Enum Optimizer 

pub enum Optimizer {
    GradientDescent {
        learning_rate: f64,
    },
    Adam {
        learning_rate: f64,
        beta1: f64,
        beta2: f64,
        epsilon: f64,
        m: Array1<f64>,
        v: Array1<f64>,
    },
    SPSA {
        learning_rate: f64,
        perturbation: f64,
    },
    QuantumNaturalGradient {
        learning_rate: f64,
        regularization: f64,
    },
    SciRS2 {
        method: String,
        config: HashMap<String, f64>,
        adam_m: Array1<f64>,
        adam_v: Array1<f64>,
        lbfgs_history: Vec<(Array1<f64>, Array1<f64>)>,
        prev_params: Option<Array1<f64>>,
        prev_gradient: Option<Array1<f64>>,
        cg_direction: Option<Array1<f64>>,
    },
}

Optimizer for quantum machine learning models

Variants

GradientDescent

Gradient descent

Fields

learning_rate: f64

Learning rate

Adam

Adam optimizer

Fields

learning_rate: f64

Learning rate

beta1: f64

Beta1 parameter

beta2: f64

Beta2 parameter

epsilon: f64

Epsilon parameter

m: Array1<f64>

First-moment (momentum) accumulator. Lazily (re)sized to match the parameter vector on the first update_parameters call.

v: Array1<f64>

Second-moment (uncentered variance) accumulator. Lazily (re)sized to match the parameter vector on the first update_parameters call.

SPSA

SPSA optimizer

Fields

learning_rate: f64

Learning rate

perturbation: f64

Perturbation size

QuantumNaturalGradient

Quantum Natural Gradient optimizer.

This variant stores the scalar hyper-parameters only. Callers are expected to pre-condition gradients through QuantumAutoDiff::natural_gradients() (which requires a circuit executor closure) and then pass the resulting natural-gradient vector to update_parameters. The regularization field is used as additive damping: Δθ_i = −lr · g_i / (1 + reg).

Fields

learning_rate: f64

Learning rate

regularization: f64

Tikhonov regularisation added to the QFIM diagonal before inversion

SciRS2

SciRS2-based optimizers: Adam, L-BFGS (two-loop recursion), and nonlinear Conjugate Gradient (Fletcher-Reeves/Polak-Ribiere), each with real per-parameter state carried between update_parameters calls.

Fields

method: String

Optimizer method: “adam”, “lbfgs”, or “cg”

config: HashMap<String, f64>

Configuration parameters

adam_m: Array1<f64>

Adam first-moment accumulator (method == “adam”)

adam_v: Array1<f64>

Adam second-moment accumulator (method == “adam”)

lbfgs_history: Vec<(Array1<f64>, Array1<f64>)>

Bounded curvature-pair history (s_k, y_k) for L-BFGS’s two-loop recursion (method == “lbfgs”), newest last, truncated to config["m"] entries.

prev_params: Option<Array1<f64>>

Previous call’s parameter vector, used by both L-BFGS and CG to form the next curvature pair / conjugate direction.

prev_gradient: Option<Array1<f64>>

Previous call’s gradient vector.

cg_direction: Option<Array1<f64>>

Previous CG search direction (method == “cg”).

Implementations

impl Optimizer

pub fn new(method: OptimizationMethod) -> Self

Creates a new optimizer with default parameters

pub fn update_parameters( &mut self, parameters: &mut Array1<f64>, gradients: &ArrayView1<'_, f64>, iteration: usize, ) -> Result<()>

Updates parameters based on gradients.

Each variant now carries and mutates its own real optimizer state (Adam’s first/second moments, L-BFGS’s curvature-pair history, CG’s previous conjugate direction, …), so this takes &mut self.

Trait Implementations

impl Clone for Optimizer

fn clone(&self) -> Optimizer

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Optimizer

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Display for Optimizer

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.