S. D. V. Stephens

notes / Studies in Algebra and Group Theory / chapter 4: Categories, Bilinear Forms and Inner Products / From my book

B9.5 Orthogonal and Unitary Operators

9.5.1 Definitions and Characterizations

Orthogonal (real) and unitary (complex) operators are the isometries of inner product spaces: they preserve lengths and angles.

Definition 9.5.1 (Orthogonal and unitary operators). Let be an inner product space. An operator is orthogonal (if is real) or unitary (if is complex) if for all .

Note. The definitions are parallel; some authors use “orthogonal” in both cases. Note that an orthogonal operator is not the same thing as an orthogonal projection.

Theorem 9.5.2 (Characterizations of orthogonal and unitary operators). For an operator on a finite-dimensional inner product space, the following are equivalent:

  1. is orthogonal (resp. unitary).
  2. , i.e., is invertible and .
  3. .
  4. for all .
  5. maps some orthonormal basis to an orthonormal basis.
  6. maps every orthonormal basis to an orthonormal basis.

Proof. (i) (ii): for all , so .

(ii) (iii): On a finite-dimensional space, a one-sided inverse of an operator is two-sided: makes injective, hence invertible with , so ; and makes surjective, hence invertible with , so .

(ii) (iv): .

(iv) (i): The inner product can be recovered from norms by the polarization identities. Over , , so Over , for we have . Multiplying by and summing over , and using and , gives In both cases the right side is unchanged when are replaced by , since . So .

(i) (vi): If is an orthonormal basis, then , so is an orthonormal list of length , hence an orthonormal basis.

(vi) (v): Orthonormal bases exist (Theorem 9.2.2).

(v) (ii): Let be an orthonormal basis with orthonormal. Then for all . Since the span , for each , so . ∎

Corollary 9.5.3 (Orthogonal and unitary operators form a group). The orthogonal (resp. unitary) operators on a finite-dimensional inner product space form a subgroup of , denoted (resp. ). If is orthogonal or unitary, then so is .

Proof. By the theorem, such operators are invertible. The identity preserves the inner product. If and do, then . And . ∎

Proposition 9.5.4 (Eigenvalues and invariant subspaces). Let be an orthogonal or unitary operator on a finite-dimensional inner product space .

  1. Every eigenvalue of satisfies . In the real case, .
  2. If is -invariant, then , and is also -invariant.

Proof. (i) If with , then , so .

(ii) is injective and is finite-dimensional, so . Hence , i.e., is invariant under . By Lemma 9.3.7(i) applied to (whose adjoint is ), is -invariant. ∎

Matrix Characterization

Proposition 9.5.5 (Orthogonal and unitary matrices). Let be an orthonormal basis of , and let be the matrix of in this basis.

  1. (Real case) is orthogonal iff ; such is called an orthogonal matrix.
  2. (Complex case) is unitary iff ; such is called a unitary matrix.

Equivalently: the columns of form an orthonormal basis of with its standard inner product.

Proof. In an orthonormal basis, (Theorem 9.3.4), so iff .

Write for the columns of . The -entry of is , the standard inner product. So iff the columns are orthonormal. ∎

Corollary 9.5.6 (Rows of orthogonal and unitary matrices). If is orthogonal (resp. unitary), then the rows of also form an orthonormal basis of .

Proof. implies that is invertible with , so . Writing for the rows of (as vectors in ), the -entry of is . So the rows are orthonormal. ∎

Proposition 9.5.7 (Determinant of orthogonal and unitary operators). If is orthogonal or unitary, then . For orthogonal , .

Proof. Let be the matrix of in an orthonormal basis. Then has matrix (Theorem 9.3.4), and (Example 5.4.25; over this is , Proposition 5.4.6). Hence, by multiplicativity (Theorem 5.4.7), Over , is a real number of absolute value , so . ∎

Skew-adjoint operators give a supply of orthogonal ones.

Proposition 9.5.8 (Cayley transform). Let be skew-adjoint on a finite-dimensional real inner product space. Then and are invertible, and is orthogonal.

Proof. is invertible by Proposition 9.4.7(iii), and so is , since is also skew-adjoint. Since , we have and . The operators and commute (both are polynomials in ). Using Proposition 9.3.3, Hence and is orthogonal by Theorem 9.5.2. ∎

9.5.2 Reflections

A particularly simple class of orthogonal operators are the reflections across hyperplanes.

Definition 9.5.9 (Reflection). Let be a finite-dimensional real inner product space. A reflection is an orthogonal operator such that and .

Note. The condition says is an involution. Since is orthogonal, it is also self-adjoint (). The only possible eigenvalues of are (if with , then ), and says the -eigenspace is a line.

Proposition 9.5.10 (Formula for reflections). For a unit vector , the map is a reflection, with , and fixes every vector of the hyperplane . Conversely, every reflection is for a unit vector spanning . Geometrically, is the reflection across the hyperplane .

Proof. Linear: Clear from the formula.

Orthogonal:

Involution: , so

Eigenspaces: iff , iff (if , then ). So is a line. If , then .

Converse: Let be a reflection and a unit vector spanning , so . If , then , so . The vector then lies in and satisfies , so . Hence fixes pointwise. Writing any as with the second term in , ∎

Proposition 9.5.11 (Reflections map equal-length vectors). Let with .

  1. If , the reflection with satisfies .
  2. If and , there is a reflection with .

Proof. (i) Since , Therefore

(ii) Since , we can choose a unit vector ; then . ∎

Note. The hypothesis in (ii) is needed: if , the only reflection is , which fixes only .

Theorem 9.5.12 (Orthogonal operators as products of reflections). Let be a real inner product space of dimension . Every orthogonal operator on is a product of at most reflections. (The identity is the empty product.)

Proof. Induction on . For , the only operator is the identity, the empty product.

For , let be orthogonal and pick a unit vector . Then . If , let . Otherwise, by Proposition 9.5.11(i) there is a reflection with ; let . In both cases is orthogonal and .

Since is -invariant, so is (Proposition 9.5.4(ii)), and is an orthogonal operator on the -dimensional space . By induction, for some reflections of with .

Extend each to by setting . Then is orthogonal (because and is orthogonal on ) and . Moreover, iff and , so is a line, and is a reflection.

Now , since both sides fix and agree on . If , we are done. If , then since , a product of at most reflections. ∎

9.5.3 The Classical Matrix Groups

The orthogonal and unitary operators, written as matrices in orthonormal bases, form important matrix groups.

Definition 9.5.13 (Orthogonal and special orthogonal groups).

  • , the orthogonal group.
  • , the special orthogonal group.

Definition 9.5.14 (Unitary and special unitary groups).

  • , the unitary group.
  • , the special unitary group.

Note. By Proposition 9.5.5, choosing an orthonormal basis of an -dimensional real inner product space identifies with , and similarly in the complex case. We write for the orthogonal operators of determinant .

Proposition 9.5.15 (Index of in ). For , is a normal subgroup of of index 2. We have an exact sequence

Proof. The determinant of an orthogonal matrix is , and is multiplicative, so is a homomorphism. It is surjective because and the reflection matrix has determinant . Its kernel is , which is therefore normal, and by the First Isomorphism Theorem (Theorem 3.3.3). ∎

Note. Elements of are called rotations or proper orthogonal transformations; they are the orthogonal transformations that preserve orientation, in the sense that they have positive determinant. Elements of , such as reflections, reverse orientation.

In dimension 2 everything can be made explicit. Write

Proposition 9.5.16 (The group ). Every element of is either a rotation (determinant ) or a matrix (determinant ), for a unique . The matrix is the reflection across the line spanned by : it fixes this vector and negates the orthogonal vector . If , the rotation has no real eigenvalues.

Proof. Let . By Proposition 9.5.5, its columns are orthonormal in . Since is a unit vector, for a unique . The vector is a unit vector in the line , so . The sign gives , with , and the sign gives , with .

Let . Using and , These two vectors form an orthonormal basis, so is the reflection with (Proposition 9.5.10).

Finally, the characteristic polynomial of is , whose discriminant is negative unless , i.e., unless . ∎

Proposition 9.5.17 ( is the circle group). , and is a well-defined group isomorphism from to the circle group .

Proof. By Proposition 9.5.16, the elements of are exactly the . Since is determined by its first column , we have iff ; so the map is well defined and injective. It is surjective because every complex number of absolute value is for some . By the addition formulas for sine and cosine, which corresponds to . So the map is a homomorphism. ∎

Example 9.5.18 (Low-dimensional cases).

  • , and .
  • (Proposition 9.5.17), while consists of the rotations and the reflections .
  • , acting on by multiplication.

The structure of a single orthogonal operator is now a consequence of the existence of small invariant subspaces.

Theorem 9.5.19 (Structure of orthogonal operators). Let be an orthogonal operator on a finite-dimensional real inner product space . Then is an orthogonal direct sum (the summands are pairwise orthogonal) where , , and each is a -dimensional -invariant subspace with an orthonormal basis in which has matrix for some .

Proof. Step 1: a decomposition into small blocks. We show by induction on that is an orthogonal direct sum of -invariant subspaces, each of which is either a line on which or a plane with an orthonormal basis in which has matrix with . If there is nothing to prove.

If , then has an invariant subspace of dimension or (Theorem 7.10.7). The restriction is orthogonal for the restricted inner product, (Theorem 9.1.31), and is -invariant by Proposition 9.5.4(ii), with orthogonal. By induction, decomposes as required, so it remains to decompose .

If , then acts on by an eigenvalue, which is by Proposition 9.5.4(i). If , choose an orthonormal basis of ; the matrix of lies in , so by Proposition 9.5.16 it is or . If it is , then is the orthogonal direct sum of two lines on which acts by and . If it is with , then and is the orthogonal direct sum of two lines. If it is with , replacing by changes the matrix to with .

Step 2: identifying . Let (resp. ) be the sum of the lines in Step 1 on which (resp. ), and the planes. Then . Conversely, is injective on each of the other blocks: it is on , and on its matrix has determinant . Since preserves each block, a vector with has of each component equal to , so . Thus , and similarly , using . ∎

Note. In matrix form: in an orthonormal basis adapted to this decomposition, has the block-diagonal matrix the real canonical form of an orthogonal matrix. This is a very good way to think about a single orthogonal transformation, built from rotations of individual planes and reflections; it says little about the composition of two orthogonal transformations, whose invariant subspaces are in general unrelated.

Exercise 9.5.20 (Inner product spaces).

  1. Let be the space of polynomials of degree at most 2, with inner product . Apply Gram–Schmidt to the basis to find an orthonormal basis.
  2. Let be self-adjoint on a finite-dimensional real inner product space, and suppose with maximizes the Rayleigh quotient among unit vectors. Prove that is an eigenvector of . (Do not use the spectral theorem.)
  3. Let be an invertible linear operator on a finite-dimensional real inner product space. Show that maps the unit sphere to an ellipsoid. (Hint: consider .)
  4. For an invertible operator on a finite-dimensional real inner product space, define the dilatation . Show that equals the ratio of the largest to the smallest eigenvalue of .
  5. Let be the matrix of an inner product on with respect to the standard basis, so that . For , prove that an entry of of largest absolute value occurs on the diagonal.
Show solution

(1) Start with , , .

, so .

, so , and . So .

and , so and . So .

The orthonormal basis is .

(2) Since for , maximizes over all nonzero vectors. Let and , and put (nonzero, since ). Since is self-adjoint and the space is real, , so Thus This differentiable function of has a maximum at , so its derivative there vanishes. The derivative at is . So for all .

Thus (Corollary 9.1.32), i.e., for some .

(3) We have . Now where is self-adjoint (Proposition 9.3.3) and satisfies for .

By the spectral theorem, there exists an orthonormal eigenbasis with , where .

For : .

So . Setting , this is , an ellipsoid with semi-axes along .

(4) We have . The operator is self-adjoint, and for (since is invertible).

By the spectral theorem, let be an orthonormal eigenbasis with ; then . For with (i.e., ): This weighted average of the is at most and at least , with equality when is a corresponding eigenvector. So and .

Thus .

(5) For , let . Positive-definiteness gives and , and, evaluating at , so .

If we had and , then , a contradiction. So : an entry of largest absolute value lies on the diagonal.