volley Revolver denote matrix-encoding and slot-rotation schemes for CKKS/HEAAN-based homomorphic evaluation of convolutions and matrix multiplications (Chiang, 2022). In perturbative QCD, REvolver is a C++ library for automated RG evolution, threshold matching, and mass-scheme conversion (Hoang et al., 2021).

Several further uses are descriptive rather than acronymic. A revolver silica fiber is a single-ring, negative-curvature, hollow-core fiber in the anti-resonant class (Kolyadin et al., 2018). The Revolver target concept in inertial-confinement fusion is a direct-drive, triple-shell architecture (Scheiner et al., 2019). An optical revolver is a dark trapping geometry produced by interfering counter-propagating Laguerre–Gaussian beams (Arnold, 2012). In graph analytics, Revolver is an asynchronous graph partitioner based on learning automata and label propagation (Mofrad et al., 2019). The name therefore functions as a cross-domain label rather than a discipline-specific term of art.

2. REvolveR in robot-to-robot policy transfer

In robotics, REvolveR addresses transfer between source and target robots that differ substantially in morphology and kinematics. The central failure mode identified in the paper is that action-distribution matching or state-alignment imitation learning becomes harmful when optimal action and state distributions are mismatched across robots, even if both systems are embedded in a common state and action space (Liu et al., 2022). The paper explicitly reports that State-Only Imitation Learning can perform worse than direct fine-tuning on Ant-leg-emerge because the source policy’s behavior for the “emerging” legs is essentially random and harmful when transferred to target legs that actually affect dynamics (Liu et al., 2022).

The method formalizes source and target robots as MDPs

$$M_S = (S, A, T_S, R, \gamma), \qquad M_T = (S, A, T_T, R, \gamma),$$

and introduces a continuous evolution

$E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$

After morphology matching, robot parameters share coordinates $\theta_S,\theta_T \in \mathbb{R}^P$, and the evolving robot is parameterized by

$$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$

The paper uses linear interpolation, $f(\alpha)=\alpha$, so

$$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$

Training proceeds through a sequence of evolution parameters $\alpha_0=0$, $\alpha_{k+1}=\alpha_k+l_k$, with policy optimization on successive intermediate robots rather than a direct jump from $M_S$ to $M_T$ (Liu et al., 2022).

Two stabilizing devices are central. Local randomized evolution samples $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$0 uniformly from a moving window $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$1, so the update objective averages over nearby intermediate robots. Evolution-aware reward shaping rescales rewards as

$E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$2

with shaping factor $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$3. Under the theorem stated in the paper, if the optimizer of the shaped objective coincides with the optimal policy for some $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$4 in the local window, then as $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$5,

$E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$6

so positive $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$7 shifts the effective target toward more evolved robots (Liu et al., 2022).

The experimental program spans MuJoCo Gym and ADROIT Hand Manipulation Suite tasks. On MuJoCo Gym, REvolveR consistently improves return relative to From Scratch, Direct Transfer, and SOIL; for example, on Humanoid-length-mass with SAC it reports $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$8 versus $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$9 for Direct Transfer and $\theta_S,\theta_T \in \mathbb{R}^P$0 for SOIL (Liu et al., 2022). On sparse-reward ADROIT tasks, all baselines fail to obtain reward, reported as effectively infinite iterations, whereas REvolveR reaches $\theta_S,\theta_T \in \mathbb{R}^P$1 success with finite iterations; Hammer, Relocate, and Door require $\theta_S,\theta_T \in \mathbb{R}^P$2K, $\theta_S,\theta_T \in \mathbb{R}^P$3K, and $\theta_S,\theta_T \in \mathbb{R}^P$4K RL optimization steps, respectively (Liu et al., 2022). The paper emphasizes that this advantage derives from turning exploration over a difficult target robot into a chain of reward-bearing intermediate tasks.

The approach is explicitly limited by the assumption that morphology matching can bring source and target robots into a shared topology through zero-sized bodies and joints, by the heuristic choice of $\theta_S,\theta_T \in \mathbb{R}^P$5 and $\theta_S,\theta_T \in \mathbb{R}^P$6, and by the absence of real-hardware transfer experiments (Liu et al., 2022). The interpolation is also linear in kinematic parameters, and the theoretical statement assumes differentiability and Lipschitz continuity that may fail near contact-rich transitions.

3. REVOLVER in cosmic-web reconstruction and void finding

In cosmology, REVOLVER is a watershed-based reconstruction built on ZOBOV-like Voronoi tessellation. In the Quaia quasar analysis, it is used to identify both underdense voids and overdense clusters directly from a Voronoi-based local density estimate in comoving space, with adjacency defined by the dual Delaunay triangulation (Arsenov et al., 22 Sep 2025). The analysis uses $\theta_S,\theta_T \in \mathbb{R}^P$7 QSOs at $\theta_S,\theta_T \in \mathbb{R}^P$8 over $\theta_S,\theta_T \in \mathbb{R}^P$9, adopts a flat $$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$0CDM cosmology with $$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$1, $$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$2, $$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$3, and reconstructs densities from corrected Voronoi volumes (Arsenov et al., 22 Sep 2025).

The density estimator is

$$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$4

REVOLVER applies the watershed transform separately to minima and maxima, producing void and cluster catalogues in parallel. Structure volume is the sum of corrected member volumes, and the effective radius is

$$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$5

For voids, the catalogue includes a circumcenter and barycenter; for clusters, the circumcenter only in that run. It also reports $$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$6, $$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$7, DensRatio, Theta_eff, EdgeFlag, and the environment proxies

$$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$8

(Arsenov et al., 22 Sep 2025).

The Quaia analysis identifies 12,842 voids and 41,111 clusters, with agreement between data and 50 mocks at the 5–10\% level for radii, average inner density, and density profiles (Arsenov et al., 22 Sep 2025). Voids have median $$\theta(\alpha) = (1-f(\alpha))\theta_S + f(\alpha)\theta_T.$$9 and clusters $f(\alpha)=\alpha$0; the largest structures reach $f(\alpha)=\alpha$1 for voids and $f(\alpha)=\alpha$2 for clusters, with no evidence for ultra-large structures beyond the mocks (Arsenov et al., 22 Sep 2025). VTFE stacked density profiles are measured in 25 radial bins out to $f(\alpha)=\alpha$3 using

$f(\alpha)=\alpha$4

A distinct but related usage appears in the VAST implementation discussed for SDSS DR7. There, REVOLVER is the single-zone pruning of a ZOBOV-style watershed finder: voids are taken to be individual zones that pass a minimum effective-radius threshold, without hierarchical merging (Douglass et al., 2022). In the updated SDSS DR7 catalog, REVOLVER identifies 518 voids for the Planck 2018 cosmology and 518 for WMAP5, with median effective radius $f(\alpha)=\alpha$5 and maximum $f(\alpha)=\alpha$6 for Planck (Douglass et al., 2022). The same study reports that REVOLVER voids occupy $f(\alpha)=\alpha$7 of the survey volume and contain $f(\alpha)=\alpha$8 of galaxies for Planck 2018, illustrating how single-zone watershed voids differ from VoidFinder’s explicitly spherical regions (Douglass et al., 2022).

This difference is not merely classificatory. In the AP-effect analysis of void–galaxy cross-correlations, REVOLVER void centers do not obey simple Alcock–Paczyński scaling because “AP distort the tracer field” and “find voids” do not commute (Radinović et al., 2024). The paper shows that naïvely rescaling void centers overpredicts anisotropy in the quadrupole; using the anisotropic real-space CCF measured from the same void and tracer catalog recovers the correct $f(\alpha)=\alpha$9 within $$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$0 for REVOLVER across the tested cosmologies (Radinović et al., 2024). In the environmental-effects thesis, V$$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$1 REVOLVER pruning yields much weaker void–wall contrasts than VoidFinder; for DESI DR1 BGS, it classifies 812,160 void galaxies and 69,378 wall galaxies, and the nonparametric Bayes factors for color and H$$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$2 equivalent width are much closer to zero than the corresponding VoidFinder results (Mu, 23 Jul 2025). A common misconception is therefore that “void” is a method-independent label. The cited studies show the opposite: the object class depends strongly on the pruning rule, center definition, and boundary treatment.

4. Volley Revolver and Double Volley Revolver in homomorphic deep learning

In privacy-preserving machine learning, Volley Revolver is a matrix-encoding strategy for CKKS/HEAAN-style encrypted inference and training. The basic idea is to encrypt one matrix in row-major form and the transpose of the other so that matrix multiplication can be realized by slotwise multiplication, rotations, and blockwise reductions (Chiang, 2022). For inference, the original paper states that “for two matrices $$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$3 and $$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$4 to perform homomorphic multiplication, the main idea behind it, in a simple version, is to encrypt matrix $$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$5 and the transpose of matrix $$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$6 into two ciphertexts respectively” (Chiang, 2022). In a canonical slot mapping,

$$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$7

and packed images are processed with rotations, plaintext masks, and accumulation operators such as SumColVec and SumForConv (Chiang, 2022).

For CNN inference on MNIST, the method spans each convolution kernel into image-sized masks and accumulates the resulting partial products. The reported implementation processes 32 encrypted images of size $$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$8 simultaneously, computing ten likelihoods in $$\theta(\alpha) = (1-\alpha)\theta_S + \alpha \theta_T.$$9 seconds on 40 vCPUs, while the data owner uploads one ciphertext ($\alpha_0=0$0 MB) (Chiang, 2022). The same paper reports 98.61\% accuracy under encryption versus 98.66\% in plaintext (Chiang, 2022). The computational bottlenecks are ciphertext rotations and reductions rather than communication volume.

Subsequent work extended the method from inference to training. In the transfer-learning paper, REGNET_X_400MF is used as a frozen feature extractor, reducing the homomorphic task to multiclass logistic regression over encrypted features (Chiang, 2023). The training pipeline combines Volley Revolver with a cubic sigmoid approximation,

$\alpha_0=0$1

a Squared Likelihood Error loss, and a Quadratic Gradient update with encrypted scaling matrix $\alpha_0=0$2 (Chiang, 2023). The client uploads only 6 ciphertexts—$\alpha_0=0$3 (2), $\alpha_0=0$4 (2), $\alpha_0=0$5 (1), and $\alpha_0=0$6 (1)—and 2 iterations take $\alpha_0=0$7 minutes on a cloud with 64 vCPUs, reaching 21.49\% precision on MNIST (Chiang, 2023). The paper’s significance claim is narrow and explicit: it presents the first working homomorphic CNN training pipeline in that transfer-learning setting.

The later “Two Hidden Layers” paper introduces Double Volley Revolver, which partitions operands into multiple ciphertext blocks and alternates over block pairs to improve the computation–memory trade-off (Chiang, 17 Apr 2025). It also replaces Softmax with Sigmoid plus BCE, arguing that SLE and the 2019 CVPR Workshop loss suffer vanishing gradients as depth increases (Chiang, 17 Apr 2025). The reported CKKS parameters are $\alpha_0=0$8, $\alpha_0=0$9, $\alpha_{k+1}=\alpha_k+l_k$0, and 32768 slots, without bootstrapping; the experiment again runs for 2 iterations, at $\alpha_{k+1}=\alpha_k+l_k$1 minutes per iteration, with $\alpha_{k+1}=\alpha_k+l_k$2 GB peak memory and 21.49\% precision on MNIST (Chiang, 17 Apr 2025).

Two 2025 developments generalize the idea further. Inference++ removes the requirement that a whole image fit in a single ciphertext, partitioning an image $\alpha_{k+1}=\alpha_k+l_k$3 into $\alpha_{k+1}=\alpha_k+l_k$4 ciphertexts, one per column, and preserving convolution structure across ciphertext boundaries (Chiang, 21 Dec 2025). For a single-channel image, the partition operator is

$\alpha_{k+1}=\alpha_k+l_k$5

and the single-channel valid convolution is evaluated columnwise through vertical rotations and horizontal ciphertext selection (Chiang, 21 Dec 2025). CryptoUNets applies Double Volley REVOLVER to U-Net inference, using square activations, mean pooling, and ConvTranspose2d with kernel size 2 and stride 2; the reported encrypted inference takes ~11 minutes per image with ~18 GB memory (Chiang, 30 Apr 2025).

A recurring misconception is that REVOLVER in this context is a cryptographic scheme. It is not. The underlying cryptosystem is CKKS/HEAAN; REVOLVER is a data layout and operation schedule for ciphertext slots (Chiang, 2022). Its importance lies in minimizing rotations, preserving spatial structure, and making matrix multiplication and convolution compatible with encrypted SIMD semantics.

5. Revolver in graph partitioning and REvolver in perturbative QCD

The graph-partitioning paper “Partitioning Graphs for the Cloud using Reinforcement Learning” uses Revolver for an asynchronous, single-node, shared-memory partitioner. Each vertex is controlled by a learning automaton, and normalized label-propagation scores provide the per-partition signal

$\alpha_{k+1}=\alpha_k+l_k$6

with locality measured by

$\alpha_{k+1}=\alpha_k+l_k$7

The system is evaluated on nine real-world graphs, for $\alpha_{k+1}=\alpha_k+l_k$8, with $\alpha_{k+1}=\alpha_k+l_k$9, $M_S$0, and $M_S$1 (Mofrad et al., 2019). The paper reports that Revolver can achieve up to 28× improvement in Max Normalized Load on EU relative to baselines while producing comparable localized partitions (Mofrad et al., 2019). The paper explicitly does not provide a single formal global optimization objective, nor a convergence proof; its guarantees are empirical (Mofrad et al., 2019).

REvolver in QCD is conceptually unrelated. It is a software library for high-order renormalization-group evolution, flavor-threshold matching, and quark-mass conversions (Hoang et al., 2021). The library implements the QCD coupling evolution

$M_S$2

the $M_S$3 mass anomalous-dimension equation

$M_S$4

and the linear-scale MSR mass evolution

$M_S$5

The implementation supports 5-loop running for $M_S$6 and $M_S$7 masses, 4-loop threshold matching, Mathematica and Python interfaces, and complex renormalization scales, with numerical solutions to machine precision (Hoang et al., 2021).

The QCD library also automates threshold matching and mass-scheme conversions among MSR, 1S, kinetic, PS, RS, and pole mass conventions (Hoang et al., 2021). Here the name marks a numerical infrastructure for precise perturbative calculations rather than a learning or reconstruction algorithm. A plausible implication is that the label’s recurrence across unrelated computational fields reflects mnemonic appeal rather than technical lineage.

6. Revolver configurations in photonics, fusion, and cold-atom optics

In photonics, the term refers to a revolver fiber rather than an algorithm. The fiber is a single-ring, negative-curvature, hollow-core geometry whose cladding is formed by thin-walled capillaries and whose guidance combines anti-resonant reflection with inhibited coupling (Kolyadin et al., 2018). The design study optimizes the fiber for Raman conversion from 1560 nm to 4420 nm in H$M_S$8, ultimately fabricating a silica structure with $M_S$9m, $M_T$0 capillaries, wall thickness $M_T$1m, and minimum gap $M_T$2m (Kolyadin et al., 2018). The measured losses are $M_T$3 dB/m at the pump and $M_T$4 dB/m at the Stokes wavelength, with modal glass overlap of $M_T$5, enabling 250 mW average output power at 4.42 $M_T$6m and 36\% quantum efficiency (Kolyadin et al., 2018). The method is design-driven: anti-resonant band placement, bend-aware core sizing, and suppression of surface-mode resonances determine performance.

In inertial-confinement fusion, Revolver is a direct-drive, triple-shell concept in which an outer ablator shell collides with a denser driver shell, which in turn compresses liquid DT through a high-$M_T$7 pusher (Scheiner et al., 2019). The OMEGA experiments reported in the cited paper isolate the outer two shells and measure the post-collision chromium-shell trajectory. The measured outer-surface velocity is $M_T$8 cm/$M_T$9s, compared with a HYDRA prediction of 7.27 cm/$E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$00s, and the modeled sub-scale energy transfer efficiency is $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$01 (Scheiner et al., 2019). The experiment deliberately omits the $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$02–$E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$03 mg/cc foam common in double-shell targets, uses a 7.19 g/cc Cr shell, and includes a 23 $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$04m parylene cushion layer to mediate collision pressure and reduce high-mode transfer (Scheiner et al., 2019). Here the name denotes a target architecture rather than a formal procedure.

In ultracold-atom optics, the optical revolver is a dark trapping geometry formed by interfering counter-propagating Laguerre–Gaussian beams of chosen azimuthal indices and powers (Arnold, 2012). The composite field creates an inner helical dark core together with ring-like sites and azimuthal modulation. The paper gives the approximate helical lead angle

$E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$05

and for the optical revolver example reports $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$06 mrad, axial and radial trap frequencies of $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$07 kHz and $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$08 kHz, scattering rate $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$09 Hz, and potential depth $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$10K for $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$11Rb at $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$12 nm (Arnold, 2012). A phase grating near the trap is proposed as the stabilizing element. The paper’s physical rationale is that atoms localized at dark minima experience minimal scattering and near-zero differential light shifts.

These physical uses are not “methods” in the algorithmic sense, but they preserve a common structural idea: a revolver as a multiply connected, rotationally organized configuration. That interpretation remains descriptive rather than formal; the cited papers do not claim a shared abstraction across fiber design, shell collisions, and dark optical traps.

7. Conceptual commonalities and domain-specific divergence

Across the cited literature, REVOLVER denotes one of four broad types of object. First, it can be a curriculum or interpolation strategy, as in robot policy transfer through a continuous path in morphology space (Liu et al., 2022). Second, it can be a topological reconstruction rule, as in Voronoi–watershed identification of voids and clusters, including single-zone pruning (Arsenov et al., 22 Sep 2025). Third, it can be a slot-layout and rotation schedule for encrypted linear algebra (Chiang, 2022). Fourth, it can be a physical design configuration, as in revolver fibers, shell architectures, and optical traps (Kolyadin et al., 2018).

The strongest methodological controversies arise where the name labels an object class rather than a fixed equation set. In cosmic-web studies, the choice between REVOLVER, VIDE, and VoidFinder changes void counts, effective radii, galaxy fractions, and downstream environmental inferences (Douglass et al., 2022). In AP analyses, incorrect treatment of the void-finder response biases $E : [0,1] \to M, \qquad E(0)=M_S,\; E(1)=M_T.$13 unless one propagates the anisotropic real-space CCF measured from the same catalog (Radinović et al., 2024). In robotics, the controversy is over transfer paradigms: REvolveR rejects action/state distribution matching under large dynamics mismatch and instead changes the domain itself smoothly (Liu et al., 2022). In homomorphic learning, the principal constraint is not statistical validity but multiplicative depth, rotation count, and slot capacity; REVOLVER is useful only insofar as those resource bottlenecks dominate (Chiang, 2022).

The term therefore has encyclopedic value primarily as a name family used for structurally different technical constructs. Any precise use requires the surrounding field, capitalization, and citation. Without that context, “REVOLVER method” is ambiguous between at least a continuous-evolution RL transfer procedure (Liu et al., 2022), a Voronoi-watershed void/cluster reconstructor (Arsenov et al., 22 Sep 2025), a homomorphic matrix-encoding framework (Chiang, 2022), an asynchronous graph partitioner (Mofrad et al., 2019), a QCD running-and-matching library (Hoang et al., 2021), and several physical designs in fiber optics, fusion, and cold-atom trapping (Kolyadin et al., 2018).