laddu-cpu 0.19.0

Python bindings for laddu

Amplitude analysis made short and sweet

laddu (/ˈlʌduː/) is a library for analysis of particle physics data. It is intended to be a simple and efficient alternative to some of the other tools out there. laddu is written in Rust with bindings to Python via PyO3 and maturin and is the spiritual successor to rustitude, one of my first Rust projects. The goal of this project is to allow users to perform complex amplitude analyses (like partial-wave analyses) without complex code or configuration files.

[!CAUTION] This crate is still in an early development phase, and the API is not stable. It can (and likely will) be subject to breaking changes before the 1.0.0 version release (and hopefully not many after that).

Table of Contents

• Key Features
• Installation
• Quick Start
  • Rust
    • Writing a New Amplitude
    • Calculating a Likelihood
  • Python
    • Fitting Data
    • Other Examples
• Data Format
• MPI Support
• Future Plans
• Alternatives

Key Features

• A simple interface focused on combining Amplitudes into models which can be evaluated over Datasets.
• A single Amplitude trait which makes it easy to write new amplitudes and integrate them into the library.
• Easy interfaces to precompute and cache values before the main calculation to speed up model evaluations.
• Efficient parallelism using rayon.
• Python bindings to allow users to write quick, easy-to-read code that just works.

Installation

laddu can be added to a Rust project with cargo:

cargo add laddu

The library's Python bindings are located in a library by the same name, which can be installed simply with your favorite Python package manager:

pip install laddu

Quick Start

Rust

Writing a New Amplitude

While it is probably easier for most users to skip to the Python section, there is currently no way to implement a new amplitude directly from Python. At the time of writing, Rust is not a common language used by particle physics, but this tutorial should hopefully convince the reader that they don't have to know the intricacies of Rust to write performant amplitudes. As an example, here is how one might write a Breit-Wigner, parameterized as follows:

I_{\ell}(m; m_0, \Gamma_0, m_1, m_2) =  \frac{1}{\pi}\frac{m_0 \Gamma_0 B_{\ell}(m, m_1, m_2)}{(m_0^2 - m^2) - \imath m_0 \Gamma}

where

\Gamma = \Gamma_0 \frac{m_0}{m} \frac{q(m, m_1, m_2)}{q(m_0, m_1, m_2)} \left(\frac{B_{\ell}(m, m_1, m_2)}{B_{\ell}(m_0, m_1, m_2)}\right)^2

is the relativistic width correction, $q(m_a, m_b, m_c)$ is the breakup momentum of a particle with mass $m_a$ decaying into two particles with masses $m_b$ and $m_c$, $B_{\ell}(m_a, m_b, m_c)$ is the Blatt-Weisskopf barrier factor for the same decay assuming particle $a$ has angular momentum $\ell$, $m_0$ is the mass of the resonance, $\Gamma_0$ is the nominal width of the resonance, $m_1$ and $m_2$ are the masses of the decay products, and $m$ is the "input" mass.

Although this particular amplitude is already included in laddu, let's assume it isn't and imagine how we would write it from scratch:

use laddu::{
   AmplitudeID, BarrierKind, Cache, DatasetMetadata, Event, Expression, LadduResult, Mass,
   ParameterID, Parameter, Parameters, PI, QR_DEFAULT, Resources, ScalarID, Sheet,
};
use laddu::expression::{IntoTags, Tags};
use laddu::traits::*;
use laddu::math::{blatt_weisskopf_m, q_m};
use laddu::{Deserialize, Serialize, typetag};
use num::complex::Complex64;

#[derive(Clone, Serialize, Deserialize)]
pub struct MyBreitWigner {
    tags: Tags,
    mass: Parameter,
    width: Parameter,
    pid_mass: ParameterID,
    pid_width: ParameterID,
    l: usize,
    daughter_1_mass: Mass,
    daughter_2_mass: Mass,
    resonance_mass: Mass,
    daughter_1_mass_id: ScalarID,
    daughter_2_mass_id: ScalarID,
    resonance_mass_id: ScalarID,
}
impl MyBreitWigner {
    pub fn new(
        tags: impl IntoTags,
        mass: Parameter,
        width: Parameter,
        l: usize,
        daughter_1_mass: &Mass,
        daughter_2_mass: &Mass,
        resonance_mass: &Mass,
    ) -> LadduResult<Expression> {
        Self {
            tags: tags.into_tags(),
            mass,
            width,
            pid_mass: ParameterID::default(),
            pid_width: ParameterID::default(),
            l,
            daughter_1_mass: daughter_1_mass.clone(),
            daughter_2_mass: daughter_2_mass.clone(),
            resonance_mass: resonance_mass.clone(),
            daughter_1_mass_id: ScalarID::default(),
            daughter_2_mass_id: ScalarID::default(),
            resonance_mass_id: ScalarID::default(),
        }
        .into_expression()
    }
}

#[typetag::serde]
impl Amplitude for MyBreitWigner {
    fn register(&mut self, resources: &mut Resources) -> LadduResult<AmplitudeID> {
        self.pid_mass = resources.register_parameter(&self.mass)?;
        self.pid_width = resources.register_parameter(&self.width)?;
        self.daughter_1_mass_id = resources.register_scalar(None);
        self.daughter_2_mass_id = resources.register_scalar(None);
        self.resonance_mass_id = resources.register_scalar(None);
        resources.register_amplitude(self.tags.clone())
    }

    fn bind(
        &mut self,
        metadata: &DatasetMetadata,
    ) -> LadduResult<()> {
        self.daughter_1_mass.bind(metadata)?;
        self.daughter_2_mass.bind(metadata)?;
        self.resonance_mass.bind(metadata)?;
        Ok(())
    }

    fn precompute(&self, event: &Event<'_>, cache: &mut Cache) {
        cache.store_scalar(self.daughter_1_mass_id, event.evaluate(&self.daughter_1_mass));
        cache.store_scalar(self.daughter_2_mass_id, event.evaluate(&self.daughter_2_mass));
        cache.store_scalar(self.resonance_mass_id, event.evaluate(&self.resonance_mass));
    }

    fn compute(&self, parameters: &Parameters, cache: &Cache) -> Complex64 {
        let mass = cache.get_scalar(self.resonance_mass_id);
        let mass0 = parameters.get(self.pid_mass).abs();
        let width0 = parameters.get(self.pid_width).abs();
        let mass1 = cache.get_scalar(self.daughter_1_mass_id);
        let mass2 = cache.get_scalar(self.daughter_2_mass_id);
        let q0 = q_m(mass0, mass1, mass2, Sheet::Physical);
        let q = q_m(mass, mass1, mass2, Sheet::Physical);
        let f0 = blatt_weisskopf_m(mass0, mass1, mass2, self.l, QR_DEFAULT, Sheet::Physical, BarrierKind::Full);
        let f = blatt_weisskopf_m(mass, mass1, mass2, self.l, QR_DEFAULT, Sheet::Physical, BarrierKind::Full);
        let width = width0 * (mass0 / mass) * (q / q0) * (f / f0).powi(2);
        let n = (mass0 * width0 / PI).sqrt();
        let d = Complex64::new(mass0.powi(2) - mass.powi(2), -(mass0 * width));
        Complex64::from(f * n) / d
    }
}

Calculating a Likelihood

We could then write some code to use this amplitude. For demonstration purposes, let's just calculate an extended unbinned negative log-likelihood, assuming we have some data and Monte Carlo in the proper parquet format:

use laddu::{io, Scalar, Dataset, DatasetReadOptions, Mass, NLL, parameter};
let p4_names = ["beam", "proton", "kshort1", "kshort2"];
let aux_names = ["pol_magnitude", "pol_angle"];
let options = DatasetReadOptions::default()
    .p4_names(p4_names)
    .aux_names(aux_names)
    .alias("resonance", ["kshort1", "kshort2"]);
let ds_data = io::read_parquet("test_data/data.parquet", &options).unwrap();
let ds_mc = io::read_parquet("test_data/mc.parquet", &options).unwrap();

let resonance_mass = Mass::new(["kshort1", "kshort2"]);
let p1_mass = Mass::new(["kshort1"]);
let p2_mass = Mass::new(["kshort2"]);
let bw = MyBreitWigner::new(
    "bw",
    parameter("mass"),
    parameter("width"),
    2,
    &p1_mass,
    &p2_mass,
    &resonance_mass,
).unwrap();
let mag = Scalar::new("mag", parameter("magnitude")).unwrap();
let expr = (mag * bw).norm_sqr();

let nll = NLL::new(&expr, &ds_data, &ds_mc).unwrap();
println!("Parameters names and order: {:?}", nll.parameters());
let result = nll.evaluate(&[1.27, 0.120, 100.0]);
println!("The extended negative log-likelihood is {}", result);

In practice, amplitudes can also be added together, their real and imaginary parts can be taken, and evaluators should mostly take the real part of whatever complex value comes out of the model.

Python

Fitting Data

While we cannot (yet) implement new amplitudes within the Python interface alone, it does contain all the functionality required to analyze data. Here's an example to show some of the syntax. This models includes three partial waves described by the $Z_{\ell}^m$ amplitude listed in Equation (D13) here[^1]. Since we take the squared norm of each individual sum, they are invariant up to a total phase, thus the S-wave was arbitrarily picked to be purely real.

import laddu as ld
import matplotlib.pyplot as plt
import numpy as np
from laddu import parameter

def main():
    p4_columns = ['beam', 'proton', 'kshort1', 'kshort2']
    aux_columns = ['pol_magnitude', 'pol_angle']
    ds_data = ld.io.read_parquet('path/to/data.parquet', p4s=p4_columns, aux=aux_columns)
    ds_mc = ld.io.read_parquet('path/to/accmc.parquet', p4s=p4_columns, aux=aux_columns)
    beam = ld.Particle.stored('beam')
    target = ld.Particle.missing('target')
    kshort1 = ld.Particle.stored('kshort1')
    kshort2 = ld.Particle.stored('kshort2')
    kk = ld.Particle.composite('kk', (kshort1, kshort2))
    recoil = ld.Particle.stored('proton')
    reaction = ld.Reaction.two_to_two(beam, target, kk, recoil)
    polarization = reaction.polarization('pol_magnitude', 'pol_angle')
    angles = reaction.decay('kk').angles('kshort1')

    z00p = ld.Zlm("z00p", l=0, m=0, r="+", angles=angles, polarization=polarization)
    z00n = ld.Zlm("z00n", l=0, m=0, r="-", angles=angles, polarization=polarization)
    z22p = ld.Zlm("z22p", l=2, m=2, r="+", angles=angles, polarization=polarization)

    s0p = ld.Scalar("s0p", value=parameter("s0p"))
    s0n = ld.Scalar("s0n", value=parameter("s0n"))
    d2p = ld.ComplexScalar("d2p", re=parameter("d2 re"), im=parameter("d2 im"))

    pos_re = (s0p * z00p.real() + d2p * z22p.real()).norm_sqr()
    pos_im = (s0p * z00p.imag() + d2p * z22p.imag()).norm_sqr()
    neg_re = (s0n * z00n.real()).norm_sqr()
    neg_im = (s0n * z00n.imag()).norm_sqr()
    expr = pos_re + pos_im + neg_re + neg_im

    nll = ld.NLL(expr, ds_data, ds_mc)
    status = nll.minimize([1.0] * len(nll.parameters))
    print(status)
    fit_weights = nll.project_weights(status.x)
    s0p_weights = nll.project_weights(status.x, subset=["z00p", "s0p"])
    s0n_weights = nll.project_weights(status.x, subset=["z00n", "s0n"])
    d2p_weights = nll.project_weights(status.x, subset=["z22p", "d2p"])
    masses_mc = res_mass.value_on(ds_mc)
    masses_data = res_mass.value_on(ds_data)
    weights_data = ds_data.weights
    plt.hist(masses_data, weights=weights_data, bins=80, range=(1.0, 2.0), label="Data", histtype="step")
    plt.hist(masses_mc, weights=fit_weights, bins=80, range=(1.0, 2.0), label="Fit", histtype="step")
    plt.hist(masses_mc, weights=s0p_weights, bins=80, range=(1.0, 2.0), label="$S_0^+$", histtype="step")
    plt.hist(masses_mc, weights=s0n_weights, bins=80, range=(1.0, 2.0), label="$S_0^-$", histtype="step")
    plt.hist(masses_mc, weights=d2p_weights, bins=80, range=(1.0, 2.0), label="$D_2^+$", histtype="step")
    plt.legend()
    plt.savefig("demo.svg")


if __name__ == "__main__":
    main()

This example would probably make the most sense for a binned fit, since there isn't actually any mass dependence in any of these amplitudes (so it will just plot the relative amount of each wave over the entire dataset).

Other Examples

You can find other Python examples in the py-laddu/examples folder. These scripts have inline script metadata (PEP 723) which allows them to be run with uv run (which will automatically install the necessary dependencies).

Example 1

The first example script uses data generated with gen_amp. These data consist of a data file with two resonances, an $f_0(1500)$ modeled as a Breit-Wigner with a mass of $1506\text{ MeV}/c^2$ and a width of $112\text{ MeV}/c^2$ and an $f_2'(1525)$, also modeled as a Breit-Wigner, with a mass of $1517\text{ MeV}/c^2$ and a width of $86\text{ MeV}/c^2$, as per the PDG. These were generated to decay to pairs of $K_S^0$s and are produced via photoproduction off a proton target (as in the GlueX experiment). The beam photon is polarized with an angle of $0$ degrees relative to the production plane and a polarization magnitude of $0.3519$ (out of unity). The configuration file used to generate the corresponding data and Monte Carlo files can also be found in the python_examples, and the datasets contain $100,000$ data events and $1,000,000$ Monte Carlo events (generated with the -f argument to create a Monte Carlo file without resonances). The result of this fit can be seen in the following image (using the default 50 bins):

Additionally, this example has an optional MCMC analysis complete with a custom observer to monitor convergence based on the integrated autocorrelation time. This can be run using example_1_mcmc.py script after example_1.py has completed, as it uses data stored during the execution of example_1.py to initialize the walkers. A word of warning, this analysis takes a long time, but is meant more as a demonstration of what's possible with the current system. The custom autocorrelation observer plays an important role in choosing convergence criteria, since this problem has an implicit symmetry (the absolute phase between the two waves matters, but the sign is ambiguous) which cause the posterior distributions to sometimes be multimodal, which can lead to long IATs. Instead, the custom implementation projects the walkers' positions onto the two waves and uses those projections as a proxy to the real chain. This proxy is unimodal by definition, so the IATs calculated from it are much smaller and more realistically describe convergence.

Some example plots can be seen below for the first data bin:

Example 2

The second example uses linear algebra to calculate unpolarized and polarized moments $H(\ell, m)$ from the same data used in the first example. While the results are not as informative, it is a good demonstration of how laddu can also be used outside of the context of a maximum likelihood fit. The mechanism for obtaining moments could also be written with numpy and scipy alone, but laddu's data format, variables, and amplitude evaluation make it easy to write the same code in a very straightforward way which will efficiently run in parallel or even over an MPI instance.

A moment analysis is similar to a partial-wave analysis, except the target observables are coefficients attached to spherical harmonics in a standard sum rather than a coherent sum. This makes it difficult to extract these observables from a maximum likelihood fit, since the corresponding intensity function may be negative for some parameter values, making the logarithm of that intensity undefined. However, similar to a discrete Fourier transform, a moment analysis can be performed by simply summing the spherical harmonic evaluated on each event, and this is what the example does. However, there are two additional considerations. First, since the example data contains a polarized photon beam, we can extract polarized moments which span a basis of not only spherical harmonics of decay angles, but also sine and cosine functions of the polarization angle. Second, we have a Monte Carlo dataset which models the detector efficiency, so we have to construct a matrix of normalization integrals (similar to what we do in an extended maximum likelihood fit). Any non-unitary acceptance function will cause mixing between all of the moments, so we invert the normalization integral matrix and use the matrix-vector product to transform measured moments into the true physical moments.

The resulting unpolarized moments are shown below:

Data Format

laddu focuses on a column layout rather than a specific file container. Each particle is represented by four floating-point columns (_px, _py, _pz, _e), auxiliary scalars keep their explicit names (e.g. pol_magnitude), and an optional weight column stores per-event weights. As long as those columns exist, the physical storage format may vary. Today, Parquet is the preferred container because it is small, language-agnostic, and easy to stream, but the Rust core can also read ROOT TTrees via oxyroot, and the Python bindings add an AmpTools-aware backend on top of uproot. When the weight column is omitted, both the Rust and Python loaders fill it with ones so that unweighted samples continue to work without any extra preprocessing.

For example, the following columns describe a dataset with four particles, the first of which is a polarized photon beam as in the GlueX experiment:

Column name	Data Type	Interpretation
beam_px	Float32 or Float64	Beam momentum (x-component)
beam_py	Float32 or Float64	Beam momentum (y-component)
beam_pz	Float32 or Float64	Beam momentum (z-component)
beam_e	Float32 or Float64	Beam energy
pol_magnitude	Float32 or Float64	Beam polarization magnitude
pol_angle	Float32 or Float64	Beam polarization angle
proton_px	Float32 or Float64	Recoil proton momentum (x-component)
proton_py	Float32 or Float64	Recoil proton momentum (y-component)
proton_pz	Float32 or Float64	Recoil proton momentum (z-component)
proton_e	Float32 or Float64	Recoil proton energy
kshort1_px	Float32 or Float64	Decay product 1 momentum (x-component)
kshort1_py	Float32 or Float64	Decay product 1 momentum (y-component)
kshort1_pz	Float32 or Float64	Decay product 1 momentum (z-component)
kshort1_e	Float32 or Float64	Decay product 1 energy
kshort2_px	Float32 or Float64	Decay product 2 momentum (x-component)
kshort2_py	Float32 or Float64	Decay product 2 momentum (y-component)
kshort2_pz	Float32 or Float64	Decay product 2 momentum (z-component)
kshort2_e	Float32 or Float64	Decay product 2 energy
weight	Float32 or Float64	Event weight

AmpTools-format ROOT files can be directly read as Dataset objects (this is currently implemented in the Python bindings; the Rust API still targets Parquet/standard ROOT TTrees):

import laddu as ld

dataset = ld.io.read_amptools(
    'example_amp.root',
    pol_in_beam=True,
)

This loads the ROOT tree, infers the particle names, and exposes the result through the same Dataset interface used for Parquet inputs.

MPI Support

The latest version of laddu supports the Message Passing Interface (MPI) protocol for distributed computing. MPI-compatible versions of the core laddu methods have been written behind the mpi feature gate. To build laddu with MPI compatibility, it can be added with the mpi feature via cargo add laddu --features mpi. Note that this requires a working MPI installation, and OpenMPI or MPICH are recommended, as well as LLVM/Clang. The installation of these packages differs by system, but are generally available via each system's package manager. The Python implementation of laddu contains a library laddu-cpu and may be optionally installed with a dependency laddu-mpi. If the latter is available, it will be used at runtime unless otherwise specified. Note that this just selects the backend, and doesn't actually use MPI at all, it just gives the option to use it. You can install the optional dependency automatically with pip install 'laddu[mpi]', which builds laddu-mpi against the local system MPI installation.

To use MPI in Rust, one must simply surround their main analysis code with a call to laddu::mpi::use_mpi(true) and laddu::mpi::finalize_mpi(). The first method has a boolean flag which allows for runtime switching of MPI use (for example, disabling MPI with an environment variable). These same methods exist in Python as laddu.mpi.use_mpi(trigger=true) and laddu.mpi.finalize_mpi(), and an additional context manager, laddu.mpi.MPI(trigger=true), can be used to quickly wrap a main() function. See the documentation for more details.

In Python, the default Dataset interface keeps its usual global sequence semantics when MPI is enabled. That means len(dataset), dataset[i], for event in dataset, list(dataset), dataset.events, dataset.weights, and dataset.n_events_weighted all behave as if the full dataset were present locally. Use dataset.iter_local(), dataset.events_local, dataset.weights_local, dataset.n_events_local, and dataset.n_events_weighted_local only when code explicitly wants rank-local ownership semantics. In non-MPI runs, local and global access behave the same way.

The intended guarantees are:

• Default/global access preserves the same dataset-wide ordering and indexing semantics in MPI and non-MPI runs.
• *_local access is always limited to the current rank's ownership and should only be used for explicitly rank-local code paths.
• iter_global() and the default iterator may fetch remote events under MPI; iter_local() never does.
• Global counts and weighted sums (n_events, n_events_weighted, weights) remain valid for whole-dataset reporting under MPI.
• Local/global results are identical in non-MPI runs because there is only one rank.

Future Plans

• GPU integration (this is incredibly difficult to do right now, but it's something I'm looking into).
• As always, more tests and documentation.

Alternatives

While this is likely the first Rust project (aside from my previous attempt, rustitude), there are several other amplitude analysis programs out there at time of writing. This library is a rewrite of rustitude which was written when I was just learning Rust and didn't have a firm grasp of a lot of the core concepts that are required to make the analysis pipeline memory- and CPU-efficient. In particular, rustitude worked well, but ate up a ton of memory and did not handle precalculation as nicely.

AmpTools

The main inspiration for this project is the library most of my collaboration uses, AmpTools. AmpTools has several advantages over laddu: it's probably faster for almost every use case, but this is mainly because it is fully integrated with MPI and GPU support. I'm not actually sure if there's a fair benchmark between the two libraries, but I'd wager AmpTools would still win. AmpTools is a much older, more developed project, dating back to 2010. However, it does have its disadvantages. First and foremost, the primary interaction with the library is through configuration files which are not really code and sort of represent a domain specific language. As such, there isn't really a way to check if a particular config will work before running it. Users could technically code up their analyses in C++ as well, but I think this would generally be more work for very little benefit. AmpTools primarily interacts with Minuit, so there aren't simple ways to perform alternative optimization algorithms, and the outputs are a file which must also be parsed by code written by the user. This usually means some boilerplate setup for each analysis, a slew of input and output files, and, since it doesn't ship with any amplitudes, integration with other libraries. The data format is also very rigid, to the point where including beam polarization information feels hacked on (see the Zlm implementation here which requires the event-by-event polarization to be stored in the beam's four-momentum). While there isn't an official Python interface, Lawrence Ng has made some progress porting the code here.

PyPWA

PyPWA is a library written in pure Python. While this might seem like an issue for performance (and it sort of is), the library has several features which encourage the use of JIT compilers. The upside is that analyses can be quickly prototyped and run with very few dependencies, it can even run on GPUs and use multiprocessing. The downside is that recent development has been slow and the actual implementation of common amplitudes is, in my opinion, messy. I don't think that's a reason to not use it, but it does make it difficult for new users to get started.

ComPWA

ComPWA is a newcomer to the field. It's also a pure Python implementation and is comprised of three separate libraries. QRules can be used to validate and generate particle reaction topologies using conservation rules. AmpForm uses SymPy to transform these topologies into mathematical expressions, and it can also simplify the mathematical forms through the built-in CAS of SymPy. Finally, TensorWaves connects AmpForm to various fitting methods. In general, these libraries have tons of neat features, are well-documented, and are really quite nice to use. I would like to eventually see laddu as a companion to ComPWA (rather than direct competition), but I don't really know enough about the libraries to say much more than that.

Others

It could be the case that I am leaving out software with which I am not familiar. If so, I'd love to include it here for reference. I don't think that laddu will ever be the end-all-be-all of amplitude analysis, just an alternative that might improve on existing systems. It is important for physicists to be aware of these alternatives. For example, if you really don't want to learn Rust but need to implement an amplitude which isn't already included here, laddu isn't for you, and one of these alternatives might be best.

[^1]: Mathieu, V., Albaladejo, M., Fernández-Ramírez, C., Jackura, A. W., Mikhasenko, M., Pilloni, A., & Szczepaniak, A. P. (2019). Moments of angular distribution and beam asymmetries in $\eta\pi^0$ photoproduction at GlueX. Physical Review D, 100(5). doi:10.1103/physrevd.100.054017