Adjoints and self-adjoint operators
In R^n, the identity (Ax)·y = x·(A^T y) moves a matrix across the dot product. In an abstract inner product space there are no entries to flip, so which map moves an operator across the inner product, and which operators are their own mirror image?
Where you are 5 in Inner product spaces
Why this exists
For a real matrix, (A\mathbf{x}) \cdot \mathbf{y} = \mathbf{x} \cdot (A^{T}\mathbf{y}) (The transpose and symmetric matrices): the transpose is what a matrix becomes when it crosses the dot product. A linear map between inner product spaces has no entries to flip, and its matrix changes with the basis.
The Riesz representation theorem of Orthonormal bases and Gram-Schmidt, proved gives every linear map a partner that moves it across the inner product. Operators equal to their partner generalize symmetric matrices; over \mathbb{C}, those that commute with it are the ones the spectral theorem diagonalizes.
Definition
Let V and W be finite-dimensional inner product spaces over \mathbb{F} and T \in \mathcal{L}(V, W). The adjoint of T is the function T^{*}: W \to V with
for all \mathbf{v} \in V and \mathbf{w} \in W (the left side is computed in W). Result 1 shows T^{*} exists, is unique and is linear. The field \mathbb{F} is itself an inner product space with \langle a, b \rangle = a\bar{b}.
An operator T \in \mathcal{L}(V) is self-adjoint when T^{*} = T, that is, \langle T\mathbf{v}, \mathbf{w} \rangle = \langle \mathbf{v}, T\mathbf{w} \rangle for all \mathbf{v}, \mathbf{w}. It is normal when TT^{*} = T^{*}T.
For A \in \mathbb{F}^{m \times n}, the conjugate transpose is A^{*} = \bar{A}^{T}: conjugate every entry, then flip. A square matrix with A^{*} = A is Hermitian; over \mathbb{R}, Hermitian means symmetric. The adjoint is not the dual map T' of Dual spaces, which acts on functionals and needs no inner product.
Intuition
The number \langle T\mathbf{v}, \mathbf{w} \rangle asks, in W, how far T\mathbf{v} points along \mathbf{w}; T^{*}\mathbf{w} is the vector of V that gives the same answer for every \mathbf{v} at once. In particular \mathbf{w} is orthogonal to everything T produces exactly when T^{*}\mathbf{w} = \mathbf{0}: the null space of T^{*} is the orthogonal complement of the range of T, as the figure shows.
Results
Throughout, U, V and W are finite-dimensional inner product spaces over \mathbb{F}.
Result 1 (the adjoint exists). For every T \in \mathcal{L}(V, W) there is exactly one function T^{*}: W \to V with \langle T\mathbf{v}, \mathbf{w} \rangle = \langle \mathbf{v}, T^{*}\mathbf{w} \rangle for all \mathbf{v}, \mathbf{w}, and T^{*} \in \mathcal{L}(W, V).
Result 2 (rules). For S, T \in \mathcal{L}(V, W), R \in \mathcal{L}(W, U) and \lambda \in \mathbb{F}:
If T is invertible, so is T^{*}, and (T^{*})^{-1} = (T^{-1})^{*}.
Result 3 (the matrix). Let \mathcal{B} be an orthonormal basis of V and \mathcal{C} an orthonormal basis of W. Then
Result 4 (null spaces and ranges). For T \in \mathcal{L}(V, W),
So \operatorname{rank} T^{*} = \operatorname{rank} T, and T is injective exactly when T^{*} is surjective.
Result 5 (a zero quadratic form). Let T \in \mathcal{L}(V) with \langle T\mathbf{v}, \mathbf{v} \rangle = 0 for every \mathbf{v} \in V. If \mathbb{F} = \mathbb{C}, then T = 0. If \mathbb{F} = \mathbb{R} and T is self-adjoint, then T = 0.
Result 6 (self-adjoint operators). Let T \in \mathcal{L}(V) be self-adjoint. Then \langle T\mathbf{v}, \mathbf{v} \rangle is real for every \mathbf{v}, and every eigenvalue of T is real. Conversely, if \mathbb{F} = \mathbb{C} and \langle T\mathbf{v}, \mathbf{v} \rangle \in \mathbb{R} for every \mathbf{v}, then T is self-adjoint.
Result 7 (normal operators). T \in \mathcal{L}(V) is normal if and only if \lVert T\mathbf{v} \rVert = \lVert T^{*}\mathbf{v} \rVert for every \mathbf{v}. If T is normal, then T\mathbf{v} = \lambda\mathbf{v} exactly when T^{*}\mathbf{v} = \bar{\lambda}\mathbf{v}, and eigenvectors of T for distinct eigenvalues are orthogonal. Every self-adjoint operator is normal, so all of this holds for it.
Why it is true
One fact is used again and again: if \langle \mathbf{v}, \mathbf{u} \rangle = \langle \mathbf{v}, \mathbf{u}' \rangle for every \mathbf{v}, then \mathbf{u} = \mathbf{u}' (take \mathbf{v} = \mathbf{u} - \mathbf{u}', which then has norm 0). So to identify T^{*}\mathbf{w} it is enough to show some vector satisfies the defining identity.
Result 1. Fix \mathbf{w} \in W. The map \mathbf{v} \mapsto \langle T\mathbf{v}, \mathbf{w} \rangle is a linear functional on V, since T is linear and the inner product is linear in the first slot. By the Riesz representation theorem (Result 7 of Orthonormal bases and Gram-Schmidt, proved) exactly one vector of V represents it; call it T^{*}\mathbf{w}. For linearity, conjugate homogeneity in the second slot (Result 2 of Inner products and norms) gives, for all \mathbf{v},
so T^{*}(\mathbf{w}_1 + \lambda\mathbf{w}_2) = T^{*}\mathbf{w}_1 + \lambda T^{*}\mathbf{w}_2.
Result 2. Each right side satisfies the defining identity, so it is the adjoint. For instance \langle (\lambda T)\mathbf{v}, \mathbf{w} \rangle = \lambda\langle \mathbf{v}, T^{*}\mathbf{w} \rangle = \langle \mathbf{v}, \bar{\lambda}T^{*}\mathbf{w} \rangle, and \langle RT\mathbf{v}, \mathbf{u} \rangle = \langle T\mathbf{v}, R^{*}\mathbf{u} \rangle = \langle \mathbf{v}, T^{*}R^{*}\mathbf{u} \rangle. For (T^{*})^{*}, conjugate symmetry gives
Sums and I^{*} = I are the same one-line check. If T is invertible, apply the product rule to T^{-1}T = I and TT^{-1} = I: T^{*}(T^{-1})^{*} = I and (T^{-1})^{*}T^{*} = I.
Result 3. Write \mathcal{B} = (\mathbf{b}_1, \dots, \mathbf{b}_n) and \mathcal{C} = (\mathbf{c}_1, \dots, \mathbf{c}_m). Coordinates in an orthonormal basis are inner products (Result 3 of Orthonormal bases and Gram-Schmidt, proved), so entry (j, k) of [T]_{\mathcal{C} \leftarrow \mathcal{B}} is \langle T\mathbf{b}_k, \mathbf{c}_j \rangle, and entry (k, j) of [T^{*}]_{\mathcal{B} \leftarrow \mathcal{C}} is
Moving entry (j, k) to place (k, j) and conjugating is the conjugate transpose.
Result 4. \mathbf{w} \in \ker T^{*} means T^{*}\mathbf{w} = \mathbf{0}, which holds exactly when \langle \mathbf{v}, T^{*}\mathbf{w} \rangle = 0 for all \mathbf{v}, that is, \langle T\mathbf{v}, \mathbf{w} \rangle = 0 for all \mathbf{v}: \mathbf{w} \in (\operatorname{im} T)^{\perp}. Applying this to T^{*} and using (T^{*})^{*} = T gives \ker T = (\operatorname{im} T^{*})^{\perp}. Taking orthogonal complements of both equations and using (X^{\perp})^{\perp} = X for subspaces of a finite-dimensional space (Result 6 of Orthonormal bases and Gram-Schmidt, proved) gives the other two. Then \dim \operatorname{im} T^{*} = \dim V - \dim \ker T, which is \operatorname{rank} T by the fundamental theorem of linear maps (Null space, range and the fundamental theorem). Finally \ker T = \{\mathbf{0}\} exactly when \operatorname{im} T^{*} = \{\mathbf{0}\}^{\perp} = V.
Result 5. Over \mathbb{C}, expanding each term by additivity and conjugate homogeneity checks that for all \mathbf{u}, \mathbf{w},
(The first two terms contribute 2\langle T\mathbf{u}, \mathbf{w} \rangle + 2\langle T\mathbf{w}, \mathbf{u} \rangle, the last two 2\langle T\mathbf{u}, \mathbf{w} \rangle - 2\langle T\mathbf{w}, \mathbf{u} \rangle.) Every term on the right is 0, so \langle T\mathbf{u}, \mathbf{w} \rangle = 0; with \mathbf{w} = T\mathbf{u}, T\mathbf{u} = \mathbf{0}. Over \mathbb{R} with T self-adjoint, \langle T\mathbf{w}, \mathbf{u} \rangle = \langle \mathbf{w}, T\mathbf{u} \rangle = \langle T\mathbf{u}, \mathbf{w} \rangle, so the first two terms alone equal 4\langle T\mathbf{u}, \mathbf{w} \rangle, and the same ending works.
Result 6. \overline{\langle T\mathbf{v}, \mathbf{v} \rangle} = \langle \mathbf{v}, T\mathbf{v} \rangle = \langle T\mathbf{v}, \mathbf{v} \rangle, so the number is real. If T\mathbf{v} = \lambda\mathbf{v} with \mathbf{v} \neq \mathbf{0}, then \lambda\lVert \mathbf{v} \rVert^2 = \langle T\mathbf{v}, \mathbf{v} \rangle is real, so \lambda is. Conversely, over \mathbb{C}, \overline{\langle T\mathbf{v}, \mathbf{v} \rangle} = \langle \mathbf{v}, T\mathbf{v} \rangle = \langle T^{*}\mathbf{v}, \mathbf{v} \rangle, so if every \langle T\mathbf{v}, \mathbf{v} \rangle is real, then \langle (T - T^{*})\mathbf{v}, \mathbf{v} \rangle = 0 for every \mathbf{v}, and Result 5 gives T = T^{*}.
Result 7. For every \mathbf{v},
The operator T^{*}T - TT^{*} is self-adjoint by Result 2, so by Result 5 (over either field) the left side vanishes for all \mathbf{v} exactly when T^{*}T = TT^{*}. If T is normal, so is T - \lambda I, whose adjoint is T^{*} - \bar{\lambda}I: expanding, both products equal T^{*}T - \lambda T^{*} - \bar{\lambda}T + |\lambda|^2 I. Hence \lVert (T - \lambda I)\mathbf{v} \rVert = \lVert (T^{*} - \bar{\lambda}I)\mathbf{v} \rVert, and one side is 0 exactly when the other is. If T\mathbf{u} = \alpha\mathbf{u} and T\mathbf{v} = \beta\mathbf{v} with \alpha \neq \beta, then
so \langle \mathbf{u}, \mathbf{v} \rangle = 0. A self-adjoint T has TT^{*} = T^2 = T^{*}T.
Worked examples
An adjoint from the definition
Let T: \mathbb{C}^2 \to \mathbb{C}^2 send (z_1, z_2) to (2z_1 + iz_2,\ (1 - i)z_1 + 3z_2), with the Euclidean inner product. Find T^{*} from the definition, and confirm it with Result 3.
SolutionFor \mathbf{w} = (w_1, w_2), expand and collect the coefficients of z_1 and z_2:
The brackets are the conjugates of 2w_1 + (1 + i)w_2 and -iw_1 + 3w_2, so the sum is \langle \mathbf{z}, T^{*}\mathbf{w} \rangle with
The standard basis is orthonormal, so by Result 3 the matrix of T^{*} is A^{*} for A = \begin{bmatrix} 2 & i \\ 1 - i & 3 \end{bmatrix}. Conjugating and flipping gives A^{*} = \begin{bmatrix} 2 & 1 + i \\ -i & 3 \end{bmatrix}, the same map. T is not self-adjoint: the (1, 2) entry of A is i, that of A^{*} is 1 + i.
When the basis is not orthonormal
On \mathbb{R}^2 take \langle \mathbf{x}, \mathbf{y} \rangle = 2x_1y_1 + x_2y_2, and T\mathbf{x} = A\mathbf{x} with A = \begin{bmatrix} 1 & 1 \\ 0 & 1 \end{bmatrix}. Find T^{*}. Is its standard matrix A^{T}?
SolutionThe weights are positive, so this is an inner product. From the definition,
That is 2x_1u_1 + x_2u_2 = \langle \mathbf{x}, \mathbf{u} \rangle for \mathbf{u} = (y_1, 2y_1 + y_2). So T^{*}(y_1, y_2) = (y_1, 2y_1 + y_2), with standard matrix \begin{bmatrix} 1 & 0 \\ 2 & 1 \end{bmatrix}, not A^{T} = \begin{bmatrix} 1 & 0 \\ 1 & 1 \end{bmatrix}.
Result 3 is not broken: \displaystyle \lVert \mathbf{e}_1 \rVert = \sqrt{2} here, so the standard basis is not orthonormal. In the orthonormal basis \displaystyle \mathcal{B} = (\mathbf{e}_1/\sqrt{2}, \mathbf{e}_2), \displaystyle T(\mathbf{e}_1/\sqrt{2}) = \mathbf{e}_1/\sqrt{2} and \displaystyle T\mathbf{e}_2 = (1, 1) = \sqrt{2}\,(\mathbf{e}_1/\sqrt{2}) + \mathbf{e}_2, so
Its first column says T^{*}\mathbf{e}_1 = \mathbf{e}_1 + 2\mathbf{e}_2, as found above.
The adjoint of differentiation
On \mathcal{P}_2(\mathbb{R}) with \langle p, q \rangle = \int_0^1 p(x)q(x)\,dx, let Dp = p'. Find the matrix of D^{*} in an orthonormal basis, and D^{*}1.
SolutionExample 1 of Orthonormal bases and Gram-Schmidt, proved gives the orthonormal basis q_1 = 1, \displaystyle q_2 = \sqrt{3}(2x - 1), \displaystyle q_3 = \sqrt{5}(6x^2 - 6x + 1). Differentiating, Dq_1 = 0, \displaystyle Dq_2 = 2\sqrt{3}\,q_1, and \displaystyle Dq_3 = 6\sqrt{5}(2x - 1) = 2\sqrt{15}\,q_2. So by Result 3, with all entries real,
The first column of [D^{*}] gives \displaystyle D^{*}1 = 2\sqrt{3}\,q_2 = 12x - 6. Check: by the The fundamental theorem of calculus, \langle Dp, 1 \rangle = p(1) - p(0), which is 0, 1, 1 for p = 1, x, x^2, and \int_0^1 (12x - 6)p\,dx is 0, 4 - 3 = 1 and 3 - 2 = 1. Since D^{*}1 \neq 0 = D1, D is not self-adjoint.
A Hermitian matrix
Let A = \begin{bmatrix} 2 & 1 - i \\ 1 + i & 3 \end{bmatrix} act on \mathbb{C}^2. Show that A is Hermitian, find its eigenvalues and eigenvectors, and check Results 6 and 7.
SolutionConjugating gives \begin{bmatrix} 2 & 1 + i \\ 1 - i & 3 \end{bmatrix}, and flipping returns A, so A^{*} = A and \mathbf{z} \mapsto A\mathbf{z} is self-adjoint (Result 3). The The characteristic equation and finding eigenvalues is
so \lambda = 1 or \lambda = 4, both real, as Result 6 requires. Row reducing A - I gives the eigenvector \mathbf{v}_1 = (1 - i, -1); row reducing A - 4I gives \mathbf{v}_4 = (1 - i, 2). They are orthogonal (Result 7):
The cross terms (1 - i)z_2\bar{z}_1 and (1 + i)z_1\bar{z}_2 of the quadratic form are conjugates, so it is real:
Dividing by the norms \displaystyle \sqrt{3} and \displaystyle \sqrt{6} gives an orthonormal basis in which the matrix is \operatorname{diag}(1, 4).
Self-adjoint from any map
Let T \in \mathcal{L}(V, W). Prove that T^{*}T is self-adjoint, that \langle T^{*}T\mathbf{v}, \mathbf{v} \rangle \ge 0 for every \mathbf{v}, and that \ker T^{*}T = \ker T, so \operatorname{rank} T^{*}T = \operatorname{rank} T.
SolutionBy Result 2, (T^{*}T)^{*} = T^{*}(T^{*})^{*} = T^{*}T. By the definition of T^{*},
If T\mathbf{v} = \mathbf{0}, then T^{*}T\mathbf{v} = \mathbf{0}. Conversely, if T^{*}T\mathbf{v} = \mathbf{0}, then \lVert T\mathbf{v} \rVert^2 = \langle T^{*}T\mathbf{v}, \mathbf{v} \rangle = 0, so T\mathbf{v} = \mathbf{0}. Equal null spaces in the same domain V give equal ranks by the fundamental theorem of linear maps. Also, if T^{*}T\mathbf{v} = \lambda\mathbf{v} with \mathbf{v} \neq \mathbf{0}, then \lambda\lVert \mathbf{v} \rVert^2 = \lVert T\mathbf{v} \rVert^2, so every eigenvalue of T^{*}T is real and at least 0.
A looper that is normal
Leah and Diego program a looper, a box that replays a loop of four samples \mathbf{x} = (x_1, x_2, x_3, x_4) \in \mathbb{R}^4 over and over, so x_1 follows x_4. Their effect adds half of the previous sample: \displaystyle E\mathbf{x} = \mathbf{x} + \frac{1}{2}S\mathbf{x} with S(x_1, x_2, x_3, x_4) = (x_4, x_1, x_2, x_3). With the dot product, find E^{*} and decide whether E is self-adjoint and whether it is normal. Then do the same for the effect without the loop,
From the definition,
so S^{*}(y_1, y_2, y_3, y_4) = (y_2, y_3, y_4, y_1), the shift the other way, and S^{*}S = SS^{*} = I. By Result 2, \displaystyle E^{*} = I + \frac{1}{2}S^{*}: the adjoint adds half of the next sample. It is not self-adjoint, since \displaystyle E\mathbf{e}_1 = (1, \frac{1}{2}, 0, 0) while \displaystyle E^{*}\mathbf{e}_1 = (1, 0, 0, \frac{1}{2}). It is normal, because both products expand the same way:
As Result 7 predicts, E\mathbf{e}_1 and E^{*}\mathbf{e}_1 differ but both have norm \displaystyle \frac{\sqrt{5}}{2}.
The matrix of F has 1 on the diagonal and \displaystyle \frac{1}{2} just below it, so its transpose gives
Now F\mathbf{e}_4 = (0, 0, 0, 1) has norm 1 and \displaystyle F^{*}\mathbf{e}_4 = (0, 0, \frac{1}{2}, 1) has norm \displaystyle \frac{\sqrt{5}}{2}, so by Result 7 F is not normal.
Boundaries
- Finite dimension is used. On \mathcal{P}(\mathbb{R}) with \langle p, q \rangle = \int_0^1 pq\,dx, let Tp be the constant polynomial p(0). Then \langle Tp, 1 \rangle = p(0), so T^{*}1 would represent evaluation at 0, which Exercise 17 of Orthonormal bases and Gram-Schmidt, proved rules out. This T has no adjoint.
- Result 3 needs orthonormal bases (Example 2), and being self-adjoint depends on the inner product: with Example 2’s inner product, the symmetric matrix \begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix} has adjoint \displaystyle (y_1, y_2) \mapsto (\frac{1}{2}y_2, 2y_1). Over \mathbb{C} the test is A^{*} = A, not A^{T} = A (Exercise 6).
- Over \mathbb{R}, Result 5 needs self-adjointness, and the converse in Result 6 fails. The rotation T(x, y) = (-y, x) of \mathbb{R}^2 has \langle T\mathbf{v}, \mathbf{v} \rangle = 0 for every \mathbf{v}, yet T \neq 0.
Where this goes next
The The spectral theorem proves that a self-adjoint operator on a real space, or a normal one on a complex space, has an orthonormal basis of eigenvectors, as Example 4 found by hand. It uses Results 6 and 7 and Exercise 13. The Singular value decomposition starts from Example 5: the eigenvalues of T^{*}T are at least 0, and their square roots are the singular values. Result 4 is also why Least squares: the best approximate solution of Ax = b solve A^{*}A\mathbf{x} = A^{*}\mathbf{b}: the error A\mathbf{x} - \mathbf{b} must lie in (\operatorname{im} A)^{\perp} = \ker A^{*}.
Exercises
Computing adjoints
- Let T: \mathbb{C}^2 \to \mathbb{C}^3 be T(z_1, z_2) = (z_1 + iz_2,\ 2z_2,\ (1 + i)z_1), with Euclidean inner products. Find T^{*}.
- Let \varphi: \mathbb{C}^3 \to \mathbb{C} be \varphi(\mathbf{z}) = z_1 + iz_3. Find \varphi^{*}: \mathbb{C} \to \mathbb{C}^3.
- On \mathbb{R}^2 with \langle \mathbf{x}, \mathbf{y} \rangle = x_1y_1 + 3x_2y_2, let T\mathbf{x} = A\mathbf{x} with A = \begin{bmatrix} 1 & 2 \\ 0 & 1 \end{bmatrix}. Find the standard matrix of T^{*}.
- On \mathcal{P}_1(\mathbb{R}) with \langle p, q \rangle = \int_{-1}^{1} p(x)q(x)\,dx, let T(a + bx) = b + ax. Find T^{*}. Is T self-adjoint?
- On \mathcal{P}_2(\mathbb{R}) with \langle p, q \rangle = p(-1)q(-1) + p(0)q(0) + p(1)q(1), let Dp = p'. Find D^{*}1 and D^{*}x.
Self-adjoint and normal operators
- For each matrix acting on \mathbb{C}^2, decide whether it is self-adjoint and whether it is normal: A = \begin{bmatrix} 1 & 2i \\ -2i & 4 \end{bmatrix}, B = \begin{bmatrix} 1 & 2i \\ 2i & 4 \end{bmatrix}, C = \begin{bmatrix} 1 & -1 \\ 1 & 1 \end{bmatrix}.
- Find the eigenvalues and eigenvectors of \begin{bmatrix} 3 & 2i \\ -2i & 0 \end{bmatrix} and check that the eigenvectors are orthogonal.
- Show that A = \begin{bmatrix} 2 & -3 \\ 3 & 2 \end{bmatrix} is normal on \mathbb{C}^2, that (1, -i) is an eigenvector of A, and that it is an eigenvector of A^{*} for the conjugate eigenvalue.
- For which real a and b is \begin{bmatrix} 1 & a \\ b & 1 \end{bmatrix} normal on \mathbb{R}^2?
Null spaces and ranges
- Let T: \mathbb{C}^3 \to \mathbb{C}^2 have matrix \begin{bmatrix} 1 & i & 0 \\ 0 & 1 & 1 + i \end{bmatrix}. Find \ker T and \operatorname{im} T^{*} and check Result 4. Is \ker T orthogonal to the rows of the matrix without conjugating them?
- For D of Example 3, find \ker D^{*} from Result 4 without computing D^{*}, then confirm it with the matrix.
- Let T\mathbf{x} = A\mathbf{x} on \mathbb{R}^2 with A = \begin{bmatrix} 2 & -1 \\ -6 & 3 \end{bmatrix}. Find \operatorname{im} T and \ker T^{*}, and write (1, 1) as a vector of \operatorname{im} T plus one of \ker T^{*}.
Proofs
In these, V is a finite-dimensional inner product space.
- Let T \in \mathcal{L}(V) be self-adjoint and U a subspace invariant under T. Prove that U^{\perp} is invariant under T.
- Let V be real. Prove that \langle T\mathbf{v}, \mathbf{v} \rangle = 0 for every \mathbf{v} if and only if T^{*} = -T.
- Prove that a normal T has \ker T = \ker T^{*} and \operatorname{im} T = \operatorname{im} T^{*}.
- Prove that \lambda is an eigenvalue of T \in \mathcal{L}(V) if and only if \bar{\lambda} is an eigenvalue of T^{*}.
- Let P \in \mathcal{L}(V) with P^2 = P. Prove that P = \operatorname{proj}_{\operatorname{im} P} if and only if P is self-adjoint.
- Let G be a symmetric n \times n real matrix for which \langle \mathbf{x}, \mathbf{y} \rangle = \mathbf{y}^{T}G\mathbf{x} is an inner product on \mathbb{R}^n. Prove that the adjoint of \mathbf{x} \mapsto A\mathbf{x} is \mathbf{y} \mapsto G^{-1}A^{T}G\mathbf{y}.
- Prove that the self-adjoint operators on V form a subspace of \mathcal{L}(V) over \mathbb{R}, that over \mathbb{C} with V \neq \{\mathbf{0}\} they are not closed under multiplication by i, and that for self-adjoint S and T, ST is self-adjoint if and only if ST = TS.
Answers
- The standard bases are orthonormal, so by Result 3 the matrix of T^{*} is the conjugate transpose of \begin{bmatrix} 1 & i \\ 0 & 2 \\ 1 + i & 0 \end{bmatrix}, namely \begin{bmatrix} 1 & 0 & 1 - i \\ -i & 2 & 0 \end{bmatrix}. So T^{*} sends (w_1, w_2, w_3) to (w_1 + (1 - i)w_3,\ -iw_1 + 2w_2).
- \varphi^{*}(c) = (c, 0, -ic). Indeed \langle \varphi(\mathbf{z}), c \rangle = (z_1 + iz_3)\bar{c}, which equals z_1\bar{c} + z_3\,\overline{-ic}, that is, \langle \mathbf{z}, (c, 0, -ic) \rangle.
- \langle A\mathbf{x}, \mathbf{y} \rangle = (x_1 + 2x_2)y_1 + 3x_2y_2, which is \displaystyle x_1y_1 + 3x_2\big(\frac{2}{3}y_1 + y_2\big). So \displaystyle T^{*}(y_1, y_2) = (y_1, \frac{2}{3}y_1 + y_2), with matrix \displaystyle \begin{bmatrix} 1 & 0 \\ \frac{2}{3} & 1 \end{bmatrix}, not A^{T}.
- An orthonormal basis is \displaystyle q_1 = \frac{1}{\sqrt{2}}, \displaystyle q_2 = \sqrt{\frac{3}{2}}\,x. Then \displaystyle Tq_1 = \frac{x}{\sqrt{2}} = \frac{1}{\sqrt{3}}q_2 and \displaystyle Tq_2 = \sqrt{\frac{3}{2}} = \sqrt{3}\,q_1, so \displaystyle [T] = \begin{bmatrix} 0 & \sqrt{3} \\ \frac{1}{\sqrt{3}} & 0 \end{bmatrix} and [T^{*}] is its transpose. This gives T^{*}1 = 3x and \displaystyle T^{*}x = \frac{1}{3}, so \displaystyle T^{*}(a + bx) = \frac{b}{3} + 3ax. Not self-adjoint: T^{*}1 = 3x \neq x = T1.
- As in Exercise 3 of Orthonormal bases and Gram-Schmidt, proved, \displaystyle q_1 = \frac{1}{\sqrt{3}}, \displaystyle q_2 = \frac{x}{\sqrt{2}}, \displaystyle q_3 = \frac{3x^2 - 2}{\sqrt{6}} is orthonormal. \displaystyle Dq_2 = \sqrt{\frac{3}{2}}\,q_1 and \displaystyle Dq_3 = \sqrt{6}\,x = 2\sqrt{3}\,q_2, so \displaystyle D^{*}q_1 = \sqrt{\frac{3}{2}}\,q_2 and \displaystyle D^{*}q_2 = 2\sqrt{3}\,q_3. Then \displaystyle D^{*}1 = \sqrt{3}\,D^{*}q_1 = \frac{3}{2}x and \displaystyle D^{*}x = \sqrt{2}\,D^{*}q_2 = 6x^2 - 4.
- A^{*} = A: self-adjoint, hence normal. B^{*} = \begin{bmatrix} 1 & -2i \\ -2i & 4 \end{bmatrix} \neq B, and BB^{*} has (1, 2) entry 6i while B^{*}B has -6i: neither. C^{*} = C^{T} \neq C but CC^{T} = C^{T}C = 2I: normal, not self-adjoint.
- The characteristic equation is \lambda^2 - 3\lambda - 4 = 0, so \lambda = 4 with eigenvector (2i, 1) and \lambda = -1 with eigenvector (1, 2i). Their inner product is 2i \cdot 1 + 1 \cdot \overline{2i} = 2i - 2i = 0.
- AA^{T} = A^{T}A = 13I. Next A(1, -i) = (2 + 3i, 3 - 2i), which is (2 + 3i)(1, -i). With A^{*} = A^{T}, A^{T}(1, -i) = (2 - 3i, -3 - 2i), which is (2 - 3i)(1, -i).
- AA^{T} - A^{T}A = \begin{bmatrix} a^2 - b^2 & 0 \\ 0 & b^2 - a^2 \end{bmatrix}, so it is normal exactly when b = a (symmetric) or b = -a.
- \ker T = \operatorname{span}\{(-1 + i, -1 - i, 1)\}, from z_2 = -(1 + i)z_3 and z_1 = -iz_2. \operatorname{im} T^{*} is spanned by the columns of A^{*}, (1, -i, 0) and (0, 1, 1 - i), and their inner products with the null vector are (-1 + i) + (-1 - i)i = 0 and (-1 - i) + (1 + i) = 0. Both \operatorname{im} T^{*} and (\ker T)^{\perp} have dimension 2, so they are equal. Without conjugating, the first inner product becomes -2 + 2i \neq 0.
- \operatorname{im} D = \mathcal{P}_1(\mathbb{R}), since D(x) = 1 and \displaystyle D(\frac{1}{2}x^2) = x. So \ker D^{*} = \mathcal{P}_1(\mathbb{R})^{\perp} = \operatorname{span}\{6x^2 - 6x + 1\}. In [D^{*}] the third column is zero, so D^{*}q_3 = 0, and its rank is 2.
- \operatorname{im} T = \operatorname{span}\{(1, -3)\} and \ker T^{*} = \ker A^{T} = \operatorname{span}\{(3, 1)\}. Projecting, \displaystyle (1, 1) = \left(-\frac{1}{5}, \frac{3}{5}\right) + \left(\frac{6}{5}, \frac{2}{5}\right).
- Let \mathbf{w} \in U^{\perp} and \mathbf{u} \in U. Then T\mathbf{u} \in U, so \langle T\mathbf{w}, \mathbf{u} \rangle = \langle \mathbf{w}, T\mathbf{u} \rangle = 0. As \mathbf{u} was arbitrary, T\mathbf{w} \in U^{\perp}.
- A real inner product is symmetric. If T^{*} = -T, then \langle T\mathbf{v}, \mathbf{v} \rangle = \langle \mathbf{v}, -T\mathbf{v} \rangle = -\langle T\mathbf{v}, \mathbf{v} \rangle, so it is 0. Conversely, S = T + T^{*} is self-adjoint by Result 2, and \langle S\mathbf{v}, \mathbf{v} \rangle = \langle T\mathbf{v}, \mathbf{v} \rangle + \langle \mathbf{v}, T\mathbf{v} \rangle = 0, so S = 0 by Result 5.
- By Result 7, \lVert T\mathbf{v} \rVert = \lVert T^{*}\mathbf{v} \rVert, so T\mathbf{v} = \mathbf{0} exactly when T^{*}\mathbf{v} = \mathbf{0}. Then Result 4 gives \operatorname{im} T = (\ker T^{*})^{\perp} = (\ker T)^{\perp} = \operatorname{im} T^{*}.
- By Result 2, (T - \lambda I)^{*} = T^{*} - \bar{\lambda}I, and by Result 4 the two have the same rank. An operator on finite-dimensional V is injective exactly when its rank is \dim V (fundamental theorem of linear maps). So T - \lambda I fails to be injective exactly when T^{*} - \bar{\lambda}I does.
- Let U = \operatorname{im} P. If P = \operatorname{proj}_U, then for all \mathbf{u}, \mathbf{v}, since \mathbf{v} - P\mathbf{v} \in U^{\perp} and P\mathbf{u} \in U, \langle P\mathbf{u}, \mathbf{v} \rangle = \langle P\mathbf{u}, P\mathbf{v} \rangle; likewise \langle \mathbf{u}, P\mathbf{v} \rangle = \langle P\mathbf{u}, P\mathbf{v} \rangle, so P^{*} = P. Conversely, let P^{*} = P. Write \mathbf{v} = P\mathbf{v} + (\mathbf{v} - P\mathbf{v}). Here P\mathbf{v} \in U and P(\mathbf{v} - P\mathbf{v}) = P\mathbf{v} - P^2\mathbf{v} = \mathbf{0}, so \mathbf{v} - P\mathbf{v} \in \ker P = \ker P^{*} = U^{\perp} by Result 4. The splitting V = U \oplus U^{\perp} is unique (Result 6 of Orthonormal bases and Gram-Schmidt, proved), so P\mathbf{v} = \operatorname{proj}_U \mathbf{v}.
- G is invertible: if G\mathbf{y} = \mathbf{0}, then \langle \mathbf{y}, \mathbf{y} \rangle = \mathbf{y}^{T}G\mathbf{y} = 0, so \mathbf{y} = \mathbf{0}. Let B = G^{-1}A^{T}G. As G^{T} = G and (G^{-1})^{T} = G^{-1} (The transpose and symmetric matrices), B^{T} = GAG^{-1}, so \langle \mathbf{x}, B\mathbf{y} \rangle = \mathbf{y}^{T}B^{T}G\mathbf{x} = \mathbf{y}^{T}GA\mathbf{x} = \langle A\mathbf{x}, \mathbf{y} \rangle. By the uniqueness in Result 1, \mathbf{y} \mapsto B\mathbf{y} is the adjoint.
- 0^{*} = 0, and for self-adjoint S, T and real a, b, Result 2 gives (aS + bT)^{*} = \bar{a}S + \bar{b}T = aS + bT. Over \mathbb{C}, I is self-adjoint but (iI)^{*} = -iI \neq iI. Finally (ST)^{*} = T^{*}S^{*} = TS, which equals ST exactly when S and T commute.