apex_solver/optimizer/gauss_newton.rs
1//! Gauss-Newton optimization algorithm implementation.
2//!
3//! The Gauss-Newton method is a fundamental iterative algorithm for solving nonlinear least squares problems
4//! of the form:
5//!
6//! ```text
7//! min f(x) = ½||r(x)||² = ½Σᵢ rᵢ(x)²
8//! ```
9//!
10//! where `r: ℝⁿ → ℝᵐ` is the residual vector function.
11//!
12//! # Algorithm Overview
13//!
14//! The Gauss-Newton method solves the normal equations at each iteration:
15//!
16//! ```text
17//! J^T·J·h = -J^T·r
18//! ```
19//!
20//! where:
21//! - `J` is the Jacobian matrix (m × n) of partial derivatives ∂rᵢ/∂xⱼ
22//! - `r` is the residual vector (m × 1)
23//! - `h` is the step vector (n × 1)
24//!
25//! The approximated Hessian `H ≈ J^T·J` replaces the true Hessian `∇²f = J^T·J + Σᵢ rᵢ·∇²rᵢ`,
26//! which works well when residuals are small or nearly linear.
27//!
28//! ## Convergence Properties
29//!
30//! - **Quadratic convergence** near the solution when the Gauss-Newton approximation is valid
31//! - **May diverge** if the initial guess is far from the optimum or the problem is ill-conditioned
32//! - **No step size control** - always takes the full Newton step without damping
33//!
34//! ## When to Use
35//!
36//! Gauss-Newton is most effective when:
37//! - The problem is well-conditioned with `J^T·J` having good numerical properties
38//! - The initial parameter guess is close to the solution
39//! - Fast convergence is prioritized over robustness
40//! - Residuals at the solution are expected to be small
41//!
42//! For ill-conditioned problems or poor initial guesses, consider:
43//! - [`LevenbergMarquardt`](crate::optimizer::LevenbergMarquardt) for adaptive damping
44//! - [`DogLeg`](crate::optimizer::DogLeg) for trust region control
45//!
46//! # Implementation Features
47//!
48//! - **Sparse matrix support**: Efficient handling of large-scale problems via `faer` sparse library
49//! - **Robust linear solvers**: Choice between Cholesky (fast) and QR (stable) factorizations
50//! - **Jacobi scaling**: Optional diagonal preconditioning to improve conditioning
51//! - **Manifold operations**: Support for optimization on Lie groups (SE2, SE3, SO2, SO3)
52//! - **Comprehensive diagnostics**: Detailed convergence and performance summaries
53//!
54//! # Mathematical Background
55//!
56//! At each iteration k, the algorithm:
57//!
58//! 1. **Linearizes** the problem around current estimate xₖ: `r(xₖ + h) ≈ r(xₖ) + J(xₖ)·h`
59//! 2. **Solves** the normal equations for step h: `J^T·J·h = -J^T·r`
60//! 3. **Updates** parameters: `xₖ₊₁ = xₖ ⊕ h` (using manifold plus operation)
61//! 4. **Checks** convergence criteria (cost, gradient, parameter change)
62//!
63//! The method terminates when cost change, gradient norm, or parameter update fall below
64//! specified tolerances, or when maximum iterations are reached.
65//!
66//! # Examples
67//!
68//! ## Basic usage
69//!
70//! ```no_run
71//! use apex_solver::optimizer::GaussNewton;
72//! use apex_solver::core::problem::Problem;
73//! use std::collections::HashMap;
74//!
75//! # fn main() -> Result<(), Box<dyn std::error::Error>> {
76//! // Create optimization problem
77//! let mut problem = Problem::new();
78//! // ... add residual blocks (factors) to problem ...
79//!
80//! // Set up initial parameter values
81//! let initial_values = HashMap::new();
82//! // ... initialize parameters ...
83//!
84//! // Create solver with default configuration
85//! let mut solver = GaussNewton::new();
86//!
87//! // Run optimization
88//! let result = solver.optimize(&problem, &initial_values)?;
89//! # Ok(())
90//! # }
91//! ```
92//!
93//! ## Advanced configuration
94//!
95//! ```no_run
96//! use apex_solver::optimizer::gauss_newton::{GaussNewtonConfig, GaussNewton};
97//! use apex_solver::linalg::LinearSolverType;
98//!
99//! # fn main() {
100//! let config = GaussNewtonConfig::new()
101//! .with_max_iterations(100)
102//! .with_cost_tolerance(1e-8)
103//! .with_parameter_tolerance(1e-8)
104//! .with_gradient_tolerance(1e-10)
105//! .with_linear_solver_type(LinearSolverType::SparseQR) // More stable
106//! .with_jacobi_scaling(true); // Improve conditioning
107//!
108//! let mut solver = GaussNewton::with_config(config);
109//! # }
110//! ```
111//!
112//! # References
113//!
114//! - Nocedal, J. & Wright, S. (2006). *Numerical Optimization* (2nd ed.). Springer. Chapter 10.
115//! - Madsen, K., Nielsen, H. B., & Tingleff, O. (2004). *Methods for Non-Linear Least Squares Problems* (2nd ed.).
116//! - Björck, Å. (1996). *Numerical Methods for Least Squares Problems*. SIAM.
117
118use crate::{core::problem, error, linalg, manifold, optimizer};
119
120use faer::sparse;
121use std::{collections, fmt, time};
122use tracing::debug;
123
124/// Summary statistics for the Gauss-Newton optimization process.
125#[derive(Debug, Clone)]
126pub struct GaussNewtonSummary {
127 /// Initial cost value
128 pub initial_cost: f64,
129 /// Final cost value
130 pub final_cost: f64,
131 /// Total number of iterations performed
132 pub iterations: usize,
133 /// Average cost reduction per iteration
134 pub average_cost_reduction: f64,
135 /// Maximum gradient norm encountered
136 pub max_gradient_norm: f64,
137 /// Final gradient norm
138 pub final_gradient_norm: f64,
139 /// Maximum parameter update norm
140 pub max_parameter_update_norm: f64,
141 /// Final parameter update norm
142 pub final_parameter_update_norm: f64,
143 /// Total time elapsed
144 pub total_time: time::Duration,
145 /// Average time per iteration
146 pub average_time_per_iteration: time::Duration,
147 /// Detailed per-iteration statistics history
148 pub iteration_history: Vec<IterationStats>,
149 /// Convergence status
150 pub convergence_status: optimizer::OptimizationStatus,
151}
152
153/// Per-iteration statistics for detailed logging (Ceres-style output).
154///
155/// Captures all relevant metrics for each optimization iteration, enabling
156/// detailed analysis and debugging of the optimization process.
157#[derive(Debug, Clone)]
158pub struct IterationStats {
159 /// Iteration number (0-indexed)
160 pub iteration: usize,
161 /// Cost function value at this iteration
162 pub cost: f64,
163 /// Change in cost from previous iteration
164 pub cost_change: f64,
165 /// L2 norm of the gradient (||J^T·r||)
166 pub gradient_norm: f64,
167 /// L2 norm of the parameter update step (||Δx||)
168 pub step_norm: f64,
169 /// Trust region ratio (not used in Gauss-Newton, always 0.0)
170 pub tr_ratio: f64,
171 /// Trust region radius (not used in Gauss-Newton, always 0.0)
172 pub tr_radius: f64,
173 /// Linear solver iterations (0 for direct solvers like Cholesky)
174 pub ls_iter: usize,
175 /// Time taken for this iteration in milliseconds
176 pub iter_time_ms: f64,
177 /// Total elapsed time since optimization started in milliseconds
178 pub total_time_ms: f64,
179 /// Whether the step was accepted (always true for Gauss-Newton)
180 pub accepted: bool,
181}
182
183impl IterationStats {
184 /// Print table header in Ceres-style format
185 pub fn print_header() {
186 debug!(
187 "{:>4} {:>13} {:>13} {:>13} {:>13} {:>11} {:>11} {:>7} {:>11} {:>13} {:>6}",
188 "iter",
189 "cost",
190 "cost_change",
191 "|gradient|",
192 "|step|",
193 "tr_ratio",
194 "tr_radius",
195 "ls_iter",
196 "iter_time",
197 "total_time",
198 "status"
199 );
200 }
201
202 pub fn print_line(&self) {
203 let status = if self.iteration == 0 {
204 "-"
205 } else if self.accepted {
206 "✓"
207 } else {
208 "✗"
209 };
210
211 debug!(
212 "{:>4} {:>13.6e} {:>13.2e} {:>13.2e} {:>13.2e} {:>11.2e} {:>11.2e} {:>7} {:>9.2}ms {:>11.2}ms {:>6}",
213 self.iteration,
214 self.cost,
215 self.cost_change,
216 self.gradient_norm,
217 self.step_norm,
218 self.tr_ratio,
219 self.tr_radius,
220 self.ls_iter,
221 self.iter_time_ms,
222 self.total_time_ms,
223 status
224 );
225 }
226}
227
228impl fmt::Display for GaussNewtonSummary {
229 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
230 // Determine if converged
231 let converged = matches!(
232 self.convergence_status,
233 optimizer::OptimizationStatus::Converged
234 | optimizer::OptimizationStatus::CostToleranceReached
235 | optimizer::OptimizationStatus::GradientToleranceReached
236 | optimizer::OptimizationStatus::ParameterToleranceReached
237 );
238
239 writeln!(f, "Gauss-Newton Final Result")?;
240
241 // Title with convergence status
242 if converged {
243 writeln!(f, "CONVERGED ({:?})", self.convergence_status)?;
244 } else {
245 writeln!(f, "DIVERGED ({:?})", self.convergence_status)?;
246 }
247
248 writeln!(f)?;
249 writeln!(f, "Cost:")?;
250 writeln!(f, " Initial: {:.6e}", self.initial_cost)?;
251 writeln!(f, " Final: {:.6e}", self.final_cost)?;
252 writeln!(
253 f,
254 " Reduction: {:.6e} ({:.2}%)",
255 self.initial_cost - self.final_cost,
256 100.0 * (self.initial_cost - self.final_cost) / self.initial_cost.max(1e-12)
257 )?;
258 writeln!(f)?;
259 writeln!(f, "Iterations:")?;
260 writeln!(f, " Total: {}", self.iterations)?;
261 writeln!(f)?;
262 writeln!(f, "Gradient:")?;
263 writeln!(f, " Max norm: {:.2e}", self.max_gradient_norm)?;
264 writeln!(f, " Final norm: {:.2e}", self.final_gradient_norm)?;
265 writeln!(f)?;
266 writeln!(f, "Parameter Update:")?;
267 writeln!(f, " Max norm: {:.2e}", self.max_parameter_update_norm)?;
268 writeln!(f, " Final norm: {:.2e}", self.final_parameter_update_norm)?;
269 writeln!(f)?;
270 writeln!(f, "Performance:")?;
271 writeln!(
272 f,
273 " Total time: {:.2}ms",
274 self.total_time.as_secs_f64() * 1000.0
275 )?;
276 writeln!(
277 f,
278 " Average per iteration: {:.2}ms",
279 self.average_time_per_iteration.as_secs_f64() * 1000.0
280 )?;
281
282 Ok(())
283 }
284}
285
286/// Configuration parameters for the Gauss-Newton optimizer.
287///
288/// Controls the behavior of the Gauss-Newton algorithm including convergence criteria,
289/// linear solver selection, and numerical stability enhancements.
290///
291/// # Builder Pattern
292///
293/// All configuration options can be set using the builder pattern:
294///
295/// ```
296/// use apex_solver::optimizer::gauss_newton::GaussNewtonConfig;
297/// use apex_solver::linalg::LinearSolverType;
298///
299/// let config = GaussNewtonConfig::new()
300/// .with_max_iterations(50)
301/// .with_cost_tolerance(1e-6)
302/// .with_linear_solver_type(LinearSolverType::SparseQR);
303/// ```
304///
305/// # Convergence Criteria
306///
307/// The optimizer terminates when ANY of the following conditions is met:
308///
309/// - **Cost tolerance**: `|cost_k - cost_{k-1}| < cost_tolerance`
310/// - **Parameter tolerance**: `||step|| < parameter_tolerance`
311/// - **Gradient tolerance**: `||J^T·r|| < gradient_tolerance`
312/// - **Maximum iterations**: `iteration >= max_iterations`
313/// - **Timeout**: `elapsed_time >= timeout`
314///
315/// # See Also
316///
317/// - [`GaussNewton`] - The solver that uses this configuration
318/// - [`LevenbergMarquardtConfig`](crate::optimizer::LevenbergMarquardtConfig) - For adaptive damping
319/// - [`DogLegConfig`](crate::optimizer::DogLegConfig) - For trust region methods
320#[derive(Clone)]
321pub struct GaussNewtonConfig {
322 /// Type of linear solver for the linear systems
323 pub linear_solver_type: linalg::LinearSolverType,
324 /// Maximum number of iterations
325 pub max_iterations: usize,
326 /// Convergence tolerance for cost function
327 pub cost_tolerance: f64,
328 /// Convergence tolerance for parameter updates
329 pub parameter_tolerance: f64,
330 /// Convergence tolerance for gradient norm
331 pub gradient_tolerance: f64,
332 /// Timeout duration
333 pub timeout: Option<time::Duration>,
334 /// Use Jacobi column scaling (preconditioning)
335 ///
336 /// When enabled, normalizes Jacobian columns by their L2 norm before solving.
337 /// This can improve convergence for problems with mixed parameter scales
338 /// (e.g., positions in meters + angles in radians) but adds ~5-10% overhead.
339 ///
340 /// Default: false (Gauss-Newton is typically used on well-conditioned problems)
341 pub use_jacobi_scaling: bool,
342 /// Small regularization to ensure J^T·J is positive definite
343 ///
344 /// Pure Gauss-Newton (λ=0) can fail when J^T·J is singular or near-singular.
345 /// Adding a tiny diagonal regularization (e.g., 1e-10) ensures numerical stability
346 /// while maintaining the fast convergence of Gauss-Newton.
347 ///
348 /// Default: 1e-10 (very small, practically identical to pure Gauss-Newton)
349 pub min_diagonal: f64,
350
351 /// Minimum objective function cutoff (optional early termination)
352 ///
353 /// If set, optimization terminates when cost falls below this threshold.
354 /// Useful for early stopping when a "good enough" solution is acceptable.
355 ///
356 /// Default: None (disabled)
357 pub min_cost_threshold: Option<f64>,
358
359 /// Maximum condition number for Jacobian matrix (optional check)
360 ///
361 /// If set, the optimizer checks if condition_number(J^T*J) exceeds this
362 /// threshold and terminates with IllConditionedJacobian status.
363 /// Note: Computing condition number is expensive, so this is disabled by default.
364 ///
365 /// Default: None (disabled)
366 pub max_condition_number: Option<f64>,
367
368 /// Compute per-variable covariance matrices (uncertainty estimation)
369 ///
370 /// When enabled, computes covariance by inverting the Hessian matrix after
371 /// convergence. The full covariance matrix is extracted into per-variable
372 /// blocks stored in both Variable structs and optimier::SolverResult.
373 ///
374 /// Default: false (to avoid performance overhead)
375 pub compute_covariances: bool,
376
377 /// Enable real-time visualization (graphical debugging).
378 ///
379 /// When enabled, optimization progress is logged to a Rerun viewer.
380 /// **Note:** Requires the `visualization` feature to be enabled in `Cargo.toml`.
381 ///
382 /// Default: false
383 #[cfg(feature = "visualization")]
384 pub enable_visualization: bool,
385}
386
387impl Default for GaussNewtonConfig {
388 fn default() -> Self {
389 Self {
390 linear_solver_type: linalg::LinearSolverType::default(),
391 // Ceres Solver default: 50 (changed from 100 for compatibility)
392 max_iterations: 50,
393 // Ceres Solver default: 1e-6 (changed from 1e-8 for compatibility)
394 cost_tolerance: 1e-6,
395 // Ceres Solver default: 1e-8 (unchanged)
396 parameter_tolerance: 1e-8,
397 // Ceres Solver default: 1e-10 (changed from 1e-8 for compatibility)
398 gradient_tolerance: 1e-10,
399 timeout: None,
400 use_jacobi_scaling: false,
401 min_diagonal: 1e-10,
402 // New Ceres-compatible termination parameters
403 min_cost_threshold: None,
404 max_condition_number: None,
405 compute_covariances: false,
406 #[cfg(feature = "visualization")]
407 enable_visualization: false,
408 }
409 }
410}
411
412impl GaussNewtonConfig {
413 /// Create a new Gauss-Newton configuration with default values.
414 pub fn new() -> Self {
415 Self::default()
416 }
417
418 /// Set the linear solver type
419 pub fn with_linear_solver_type(mut self, linear_solver_type: linalg::LinearSolverType) -> Self {
420 self.linear_solver_type = linear_solver_type;
421 self
422 }
423
424 /// Set the maximum number of iterations
425 pub fn with_max_iterations(mut self, max_iterations: usize) -> Self {
426 self.max_iterations = max_iterations;
427 self
428 }
429
430 /// Set the cost tolerance
431 pub fn with_cost_tolerance(mut self, cost_tolerance: f64) -> Self {
432 self.cost_tolerance = cost_tolerance;
433 self
434 }
435
436 /// Set the parameter tolerance
437 pub fn with_parameter_tolerance(mut self, parameter_tolerance: f64) -> Self {
438 self.parameter_tolerance = parameter_tolerance;
439 self
440 }
441
442 /// Set the gradient tolerance
443 pub fn with_gradient_tolerance(mut self, gradient_tolerance: f64) -> Self {
444 self.gradient_tolerance = gradient_tolerance;
445 self
446 }
447
448 /// Set the timeout duration
449 pub fn with_timeout(mut self, timeout: time::Duration) -> Self {
450 self.timeout = Some(timeout);
451 self
452 }
453
454 /// Enable or disable Jacobi column scaling (preconditioning).
455 ///
456 /// When enabled, normalizes Jacobian columns by their L2 norm before solving.
457 /// Can improve convergence for mixed-scale problems but adds ~5-10% overhead.
458 pub fn with_jacobi_scaling(mut self, use_jacobi_scaling: bool) -> Self {
459 self.use_jacobi_scaling = use_jacobi_scaling;
460 self
461 }
462
463 /// Set the minimum diagonal regularization for numerical stability.
464 ///
465 /// A small value (e.g., 1e-10) ensures J^T·J is positive definite while
466 /// maintaining the fast convergence of pure Gauss-Newton.
467 pub fn with_min_diagonal(mut self, min_diagonal: f64) -> Self {
468 self.min_diagonal = min_diagonal;
469 self
470 }
471
472 /// Set minimum objective function cutoff for early termination.
473 ///
474 /// When set, optimization terminates with MinCostThresholdReached status
475 /// if the cost falls below this threshold. Useful for early stopping when
476 /// a "good enough" solution is acceptable.
477 pub fn with_min_cost_threshold(mut self, min_cost: f64) -> Self {
478 self.min_cost_threshold = Some(min_cost);
479 self
480 }
481
482 /// Set maximum condition number for Jacobian matrix.
483 ///
484 /// If set, the optimizer checks if condition_number(J^T*J) exceeds this
485 /// threshold and terminates with IllConditionedJacobian status.
486 /// Note: Computing condition number is expensive, disabled by default.
487 pub fn with_max_condition_number(mut self, max_cond: f64) -> Self {
488 self.max_condition_number = Some(max_cond);
489 self
490 }
491
492 /// Enable or disable covariance computation (uncertainty estimation).
493 ///
494 /// When enabled, computes the full covariance matrix by inverting the Hessian
495 /// after convergence, then extracts per-variable covariance blocks.
496 pub fn with_compute_covariances(mut self, compute_covariances: bool) -> Self {
497 self.compute_covariances = compute_covariances;
498 self
499 }
500
501 /// Enable real-time visualization.
502 ///
503 /// **Note:** Requires the `visualization` feature to be enabled in `Cargo.toml`.
504 ///
505 /// # Arguments
506 ///
507 /// * `enable` - Whether to enable visualization
508 #[cfg(feature = "visualization")]
509 pub fn with_visualization(mut self, enable: bool) -> Self {
510 self.enable_visualization = enable;
511 self
512 }
513
514 /// Print configuration parameters (info level logging)
515 pub fn print_configuration(&self) {
516 debug!(
517 "\nConfiguration:\n Solver: Gauss-Newton\n Linear solver: {:?}\n Convergence Criteria:\n Max iterations: {}\n Cost tolerance: {:.2e}\n Parameter tolerance: {:.2e}\n Gradient tolerance: {:.2e}\n Timeout: {:?}\n Numerical Settings:\n Jacobi scaling: {}\n Compute covariances: {}",
518 self.linear_solver_type,
519 self.max_iterations,
520 self.cost_tolerance,
521 self.parameter_tolerance,
522 self.gradient_tolerance,
523 self.timeout,
524 if self.use_jacobi_scaling {
525 "enabled"
526 } else {
527 "disabled"
528 },
529 if self.compute_covariances {
530 "enabled"
531 } else {
532 "disabled"
533 }
534 );
535 }
536}
537
538/// State for optimization iteration
539struct LinearizerResult {
540 variables: collections::HashMap<String, problem::VariableEnum>,
541 variable_index_map: collections::HashMap<String, usize>,
542 sorted_vars: Vec<String>,
543 symbolic_structure: problem::SymbolicStructure,
544 current_cost: f64,
545 initial_cost: f64,
546}
547
548/// Result from step computation
549struct StepResult {
550 step: faer::Mat<f64>,
551 gradient_norm: f64,
552}
553
554/// Result from cost evaluation
555struct CostEvaluation {
556 new_cost: f64,
557 cost_reduction: f64,
558}
559
560/// Gauss-Newton solver for nonlinear least squares optimization.
561///
562/// Implements the classical Gauss-Newton algorithm which solves `J^T·J·h = -J^T·r` at each
563/// iteration to find the step `h`. This provides fast quadratic convergence near the solution
564/// but may diverge for poor initial guesses or ill-conditioned problems.
565///
566/// # Algorithm
567///
568/// At each iteration k:
569/// 1. Compute residual `r(xₖ)` and Jacobian `J(xₖ)`
570/// 2. Form normal equations: `(J^T·J)·h = -J^T·r`
571/// 3. Solve for step `h` using Cholesky or QR factorization
572/// 4. Update parameters: `xₖ₊₁ = xₖ ⊕ h` (manifold plus operation)
573/// 5. Check convergence criteria
574///
575/// # Examples
576///
577/// ```no_run
578/// use apex_solver::optimizer::GaussNewton;
579/// use apex_solver::core::problem::Problem;
580/// use std::collections::HashMap;
581///
582/// # fn main() -> Result<(), Box<dyn std::error::Error>> {
583/// let mut problem = Problem::new();
584/// // ... add factors to problem ...
585///
586/// let initial_values = HashMap::new();
587/// // ... initialize parameters ...
588///
589/// let mut solver = GaussNewton::new();
590/// let result = solver.optimize(&problem, &initial_values)?;
591/// # Ok(())
592/// # }
593/// ```
594///
595/// # See Also
596///
597/// - [`GaussNewtonConfig`] - Configuration options
598/// - [`LevenbergMarquardt`](crate::optimizer::LevenbergMarquardt) - For adaptive damping
599/// - [`DogLeg`](crate::optimizer::DogLeg) - For trust region control
600pub struct GaussNewton {
601 config: GaussNewtonConfig,
602 jacobi_scaling: Option<sparse::SparseColMat<usize, f64>>,
603 observers: optimizer::OptObserverVec,
604}
605
606impl Default for GaussNewton {
607 fn default() -> Self {
608 Self::new()
609 }
610}
611
612impl GaussNewton {
613 /// Create a new Gauss-Newton solver with default configuration.
614 pub fn new() -> Self {
615 Self::with_config(GaussNewtonConfig::default())
616 }
617
618 /// Create a new Gauss-Newton solver with the given configuration.
619 pub fn with_config(config: GaussNewtonConfig) -> Self {
620 Self {
621 config,
622 jacobi_scaling: None,
623 observers: optimizer::OptObserverVec::new(),
624 }
625 }
626
627 /// Add an observer to the solver.
628 ///
629 /// Observers are notified at each iteration with the current variable values.
630 /// This enables real-time visualization, logging, metrics collection, etc.
631 ///
632 /// # Examples
633 ///
634 /// ```no_run
635 /// use apex_solver::optimizer::GaussNewton;
636 /// # use apex_solver::optimizer::OptObserver;
637 /// # use std::collections::HashMap;
638 /// # use apex_solver::core::problem::VariableEnum;
639 ///
640 /// # struct MyObserver;
641 /// # impl OptObserver for MyObserver {
642 /// # fn on_step(&self, _: &HashMap<String, VariableEnum>, _: usize) {}
643 /// # }
644 /// let mut solver = GaussNewton::new();
645 /// solver.add_observer(MyObserver);
646 /// ```
647 pub fn add_observer(&mut self, observer: impl optimizer::OptObserver + 'static) {
648 self.observers.add(observer);
649 }
650
651 /// Create the appropriate linear solver based on configuration
652 fn create_linear_solver(&self) -> Box<dyn linalg::SparseLinearSolver> {
653 match self.config.linear_solver_type {
654 linalg::LinearSolverType::SparseCholesky => {
655 Box::new(linalg::SparseCholeskySolver::new())
656 }
657 linalg::LinearSolverType::SparseQR => Box::new(linalg::SparseQRSolver::new()),
658 linalg::LinearSolverType::SparseSchurComplement => {
659 // Schur complement solver requires special handling - fallback to Cholesky
660 Box::new(linalg::SparseCholeskySolver::new())
661 }
662 }
663 }
664
665 /// Check convergence criteria
666 /// Check convergence using comprehensive termination criteria.
667 ///
668 /// Implements 8 termination criteria following Ceres Solver standards for Gauss-Newton:
669 ///
670 /// 1. **Gradient Norm (First-Order Optimality)**: ||g||∞ ≤ gradient_tolerance
671 /// 2. **Parameter Change Tolerance**: ||h|| ≤ parameter_tolerance * (||x|| + parameter_tolerance)
672 /// 3. **Function Value Change Tolerance**: |ΔF| < cost_tolerance * F
673 /// 4. **Objective Function Cutoff**: F_new < min_cost_threshold (optional)
674 /// 5. **Singular/Ill-Conditioned Jacobian**: Detected during linear solve
675 /// 6. **Invalid Numerical Values**: NaN or Inf in cost or parameters
676 /// 7. **Maximum Iterations**: iteration >= max_iterations
677 /// 8. **Timeout**: elapsed >= timeout
678 ///
679 /// **Note**: Trust region radius termination is NOT applicable to Gauss-Newton
680 /// as it is a line search method, not a trust region method.
681 ///
682 /// # Arguments
683 ///
684 /// * `iteration` - Current iteration number
685 /// * `current_cost` - Cost before applying the step
686 /// * `new_cost` - Cost after applying the step
687 /// * `parameter_norm` - L2 norm of current parameter vector ||x||
688 /// * `parameter_update_norm` - L2 norm of parameter update step ||h||
689 /// * `gradient_norm` - Infinity norm of gradient ||g||∞
690 /// * `elapsed` - Elapsed time since optimization start
691 ///
692 /// # Returns
693 ///
694 /// `Some(OptimizationStatus)` if any termination criterion is satisfied, `None` otherwise.
695 ///
696 /// # Design Note
697 /// This internal convergence checker requires multiple scalar metrics for comprehensive
698 /// termination criteria. Grouping into a struct would reduce clarity without performance benefit.
699 #[allow(clippy::too_many_arguments)]
700 fn check_convergence(
701 &self,
702 iteration: usize,
703 current_cost: f64,
704 new_cost: f64,
705 parameter_norm: f64,
706 parameter_update_norm: f64,
707 gradient_norm: f64,
708 elapsed: time::Duration,
709 ) -> Option<optimizer::OptimizationStatus> {
710 // ========================================================================
711 // CRITICAL SAFETY CHECKS (perform first, before convergence checks)
712 // ========================================================================
713
714 // CRITERION 6: Invalid Numerical Values (NaN/Inf)
715 // Always check for numerical instability first
716 if !new_cost.is_finite() || !parameter_update_norm.is_finite() || !gradient_norm.is_finite()
717 {
718 return Some(optimizer::OptimizationStatus::InvalidNumericalValues);
719 }
720
721 // CRITERION 8: Timeout
722 // Check wall-clock time limit
723 if let Some(timeout) = self.config.timeout
724 && elapsed >= timeout
725 {
726 return Some(optimizer::OptimizationStatus::Timeout);
727 }
728
729 // CRITERION 7: Maximum Iterations
730 // Check iteration count limit
731 if iteration >= self.config.max_iterations {
732 return Some(optimizer::OptimizationStatus::MaxIterationsReached);
733 }
734
735 // ========================================================================
736 // CONVERGENCE CRITERIA
737 // ========================================================================
738 // Note: Gauss-Newton always accepts the computed step (no step acceptance check)
739
740 // CRITERION 1: Gradient Norm (First-Order Optimality)
741 // Check if gradient infinity norm is below threshold
742 // This indicates we're at a critical point (local minimum, saddle, or maximum)
743 if gradient_norm < self.config.gradient_tolerance {
744 return Some(optimizer::OptimizationStatus::GradientToleranceReached);
745 }
746
747 // Only check parameter and cost criteria after first iteration
748 if iteration > 0 {
749 // CRITERION 2: Parameter Change Tolerance (xtol)
750 // Ceres formula: ||h|| ≤ ε_param * (||x|| + ε_param)
751 // This is a relative measure that scales with parameter magnitude
752 let relative_step_tolerance = self.config.parameter_tolerance
753 * (parameter_norm + self.config.parameter_tolerance);
754
755 if parameter_update_norm <= relative_step_tolerance {
756 return Some(optimizer::OptimizationStatus::ParameterToleranceReached);
757 }
758
759 // CRITERION 3: Function Value Change Tolerance (ftol)
760 // Ceres formula: |ΔF| < ε_cost * F
761 // Check relative cost change (not absolute)
762 let cost_change = (current_cost - new_cost).abs();
763 let relative_cost_change = cost_change / current_cost.max(1e-10); // Avoid division by zero
764
765 if relative_cost_change < self.config.cost_tolerance {
766 return Some(optimizer::OptimizationStatus::CostToleranceReached);
767 }
768 }
769
770 // CRITERION 4: Objective Function Cutoff (optional early stopping)
771 // Useful for "good enough" solutions
772 if let Some(min_cost) = self.config.min_cost_threshold
773 && new_cost < min_cost
774 {
775 return Some(optimizer::OptimizationStatus::MinCostThresholdReached);
776 }
777
778 // CRITERION 5: Singular/Ill-Conditioned Jacobian
779 // Note: This is typically detected during the linear solve and handled there
780 // The max_condition_number check would be expensive to compute here
781 // If linear solve fails, it returns an error that's converted to NumericalFailure
782
783 // No termination criterion satisfied
784 None
785 }
786
787 /// Compute total parameter vector norm ||x||.
788 ///
789 /// Computes the L2 norm of all parameter vectors concatenated together.
790 /// This is used in the relative parameter tolerance check.
791 ///
792 /// # Arguments
793 ///
794 /// * `variables` - Map of variable names to their current values
795 ///
796 /// # Returns
797 ///
798 /// The L2 norm of the concatenated parameter vector
799 fn compute_parameter_norm(
800 variables: &collections::HashMap<String, problem::VariableEnum>,
801 ) -> f64 {
802 variables
803 .values()
804 .map(|v| {
805 let vec = v.to_vector();
806 vec.norm_squared()
807 })
808 .sum::<f64>()
809 .sqrt()
810 }
811
812 /// Create Jacobi scaling matrix from Jacobian
813 fn create_jacobi_scaling(
814 &self,
815 jacobian: &sparse::SparseColMat<usize, f64>,
816 ) -> Result<sparse::SparseColMat<usize, f64>, optimizer::OptimizerError> {
817 use faer::sparse::Triplet;
818
819 let cols = jacobian.ncols();
820 let jacobi_scaling_vec: Vec<Triplet<usize, usize, f64>> = (0..cols)
821 .map(|c| {
822 // Compute column norm: sqrt(sum(J_col^2))
823 let col_norm_squared: f64 = jacobian
824 .triplet_iter()
825 .filter(|t| t.col == c)
826 .map(|t| t.val * t.val)
827 .sum();
828 let col_norm = col_norm_squared.sqrt();
829 // Scaling factor: 1.0 / (1.0 + col_norm)
830 let scaling = 1.0 / (1.0 + col_norm);
831 Triplet::new(c, c, scaling)
832 })
833 .collect();
834
835 sparse::SparseColMat::try_new_from_triplets(cols, cols, &jacobi_scaling_vec).map_err(|e| {
836 optimizer::OptimizerError::JacobiScalingCreation(e.to_string()).log_with_source(e)
837 })
838 }
839
840 /// Initialize optimization state from problem and initial parameters
841 fn initialize_optimization_state(
842 &self,
843 problem: &problem::Problem,
844 initial_params: &collections::HashMap<
845 String,
846 (manifold::ManifoldType, nalgebra::DVector<f64>),
847 >,
848 ) -> Result<LinearizerResult, error::ApexSolverError> {
849 // Initialize variables from initial values
850 let variables = problem.initialize_variables(initial_params);
851
852 // Create column mapping for variables
853 let mut variable_index_map = collections::HashMap::new();
854 let mut col_offset = 0;
855 let mut sorted_vars: Vec<String> = variables.keys().cloned().collect();
856 sorted_vars.sort();
857
858 for var_name in &sorted_vars {
859 variable_index_map.insert(var_name.clone(), col_offset);
860 col_offset += variables[var_name].get_size();
861 }
862
863 // Build symbolic structure for sparse operations
864 let symbolic_structure =
865 problem.build_symbolic_structure(&variables, &variable_index_map, col_offset)?;
866
867 // Initial cost evaluation (residual only, no Jacobian needed)
868 let residual = problem.compute_residual_sparse(&variables)?;
869 let current_cost = optimizer::compute_cost(&residual);
870 let initial_cost = current_cost;
871
872 Ok(LinearizerResult {
873 variables,
874 variable_index_map,
875 sorted_vars,
876 symbolic_structure,
877 current_cost,
878 initial_cost,
879 })
880 }
881
882 /// Process Jacobian by creating and applying Jacobi scaling if enabled
883 fn process_jacobian(
884 &mut self,
885 jacobian: &sparse::SparseColMat<usize, f64>,
886 iteration: usize,
887 ) -> Result<sparse::SparseColMat<usize, f64>, optimizer::OptimizerError> {
888 // Create Jacobi scaling on first iteration if enabled
889 if iteration == 0 {
890 let scaling = self.create_jacobi_scaling(jacobian)?;
891 self.jacobi_scaling = Some(scaling);
892 }
893 let scaling = self
894 .jacobi_scaling
895 .as_ref()
896 .ok_or_else(|| optimizer::OptimizerError::JacobiScalingNotInitialized.log())?;
897 Ok(jacobian * scaling)
898 }
899
900 /// Compute Gauss-Newton step by solving the normal equations
901 fn compute_gauss_newton_step(
902 &self,
903 residuals: &faer::Mat<f64>,
904 scaled_jacobian: &sparse::SparseColMat<usize, f64>,
905 linear_solver: &mut Box<dyn linalg::SparseLinearSolver>,
906 ) -> Option<StepResult> {
907 // Solve the Gauss-Newton equation: J^T·J·Δx = -J^T·r
908 // Use min_diagonal for numerical stability (tiny regularization)
909 let residuals_owned = residuals.as_ref().to_owned();
910 let scaled_step = linear_solver
911 .solve_normal_equation(&residuals_owned, scaled_jacobian)
912 .ok()?;
913
914 // Get gradient from the solver (J^T * r)
915 let gradient = linear_solver.get_gradient()?;
916 // Compute gradient norm for convergence check
917 let gradient_norm = gradient.norm_l2();
918
919 // Apply inverse Jacobi scaling to get final step (if enabled)
920 let step = if self.config.use_jacobi_scaling {
921 let scaling = self.jacobi_scaling.as_ref()?;
922 &scaled_step * scaling
923 } else {
924 scaled_step
925 };
926
927 Some(StepResult {
928 step,
929 gradient_norm,
930 })
931 }
932
933 /// Apply step to parameters and evaluate new cost
934 fn apply_step_and_evaluate_cost(
935 &self,
936 step_result: &StepResult,
937 state: &mut LinearizerResult,
938 problem: &problem::Problem,
939 ) -> error::ApexSolverResult<CostEvaluation> {
940 // Apply parameter updates using manifold operations
941 let _step_norm = optimizer::apply_parameter_step(
942 &mut state.variables,
943 step_result.step.as_ref(),
944 &state.sorted_vars,
945 );
946
947 // Compute new cost (residual only, no Jacobian needed for step evaluation)
948 let new_residual = problem.compute_residual_sparse(&state.variables)?;
949 let new_cost = optimizer::compute_cost(&new_residual);
950
951 // Compute cost reduction
952 let cost_reduction = state.current_cost - new_cost;
953
954 // Update current cost
955 state.current_cost = new_cost;
956
957 Ok(CostEvaluation {
958 new_cost,
959 cost_reduction,
960 })
961 }
962
963 /// Create optimization summary
964 ///
965 /// # Design Note
966 /// This internal summary builder accepts individual scalar results from the optimization loop.
967 /// A struct parameter would not improve readability for this private method.
968 #[allow(clippy::too_many_arguments)]
969 fn create_summary(
970 &self,
971 initial_cost: f64,
972 final_cost: f64,
973 iterations: usize,
974 max_gradient_norm: f64,
975 final_gradient_norm: f64,
976 max_parameter_update_norm: f64,
977 final_parameter_update_norm: f64,
978 total_cost_reduction: f64,
979 total_time: time::Duration,
980 iteration_history: Vec<IterationStats>,
981 convergence_status: optimizer::OptimizationStatus,
982 ) -> GaussNewtonSummary {
983 GaussNewtonSummary {
984 initial_cost,
985 final_cost,
986 iterations,
987 average_cost_reduction: if iterations > 0 {
988 total_cost_reduction / iterations as f64
989 } else {
990 0.0
991 },
992 max_gradient_norm,
993 final_gradient_norm,
994 max_parameter_update_norm,
995 final_parameter_update_norm,
996 total_time,
997 average_time_per_iteration: if iterations > 0 {
998 total_time / iterations as u32
999 } else {
1000 time::Duration::from_secs(0)
1001 },
1002 iteration_history,
1003 convergence_status,
1004 }
1005 }
1006
1007 pub fn optimize(
1008 &mut self,
1009 problem: &problem::Problem,
1010 initial_params: &collections::HashMap<
1011 String,
1012 (manifold::ManifoldType, nalgebra::DVector<f64>),
1013 >,
1014 ) -> Result<
1015 optimizer::SolverResult<collections::HashMap<String, problem::VariableEnum>>,
1016 error::ApexSolverError,
1017 > {
1018 let start_time = time::Instant::now();
1019 let mut iteration = 0;
1020 let mut cost_evaluations = 1; // Initial cost evaluation
1021 let mut jacobian_evaluations = 0;
1022
1023 // Initialize optimization state
1024 let mut state = self.initialize_optimization_state(problem, initial_params)?;
1025
1026 // Create linear solver
1027 let mut linear_solver = self.create_linear_solver();
1028
1029 // Initialize summary tracking variables
1030 let mut max_gradient_norm: f64 = 0.0;
1031 let mut max_parameter_update_norm: f64 = 0.0;
1032 let mut total_cost_reduction = 0.0;
1033 let mut final_gradient_norm;
1034 let mut final_parameter_update_norm;
1035
1036 // Initialize iteration statistics tracking
1037 let mut iteration_stats = Vec::with_capacity(self.config.max_iterations);
1038 let mut previous_cost = state.current_cost;
1039
1040 // Print configuration and header if debug level is enabled
1041 if tracing::enabled!(tracing::Level::DEBUG) {
1042 self.config.print_configuration();
1043 IterationStats::print_header();
1044 }
1045
1046 // Main optimization loop
1047 loop {
1048 let iter_start = time::Instant::now();
1049 // Evaluate residuals and Jacobian
1050 let (residuals, jacobian) = problem.compute_residual_and_jacobian_sparse(
1051 &state.variables,
1052 &state.variable_index_map,
1053 &state.symbolic_structure,
1054 )?;
1055 jacobian_evaluations += 1;
1056
1057 // Process Jacobian (apply scaling if enabled)
1058 let scaled_jacobian = if self.config.use_jacobi_scaling {
1059 self.process_jacobian(&jacobian, iteration)?
1060 } else {
1061 jacobian
1062 };
1063
1064 // Compute Gauss-Newton step
1065 let step_result = match self.compute_gauss_newton_step(
1066 &residuals,
1067 &scaled_jacobian,
1068 &mut linear_solver,
1069 ) {
1070 Some(result) => result,
1071 None => {
1072 return Err(optimizer::OptimizerError::LinearSolveFailed(
1073 "Linear solver failed to solve Gauss-Newton system".to_string(),
1074 )
1075 .into());
1076 }
1077 };
1078
1079 // Update tracking variables
1080 max_gradient_norm = max_gradient_norm.max(step_result.gradient_norm);
1081 final_gradient_norm = step_result.gradient_norm;
1082 let step_norm = step_result.step.norm_l2();
1083 max_parameter_update_norm = max_parameter_update_norm.max(step_norm);
1084 final_parameter_update_norm = step_norm;
1085
1086 // Capture cost before applying step (for convergence check)
1087 let cost_before_step = state.current_cost;
1088
1089 // Apply step and evaluate new cost
1090 let cost_eval = self.apply_step_and_evaluate_cost(&step_result, &mut state, problem)?;
1091 cost_evaluations += 1;
1092 total_cost_reduction += cost_eval.cost_reduction;
1093
1094 // OPTIMIZATION: Only collect iteration statistics if debug level is enabled
1095 // This eliminates ~2-5ms overhead per iteration for non-debug optimization
1096 if tracing::enabled!(tracing::Level::DEBUG) {
1097 let iter_elapsed_ms = iter_start.elapsed().as_secs_f64() * 1000.0;
1098 let total_elapsed_ms = start_time.elapsed().as_secs_f64() * 1000.0;
1099
1100 let stats = IterationStats {
1101 iteration,
1102 cost: state.current_cost,
1103 cost_change: previous_cost - state.current_cost,
1104 gradient_norm: step_result.gradient_norm,
1105 step_norm,
1106 tr_ratio: 0.0, // Not used in Gauss-Newton
1107 tr_radius: 0.0, // Not used in Gauss-Newton
1108 ls_iter: 0, // Direct solver (Cholesky) has no iterations
1109 iter_time_ms: iter_elapsed_ms,
1110 total_time_ms: total_elapsed_ms,
1111 accepted: true, // Gauss-Newton always accepts steps
1112 };
1113
1114 iteration_stats.push(stats.clone());
1115 stats.print_line();
1116 }
1117
1118 previous_cost = state.current_cost;
1119
1120 // Notify all observers with current state
1121 // First set metrics data, then notify observers
1122 self.observers.set_iteration_metrics(
1123 state.current_cost,
1124 step_result.gradient_norm,
1125 None, // Gauss-Newton doesn't use damping
1126 step_norm,
1127 None, // Gauss-Newton doesn't use step quality
1128 );
1129
1130 // Set matrix data if available and there are observers
1131 if !self.observers.is_empty()
1132 && let (Some(hessian), Some(gradient)) =
1133 (linear_solver.get_hessian(), linear_solver.get_gradient())
1134 {
1135 self.observers
1136 .set_matrix_data(Some(hessian.clone()), Some(gradient.clone()));
1137 }
1138
1139 // Notify observers with current variable values and iteration number
1140 self.observers.notify(&state.variables, iteration);
1141
1142 // Compute parameter norm for convergence check
1143 let parameter_norm = Self::compute_parameter_norm(&state.variables);
1144
1145 // Check convergence using comprehensive termination criteria
1146 let elapsed = start_time.elapsed();
1147 if let Some(status) = self.check_convergence(
1148 iteration,
1149 cost_before_step,
1150 cost_eval.new_cost,
1151 parameter_norm,
1152 step_norm,
1153 step_result.gradient_norm,
1154 elapsed,
1155 ) {
1156 let summary = self.create_summary(
1157 state.initial_cost,
1158 state.current_cost,
1159 iteration + 1,
1160 max_gradient_norm,
1161 final_gradient_norm,
1162 max_parameter_update_norm,
1163 final_parameter_update_norm,
1164 total_cost_reduction,
1165 elapsed,
1166 iteration_stats.clone(),
1167 status.clone(),
1168 );
1169
1170 // Print summary only if debug level is enabled
1171 if tracing::enabled!(tracing::Level::DEBUG) {
1172 debug!("{}", summary);
1173 }
1174
1175 // Compute covariances if enabled
1176 let covariances = if self.config.compute_covariances {
1177 problem.compute_and_set_covariances(
1178 &mut linear_solver,
1179 &mut state.variables,
1180 &state.variable_index_map,
1181 )
1182 } else {
1183 None
1184 };
1185
1186 return Ok(optimizer::SolverResult {
1187 status,
1188 iterations: iteration + 1,
1189 initial_cost: state.initial_cost,
1190 final_cost: state.current_cost,
1191 parameters: state.variables.into_iter().collect(),
1192 elapsed_time: elapsed,
1193 convergence_info: Some(optimizer::ConvergenceInfo {
1194 final_gradient_norm,
1195 final_parameter_update_norm,
1196 cost_evaluations,
1197 jacobian_evaluations,
1198 }),
1199 covariances,
1200 });
1201 }
1202
1203 iteration += 1;
1204 }
1205 }
1206}
1207
1208impl optimizer::Solver for GaussNewton {
1209 type Config = GaussNewtonConfig;
1210 type Error = error::ApexSolverError;
1211
1212 fn new() -> Self {
1213 Self::default()
1214 }
1215
1216 fn optimize(
1217 &mut self,
1218 problem: &problem::Problem,
1219 initial_params: &std::collections::HashMap<
1220 String,
1221 (manifold::ManifoldType, nalgebra::DVector<f64>),
1222 >,
1223 ) -> Result<
1224 optimizer::SolverResult<std::collections::HashMap<String, problem::VariableEnum>>,
1225 Self::Error,
1226 > {
1227 self.optimize(problem, initial_params)
1228 }
1229}
1230
1231#[cfg(test)]
1232mod tests {
1233 use crate::{core::problem, factors, manifold, optimizer};
1234 use nalgebra::dvector;
1235 use std::collections;
1236
1237 /// Custom Rosenbrock Factor 1: r1 = 10(x2 - x1²)
1238 /// Demonstrates extensibility - custom factors can be defined outside of factors.rs
1239 #[derive(Debug, Clone)]
1240 struct RosenbrockFactor1;
1241
1242 impl factors::Factor for RosenbrockFactor1 {
1243 fn linearize(
1244 &self,
1245 params: &[nalgebra::DVector<f64>],
1246 compute_jacobian: bool,
1247 ) -> (nalgebra::DVector<f64>, Option<nalgebra::DMatrix<f64>>) {
1248 let x1 = params[0][0];
1249 let x2 = params[1][0];
1250
1251 // Residual: r1 = 10(x2 - x1²)
1252 let residual = nalgebra::dvector![10.0 * (x2 - x1 * x1)];
1253
1254 // Jacobian: ∂r1/∂x1 = -20*x1, ∂r1/∂x2 = 10
1255 let jacobian = if compute_jacobian {
1256 let mut jac = nalgebra::DMatrix::zeros(1, 2);
1257 jac[(0, 0)] = -20.0 * x1;
1258 jac[(0, 1)] = 10.0;
1259 Some(jac)
1260 } else {
1261 None
1262 };
1263
1264 (residual, jacobian)
1265 }
1266
1267 fn get_dimension(&self) -> usize {
1268 1
1269 }
1270 }
1271
1272 /// Custom Rosenbrock Factor 2: r2 = 1 - x1
1273 /// Demonstrates extensibility - custom factors can be defined outside of factors.rs
1274 #[derive(Debug, Clone)]
1275 struct RosenbrockFactor2;
1276
1277 impl factors::Factor for RosenbrockFactor2 {
1278 fn linearize(
1279 &self,
1280 params: &[nalgebra::DVector<f64>],
1281 compute_jacobian: bool,
1282 ) -> (nalgebra::DVector<f64>, Option<nalgebra::DMatrix<f64>>) {
1283 let x1 = params[0][0];
1284
1285 // Residual: r2 = 1 - x1
1286 let residual = nalgebra::dvector![1.0 - x1];
1287
1288 // Jacobian: ∂r2/∂x1 = -1
1289 let jacobian = if compute_jacobian {
1290 Some(nalgebra::DMatrix::from_element(1, 1, -1.0))
1291 } else {
1292 None
1293 };
1294
1295 (residual, jacobian)
1296 }
1297
1298 fn get_dimension(&self) -> usize {
1299 1
1300 }
1301 }
1302
1303 #[test]
1304 fn test_rosenbrock_optimization() -> Result<(), Box<dyn std::error::Error>> {
1305 // Rosenbrock function test:
1306 // Minimize: r1² + r2² where
1307 // r1 = 10(x2 - x1²)
1308 // r2 = 1 - x1
1309 // Starting point: [-1.2, 1.0]
1310 // Expected minimum: [1.0, 1.0]
1311
1312 let mut problem = problem::Problem::new();
1313 let mut initial_values = collections::HashMap::new();
1314
1315 // Add variables using Rn manifold (Euclidean space)
1316 initial_values.insert(
1317 "x1".to_string(),
1318 (manifold::ManifoldType::RN, dvector![-1.2]),
1319 );
1320 initial_values.insert(
1321 "x2".to_string(),
1322 (manifold::ManifoldType::RN, dvector![1.0]),
1323 );
1324
1325 // Add custom factors (demonstrates extensibility!)
1326 problem.add_residual_block(&["x1", "x2"], Box::new(RosenbrockFactor1), None);
1327 problem.add_residual_block(&["x1"], Box::new(RosenbrockFactor2), None);
1328
1329 // Configure Gauss-Newton optimizer
1330 let config = optimizer::gauss_newton::GaussNewtonConfig::new()
1331 .with_max_iterations(100)
1332 .with_cost_tolerance(1e-8)
1333 .with_parameter_tolerance(1e-8)
1334 .with_gradient_tolerance(1e-10);
1335
1336 let mut solver = optimizer::GaussNewton::with_config(config);
1337 let result = solver.optimize(&problem, &initial_values)?;
1338
1339 // Extract final values
1340 let x1_final = result
1341 .parameters
1342 .get("x1")
1343 .ok_or("x1 not found")?
1344 .to_vector()[0];
1345 let x2_final = result
1346 .parameters
1347 .get("x2")
1348 .ok_or("x2 not found")?
1349 .to_vector()[0];
1350
1351 // Verify convergence to [1.0, 1.0]
1352 assert!(
1353 matches!(
1354 result.status,
1355 optimizer::OptimizationStatus::Converged
1356 | optimizer::OptimizationStatus::CostToleranceReached
1357 | optimizer::OptimizationStatus::ParameterToleranceReached
1358 | optimizer::OptimizationStatus::GradientToleranceReached
1359 ),
1360 "Optimization should converge"
1361 );
1362 assert!(
1363 (x1_final - 1.0).abs() < 1e-4,
1364 "x1 should converge to 1.0, got {}",
1365 x1_final
1366 );
1367 assert!(
1368 (x2_final - 1.0).abs() < 1e-4,
1369 "x2 should converge to 1.0, got {}",
1370 x2_final
1371 );
1372 assert!(
1373 result.final_cost < 1e-6,
1374 "Final cost should be near zero, got {}",
1375 result.final_cost
1376 );
1377 Ok(())
1378 }
1379}