nalgebra/src/linalg/svd.rs
2020-04-05 18:49:48 +02:00

719 lines
25 KiB
Rust
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#[cfg(feature = "serde-serialize")]
use serde::{Deserialize, Serialize};
use approx::AbsDiffEq;
use num::{One, Zero};
use crate::allocator::Allocator;
use crate::base::{DefaultAllocator, Matrix, Matrix2x3, MatrixMN, Vector2, VectorN};
use crate::constraint::{SameNumberOfRows, ShapeConstraint};
use crate::dimension::{Dim, DimDiff, DimMin, DimMinimum, DimSub, U1, U2};
use crate::storage::Storage;
use simba::scalar::{ComplexField, RealField};
use crate::linalg::givens::GivensRotation;
use crate::linalg::symmetric_eigen;
use crate::linalg::Bidiagonal;
/// Singular Value Decomposition of a general matrix.
#[cfg_attr(feature = "serde-serialize", derive(Serialize, Deserialize))]
#[cfg_attr(
feature = "serde-serialize",
serde(bound(
serialize = "DefaultAllocator: Allocator<N::RealField, DimMinimum<R, C>> +
Allocator<N, DimMinimum<R, C>, C> +
Allocator<N, R, DimMinimum<R, C>>,
MatrixMN<N, R, DimMinimum<R, C>>: Serialize,
MatrixMN<N, DimMinimum<R, C>, C>: Serialize,
VectorN<N::RealField, DimMinimum<R, C>>: Serialize"
))
)]
#[cfg_attr(
feature = "serde-serialize",
serde(bound(
deserialize = "DefaultAllocator: Allocator<N::RealField, DimMinimum<R, C>> +
Allocator<N, DimMinimum<R, C>, C> +
Allocator<N, R, DimMinimum<R, C>>,
MatrixMN<N, R, DimMinimum<R, C>>: Deserialize<'de>,
MatrixMN<N, DimMinimum<R, C>, C>: Deserialize<'de>,
VectorN<N::RealField, DimMinimum<R, C>>: Deserialize<'de>"
))
)]
#[derive(Clone, Debug)]
pub struct SVD<N: ComplexField, R: DimMin<C>, C: Dim>
where
DefaultAllocator: Allocator<N, DimMinimum<R, C>, C>
+ Allocator<N, R, DimMinimum<R, C>>
+ Allocator<N::RealField, DimMinimum<R, C>>,
{
/// The left-singular vectors `U` of this SVD.
pub u: Option<MatrixMN<N, R, DimMinimum<R, C>>>,
/// The right-singular vectors `V^t` of this SVD.
pub v_t: Option<MatrixMN<N, DimMinimum<R, C>, C>>,
/// The singular values of this SVD.
pub singular_values: VectorN<N::RealField, DimMinimum<R, C>>,
}
impl<N: ComplexField, R: DimMin<C>, C: Dim> Copy for SVD<N, R, C>
where
DefaultAllocator: Allocator<N, DimMinimum<R, C>, C>
+ Allocator<N, R, DimMinimum<R, C>>
+ Allocator<N::RealField, DimMinimum<R, C>>,
MatrixMN<N, R, DimMinimum<R, C>>: Copy,
MatrixMN<N, DimMinimum<R, C>, C>: Copy,
VectorN<N::RealField, DimMinimum<R, C>>: Copy,
{
}
impl<N: ComplexField, R: DimMin<C>, C: Dim> SVD<N, R, C>
where
DimMinimum<R, C>: DimSub<U1>, // for Bidiagonal.
DefaultAllocator: Allocator<N, R, C>
+ Allocator<N, C>
+ Allocator<N, R>
+ Allocator<N, DimDiff<DimMinimum<R, C>, U1>>
+ Allocator<N, DimMinimum<R, C>, C>
+ Allocator<N, R, DimMinimum<R, C>>
+ Allocator<N, DimMinimum<R, C>>
+ Allocator<N::RealField, DimMinimum<R, C>>
+ Allocator<N::RealField, DimDiff<DimMinimum<R, C>, U1>>,
{
/// Computes the Singular Value Decomposition of `matrix` using implicit shift.
pub fn new(matrix: MatrixMN<N, R, C>, compute_u: bool, compute_v: bool) -> Self {
Self::try_new(
matrix,
compute_u,
compute_v,
N::RealField::default_epsilon(),
0,
)
.unwrap()
}
/// Attempts to compute the Singular Value Decomposition of `matrix` using implicit shift.
///
/// # Arguments
///
/// * `compute_u` set this to `true` to enable the computation of left-singular vectors.
/// * `compute_v` set this to `true` to enable the computation of left-singular vectors.
/// * `eps` tolerance used to determine when a value converged to 0.
/// * `max_niter` maximum total number of iterations performed by the algorithm. If this
/// number of iteration is exceeded, `None` is returned. If `niter == 0`, then the algorithm
/// continues indefinitely until convergence.
pub fn try_new(
mut matrix: MatrixMN<N, R, C>,
compute_u: bool,
compute_v: bool,
eps: N::RealField,
max_niter: usize,
) -> Option<Self> {
assert!(
matrix.len() != 0,
"Cannot compute the SVD of an empty matrix."
);
let (nrows, ncols) = matrix.data.shape();
let min_nrows_ncols = nrows.min(ncols);
let dim = min_nrows_ncols.value();
let m_amax = matrix.camax();
if !m_amax.is_zero() {
matrix.unscale_mut(m_amax);
}
let b = Bidiagonal::new(matrix);
let mut u = if compute_u { Some(b.u()) } else { None };
let mut v_t = if compute_v { Some(b.v_t()) } else { None };
let mut diagonal = b.diagonal();
let mut off_diagonal = b.off_diagonal();
let mut niter = 0;
let (mut start, mut end) = Self::delimit_subproblem(
&mut diagonal,
&mut off_diagonal,
&mut u,
&mut v_t,
b.is_upper_diagonal(),
dim - 1,
eps,
);
while end != start {
let subdim = end - start + 1;
// Solve the subproblem.
if subdim > 2 {
let m = end - 1;
let n = end;
let mut vec;
{
let dm = diagonal[m];
let dn = diagonal[n];
let fm = off_diagonal[m];
let tmm = dm * dm + off_diagonal[m - 1] * off_diagonal[m - 1];
let tmn = dm * fm;
let tnn = dn * dn + fm * fm;
let shift = symmetric_eigen::wilkinson_shift(tmm, tnn, tmn);
vec = Vector2::new(
diagonal[start] * diagonal[start] - shift,
diagonal[start] * off_diagonal[start],
);
}
for k in start..n {
let m12 = if k == n - 1 {
N::RealField::zero()
} else {
off_diagonal[k + 1]
};
let mut subm = Matrix2x3::new(
diagonal[k],
off_diagonal[k],
N::RealField::zero(),
N::RealField::zero(),
diagonal[k + 1],
m12,
);
if let Some((rot1, norm1)) = GivensRotation::cancel_y(&vec) {
rot1.inverse()
.rotate_rows(&mut subm.fixed_columns_mut::<U2>(0));
let rot1 = GivensRotation::new_unchecked(rot1.c(), N::from_real(rot1.s()));
if k > start {
// This is not the first iteration.
off_diagonal[k - 1] = norm1;
}
let v = Vector2::new(subm[(0, 0)], subm[(1, 0)]);
// FIXME: does the case `v.y == 0` ever happen?
let (rot2, norm2) = GivensRotation::cancel_y(&v)
.unwrap_or((GivensRotation::identity(), subm[(0, 0)]));
rot2.rotate(&mut subm.fixed_columns_mut::<U2>(1));
let rot2 = GivensRotation::new_unchecked(rot2.c(), N::from_real(rot2.s()));
subm[(0, 0)] = norm2;
if let Some(ref mut v_t) = v_t {
if b.is_upper_diagonal() {
rot1.rotate(&mut v_t.fixed_rows_mut::<U2>(k));
} else {
rot2.rotate(&mut v_t.fixed_rows_mut::<U2>(k));
}
}
if let Some(ref mut u) = u {
if b.is_upper_diagonal() {
rot2.inverse()
.rotate_rows(&mut u.fixed_columns_mut::<U2>(k));
} else {
rot1.inverse()
.rotate_rows(&mut u.fixed_columns_mut::<U2>(k));
}
}
diagonal[k + 0] = subm[(0, 0)];
diagonal[k + 1] = subm[(1, 1)];
off_diagonal[k + 0] = subm[(0, 1)];
if k != n - 1 {
off_diagonal[k + 1] = subm[(1, 2)];
}
vec.x = subm[(0, 1)];
vec.y = subm[(0, 2)];
} else {
break;
}
}
} else if subdim == 2 {
// Solve the remaining 2x2 subproblem.
let (u2, s, v2) = compute_2x2_uptrig_svd(
diagonal[start],
off_diagonal[start],
diagonal[start + 1],
compute_u && b.is_upper_diagonal() || compute_v && !b.is_upper_diagonal(),
compute_v && b.is_upper_diagonal() || compute_u && !b.is_upper_diagonal(),
);
let u2 = u2.map(|u2| GivensRotation::new_unchecked(u2.c(), N::from_real(u2.s())));
let v2 = v2.map(|v2| GivensRotation::new_unchecked(v2.c(), N::from_real(v2.s())));
diagonal[start + 0] = s[0];
diagonal[start + 1] = s[1];
off_diagonal[start] = N::RealField::zero();
if let Some(ref mut u) = u {
let rot = if b.is_upper_diagonal() {
u2.unwrap()
} else {
v2.unwrap()
};
rot.rotate_rows(&mut u.fixed_columns_mut::<U2>(start));
}
if let Some(ref mut v_t) = v_t {
let rot = if b.is_upper_diagonal() {
v2.unwrap()
} else {
u2.unwrap()
};
rot.inverse().rotate(&mut v_t.fixed_rows_mut::<U2>(start));
}
end -= 1;
}
// Re-delimit the subproblem in case some decoupling occurred.
let sub = Self::delimit_subproblem(
&mut diagonal,
&mut off_diagonal,
&mut u,
&mut v_t,
b.is_upper_diagonal(),
end,
eps,
);
start = sub.0;
end = sub.1;
niter += 1;
if niter == max_niter {
return None;
}
}
diagonal *= m_amax;
// Ensure all singular value are non-negative.
for i in 0..dim {
let sval = diagonal[i];
if sval < N::RealField::zero() {
diagonal[i] = -sval;
if let Some(ref mut u) = u {
u.column_mut(i).neg_mut();
}
}
}
Some(Self {
u,
v_t,
singular_values: diagonal,
})
}
/*
fn display_bidiag(b: &Bidiagonal<N, R, C>, begin: usize, end: usize) {
for i in begin .. end {
for k in begin .. i {
print!(" ");
}
println!("{} {}", b.diagonal[i], b.off_diagonal[i]);
}
for k in begin .. end {
print!(" ");
}
println!("{}", b.diagonal[end]);
}
*/
fn delimit_subproblem(
diagonal: &mut VectorN<N::RealField, DimMinimum<R, C>>,
off_diagonal: &mut VectorN<N::RealField, DimDiff<DimMinimum<R, C>, U1>>,
u: &mut Option<MatrixMN<N, R, DimMinimum<R, C>>>,
v_t: &mut Option<MatrixMN<N, DimMinimum<R, C>, C>>,
is_upper_diagonal: bool,
end: usize,
eps: N::RealField,
) -> (usize, usize) {
let mut n = end;
while n > 0 {
let m = n - 1;
if off_diagonal[m].is_zero()
|| off_diagonal[m].norm1() <= eps * (diagonal[n].norm1() + diagonal[m].norm1())
{
off_diagonal[m] = N::RealField::zero();
} else if diagonal[m].norm1() <= eps {
diagonal[m] = N::RealField::zero();
Self::cancel_horizontal_off_diagonal_elt(
diagonal,
off_diagonal,
u,
v_t,
is_upper_diagonal,
m,
m + 1,
);
if m != 0 {
Self::cancel_vertical_off_diagonal_elt(
diagonal,
off_diagonal,
u,
v_t,
is_upper_diagonal,
m - 1,
);
}
} else if diagonal[n].norm1() <= eps {
diagonal[n] = N::RealField::zero();
Self::cancel_vertical_off_diagonal_elt(
diagonal,
off_diagonal,
u,
v_t,
is_upper_diagonal,
m,
);
} else {
break;
}
n -= 1;
}
if n == 0 {
return (0, 0);
}
let mut new_start = n - 1;
while new_start > 0 {
let m = new_start - 1;
if off_diagonal[m].norm1() <= eps * (diagonal[new_start].norm1() + diagonal[m].norm1())
{
off_diagonal[m] = N::RealField::zero();
break;
}
// FIXME: write a test that enters this case.
else if diagonal[m].norm1() <= eps {
diagonal[m] = N::RealField::zero();
Self::cancel_horizontal_off_diagonal_elt(
diagonal,
off_diagonal,
u,
v_t,
is_upper_diagonal,
m,
n,
);
if m != 0 {
Self::cancel_vertical_off_diagonal_elt(
diagonal,
off_diagonal,
u,
v_t,
is_upper_diagonal,
m - 1,
);
}
break;
}
new_start -= 1;
}
(new_start, n)
}
// Cancels the i-th off-diagonal element using givens rotations.
fn cancel_horizontal_off_diagonal_elt(
diagonal: &mut VectorN<N::RealField, DimMinimum<R, C>>,
off_diagonal: &mut VectorN<N::RealField, DimDiff<DimMinimum<R, C>, U1>>,
u: &mut Option<MatrixMN<N, R, DimMinimum<R, C>>>,
v_t: &mut Option<MatrixMN<N, DimMinimum<R, C>, C>>,
is_upper_diagonal: bool,
i: usize,
end: usize,
) {
let mut v = Vector2::new(off_diagonal[i], diagonal[i + 1]);
off_diagonal[i] = N::RealField::zero();
for k in i..end {
if let Some((rot, norm)) = GivensRotation::cancel_x(&v) {
let rot = GivensRotation::new_unchecked(rot.c(), N::from_real(rot.s()));
diagonal[k + 1] = norm;
if is_upper_diagonal {
if let Some(ref mut u) = *u {
rot.inverse()
.rotate_rows(&mut u.fixed_columns_with_step_mut::<U2>(i, k - i));
}
} else if let Some(ref mut v_t) = *v_t {
rot.rotate(&mut v_t.fixed_rows_with_step_mut::<U2>(i, k - i));
}
if k + 1 != end {
v.x = -rot.s().real() * off_diagonal[k + 1];
v.y = diagonal[k + 2];
off_diagonal[k + 1] *= rot.c();
}
} else {
break;
}
}
}
// Cancels the i-th off-diagonal element using givens rotations.
fn cancel_vertical_off_diagonal_elt(
diagonal: &mut VectorN<N::RealField, DimMinimum<R, C>>,
off_diagonal: &mut VectorN<N::RealField, DimDiff<DimMinimum<R, C>, U1>>,
u: &mut Option<MatrixMN<N, R, DimMinimum<R, C>>>,
v_t: &mut Option<MatrixMN<N, DimMinimum<R, C>, C>>,
is_upper_diagonal: bool,
i: usize,
) {
let mut v = Vector2::new(diagonal[i], off_diagonal[i]);
off_diagonal[i] = N::RealField::zero();
for k in (0..i + 1).rev() {
if let Some((rot, norm)) = GivensRotation::cancel_y(&v) {
let rot = GivensRotation::new_unchecked(rot.c(), N::from_real(rot.s()));
diagonal[k] = norm;
if is_upper_diagonal {
if let Some(ref mut v_t) = *v_t {
rot.rotate(&mut v_t.fixed_rows_with_step_mut::<U2>(k, i - k));
}
} else if let Some(ref mut u) = *u {
rot.inverse()
.rotate_rows(&mut u.fixed_columns_with_step_mut::<U2>(k, i - k));
}
if k > 0 {
v.x = diagonal[k - 1];
v.y = rot.s().real() * off_diagonal[k - 1];
off_diagonal[k - 1] *= rot.c();
}
} else {
break;
}
}
}
/// Computes the rank of the decomposed matrix, i.e., the number of singular values greater
/// than `eps`.
pub fn rank(&self, eps: N::RealField) -> usize {
assert!(
eps >= N::RealField::zero(),
"SVD rank: the epsilon must be non-negative."
);
self.singular_values.iter().filter(|e| **e > eps).count()
}
/// Rebuild the original matrix.
///
/// This is useful if some of the singular values have been manually modified.
/// Returns `Err` if the right- and left- singular vectors have not been
/// computed at construction-time.
pub fn recompose(self) -> Result<MatrixMN<N, R, C>, &'static str> {
match (self.u, self.v_t) {
(Some(mut u), Some(v_t)) => {
for i in 0..self.singular_values.len() {
let val = self.singular_values[i];
u.column_mut(i).scale_mut(val);
}
Ok(u * v_t)
}
(None, None) => Err("SVD recomposition: U and V^t have not been computed."),
(None, _) => Err("SVD recomposition: U has not been computed."),
(_, None) => Err("SVD recomposition: V^t has not been computed."),
}
}
/// Computes the pseudo-inverse of the decomposed matrix.
///
/// Any singular value smaller than `eps` is assumed to be zero.
/// Returns `Err` if the right- and left- singular vectors have not
/// been computed at construction-time.
pub fn pseudo_inverse(mut self, eps: N::RealField) -> Result<MatrixMN<N, C, R>, &'static str>
where
DefaultAllocator: Allocator<N, C, R>,
{
if eps < N::RealField::zero() {
Err("SVD pseudo inverse: the epsilon must be non-negative.")
} else {
for i in 0..self.singular_values.len() {
let val = self.singular_values[i];
if val > eps {
self.singular_values[i] = N::RealField::one() / val;
} else {
self.singular_values[i] = N::RealField::zero();
}
}
self.recompose().map(|m| m.adjoint())
}
}
/// Solves the system `self * x = b` where `self` is the decomposed matrix and `x` the unknown.
///
/// Any singular value smaller than `eps` is assumed to be zero.
/// Returns `Err` if the singular vectors `U` and `V` have not been computed.
// FIXME: make this more generic wrt the storage types and the dimensions for `b`.
pub fn solve<R2: Dim, C2: Dim, S2>(
&self,
b: &Matrix<N, R2, C2, S2>,
eps: N::RealField,
) -> Result<MatrixMN<N, C, C2>, &'static str>
where
S2: Storage<N, R2, C2>,
DefaultAllocator: Allocator<N, C, C2> + Allocator<N, DimMinimum<R, C>, C2>,
ShapeConstraint: SameNumberOfRows<R, R2>,
{
if eps < N::RealField::zero() {
Err("SVD solve: the epsilon must be non-negative.")
} else {
match (&self.u, &self.v_t) {
(Some(u), Some(v_t)) => {
let mut ut_b = u.ad_mul(b);
for j in 0..ut_b.ncols() {
let mut col = ut_b.column_mut(j);
for i in 0..self.singular_values.len() {
let val = self.singular_values[i];
if val > eps {
col[i] = col[i].unscale(val);
} else {
col[i] = N::zero();
}
}
}
Ok(v_t.ad_mul(&ut_b))
}
(None, None) => Err("SVD solve: U and V^t have not been computed."),
(None, _) => Err("SVD solve: U has not been computed."),
(_, None) => Err("SVD solve: V^t has not been computed."),
}
}
}
}
impl<N: ComplexField, R: DimMin<C>, C: Dim, S: Storage<N, R, C>> Matrix<N, R, C, S>
where
DimMinimum<R, C>: DimSub<U1>, // for Bidiagonal.
DefaultAllocator: Allocator<N, R, C>
+ Allocator<N, C>
+ Allocator<N, R>
+ Allocator<N, DimDiff<DimMinimum<R, C>, U1>>
+ Allocator<N, DimMinimum<R, C>, C>
+ Allocator<N, R, DimMinimum<R, C>>
+ Allocator<N, DimMinimum<R, C>>
+ Allocator<N::RealField, DimMinimum<R, C>>
+ Allocator<N::RealField, DimDiff<DimMinimum<R, C>, U1>>,
{
/// Computes the Singular Value Decomposition using implicit shift.
pub fn svd(self, compute_u: bool, compute_v: bool) -> SVD<N, R, C> {
SVD::new(self.into_owned(), compute_u, compute_v)
}
/// Attempts to compute the Singular Value Decomposition of `matrix` using implicit shift.
///
/// # Arguments
///
/// * `compute_u` set this to `true` to enable the computation of left-singular vectors.
/// * `compute_v` set this to `true` to enable the computation of left-singular vectors.
/// * `eps` tolerance used to determine when a value converged to 0.
/// * `max_niter` maximum total number of iterations performed by the algorithm. If this
/// number of iteration is exceeded, `None` is returned. If `niter == 0`, then the algorithm
/// continues indefinitely until convergence.
pub fn try_svd(
self,
compute_u: bool,
compute_v: bool,
eps: N::RealField,
max_niter: usize,
) -> Option<SVD<N, R, C>> {
SVD::try_new(self.into_owned(), compute_u, compute_v, eps, max_niter)
}
/// Computes the singular values of this matrix.
pub fn singular_values(&self) -> VectorN<N::RealField, DimMinimum<R, C>> {
SVD::new(self.clone_owned(), false, false).singular_values
}
/// Computes the rank of this matrix.
///
/// All singular values below `eps` are considered equal to 0.
pub fn rank(&self, eps: N::RealField) -> usize {
let svd = SVD::new(self.clone_owned(), false, false);
svd.rank(eps)
}
/// Computes the pseudo-inverse of this matrix.
///
/// All singular values below `eps` are considered equal to 0.
pub fn pseudo_inverse(self, eps: N::RealField) -> Result<MatrixMN<N, C, R>, &'static str>
where
DefaultAllocator: Allocator<N, C, R>,
{
SVD::new(self.clone_owned(), true, true).pseudo_inverse(eps)
}
}
// Explicit formulae inspired from the paper "Computing the Singular Values of 2-by-2 Complex
// Matrices", Sanzheng Qiao and Xiaohong Wang.
// http://www.cas.mcmaster.ca/sqrl/papers/sqrl5.pdf
fn compute_2x2_uptrig_svd<N: RealField>(
m11: N,
m12: N,
m22: N,
compute_u: bool,
compute_v: bool,
) -> (
Option<GivensRotation<N>>,
Vector2<N>,
Option<GivensRotation<N>>,
) {
let two: N::RealField = crate::convert(2.0f64);
let half: N::RealField = crate::convert(0.5f64);
let denom = (m11 + m22).hypot(m12) + (m11 - m22).hypot(m12);
// NOTE: v1 is the singular value that is the closest to m22.
// This prevents cancellation issues when constructing the vector `csv` below. If we chose
// otherwise, we would have v1 ~= m11 when m12 is small. This would cause catastrophic
// cancellation on `v1 * v1 - m11 * m11` below.
let mut v1 = m11 * m22 * two / denom;
let mut v2 = half * denom;
let mut u = None;
let mut v_t = None;
if compute_u || compute_v {
let (csv, sgn_v) = GivensRotation::new(m11 * m12, v1 * v1 - m11 * m11);
v1 *= sgn_v;
v2 *= sgn_v;
if compute_v {
v_t = Some(csv);
}
if compute_u {
let cu = (m11.scale(csv.c()) + m12 * csv.s()) / v1;
let su = (m22 * csv.s()) / v1;
let (csu, sgn_u) = GivensRotation::new(cu, su);
v1 *= sgn_u;
v2 *= sgn_u;
u = Some(csu);
}
}
(u, Vector2::new(v1, v2), v_t)
}