Analysis

Operator theory

Studies bounded and unbounded linear operators on Hilbert spaces, including their spectra.

IntuitionOperators as infinite-dimensional matrices

A matrix takes a finite list of numbers and outputs another finite list, by mixing the entries linearly. An operator does exactly the same job for functions: differentiation $Df=f'$ takes a function and linearly produces another function; multiplying by $x$ , convolving with a fixed kernel, or applying a Fourier transform are all operators. Operator theory studies these maps on a Hilbert space $H$ the way linear algebra studies matrices on $\mathbb{R}^n$ : what are their eigenvalues (now called the spectrum), which operators are "nice" (self-adjoint, like a symmetric matrix), and — a genuinely new infinite-dimensional phenomenon — which operators are even defined on the whole space at all, since differentiation, unlike matrix multiplication, is not defined for every function in $H$ .

A finite matrix at work: the symmetric matrix

A=\begin{pmatrix}2&1\\1&-1\end{pmatrix}

(with

A=A^{\mathsf T}

, the finite-dimensional prototype of self-adjointness

T=T^*

) maps the unit square to the shaded parallelogram of area

|\det A|=3

. The dashed lines are the eigen-directions — the axes the operator stretches without rotating.

UndergraduateBounded operators, adjoints, and self-adjointness

Definition: Bounded linear operator and operator norm

A linear map $T$ from a Hilbert space $H$ to itself is bounded if it does not stretch vectors by an unlimited factor: there is a constant $c$ with $\|Tx\|\le c\|x\|$ for all $x\in H$ . The smallest such $c$ is the operator norm, $\|T\|=\sup_{\|x\|\le1}\|Tx\|$ — the largest factor by which $T$ can stretch a unit vector. Boundedness is exactly equivalent to continuity for a linear map, so "bounded operator" and "continuous operator" mean the same thing.

Written as a formula, the operator norm is the supremum of how much $T$ can stretch a unit vector:

\|T\|=\sup_{\|x\|\le1}\|Tx\|

Every bounded $T$ on $H$ has a unique adjoint operator $T^*$ , defined by $\langle Tx,y\rangle=\langle x,T^*y\rangle$ for all $x,y\in H$ — it exists by the Riesz representation theorem, applied to the functional $x\mapsto\langle Tx,y\rangle$ for each fixed $y$ . In finite dimensions, taking the adjoint is exactly conjugate-transposing a matrix. $T$ is called self-adjoint (or Hermitian) if $T=T^*$ , normal if $TT^*=T^*T$ , and unitary if $T^*T=TT^*=I$ — the operator analogues of a symmetric matrix, a matrix that commutes with its transpose, and a rotation matrix, respectively.

Self-adjointness is the single most important property in the theory, because it is exactly the condition that forces the spectrum to be real and gives access to a spectral theorem (proved below). Written out, $T$ is self-adjoint precisely when moving $T$ from one side of the inner product to the other never changes anything:

\langle Tx,y\rangle=\langle x,Ty\rangle\quad\forall x,y\in H

Compare this with a real symmetric matrix $A=A^{\mathsf T}$ , for which $\langle Ax,y\rangle=\langle x,Ay\rangle$ holds automatically — self-adjointness is exactly the infinite-dimensional version of matrix symmetry. Different operator classes have sharply different spectra, summarized below.

Operator classes and their spectra
Class	Defining condition	Spectrum $\sigma(T)$
Self-adjoint	$T=T^*$	Real: $\sigma(T)\subset\mathbb{R}$
Unitary	$T^T=TT^=I$	Unit circle: $\sigma(T)\subset\{\|z\|=1\}$
Compact self-adjoint	$T=T^*$ , $T$ compact	Discrete, $\lambda_n\to0$
Multiplication by $x$ on $L^2[0,1]$	$Tf(x)=xf(x)$	Continuous: $\sigma(T)=[0,1]$ , no eigenvalues

UndergraduateTwo pillars: the spectral theorem and the Hellinger–Toeplitz theorem

Theorem: Spectral theorem for compact self-adjoint operators

Let $T$ be a compact self-adjoint operator on a Hilbert space $H$ . Then there is an orthonormal basis of $H$ consisting of eigenvectors $e_1,e_2,\dots$ of $T$ , with real eigenvalues $\lambda_1,\lambda_2,\dots$ such that $\lambda_n\to0$ if $H$ is infinite-dimensional, and $Tx=\sum_n\lambda_n\langle x,e_n\rangle e_n$ for every $x\in H$ .

Why is it true?

It says a compact self-adjoint operator, however complicated it looks, is secretly a diagonal matrix in the right orthonormal basis — exactly as a symmetric matrix in linear algebra is always diagonalizable by an orthonormal eigenbasis. This is what makes it possible to define functions of the operator, solve $Tx=y$ , and decompose signals or images by their dominant eigen-directions (principal component analysis is this theorem in disguise).

Proof

First, every eigenvalue of a self-adjoint operator is real: if $Tx=\lambda x$ with $x\ne0$ , then $\lambda\|x\|^2=\langle Tx,x\rangle=\langle x,Tx\rangle=\bar\lambda\|x\|^2$ (using self-adjointness on the middle step), so $\lambda=\bar\lambda$ . Similarly, eigenvectors for distinct eigenvalues $\lambda\ne\mu$ are orthogonal: $\lambda\langle x,y\rangle=\langle Tx,y\rangle=\langle x,Ty\rangle=\mu\langle x,y\rangle$ forces $\langle x,y\rangle=0$ .

Next, compactness guarantees an eigenvalue of maximal absolute value actually exists: the operator norm satisfies $\|T\|=\sup_{\|x\|=1}|\langle Tx,x\rangle|$ for self-adjoint $T$ , and compactness lets one extract a convergent subsequence from a maximizing sequence for $|\langle Tx,x\rangle|$ , producing a unit vector $e_1$ with $Te_1=\lambda_1 e_1$ where $|\lambda_1|=\|T\|$ .

Now induct: having found orthonormal eigenvectors $e_1,\dots,e_{n-1}$ with eigenvalues $\lambda_1,\dots,\lambda_{n-1}$ , restrict $T$ to the closed subspace $H_n=\{e_1,\dots,e_{n-1}\}^\perp$ . Because $T$ maps $H_n$ into itself (self-adjointness makes the orthogonal complement of an invariant subspace invariant too) and remains compact and self-adjoint there, the same maximal-eigenvalue argument produces the next eigenvector $e_n\in H_n$ with $|\lambda_n|\le|\lambda_{n-1}|$ .

Finally, if this process does not terminate, $\lambda_n\to0$ : otherwise infinitely many $|\lambda_n|$ would stay above some $\delta>0$ , but then $\{Te_n\}=\{\lambda_ne_n\}$ would have no convergent subsequence (since $\|\lambda_ne_n-\lambda_me_n\|^2\ge\delta^2\cdot2$ for $n\ne m$ by orthonormality), contradicting compactness of $T$ . One then checks that the closed span of $\{e_n\}$ together with $\ker T$ exhausts $H$ , and that $T$ acts as $0$ on $\ker T$ , giving the eigen-expansion $Tx=\sum_n\lambda_n\langle x,e_n\rangle e_n$ for all $x\in H$ .

Theorem: Hellinger–Toeplitz theorem

Let $T$ be a linear operator defined on the entire Hilbert space $H$ (not just a dense subspace) that is symmetric, meaning $\langle Tx,y\rangle=\langle x,Ty\rangle\quad\forall x,y\in H$ . Then $T$ is automatically bounded.

Why is it true?

This is the theorem that explains why unbounded operators are unavoidable in quantum mechanics: physically important symmetric operators like position, momentum, and the Hamiltonian genuinely cannot be defined on every vector of $H$ (only on a dense domain), because if they were, this theorem would force them to be bounded — but they demonstrably are not.

Proof

We use the closed graph theorem (a standard consequence of the Baire category theorem): a linear operator defined on all of a Hilbert space is bounded if and only if its graph $\{(x,Tx):x\in H\}$ is closed in $H\times H$ , i.e. whenever $x_n\to x$ and $Tx_n\to y$ , we must have $y=Tx$ .

Suppose $x_n\to x$ and $Tx_n\to y$ . We must show $y=Tx$ . For any fixed $z\in H$ , symmetry gives $\langle Tx_n,z\rangle=\langle x_n,Tz\rangle$ for every $n$ .

Taking $n\to\infty$ on both sides: the left side $\langle Tx_n,z\rangle\to\langle y,z\rangle$ because $Tx_n\to y$ and the inner product is continuous; the right side $\langle x_n,Tz\rangle\to\langle x,Tz\rangle$ because $x_n\to x$ . So $\langle y,z\rangle=\langle x,Tz\rangle=\langle Tx,z\rangle$ (using symmetry once more on the right).

Since $\langle y,z\rangle=\langle Tx,z\rangle$ holds for every $z\in H$ , we get $\langle y-Tx,z\rangle=0$ for all $z$ ; taking $z=y-Tx$ gives $\|y-Tx\|^2=0$ , so $y=Tx$ . The graph of $T$ is therefore closed, and the closed graph theorem concludes $T$ is bounded.

UndergraduateReal-World Applications and Worked Examples

Operator theory is the mathematical language of quantum mechanics (observables are self-adjoint operators, energies are eigenvalues of the Hamiltonian), of vibration and stability analysis in engineering (natural frequencies are eigenvalues of a compact operator), and of data science (principal component analysis diagonalizes a compact self-adjoint covariance operator). It is also where subtle mathematical care becomes physically essential: unbounded operators cannot be handled as casually as matrices.

Example: Engineering and data science: diagonalizing a symmetric matrix

A $2\times2$ symmetric matrix, seen as a compact self-adjoint operator on $\mathbb{R}^2$ , is a coupling between two identical masses connected by springs: $A=\begin{pmatrix}2&1\\1&2\end{pmatrix}$ . Find its eigenvalues, i.e. its normal-mode frequencies (in this toy stiffness-matrix model).

Solution

Eigenvalues solve the characteristic equation $\det(A-\lambda I)=0$ . With $A=\begin{pmatrix}2&1\\1&2\end{pmatrix}$ , this is $\det\begin{pmatrix}2-\lambda&1\\1&2-\lambda\end{pmatrix}=0$ .

Expanding the determinant of this $2\times2$ matrix gives $(2-\lambda)(2-\lambda)-(1)(1)=0$ , i.e. $(2-\lambda)^2-1=0$ .

This factors as $[(2-\lambda)-1][(2-\lambda)+1]=0$ , i.e. $(1-\lambda)(3-\lambda)=0$ , so $\lambda_1=1$ and $\lambda_2=3$ .

As the spectral theorem predicts, these are both real (the matrix is symmetric), and the corresponding eigenvectors $(1,-1)/\sqrt2$ and $(1,1)/\sqrt2$ are orthogonal: the two normal modes are the masses swinging out of phase (lower frequency, weaker effective stiffness $\lambda_1=1$ ) and in phase (higher frequency, stronger effective stiffness $\lambda_2=3$ ).

Example: Physics and signal processing: why is differentiation unbounded?

On $L^2[0,1]$ , consider the differentiation operator $Df=f'$ on smooth functions vanishing at the endpoints. Using the test functions $f_n(x)=\sin(n\pi x)$ for $n=1,2,3,\dots$ , show that $T$ cannot satisfy a bound $\|Df\|_2\le c\|f\|_2$ for any constant $c$ — i.e. $D$ is unbounded, consistent with the Hellinger–Toeplitz theorem (since $D$ is only densely defined, not on all of $L^2[0,1]$ ).

Solution

First compute $\|f_n\|_2^2=\int_0^1\sin^2(n\pi x)\,dx=\frac12$ for every $n$ (the average of $\sin^2$ over a whole number of periods is $\tfrac12$ ), so $\|f_n\|_2=1/\sqrt2$ stays constant as $n$ grows.

Next, differentiate: $Df_n(x)=f_n'(x)=n\pi\cos(n\pi x)$ . Its norm squared is $\|Df_n\|_2^2=\int_0^1(n\pi)^2\cos^2(n\pi x)\,dx=(n\pi)^2\cdot\frac12$ , so $\|Df_n\|_2=\frac{n\pi}{\sqrt2}$ , growing linearly in $n$ with no upper bound.

Suppose toward contradiction that $\|Df\|_2\le c\|f\|_2$ held for some fixed $c$ and all such $f$ . Applying it to $f_n$ gives $\frac{n\pi}{\sqrt2}\le c\cdot\frac{1}{\sqrt2}$ , i.e. $n\pi\le c$ for every $n=1,2,3,\dots$ — impossible, since the left side grows without bound while $c$ is fixed.

So no such $c$ exists: $D$ is unbounded, exactly as the Hellinger–Toeplitz theorem forces for a symmetric operator that is only densely defined. Physically, this is why the momentum operator $\hat p=-i\hbar\dfrac{d}{dx}$ in quantum mechanics — differentiation up to a constant — has no universal bound on how much it can amplify a wavefunction's higher-frequency components.

Common mistake. Do not conflate symmetric with self-adjoint for unbounded operators. On a dense domain

D(T)\subsetneq H

T

is symmetric if

\langle Tx,y\rangle=\langle x,Ty\rangle\quad\forall x,y\in H

holds for

x,y\in D(T)

; it is truly self-adjoint only if, in addition, the domain of the adjoint

T^*

exactly equals

D(T)

(not merely contains it). For bounded operators the two notions coincide, which is why this distinction is easy to miss — but for unbounded operators like the momentum or position operator, getting the domain wrong can silently produce a symmetric-but-not-self-adjoint operator with no real spectral theorem, spurious "eigenvalues", or a physically meaningless time evolution. Von Neumann's theory of self-adjoint extensions (via deficiency indices) exists precisely to detect and repair this gap.

Historical note

David Hilbert's early-1900s theory of eigenvalues for integral-equation operators already contained the germ of spectral theory, but it was carried out concretely, operator by operator, without a unifying abstract framework. That framework arrived with John von Neumann's 1929–1932 work: in axiomatizing Hilbert space to give quantum mechanics a rigorous foundation, he also had to make precise, for the first time, what it means for an unbounded operator (like the Hamiltonian or momentum operator of an unconfined particle) to be self-adjoint, introducing the notion of an operator's domain as part of its identity and proving the spectral theorem in its modern, fully general form covering both bounded and unbounded self-adjoint operators. This resolved a genuine physical puzzle of the time: Heisenberg's matrix mechanics and Schrödinger's wave mechanics looked completely different but, once phrased as self-adjoint operators on a common Hilbert space, were proved to be one and the same theory.

John von Neumann David Hilbert

ResearchOpen frontier: spectral theory of random and many-body operators

Research frontier as of 2026

A major active area is the spectral theory of random Schrödinger operators $H_\omega=-\Delta+V_\omega$ on $\ell^2$ $(\mathbb{Z}^d)$ , where $V_\omega$ is a random potential modeling a disordered medium. Anderson localization — the phenomenon where disorder freezes out quantum transport, forcing the spectrum to be pure point with exponentially decaying eigenfunctions — is fully understood in one dimension and at strong disorder or spectral edges in any dimension (via multiscale analysis, Fröhlich–Spencer 1983, and the fractional moment method, Aizenman–Molchanov 1993), but the existence of an extended-states regime with purely absolutely continuous spectrum for weak disorder in dimension $d\ge3$ — the conjectured "mobility edge" — remains a major open problem. Progress continues on delicate edge cases, such as the 2020 resolution by Ding and Smart of localization for the two-dimensional Anderson–Bernoulli model with singular (two-valued) disorder. A further active direction pushes single-particle localization results toward many-body localization for interacting disordered quantum spin chains, where even the right definitions of a rigorous spectral statement remain a topic of ongoing research as of 2026.

If $T$ is a self-adjoint operator on a Hilbert space $H$ and $Tx=\lambda x$ for some $x\ne0$ , what must be true of $\lambda$ ?

What are the eigenvalues of the Hermitian matrix $A=\begin{pmatrix}2&1\\1&2\end{pmatrix}$ ?

What does the Hellinger–Toeplitz theorem say about a symmetric linear operator $T$ defined on the entire Hilbert space $H$ ?

In the $f_n(x)=\sin(n\pi x)$ example, why does the ratio $\|Df_n\|_2/\|f_n\|_2$ grow without bound as $n\to\infty$ , showing the differentiation operator is unbounded?

References

Michael Reed, Barry Simon (1980). Methods of Modern Mathematical Physics I: Functional Analysis
John B. Conway (2000). A Course in Operator Theory
Werner Kirsch (2008). An Invitation to Random Schrödinger Operators · arXiv:0709.3707