Linear Operators and Adjoints

In an inner-product space, the vectors are the things — states, functions, arrows. The operators are the verbs: the machines that act on those things and turn one vector into another. Rotate a state, differentiate a function, multiply by position, measure an energy — in the language of Hilbert space every one of these is a linear operator. If vectors are the nouns of quantum mechanics, operators are its physics: every observable quantity you can measure is an operator, and this page builds the one property that makes an operator measurable.

That property hinges on a beautiful partner construction. Every operator A has a shadow twin called its adjoint A^\dagger ("A-dagger"), defined entirely through the inner product. When an operator equals its own adjoint — A = A^\dagger, called self-adjoint or Hermitian — something magical happens: its eigenvalues are forced to be real numbers. Since a measurement in the lab always reads out a real number, this is exactly why the observables of quantum mechanics are Hermitian operators. Let us build it.

What a linear operator is

A linear operator A on a space V is a map V \to V that respects the vector-space structure — it commutes with adding and scaling:

A(\alpha u + \beta v) = \alpha\,Au + \beta\,Av.

In finite dimensions, once you fix a basis, an operator is just a matrix and Av is matrix-times-vector — nothing new. The word "operator" is what we use when we want to keep the coordinate-free viewpoint, or when the space is infinite-dimensional and no finite matrix exists. Two operators worth meeting from quantum mechanics: the position operator \hat x that multiplies a wavefunction by x, and the momentum operator \hat p = -i\hbar\,\tfrac{d}{dx} that differentiates it. Both are linear; neither is a finite matrix.

The adjoint: sliding the operator across the inner product

Here is the key move, and it is worth reading slowly. Given an operator A, its adjoint A^\dagger is the unique operator that lets you move A from one side of an inner product to the other:

\langle A u,\, v\rangle = \langle u,\, A^\dagger v\rangle \quad\text{for all } u, v.

Read it as a rule for "unloading" A: whatever A does to the left slot, its adjoint A^\dagger does the equivalent thing from the right slot. The inner product is the referee that defines what "equivalent" means. In finite dimensions with the complex dot product, this abstract definition has a wonderfully concrete answer: the adjoint is the conjugate transpose — transpose the matrix and conjugate every entry,

A^\dagger = \overline{A^{\mathsf T}} = (\bar A)^{\mathsf T}.

(For a real matrix there is no conjugation to do, so the adjoint is simply the transpose A^{\mathsf T}.) The dagger just packages "transpose and conjugate" into one symbol, and the inner-product equation above is the reason those two operations belong together.

Self-adjoint (Hermitian) operators

An operator that is its own adjoint is the star of the show:

A is self-adjoint (or Hermitian) when A = A^\dagger, i.e. \langle A u, v\rangle = \langle u, A v\rangle for all u, v.
In matrix terms this is A = \overline{A^{\mathsf T}}: the entries satisfy a_{ij} = \overline{a_{ji}}, so the matrix is a mirror image of itself across the diagonal, with conjugation. In particular the diagonal entries a_{ii} = \overline{a_{ii}} must be real.
For real matrices, "Hermitian" is just symmetric, A = A^{\mathsf T}.

The 2×2 grid shows a matrix and its conjugate transpose side by side. Toggle the entries and watch the reflect-and-conjugate operation. When the "before" and "after" grids match, the matrix is Hermitian — its diagonal is real, and its off-diagonal entries are conjugate mirror partners.

Left: a 2×2 complex matrix A. Right: its adjoint A^\dagger, the transpose with every entry conjugated. The matrix is Hermitian exactly when the two grids are identical — real diagonal, mirror-conjugate off-diagonal.

The punchline: Hermitian ⟹ real eigenvalues

Why do physicists insist that observables be Hermitian? Because of a short, elegant argument that forces the eigenvalues to be real. Suppose A is Hermitian and Av = \lambda v for some non-zero eigenvector v. Compute \langle A v, v\rangle two ways.

Down the left slot, using Av = \lambda v and that the first slot is conjugate-linear: \langle Av, v\rangle = \langle \lambda v, v\rangle = \overline{\lambda}\,\langle v, v\rangle.

Down the right slot, using self-adjointness \langle Av, v\rangle = \langle v, Av\rangle and then Av = \lambda v: \langle v, Av\rangle = \langle v, \lambda v\rangle = \lambda\,\langle v, v\rangle.

The two results must agree, and since \langle v, v\rangle = \lVert v\rVert^2 > 0 (the eigenvector is non-zero), we can divide it out:

\overline{\lambda} = \lambda \quad\Longrightarrow\quad \lambda \text{ is real.}

A number equal to its own conjugate has no imaginary part. So every eigenvalue of a Hermitian operator is a real number — which is precisely the guarantee a measurement needs. There is a companion fact, equally important, that comes from the same style of argument: eigenvectors of a Hermitian operator belonging to different eigenvalues are automatically orthogonal. Put the two together — real eigenvalues, orthogonal eigenvectors — and you have the spectral theorem, the crown jewel that says a Hermitian operator can be diagonalised in an orthonormal basis of its own eigenvectors.

Worked example — a Hermitian 2×2 and its real spectrum

Consider the matrix

A = \begin{bmatrix} 2 & 3 - i \\ 3 + i & 5 \end{bmatrix}.

Is it Hermitian? Take the transpose (swap off-diagonals) and conjugate every entry. The diagonal 2, 5 is real, and the off-diagonal 3 - i conjugates to 3 + i, which is exactly the mirror entry — so A^\dagger = A. Yes, Hermitian. Its eigenvalues solve \det(A - \lambda I) = 0:

(2-\lambda)(5-\lambda) - (3-i)(3+i) = 0.

Now (3-i)(3+i) = 9 - i^2 = 9 + 1 = 10 (a difference of squares — the conjugate pair multiplies to a real number, no surprise). So \lambda^2 - 7\lambda + 10 - 10 = \lambda^2 - 7\lambda = 0, giving \lambda = 0 and \lambda = 7 — both real, exactly as the theorem promised. Had the matrix not been Hermitian, the eigenvalues could easily have come out complex.

The momentum operator \hat p = -i\hbar\,\tfrac{d}{dx} looks alarming — it has an explicit i out front, so how can it possibly have real eigenvalues? The answer is a lovely use of the adjoint. Compute \langle \hat p f, g\rangle = \int \overline{(-i\hbar f')}\,g\,dx = i\hbar\int \overline{f'}\,g\,dx and integrate by parts. The boundary terms vanish for wavefunctions that decay at infinity, and the derivative moves onto g — but integration by parts flips the sign, and the conjugate on the leading i flips it back. The two sign flips conspire so that \langle \hat p f, g\rangle = \langle f, \hat p g\rangle. The -i is there precisely to make the derivative self-adjoint; without it, \tfrac{d}{dx} alone is anti-Hermitian (A^\dagger = -A) and would have imaginary eigenvalues. The factor of i is not decoration — it is what makes momentum a measurable quantity.

The adjoint of a product reverses the order. Just like the transpose and the matrix inverse, (AB)^\dagger = B^\dagger A^\dagger, not A^\dagger B^\dagger. A direct consequence bites in quantum mechanics: even when A and B are each Hermitian, their product AB is Hermitian only if they commute (because (AB)^\dagger = B^\dagger A^\dagger = BA, which equals AB only when AB = BA). Since position and momentum famously do not commute, \hat x\hat p is not itself an observable — you have to symmetrise it.

And do not confuse Hermitian (A^\dagger = A, real eigenvalues, an observable) with unitary (A^\dagger = A^{-1}, eigenvalues of modulus 1, a rotation/time-evolution). They are different conditions that happen to share the dagger. Hermitian operators are the measurements; unitary operators are the motions.