Search

The problem of obtaining spectral densities from lattice data has been receiving great attention due to its importance in our understanding of scattering processes in Quantum Field Theory, with applications both in the Standard Model and beyond. The problem is notoriously difficult as it amounts to performing an inverse Laplace transform, starting from a finite set of noisy data. Several strategies are now available to tackle this inverse problem. In this work, we discuss how Backus-Gilbert methods, in particular the variation introduced by some of the authors, relate to the solution based on Gaussian Processes. Both methods allow computing spectral densities smearing with a kernel, whose features depend on the detail of the algorithm. We will discuss such kernel, and show how Backus-Gilbert methods can be understood in a Bayesian fashion. As a consequence of this correspondence, we are able to interpret the algorithmic parameters of Backus-Gilbert methods as hyperparameters in the Bayesian language, which can be chosen by maximising a likelihood function. By performing a comparative study on lattice data, we show that, when both frameworks are set to compute the same quantity, the results are generally in agreement. Finally, we adopt a strategy to systematically validate both methodologies against pseudo-data, using covariance matrices measured from lattice simulations. In our setup, we find that the determination of the algorithmic parameters based on a stability analysis provides results that are, on average, more conservative than those based on the maximisation of a likelihood function.

We develop an optimized hardware strategy to mitigate errors in a noisy qubit. Our scheme builds on the physical principle of error filtration and exploits auxiliary qubits. Both signal and ancillas are subject to local noise, yet constructive interference (and in some cases post-selection) allows us to reduce the level of noise in the signal qubit. Seeking for the optimal unitary that makes the qubits interfere in the most effective way, we start with a set of universal gates and proceed by optimizing suitable functionals by gradient descent or stochastic approximation. We benchmark our approach against a number of figure of merits that correspond to different applications, including entanglement fidelity, quantum Fisher information (for applications in quantum sensing), and CHSH value (for applications in tests of non-locality and quantum cryptography), with one, two, and three ancillary qubits. With one and two ancillas we also provide explicit expressions from an ansatz for an optimal unitary., Comment: Comments are wellcome

We demonstrate spectroscopy of incoherent light with sub-diffraction resolution. In a proof-of-principle experiment we analyze the spectrum of a pair of incoherent point-like sources whose separation is below the diffraction limit. The two sources mimic a planetary system, with a brighter source for the star and a dimmer one for the planet. Acquiring spectral information about the secondary source is hard because the two images have a substantial overlap. This limitation is solved by leveraging a structured measurement based on spatial-mode demultiplexing, where light is first sorted in its Hermite-Gaussian components in the transverse field, then measured by photon detection. This allows us to effectively decouple the photons coming from the two sources. An application is suggested to enhance exoplanets' atmosphere spectroscopy. A number of experiments of super-resolution imaging based on spatial demultiplexing have been conducted in the past few years, with promising results. Here, for the first time to the best of our knowledge, we extend this concept to the domain of spectroscopy., Comment: 11 pages, 6 figures, comments welcome

We present a new lattice QCD calculation of the leading order hadronic vacuum polarization contribution to the muon anomalous magnetic moment $a_\mu$. We reduce uncertainties compared to our earlier computation by $40\%$, arXiv:2002.12347. We perform simulations on finer lattices allowing for an even more accurate continuum extrapolation. We also include a small, long-distance contribution obtained using input from experiments in a low-energy regime where they all agree. Combined with other standard model contributions our result leads to a prediction that differs from the measurement of $a_\mu$ by only 0.9 standard deviations. This provides a remarkable validation of the standard model to 0.37ppm., Comment: 55 pages, 31 figures

We present our work on developing and training scalable graph foundation models (GFM) using HydraGNN, a multi-headed graph convolutional neural network architecture. HydraGNN expands the boundaries of graph neural network (GNN) in both training scale and data diversity. It abstracts over message passing algorithms, allowing both reproduction of and comparison across algorithmic innovations that define convolution in GNNs. This work discusses a series of optimizations that have allowed scaling up the GFM training to tens of thousands of GPUs on datasets that consist of hundreds of millions of graphs. Our GFMs use multi-task learning (MTL) to simultaneously learn graph-level and node-level properties of atomistic structures, such as the total energy and atomic forces. Using over 150 million atomistic structures for training, we illustrate the performance of our approach along with the lessons learned on two United States Department of Energy (US-DOE) supercomputers, namely the Perlmutter petascale system at the National Energy Research Scientific Computing Center and the Frontier exascale system at Oak Ridge National Laboratory. The HydraGNN architecture enables the GFM to achieve near-linear strong scaling performance using more than 2,000 GPUs on Perlmutter and 16,000 GPUs on Frontier. Hyperparameter optimization (HPO) was performed on over 64,000 GPUs on Frontier to select GFM architectures with high accuracy. Early stopping was applied on each GFM architecture for energy awareness in performing such an extreme-scale task. The training of an ensemble of highest-ranked GFM architectures continued until convergence to establish uncertainty quantification (UQ) capabilities with ensemble learning. Our contribution opens the door for rapidly developing, training, and deploying GFMs using large-scale computational resources to enable AI-accelerated materials discovery and design., Comment: 16 pages, 13 figures

The resolution of optical imaging is classically limited by the width of the point-spread function, which in turn is determined by the Rayleigh length. Recently, spatial-mode demultiplexing (SPADE) has been proposed as a method to achieve sub-Rayleigh estimation and discrimination of natural, incoherent sources. Here we show that SPADE is optimal in the broader context of machine learning. To this goal, we introduce a hybrid quantum-classical image classifier that achieves sub-Rayleigh resolution. The algorithm includes a quantum and a classical part. In the quantum part, a physical device (demultiplexer) is used to sort the transverse field, followed by mode-wise photon detection. This part of the algorithm implements a physical pre-processing of the quantum field that cannot be classically simulated without essentially reducing the signal-to-noise ratio. In the classical part of the algorithm, the collected data are fed into an artificial neural network for training and classification. As a case study, we classify images from the MNIST dataset after severe blurring due to diffraction. Our numerical experiments demonstrate the ability to learn highly blurred images that would be otherwise indistinguishable by direct imaging without the physical pre-processing of the quantum field., Comment: 15 pages, 9 captioned figures, comments welcome

Spectral densities encode non-perturbative information that enters the calculation of a plethora of physical observables in strongly coupled field theories. Phenomenological applications encompass aspects of standard-model hadronic physics, observable at current colliders, as well as correlation functions characterizing new physics proposals, testable in future experiments. By making use of numerical data produced with lattice gauge theories, we perform a systematic study to demonstrate the effectiveness of recent technological progress in the reconstruction of spectral densities. To this purpose, we write and test new software packages that use energy-smeared spectral densities to analyze the mass spectrum of mesons. We assess the effectiveness of different smearing kernels and optimize the smearing parameters to the characteristics of available lattice ensembles. For concreteness, we analyze the Sp(4) lattice gauge theory with matter transforming in an admixture of fundamental and 2-index antisymmetric representations of the gauge group. We generate new ensembles for this theory, with lattices that have a longer extent in the time direction with respect to the spatial ones. We run our tests on these ensembles, obtaining new results about the spectrum of light mesons and their excitations. We make available our algorithm and software for the extraction of spectral densities, that can be applied to theories with other gauge groups, including the theory of strong interactions (QCD) governing hadronic physics in the standard model., Comment: 44 pages, 20 figures

Recently, acoustic tweezers based on an array of ultrasonic transducers have been reported taking inspiration from holographic optical tweezers. In the latter technique, the calibration of the optical trap is an essential procedure to obtain the trap stiffnesses. On the contrary, in the case of acoustic tweezers the calibration of the acoustic forces is seldom carried out. To cover this gap, in this work, we adapt the calibration protocols employed in optical tweezers to acoustic tweezers based on arrays of ultrasonic transducers. We measure trap stiffnesses in the mN/m range that are consistent with theoretical estimates obtained by calculations of the acoustic radiation forces based on the Gorkov potential. This work gives a common framework to the optical and acoustic manipulation communities, paving the way to a consistent calibration of hybrid acoustooptical setups., Comment: 8 pages, 6 figures

Social scientists increasingly use demographically stratified social media data to study the attitudes, beliefs, and behavior of the general public. To facilitate such analyses, we construct, validate, and release publicly the representative DADIT dataset of 30M tweets of 20k Italian Twitter users, along with their bios and profile pictures. We enrich the user data with high-quality labels for gender, age, and location. DADIT enables us to train and compare the performance of various state-of-the-art models for the prediction of the gender and age of social media users. In particular, we investigate if tweets contain valuable information for the task, since popular classifiers like M3 don't leverage them. Our best XLM-based classifier improves upon the commonly used competitor M3 by up to 53% F1. Especially for age prediction, classifiers profit from including tweets as features. We also confirm these findings on a German test set., Comment: Accepted to LREC-COLING 2024

Quantum key distribution (QKD) promises everlasting security based on the laws of physics. Most common protocols are grouped into two distinct categories based on the degrees of freedom used to carry information, which can be either discrete or continuous, each presenting unique advantages in either performance, feasibility for near-term implementation, and compatibility with existing telecommunications architectures. Recently, hybrid QKD protocols have been introduced to leverage advantages from both categories. In this work we provide a rigorous security proof for a protocol introduced by Qi in 2021, where information is encoded in discrete variables as in the widespread Bennett Brassard 1984 (BB84) protocol but decoded continuously via heterodyne detection. Security proofs for hybrid protocols inherit the same challenges associated with continuous-variable protocols due to unbounded dimensions. Here we successfully address these challenges by exploiting symmetry. Our approach enables truncation of the Hilbert space with precise control of the approximation errors and lead to a tight, semi-analytical expression for the asymptotic key rate under collective attacks. As concrete examples, we apply our theory to compute the key rates under passive attacks, linear loss, and Gaussian noise., Comment: 31 pages including appendices, 6 figures. Comments welcome

We provide a theoretical, numerical, and experimental investigation of the Kerr nonlinearity impact on the performance of a frequency-multiplexed Extreme Learning Machine (ELM). In such ELM, the neuron signals are encoded in the lines of a frequency comb. The Kerr nonlinearity facilitates the randomized neuron connections allowing for efficient information mixing. A programmable spectral filter applies the output weights. The system operates in a continuous-wave regime. Even at low input peak powers, the resulting weak Kerr nonlinearity is sufficient to significantly boost the performance on several tasks. This boost already arises when one uses only the very small Kerr nonlinearity present in a 20-meter long erbium-doped fiber amplifier. In contrast, a subsequent propagation in 540 meters of a single-mode fiber improves the performance only slightly, whereas additional information mixing with a phase modulator does not result in a further improvement at all. We introduce a model to show that, in frequency-multiplexed ELMs, the Kerr nonlinearity mixes information via four-wave mixing, rather than via self- or cross-phase modulation. At low powers, this effect is quartic in the comb-line amplitudes. Numerical simulations validate our experimental results and interpretation.

Speech recognition is a critical task in the field of artificial intelligence and has witnessed remarkable advancements thanks to large and complex neural networks, whose training process typically requires massive amounts of labeled data and computationally intensive operations. An alternative paradigm, reservoir computing, is energy efficient and is well adapted to implementation in physical substrates, but exhibits limitations in performance when compared to more resource-intensive machine learning algorithms. In this work we address this challenge by investigating different architectures of interconnected reservoirs, all falling under the umbrella of deep reservoir computing. We propose a photonic-based deep reservoir computer and evaluate its effectiveness on different speech recognition tasks. We show specific design choices that aim to simplify the practical implementation of a reservoir computer while simultaneously achieving high-speed processing of high-dimensional audio signals. Overall, with the present work we hope to help the advancement of low-power and high-performance neuromorphic hardware.

Recent advances in large language models (LLMs) like GPT-3.5 and GPT-4 promise automation with better results and less programming, opening up new opportunities for text analysis in political science. In this study, we evaluate LLMs on three original coding tasks involving typical complexities encountered in political science settings: a non-English language, legal and political jargon, and complex labels based on abstract constructs. Along the paper, we propose a practical workflow to optimize the choice of the model and the prompt. We find that the best prompting strategy consists of providing the LLMs with a detailed codebook, as the one provided to human coders. In this setting, an LLM can be as good as or possibly better than a human annotator while being much faster, considerably cheaper, and much easier to scale to large amounts of text. We also provide a comparison of GPT and popular open-source LLMs, discussing the trade-offs in the model's choice. Our software allows LLMs to be easily used as annotators and is publicly available: https://github.com/lorelupo/pappa.

In the upcoming decade, deep learning may revolutionize the natural sciences, enhancing our capacity to model and predict natural occurrences. This could herald a new era of scientific exploration, bringing significant advancements across sectors from drug development to renewable energy. To answer this call, we present DeepSpeed4Science initiative (deepspeed4science.ai) which aims to build unique capabilities through AI system technology innovations to help domain experts to unlock today's biggest science mysteries. By leveraging DeepSpeed's current technology pillars (training, inference and compression) as base technology enablers, DeepSpeed4Science will create a new set of AI system technologies tailored for accelerating scientific discoveries by addressing their unique complexity beyond the common technical approaches used for accelerating generic large language models (LLMs). In this paper, we showcase the early progress we made with DeepSpeed4Science in addressing two of the critical system challenges in structural biology research.

Heavy-tailed distributions naturally arise in several settings, from finance to telecommunications. While regret minimization under subgaussian or bounded rewards has been widely studied, learning with heavy-tailed distributions only gained popularity over the last decade. In this paper, we consider the setting in which the reward distributions have finite absolute raw moments of maximum order $1+\epsilon$, uniformly bounded by a constant $u<+\infty$, for some $\epsilon \in (0,1]$. In this setting, we study the regret minimization problem when $\epsilon$ and $u$ are unknown to the learner and it has to adapt. First, we show that adaptation comes at a cost and derive two negative results proving that the same regret guarantees of the non-adaptive case cannot be achieved with no further assumptions. Then, we devise and analyze a fully data-driven trimmed mean estimator and propose a novel adaptive regret minimization algorithm, AdaR-UCB, that leverages such an estimator. Finally, we show that AdaR-UCB is the first algorithm that, under a known distributional assumption, enjoys regret guarantees nearly matching those of the non-adaptive heavy-tailed case.

Quantum random number generators (QRNGs) based on quadrature measurements of the vacuum have so far used balanced homodyne detection to obtain a source of high entropy. Here we propose a simple direct detection measurement scheme using only a laser and a photodiode that still extracts randomness from vacuum fluctuations. We prove the security of the QRNG based on a reduced set of assumptions in comparison to previous security proofs for quadrature detection as our proof does not require the laser or electronic noise to be Gaussian. Using a low-cost setup based on a vertical-cavity surface-emitting laser we experimentally implement the QRNG scheme. We propose a system characterization method, apply it to our implementation and demonstrate a real-time randomness extraction rate of 3.41 Gbit per second. The unique combination of speed, low cost, and rigorous security proof gives our QRNG design a large potential for a wide-scale usage in a variety of applications ranging from quantum key distribution to mobile applications and internet of things., Comment: 12 pages, 5 figures

Four-dimensional gauge theories based on symplectic Lie groups provide elegant realisations of the microscopic origin of several new physics models. Numerical studies pursued on the lattice provide quantitative information necessary for phenomenological applications. To this purpose, we implemented Sp(2N) gauge theories using Monte Carlo techniques within Grid, a performant framework designed for the numerical study of quantum field theories on the lattice. We show the first results obtained using this library, focusing on the case-study provided by the Sp(4) theory coupled to Nas = 4 Wilson-Dirac fermions transforming in the 2-index antisymmetric representation. In particular, we discuss preliminary tests of the algorithm and we test some of its main functionalities., Comment: 9 pages, 7 figures, contribution for the 40th International Symposium on Lattice Field Theory (Lattice 2023), July 31st - August 4th, 2023, Fermi National Accelerator Laboratory, presenting the results of the paper: arXiv:2306.11649

Dephasing is a main noise mechanism that afflicts quantum information, it reduces visibility, and destroys coherence and entanglement. Therefore, it must be reduced, mitigated, and if possible corrected, to allow for the demonstration of quantum advantage in any application of quantum technology, from computing to sensing and communications. Here we discuss a hardware scheme of error filtration to mitigate the effects of dephasing in optical quantum metrology. The scheme uses only passive linear optics and ancillary vacuum modes, and we do not need single-photon sources or entanglement. It exploits constructive and destructive interference to partially cancel the detrimental effects of statistically independent sources of dephasing. We apply this scheme to preserve coherent states and to phase stabilize stellar interferometry, and show that a significant improvement can be obtained by using only a few ancillary modes., Comment: 12 pages, 4 figures, comments are welcome

In recent years, excitation of surface phonon polaritons (SPhPs) in van der Waals materials received wide attention from the nanophotonics community. Alpha-phase Molybdenum trioxide ($\alpha$-MoO3), a naturally occurring biaxial hyperbolic crystal, emerged as a promising polaritonic material due to its ability to support SPhPs for three orthogonal directions at different wavelength bands (range 10-20 $\mu$m). Here, we report on the fabrication and IR characterization of large-area (over 1 cm$^2$ size) $\alpha$-MoO3 polycrystalline films deposited on fused silica substrates by pulsed laser deposition. Single alpha-phase MoO3 films exhibiting a polarization-dependent reflection peak at 1006 cm$^{-1}$ with a resonance Q-factor as high as 53 were achieved. Reflection can be tuned via changing incident polarization with a dynamic range of $\Delta$R=0.3 at 45 deg. incidence angle. We also report a polarization-independent almost perfect absorption condition (R<0.01) at 972 cm$^{-1}$ which is preserved for a broad angle of incidence. The development of a low-cost polaritonic platform with high-Q resonances in the mid-infrared (mid-IR) range is crucial for a wide number of functionalities including sensors, filters, thermal emitters, and label-free biochemical sensing devices. In this framework our findings appear extremely promising for the further development of lithography-free, scalable films, for efficient and large-scale devices operating in the free space, using far-field detection setups., Comment: 17 pages, 12 figures

Interferometric methods have been recently investigated to achieve sub-Rayleigh imaging and precision measurements of faint incoherent sources up to the ultimate quantum limit. Here we consider single-photon imaging of two point-like emitters of unequal intensity. This is motivated by the fact that pairs of natural emitters will typically have unequal brightness, as for example binary star systems and exoplanets. We address the problems of estimating the transverse separation $d$ and the relative intensity $\epsilon$. Our theoretical analysis shows that the associated statistical errors are qualitatively different from the case of equal intensity. We employ multi-plane light conversion technology to experimentally implement Hermite-Gaussian (HG) spatial-mode demultiplexing (SPADE), and demonstrate sub-Rayleigh measurement of two emitters with Gaussian point-spread function. The experimental errors are comparable with the theoretical bounds. The latter are benchmarked against direct imaging, yielding a $\epsilon^{-1/2}$ improvement in the signal-to-noise ratio, which may be significant when the primary source is much brighter than the secondary one, as for example for imaging of exoplanets. However, achieving this improved scaling requires low noise in the implementation of SPADE, which is typically affected by crosstalk between HG modes., Comment: 9 pages, 8 figures, comments welcome

We frame entanglement detection as a problem of random variable inference to introduce a quantitative method to measure and understand whether entanglement witnesses lead to an efficient procedure for that task. Hence we quantify how many bits of information a family of entanglement witnesses can infer about the entanglement of a given quantum state sample. The bits are computed in terms of the mutual information and we unveil there exists hidden information not \emph{efficiently} processed. We show that there is more information in the expected value of the entanglement witnesses, i.e. $\mathbb{E}[W]=\langle W \rangle_\rho$ than in the sign of $\mathbb{E}[W]$. This suggests that an entanglement witness can provide more information about the entanglement if for our decision boundary we compute a different functional of its expectation value, rather than $\mathrm{sign}\left(\mathbb{E}\right [ W ])$.

Predicting which proteins interact together from amino-acid sequences is an important task. We develop a method to pair interacting protein sequences which leverages the power of protein language models trained on multiple sequence alignments, such as MSA Transformer and the EvoFormer module of AlphaFold. We formulate the problem of pairing interacting partners among the paralogs of two protein families in a differentiable way. We introduce a method called DiffPALM that solves it by exploiting the ability of MSA Transformer to fill in masked amino acids in multiple sequence alignments using the surrounding context. MSA Transformer encodes coevolution between functionally or structurally coupled amino acids. We show that it captures inter-chain coevolution, while it was trained on single-chain data, which means that it can be used out-of-distribution. Relying on MSA Transformer without fine-tuning, DiffPALM outperforms existing coevolution-based pairing methods on difficult benchmarks of shallow multiple sequence alignments extracted from ubiquitous prokaryotic protein datasets. It also outperforms an alternative method based on a state-of-the-art protein language model trained on single sequences. Paired alignments of interacting protein sequences are a crucial ingredient of supervised deep learning methods to predict the three-dimensional structure of protein complexes. DiffPALM substantially improves the structure prediction of some eukaryotic protein complexes by AlphaFold-Multimer, without significantly deteriorating any of those we tested. It also achieves competitive performance with using orthology-based pairing., Comment: 33 pages, 14 figures, 2 tables