rsQuick Start

§ 1Workspace at a glance

ppvm is a Cargo workspace of focused crates. A foundation trio —ppvm-traits, ppvm-pauli-word, andppvm-pauli-sum — provides Pauli propagation; everything else builds on it. Depend on what your workload actually needs — there is no umbrella crate.

ppvm-traits foundation

The trait system: the Config trait that bundles storage / coefficient type / hasher / truncation strategy, the single-qubitPauli alphabet, the gate & noise trait hierarchy (Clifford, TGate, rotations, depolarizing, loss, …), and the shared ACMap / Trace map impls. Required by every other ppvm crate.

ppvm-pauli-word foundation

Packed Pauli strings: PauliWord, PhasedPauliWord,LossyPauliWord, and PauliPattern. Built onppvm-traits.

ppvm-pauli-sum backend

The Pauli-propagation engine: the PauliSum<Config>(and LossyPauliSum) type, the truncationstrategy types, and the concrete Configbundles (fxhash, indexmap, dashmap,gxhash). The usual entry point for native Pauli propagation.

ppvm-tableau backend

Generalized stabilizer-tableau simulator with non-Clifford support via sparse coefficient tracking. Provides Tableau,GeneralizedTableau, and the SparseVector trait (indexable by usize, u128, orbnum types for > 64 qubits). Optional rayonfeature for parallel sampling.

ppvm-stim interop

Parse and execute Stim programs againstppvm-tableau. Re-exports the parser, providesparse_extended, execute, and asample helper for multi-shot loops. Depends on the Pauli crates, ppvm-tableau, and stim-parser.

ppvm-sym backend

Symbolic (parametric) Pauli propagation. Pauli sums carry coefficients of type Term — polynomials in sines and cosines of symbolic parameters — useful for analytic studies of variational circuits and Trotter schedules.

stim-parser parser

Standalone parser for the Stim circuit format. Use directly if you want to consume Stim programs without pulling in a simulator;ppvm-stim already depends on it.

§ 2Pick crates by use case

• Pauli propagation only (Heisenberg-picture observables, noisy circuits at scale): ppvm-pauli-sum.
• Stabilizer / generalized-tableau simulation(Schrödinger picture, mid-circuit measurement, loss):ppvm-pauli-sum + ppvm-tableau.
• Run Stim programs against the tableau:ppvm-pauli-sum + ppvm-tableau +ppvm-stim.
• Symbolic / parametric circuits:ppvm-pauli-sum + ppvm-sym.
• Just parse Stim, no simulator: stim-parser.

§ 3Setup

Requires Rust ≥ 1.83 (edition 2024). Add the crates relevant to your workload — the git dependency form is identical for each:

[dependencies]
ppvm-pauli-sum = { git = "https://github.com/QuEraComputing/ppvm" }
ppvm-tableau   = { git = "https://github.com/QuEraComputing/ppvm" }
ppvm-stim      = { git = "https://github.com/QuEraComputing/ppvm" }
ppvm-sym       = { git = "https://github.com/QuEraComputing/ppvm" }

On x86_64, set RUSTFLAGS="-C target-feature=+aes,+sse2" — gxhash, used by the default Config bundles, needs the AES instruction set. CI sets this automatically. On targets without AES, build with --no-default-features --features=indexmap,ahash to drop gxhash.

§ 4ppvm-pauli-sum — Pauli propagation

PauliSum<T: Config> is generic over a configuration bundle (storage, coefficient type, hasher, truncation strategy). Pick a pre-built config from ppvm_pauli_sum::config:

• indexmap::ByteFxHashF64<N> — deterministic ordering, f64 coefficients. Good default.
• dashmap::ByteFxHash<N> — concurrent map for use with rayon.
• fxhash::Byte<N> — plain HashMap; fastest single-threaded.

N is the number of bytes per Pauli word: N = ceil(n_qubits / 8). Twelve qubits → N = 2.

use ppvm_pauli_sum::prelude::*;

fn main() {
    let mut state: PauliSum<config::indexmap::ByteFxHashF64<1>> =
        PauliSum::builder().n_qubits(2).build();
    state += ("ZZ", 1.0);

    // Circuit is H(0); CNOT(0, 1) — apply in reverse for Heisenberg propagation.
    state.cnot(0, 1);
    state.h(0);

    // <0…0| state |0…0> via a pattern match.
    let zero_state: PauliPattern = "Z?*".into();
    println!("{}", state.trace(&zero_state));
}

The same circuit and the same gate vocabulary the Python bindings sit on top of — every Python call goes through this layer.

§ 5ppvm-tableau — generalized stabilizer tableau

For full-state simulation in the Schrödinger picture, useGeneralizedTableau<T, I, C>. The Iparameter is the bitstring index type for the sparse coefficient vector — pick by qubit count:

• usize — up to ~64 qubits (platform dependent).
• u128 — up to 128 qubits.
• bnum::types::U256 / U512 / U1024 — for larger systems.

Using usize past 64 qubits silently overflows — pick the next-larger type if in doubt.

use ppvm_pauli_sum::prelude::*;
use ppvm_tableau::prelude::*;

fn main() {
    // GeneralizedTableau::new takes (n_qubits, coefficient_threshold).
    let mut tab: GeneralizedTableau<config::indexmap::ByteFxHashF64<1>, usize, _>
        = GeneralizedTableau::new(2, 1e-10);

    tab.h(0);
    tab.cnot(0, 1);

    let r0 = tab.measure(0);
    let r1 = tab.measure(1);
    assert_eq!(r0, r1); // GHZ correlation
}

Gates run in forward order here — the tableau is in the Schrödinger picture, not the Heisenberg picture thatppvm-pauli-sum uses. The same trait methods (h, cnot, rx, t, …) apply to both backends; only the picture changes.

Non-Clifford gates (t, u3, rotations) andreset live on the tableau. For mid-circuit measurement, the outcome is a tri-state value (Some(true), Some(false),None for loss) — loss is first-class.

§ 6ppvm-stim — running Stim programs

ppvm-stim parses extended Stim and executes against aGeneralizedTableau. The two-stage pipeline — parse once, reuse the program across shots — matters when sampling at scale.

use ppvm_stim::{parse_extended, sample};
use ppvm_tableau::prelude::*;
use ppvm_pauli_sum::prelude::*;

fn run_shots(src: &str, n_qubits: usize) -> Result<Vec<Vec<Option<bool>>>, ppvm_stim::Error> {
    let prog = parse_extended(src)?;

    // Multi-shot: pass a factory closure — sample reuses the parsed program.
    let shots = sample(&prog, 10_000, || {
        GeneralizedTableau::<config::indexmap::ByteFxHashF64<1>, usize, _>::new(n_qubits, 1e-10)
    })?;
    Ok(shots)
}

For single-shot demos there are also run_string /run_file, but they re-parse on every call — never use them inside a sampling loop.

§ 7ppvm-sym — symbolic propagation

Substitute f64 coefficients for Term — theppvm-sym polynomial-in-sin/cos representation — to propagate Paulis through parametric circuits without committing to angle values:

use ppvm_pauli_sum::prelude::*;
use ppvm_sym::Term;

fn main() {
    let mut sum = PauliSum::<config::fxhash::Byte<2, Term>>::builder()
        .n_qubits(2)
        .build();
    sum += ("ZZ", Term::from(1.0));

    sum.rz(0, Term::var(0));
    sum.ry(0, Term::var(1));
    sum.cnot(0, 1);
    sum.rx(1, Term::var(1));

    let pat: PauliPattern = "Z?*".into();
    println!("Trace expression: {}", sum.trace(&pat));

    // Evaluate at concrete angles.
    let value = sum.trace(&pat).eval(&[0.3, 0.5]).unwrap();
    println!("Value at (0.3, 0.5): {}", value);
}

The same PauliSum, the same gate trait set — only the coefficient type changes. Term::var(i) introduces a free parameter; gate methods accept anything implementingInto<T::Coeff>. Call .eval(&[...]) on a Term to substitute numeric values at the end.

§ 8stim-parser — standalone Stim parsing

Use stim-parser when you want to consume the Stim circuit format without a simulator — e.g., to translate Stim into your own IR or to lint circuits in CI. ppvm-stim already depends on this crate and re-exports its prelude.

use stim_parser::prelude::*;

let prog = parse_extended(stim_src)?;
for inst in prog.instructions() {
    // walk gates, measurements, annotations…
}

§ 9The trait hierarchy

Gate behaviour is exposed through small, composable traits defined once in ppvm_traits::traits and implemented by every backend. Generic code that takes a T: Clifford + RotationOne bound works against PauliSum and GeneralizedTableaualike:

• Clifford / CliffordExtensions — Pauli, Hadamard, S, CNOT, CZ, CY, √X, √Y (and their adjoints).
• TGate, RotationOne, RotationTwo, U3Gate — non-Clifford rotations that branch the Pauli sum.
• Measure / LossyMeasure — Z-basis measurement, with and without loss-aware outcomes.
• Depolarizing, PauliError, LossChannel, CorrelatedLossChannel, AmplitudeDamping — noise channels.
• Reset — ground-state reset (tableau).

Adding a new gate is a matter of implementing the relevant trait for whichever simulator types you care about — see theDeveloper Guide for the end-to-end recipe.

§ 10Custom truncation strategies

PauliSum takes a truncation policy as part of itsConfig. Pre-built strategies inppvm_pauli_sum::strategy cover the common cases —CoefficientThreshold(eps), MaxPauliWeight(w),MaxLossWeight(w), CombinedStrategy(a, b). Pass one through the builder:

use ppvm_pauli_sum::{prelude::*, strategy::CoefficientThreshold};

type State = PauliSum<config::indexmap::ByteFxHashF64<2, CoefficientThreshold>>;

let mut state: State = PauliSum::builder()
    .n_qubits(12)
    .strategy(CoefficientThreshold(1e-6))
    .capacity(400)            // capacity tuning has a large perf impact
    .build();

To roll your own, implement the Strategy trait — which provides capacity (an initial-size hint for the map) and atruncate method that drops entries from the map according to your policy:

use ppvm_traits::traits::strategy::Strategy;

#[derive(Debug, Clone, Copy, Default)]
struct DropTinyHighWeight;

impl Strategy for DropTinyHighWeight {
    fn capacity(&self, n_qubits: usize) -> usize {
        16 * n_qubits
    }

    fn truncate<S, V, H, M, W>(&self, map: &mut M)
    where
        // bounds elided — see ppvm_traits::traits::strategy
    {
        map.retain(|word, coeff| {
            coeff.abs() > 1e-8 || word.weight() <= 8
        });
    }
}

Once the strategy is part of the Config, truncation runs automatically inside the gate methods. You can also callstate.truncate() manually at sync points if you want explicit control.

§ 11Next steps

• Browse the full Rust API reference, grouped by crate and kind.
• Read examples/symbolic.rs, examples/trotter.rs, and examples/msd.rs in the repository for end-to-end pipelines.
• If your user-facing surface is Python, you don't have to drop down to Rust — the Python Quick Start covers the common cases.
• For architecture details (Config-trait generics, dual-map optimisation, where to put new gates), see the Developer Guide.