An operator can stretch different directions by different amounts and then turn the resulting vectors. Polar decomposition makes this description precise. The stretching part is positive semidefinite and uniquely determined. The isometric part is uniquely determined on the directions not killed by the operator; extending it to a unitary map on the whole space introduces choices exactly when the operator is singular.
This lesson builds on adjoint identities and orthogonal complements and the real and complex spectral theorems. It develops positive square roots and polar factors from spectral diagonalization.
Conventions and plan
First use the spectral theorem to construct a positive square root. Prove its uniqueness before using it to define the stretching factor. Next construct an isometry only on the range of that factor, then extend it by zero on the kernel. Finally distinguish this canonical partial isometry from its generally nonunique unitary extensions.
Throughout, is a finite-dimensional inner-product space over or , and is linear. The symbol denotes the adjoint; in orthonormal coordinates it is the conjugate transpose. Over the real field, unitary means orthogonal.
Definition 8.1.1
(Positive semidefinite operator)
Positivity always includes self-adjointness: means that and for every . Positive definiteness requires for nonzero .
Definition 8.1.2
(Partial isometry)
A linear map is a partial isometry when it is an isometry on the orthogonal complement of its kernel (its initial subspace) and zero on its kernel.
A positive square root exists and is unique
Theorem 8.1.3
(The positive square root)
On a finite-dimensional real or complex inner-product space, every positive semidefinite self-adjoint operator has a unique positive semidefinite self-adjoint square root.
Proof
(Existence from the spectral theorem, uniqueness from commutation)
Let . The spectral theorem gives an orthonormal eigenbasis with real eigenvalues . Define to have eigenvalues in that same basis. It is self-adjoint, positive semidefinite and satisfies .
For uniqueness, suppose and . Then , because both products equal . Thus preserves every eigenspace of : if , then . The restriction of to that eigenspace is self-adjoint and positive semidefinite, so it has an orthonormal eigenbasis there. Each of its eigenvalues satisfies and , forcing . Consequently is the scalar on the whole eigenspace. This determines on every spectral summand, proving .
Proof X-ray
Where positivity and commutation enter the uniqueness proof
The uniqueness half is the delicate one, and it uses each hypothesis on exactly once. The identity alone gives only the commutation , which lets act inside each eigenspace of . Self-adjointness of is then used to diagonalize it there. Positivity of is used only at the end, to choose the nonnegative root rather than .
The argument closes because on each eigenspace of the equation has just one nonnegative solution, so is forced to be a scalar there; the spectral summands of fill the whole space. Dropping positivity breaks only that last step: the sign of is then free on each eigenspace.
Write this unique operator as . Its construction is independent of the basis chosen inside a repeated eigenspace. Without positivity, even self-adjoint square roots need not be unique: on the self-adjoint matrix squares to , as does itself, but only is positive.
Common mistake
(Taking square roots of the entries)
A positive square root is a spectral object, not an entrywise one. For the entrywise root has . The true root uses the eigenvalues with eigenvectors : .
The stretching factor has the correct kernel
Proposition 8.1.4
(The stretching factor and its kernel)
Let . Then for every , , and is an orthogonal decomposition.
Proof
(Compare norms, then use self-adjointness)
The operator is positive semidefinite, since
Define . Then
Self-adjointness and finite dimension give . Hence the space splits orthogonally as
This is the right decomposition for constructing the turning part. When is singular, a formula using its inverse on all of would be meaningless.
Construct the canonical partial isometry
Proposition 8.1.5
(The canonical partial isometry)
There is a unique partial isometry with and . It satisfies ; is the orthogonal projection onto and is the orthogonal projection onto .
Proof
(Define on and extend by zero)
For a vector in , define
This is well defined: if , then , so . The rule is linear and preserves norms because . By the inner-product polarization identity it also preserves inner products on that subspace. Its range is .
Extend by zero on . We have constructed
The kernel assertion follows because an isometry has no nonzero kernel on its initial subspace . Moreover is unique with these properties: the factorization forces its value on every , and its stated kernel forces zero on the orthogonal complementary summand.
Indeed on the initial subspace is an isometric bijection onto its range, and the adjoint is its inverse there; both maps are zero on the respective orthogonal complements. So and are projections, not necessarily the identity.
Unitary extensions and the invertible case
Theorem 8.1.6
(Polar decomposition)
There is a unitary with , and is the only positive semidefinite factor possible. If is invertible, is the unique unitary factor. If is singular, the unitary factor is not unique.
Proof
(Choose an isometric bijection between the complementary kernels)
The two spaces and have the same dimension, by rank–nullity. Choose any isometric bijection between them. Retain on and use on . The resulting operator maps an orthogonal direct sum isometrically onto another such direct sum, hence is unitary. Since takes its values in the initial subspace, we still have
Conversely, any unitary factor in this equation must agree with on , so all choices occur precisely in the isometric bijection between the two complementary spaces. The positive factor is always unique: if with unitary and , then , so square-root uniqueness gives .
If is invertible, then is positive definite and invertible. There is no kernel on which to choose an extension, and
is the unique unitary factor. Directly, ; in equal finite dimensions this implies unitarity. If is singular, the canonical partial isometry remains unique, but the full unitary factor generally does not. In a real one-dimensional kernel the extension has two sign choices; in a complex one-dimensional kernel it has a whole circle of phase choices.
Concept lensGeometric
Stretch along axes, then move rigidly
For a real invertible , take an orthonormal eigenbasis of with eigenvalues . The factor stretches the unit sphere into an ellipsoid with semi-axes of length along these eigenvectors, and the orthogonal then moves that ellipsoid rigidly. So maps the unit sphere to an ellipsoid whose semi-axis lengths are the eigenvalues of . For below, the image of the unit circle has semi-axes and , and turns it through .
The order of the factors matters. This is the right polar factorization . It does not assert . For an invertible , the left positive factor is , giving the different expression .
Counterexample mode
The unitary factor is not unique for a singular matrix
Claim: in the unitary factor is always unique. Take . Then has eigenvalues and on and , so (its eigenvalues on the same vectors). Both and are orthogonal, and , while . They differ: but , the sign choice on . Repair: assume invertible, or accept that only the partial isometry (here the projection onto ) is canonical.
A distance problem makes the factors useful
Theorem 8.1.7
(Nearest orthogonal matrix)
Let be an invertible real matrix with polar decomposition . Among orthogonal matrices , the Frobenius distance is minimized by , and only by .
Proof
(Reduce to maximizing a trace)
For a real square matrix, the Frobenius norm is , the sum of squares of its entries. Orthogonal changes of coordinates preserve it. For an orthogonal and ,
Choose an orthonormal eigenbasis of , with positive eigenvalues . In that basis is still orthogonal, and each diagonal entry is at most one. Therefore , with equality only when every diagonal entry is one. An orthogonal matrix with such diagonal entries is the identity, since a column of norm one with its diagonal entry equal to one has all other entries zero. Thus and the unique nearest orthogonal matrix is .
When some vanish, the equality argument does not constrain those directions, so the invertible theorem cannot be applied to a singular matrix.
Worked examples
Worked example 8.1.8
(An invertible matrix)
Let . Then , so , which is invertible, and , the rotation through . Check and . By the distance theorem this rotation is the unique orthogonal matrix nearest to : gives , while for one gets and .
Worked example 8.1.9
(A singular symmetric matrix)
Let , already positive semidefinite, so as in the previous section. The kernel is spanned by and the image by . The canonical partial isometry is , so is the identity on and zero on : , which is the projection onto (and here). Its unitary extensions are , namely and .
Worked example 8.1.10
(A singular matrix whose image and kernel are not orthogonal)
Let . Then and , with . The rule gives , so and . Check . Here , so the unitary extensions send , giving , and in both cases.
Theorem checklist and common errors
- A positive semidefinite self-adjoint operator on a finite-dimensional real or complex inner-product space has a unique positive semidefinite square root.
- With , the kernels of and agree. The rule , extended by zero on their common kernel, gives the unique partial isometry in with .
- For square endomorphisms, this isometry extends to a unitary factor in . That factor is unique when is invertible; the positive factor is unique in every case.
- In real Frobenius distance, the orthogonal polar factor of an invertible matrix is its unique nearest orthogonal matrix.
Common mistake
(Inverting a singular positive factor)
For the formula would require , but has no inverse. The partial isometry must be defined only on and extended separately, as above.
Do not confuse a partial isometry with a unitary map. Do not infer commutation of the two factors from their existence. All displayed adjoint formulas assume orthonormal coordinates when written as ordinary conjugate transposes.
Quick checks
Quick check
(Find the polar decomposition of .)
Solution · Answer
, so , which is invertible. Then , orthogonal, and .
Quick check
(For on , find and , and say which orthogonal matrices satisfy .)
Solution · Answer
, so , and . Every orthogonal satisfies , so the unitary factor is as nonunique as possible, while is still unique.
Quick check
(Is the positive square root of ? Why does this not contradict uniqueness?)
Solution · Answer
No. and is self-adjoint, but is not positive semidefinite because . The only positive semidefinite square root of is .