Struct Landscape 

pub struct Landscape { /* private fields */ }

A continuous minimization problem: a boxed domain, a displaced minimizer, an optional rotation, and a base function.

Build one with a named constructor and hand it to run. Every landscape declares one unstepped Distribution::Float per dimension, named x0..x{d−1} — so every dimension is CMA-ES-driven and the comparison is about the algorithm rather than about which parameters it models.

Implementations

impl Landscape

pub fn sphere(dim: usize) -> Result<Self>

The sphere, Σ (xᵢ − cᵢ)² over [-5.12, 5.12]^dim.

Smooth, convex, perfectly conditioned: the problem on which any sampler worth shipping must beat uniform random cleanly.

Errors

Error::InvalidSpace if a bound is not finite or a name repeats — which this constructor never produces.

pub fn rosenbrock(dim: usize) -> Result<Self>

The Rosenbrock banana over [-5, 5]^dim.

A narrow, curved valley: the gradient points across it almost everywhere, so progress needs a search distribution that has learned the valley’s local direction. Non-convex and non-separable — the classic case for a covariance, and (unlike the ellipsoid) one where the right covariance keeps changing as the search advances.

Errors

Error::InvalidSpace if a bound is not finite or a name repeats — which this constructor never produces.

pub fn rastrigin(dim: usize) -> Result<Self>

The Rastrigin lattice over [-5.12, 5.12]^dim.

Highly multimodal with a smooth global bowl. CMA-ES is a local-ish searcher at the default population size, so this is the honest hard case, not a showcase.

Errors

Error::InvalidSpace if a bound is not finite or a name repeats — which this constructor never produces.

pub fn ellipsoid(dim: usize, condition: f64) -> Result<Self>

The axis-aligned ellipsoid, Σ cond^(i/(d−1))·(xᵢ − cᵢ)² over [-5, 5]^dim.

Ill-conditioned but separable: an axis-aligned searcher can still make progress one coordinate at a time. It exists to isolate conditioning from rotation — the contrast that says the headline result is about the frame, not merely about the condition number.

Errors

Error::InvalidSpace if a bound is not finite or a name repeats — which this constructor never produces.

pub fn rotated_ellipsoid(dim: usize, condition: f64) -> Result<Self>

The distinguishing fixture: the same ellipsoid, rotated off the coordinate frame.

f(x) = Σ cond^(i/(d−1))·yᵢ² with y = R(x − c) for a fixed orthogonal R (rotation_matrix) that mixes every axis into every other. The spectrum is unchanged — same condition number, same difficulty for a method that understands scale — but no coordinate is now a descent direction, so a sampler that models dimensions independently is left with a problem it cannot represent. This is where covariance adaptation earns its keep or does not.

Errors

Error::InvalidSpace if a bound is not finite or a name repeats — which this constructor never produces.

pub fn name(&self) -> &str

The landscape’s label.

pub fn dimension(&self) -> usize

The number of decision variables.

pub const fn schema(&self) -> &SpaceSchema

The declared search space.

pub fn shift(&self) -> &[f64]

The global minimizer, in problem coordinates.

pub fn rotation(&self) -> Option<&[f64]>

The rotation R, row-major, or None for an axis-aligned landscape.

pub const fn condition(&self) -> f64

The condition number of the Hessian: 1 for every fixture but the ellipsoids.

Rotation does not change it — an orthogonal change of basis preserves the spectrum — which is exactly why the rotated fixture isolates the frame.

pub fn flat_direction(&self) -> Option<Vec<f64>>

The flattest direction of an ellipsoid, in problem coordinates: the axis along which the objective grows most slowly, and therefore the one a working covariance must stretch along.

y = R(x − c), so the direction that maps to y = e₀ (coefficient cond⁰ = 1) is row 0 of R — the coordinate axis e₀ itself when the landscape is unrotated. None for a non-ellipsoid.

pub fn steep_direction(&self) -> Option<Vec<f64>>

The stiffest direction of an ellipsoid (coefficient cond), the counterpart of flat_direction.

pub fn eval(&self, assignment: &Assignment) -> f64

Evaluates the objective at assignment.

Reads x0..x{d−1} in order; a missing or wrong-kinded parameter yields NaN rather than panicking (a NaN never improves a best-so-far curve, so a malformed assignment degrades instead of aborting a run).

pub fn optimum_in_unit_space(&self) -> Vec<f64>

The unit-hypercube coordinate of the global minimizer, per dimension.

Used by the fixture checks to prove the optimum is not sitting at the centre of the box — i.e. not on CMA-ES’s initial mean.

Trait Implementations

impl Clone for Landscape

fn clone(&self) -> Landscape

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Landscape

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

Formats the value using the given formatter.