Struct TCI2Options 

pub struct TCI2Options {

    pub tolerance: f64,
    pub max_iter: usize,
    pub max_bond_dim: usize,
    pub pivot_search: PivotSearchStrategy,
    pub normalize_error: bool,
    pub verbosity: usize,
    pub max_nglobal_pivot: usize,
    pub nsearch: usize,
    pub sweep_strategy: Sweep2Strategy,
    pub ncheck_history: usize,
    pub strictly_nested: bool,
    pub tol_margin_global_search: f64,
    pub seed: Option<u64>,
}

Configuration for the TCI2 algorithm (crossinterpolate2).

Controls convergence criteria, bond dimension limits, pivot search strategy, and global pivot search parameters.

Recommended starting point

For most problems the defaults work well. The fields you will most commonly adjust are:

Field	Typical range	Purpose
tolerance	1e-6 – 1e-12	Relative convergence threshold
max_bond_dim	50 – 500	Hard cap on bond dimension
max_iter	20 – 100	Maximum number of half-sweeps
seed	Some(42)	Fix seed for reproducibility

Convergence criterion

The algorithm stops when all of the following hold for ncheck_history consecutive iterations:

1. The normalized bond error is below tolerance.
2. No global pivots were added.
3. The rank is stable.

Alternatively, it stops when the rank reaches max_bond_dim.

Examples

use tensor4all_tensorci::TCI2Options;

// Default options
let opts = TCI2Options::default();
assert!((opts.tolerance - 1e-8).abs() < 1e-15);
assert_eq!(opts.max_iter, 20);
assert_eq!(opts.max_bond_dim, usize::MAX);
assert_eq!(opts.verbosity, 0);

// Custom options via struct update syntax
let custom = TCI2Options {
    tolerance: 1e-12,
    max_bond_dim: 100,
    seed: Some(42),
    ..TCI2Options::default()
};
assert!((custom.tolerance - 1e-12).abs() < 1e-20);
assert_eq!(custom.max_bond_dim, 100);
assert_eq!(custom.seed, Some(42));

Fields

tolerance: f64

Convergence tolerance (default: 1e-8).

When normalize_error is enabled (the default), this is the relative threshold: the bond error is divided by the maximum absolute function value seen so far. Typical choices are 1e-8 for moderate accuracy and 1e-12 for high accuracy.

max_iter: usize

Maximum number of half-sweep iterations (default: 20).

Each iteration performs one forward or backward two-site sweep followed by a global pivot search. Increase if the function is difficult to converge.

max_bond_dim: usize

Hard upper bound on bond dimension (default: usize::MAX, i.e. no limit).

The algorithm stops early once the rank reaches this value. For expensive functions, setting this to 50–500 avoids runaway computation.

pivot_search: PivotSearchStrategy

Pivot search strategy (default: PivotSearchStrategy::Full).

Full materializes the entire candidate matrix and finds the best pivot exactly. Rook uses lazy block-rook search and is faster for very large local dimensions but may miss some pivots.

normalize_error: bool

Whether to normalize the bond error by the maximum observed sample value (default: true).

When enabled, tolerance acts as a relative threshold. Disable to use tolerance as an absolute threshold.

verbosity: usize

Verbosity level (default: 0 = silent).

1 prints per-iteration summaries, 2 adds per-bond details.

max_nglobal_pivot: usize

Maximum number of global pivots to add per iteration (default: 5).

nsearch: usize

Number of random starting points for global pivot search (default: 5).

Each starting point undergoes local optimization to find regions of high interpolation error.

sweep_strategy: Sweep2Strategy

Sweep strategy for 2-site sweeps (default: Sweep2Strategy::BackAndForth).

ncheck_history: usize

Number of recent iterations checked for convergence (default: 3).

The algorithm requires ncheck_history consecutive converged iterations before declaring convergence.

strictly_nested: bool

Whether to use strictly nested index sets (default: false).

When false, the algorithm keeps a history of previous index sets and merges them during sweeps, which generally improves convergence.

tol_margin_global_search: f64

Tolerance margin for global pivot search (default: 10.0).

Global pivots are accepted when their error exceeds abs_tolerance * tol_margin_global_search.

seed: Option<u64>

Random seed for reproducibility (default: None = OS entropy).

Set to Some(seed) for deterministic results.

Trait Implementations

impl Clone for TCI2Options

fn clone(&self) -> TCI2Options

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TCI2Options

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

Formats the value using the given formatter.

impl Default for TCI2Options

fn default() -> Self

Returns the “default value” for a type.