Unitary Operators and Isometries

The Geometric Intuition

Imagine a rigid rotation in three-dimensional space. If you rotate two vectors, the angle between them does not change. The length of the vectors does not change. The "shape" of the triangle formed by the origin and the two vectors is preserved. In the abstract language of Hilbert spaces, the inner product encodes both the lengths of the vectors and the angle between them. A unitary operator acts like this rigid rotation, but in a potentially infinite-dimensional complex space. It is an isometry: a map that preserves distances. Because the distance between two vectors \(x\) and \(y\) is determined by \(\lVert x - y \rVert\), and \(\lVert x - y \rVert^2 = \langle x-y, x-y \rangle\), preserving the inner product is equivalent to preserving all geometric distances. This is why unitary operators are often described as "structure-preserving" maps. They shuffle the basis vectors around without stretching, squashing, or shearing the space. In physics, this corresponds to the fact that a change of reference frame or a time evolution should not alter the intrinsic geometric relationships between possible states, only their orientation relative to the observer.

Inner-Product Preservation and the Adjoint

The defining property of a unitary operator \(U\) is that it preserves the inner product. That is, for any two vectors \(x, y \in \mathcal{H}\), the inner product of their images under \(U\) must equal the original inner product:

To connect this geometric condition to the algebraic definition involving the adjoint, we use the defining property of the adjoint \(U^\dagger\). By definition, \(\langle Ux, y \rangle = \langle x, U^\dagger y \rangle\). However, we are dealing with \(\langle Ux, Uy \rangle\). We can rewrite this by treating \(Uy\) as the second argument. Let \(z = Uy\). Then \(\langle Ux, z \rangle = \langle x, U^\dagger z \rangle\). Substituting \(z = Uy\) back, we get \(\langle Ux, Uy \rangle = \langle x, U^\dagger (Uy) \rangle = \langle x, (U^\dagger U) y \rangle\).

For this to equal \(\langle x, y \rangle\) for all \(x\) and \(y\), the operator acting on \(y\) in the second slot must be the identity operator \(I\). Thus, the condition \(\langle Ux, Uy \rangle = \langle x, y \rangle\) is equivalent to \(U^\dagger U = I\). In finite-dimensional spaces, this implies that \(U\) is invertible and its inverse is its adjoint, \(U^{-1} = U^\dagger\). This also implies \(U U^\dagger = I\), so \(U\) is fully unitary.

Spectral Consequences: Eigenvalues on the Unit Circle

A profound consequence of unitarity is the restriction it places on the eigenvalues. Suppose \(U\) is a unitary operator and \(v\) is an eigenvector with eigenvalue \(\lambda\). That is, \(Uv = \lambda v\). We can analyze the norm of \(v\) before and after the transformation. Since \(U\) preserves the inner product, it preserves the norm:

Substituting the eigenvalue equation into the left side:

Using the antilinearity in the first slot and linearity in the second:

Equating the two expressions for the norm squared:

Assuming \(v\) is a non-zero vector (as required for an eigenvector), we can divide by \(\lVert v \rVert^2\):

Therefore, every eigenvalue of a unitary operator lies on the unit circle in the complex plane. We can write \(\lambda = e^{i\theta}\) for some real \(\theta\). This is a stark contrast to self-adjoint operators, whose eigenvalues must be real. In quantum mechanics, this is why the time-evolution operator \(U(t) = e^{-iHt/\hbar}\) has eigenvalues that are pure phases. The probability amplitudes acquire a phase shift, but their magnitudes—and thus the probabilities—remain unchanged.

Orthogonality of Eigenspaces

Another critical structural feature of unitary operators is the orthogonality of eigenvectors corresponding to distinct eigenvalues. Let \(U v_1 = \lambda_1 v_1\) and \(U v_2 = \lambda_2 v_2\) with \(\lambda_1 \neq \lambda_2\). Consider the inner product \(\langle v_1, v_2 \rangle\). We can manipulate this using the unitary property. Note that \(\langle v_1, v_2 \rangle = \langle U^\dagger U v_1, v_2 \rangle = \langle U^\dagger v_1, U v_2 \rangle\). This path is slightly complex. A more direct route uses the fact that \(U^\dagger = U^{-1}\). If \(U v = \lambda v\), then \(v = U^{-1} (\lambda v) = \lambda U^{-1} v\). Thus \(U^{-1} v = \frac{1}{\lambda} v\). Since \(U^\dagger = U^{-1}\), we have \(U^\dagger v = \bar{\lambda} v\).

Now consider \(\langle v_1, v_2 \rangle\). We know \(\langle U v_1, U v_2 \rangle = \langle v_1, v_2 \rangle\). Substituting the eigenvalue equations:

So, \(\bar{\lambda}_1 \lambda_2 \langle v_1, v_2 \rangle = \langle v_1, v_2 \rangle\). Rearranging:

If \(\lambda_1 \neq \lambda_2\), does \(\bar{\lambda}_1 \lambda_2 = 1\)? Since \(\lvert \lambda_1 \rvert = \lvert \lambda_2 \rvert = 1\), \(\bar{\lambda}_1 = 1/\lambda_1\). So the term is \(\lambda_2/\lambda_1\). If \(\lambda_1 \neq \lambda_2\), then \(\lambda_2/\lambda_1 \neq 1\). Therefore, the coefficient is non-zero, which forces \(\langle v_1, v_2 \rangle = 0\). Thus, eigenvectors of a unitary operator belonging to distinct eigenvalues are orthogonal. This allows us to construct an orthonormal basis of eigenvectors, which is the foundation for diagonalizing the operator.