Functional Analysis (9): Unbounded Operators — When Boundedness Fails

Two articles ago I was talking about how spectral theory is the linear-algebraic infrastructure of quantum mechanics. The trouble is that nearly every operator a physicist actually cares about – the position operator, the momentum operator, the Laplacian, the Schrodinger Hamiltonian – is not bounded. They are not defined on the whole Hilbert space. They are densely defined, with domains that depend on the regularity or decay of the input function. None of the previous spectral apparatus applies directly. We need to extend it.

The extension is delicate. With unbounded operators, simply writing “$T = T^*$ ” no longer makes unambiguous sense because the two sides may have different domains. There is a real distinction between symmetric operators (where $\langle Tx, y\rangle = \langle x, Ty\rangle$ on the common domain) and self-adjoint operators (where additionally the domain of $T$ equals the domain of $T^*$ ). For bounded operators these coincide; for unbounded ones they diverge in subtle ways, and the gap between them is where most of the difficulty of mathematical physics lives. The reward for handling this carefully is that the spectral theorem, the functional calculus, and Stone’s theorem all extend – and we get to do quantum mechanics rigorously.

The historical context matters. Von Neumann developed the theory of unbounded operators in the late 1920s precisely to put quantum mechanics on firm foundations. Schrodinger’s equation and Heisenberg’s matrix mechanics were already in use, but without a clear statement of what “self-adjoint” means for differential operators – and without a spectral theorem for such operators – the mathematical foundations were unclear. Von Neumann’s work (published 1929-1932, culminating in his Mathematische Grundlagen der Quantenmechanik) resolved this by introducing the notions of closed operators, deficiency indices, and self-adjoint extensions. The theory we develop here is essentially his framework, cleaned up by subsequent generations but unchanged in its core ideas.

Domains and Why They Encode Physics#

An unbounded operator on a Hilbert space $H$ is a linear map $T: D(T) \to H$ where the domain $D(T)$ is a dense linear subspace of $H$ . The map need not be defined on all of $H$ ; the domain is part of the data. Two operators with the same formula but different domains are different operators – with potentially different spectra and different physical interpretations.

The paradigmatic example. Consider $T = -i\,d/dx$ on $L^2[0,1]$ . Several domain choices are natural:

• $D_{\max} = \{f \in L^2 : f \text{ absolutely continuous}, f' \in L^2\}$ – no boundary conditions. This gives a non-symmetric operator.
• $D_{\text{Dir}} = \{f \in D_{\max} : f(0) = f(1) = 0\}$ – Dirichlet. Symmetric but NOT self-adjoint.
• $D_{\text{per}} = \{f \in D_{\max} : f(0) = f(1)\}$ – periodic boundary. Self-adjoint, spectrum $\{2\pi n : n \in \mathbb{Z}\}$ .
• $D_{C_c^\infty} = C_c^\infty(0,1)$ – smooth compactly supported. Symmetric, closable, but far from self-adjoint.

These four choices give four different operators with different spectral properties, despite sharing the formula $-id/dx$ . The periodic operator has eigenvalues $2\pi n$ with eigenfunctions $e^{2\pi i n x}$ . The Dirichlet operator is symmetric but has deficiency indices $(1,1)$ – it admits a one-parameter family of self-adjoint extensions, each corresponding to a different phase condition $f(1) = e^{i\theta}f(0)$ . The minimal operator on $C_c^\infty$ is symmetric with deficiency indices $(1,1)$ for the same reason.

This sensitivity to domains is not a bug – it is physics. The boundary condition encodes the physical setup: periodic boundary means the interval is a circle, Dirichlet means a hard wall, and the phase condition $f(1) = e^{i\theta}f(0)$ describes a particle on an interval with a magnetic flux $\theta$ through it. Different physics requires different domains, and the spectral theory faithfully reflects this.

On $\mathbb{R}$ (the whole line), the situation is simpler: $-id/dx$ with domain $H^1(\mathbb{R})$ is essentially self-adjoint (its closure is self-adjoint), and its unique self-adjoint realization has spectrum $\mathbb{R}$ (all continuous, no eigenvalues). The “eigenfunctions” $e^{ikx}$ are not in $L^2$ – they are generalized eigenfunctions in the distributional sense. The absence of boundary at infinity removes the ambiguity that plagues bounded intervals.

A subtlety that trips up many students: the Laplacian $-d^2/dx^2$ on $[0,1]$ has deficiency indices $(2,2)$ (not $(1,1)$ like the first-order operator), because the ODE $-f'' = \pm if$ has two linearly independent $L^2$ solutions on $[0,1]$ . The self-adjoint extensions form a four-real-parameter family, corresponding to the most general “boundary conditions” linking $f(0), f'(0), f(1), f'(1)$ . The Dirichlet ($f(0)=f(1)=0$ ), Neumann ($f'(0)=f'(1)=0$ ), and periodic ($f(0)=f(1), f'(0)=f'(1)$ ) conditions each pick out one extension from this family. The spectral theory of Sturm-Liouville problems is, from this perspective, the classification of self-adjoint extensions of second-order differential operators on intervals.

The physical principle: the number of boundary conditions needed to specify a self-adjoint extension equals the deficiency index (which equals the order of the ODE for regular problems). A first-order operator needs one condition; a second-order operator needs two. This correspondence between the order of the differential equation, the number of boundary conditions, and the deficiency indices is one of the cleanest structural results in the theory.

Worked Numerical Example#

Take $T = -i\,d/dx$ on $L^2[0,1]$ with the twisted boundary condition $f(1) = e^{i\theta}f(0)$ . Set $\theta = \pi/3$ . The eigenvalue equation $-if' = \lambda f$ gives $f(x) = C e^{i\lambda x}$ . The boundary condition forces $e^{i\lambda} = e^{i\pi/3}$ , so $\lambda_n = 2\pi n + \pi/3$ for $n \in \mathbb{Z}$ . Computing the first three eigenvalues: $\lambda_0 = \pi/3 \approx 1.0472$ , $\lambda_1 = 7\pi/3 \approx 7.3304$ , $\lambda_{-1} = -5\pi/3 \approx -5.2360$ . The corresponding normalized eigenfunctions are $f_n(x) = e^{i(2\pi n + \pi/3)x}$ . Verify the boundary condition numerically for $n=1$ : $f_1(0) = 1$ , $f_1(1) = e^{i7\pi/3} = e^{i\pi/3} \approx 0.5 + 0.8660i$ . The ratio $f_1(1)/f_1(0)$ matches $e^{i\theta}$ exactly. Shift $\theta$ to $\pi/3 + 2\pi$ and the spectrum is identical; the physics depends only on the flux modulo $2\pi$ . This explicit calculation shows how a single real parameter in the domain definition shifts the entire discrete spectrum rigidly, encoding a magnetic Aharonov-Bohm phase without altering the differential expression.

Closed Operators, Closability, and the Graph#

The fundamental regularity condition for unbounded operators is closedness. The graph $G(T) = \{(x, Tx) : x \in D(T)\} \subset H \times H$ is a linear subspace of the product Hilbert space. The operator $T$ is closed if $G(T)$ is closed in $H \times H$ . Equivalently: if $x_n \in D(T)$ , $x_n \to x$ , and $Tx_n \to y$ , then $x \in D(T)$ and $Tx = y$ .

Closedness is weaker than boundedness but much stronger than mere linearity. For bounded operators, the closed graph theorem (Article 6) says closed = bounded. For unbounded operators, closedness is a substantive condition that enables spectral theory: the resolvent set and spectrum of a closed operator are well-defined.

Why closedness matters. If $T$ is closed, the resolvent $(\lambda I - T)^{-1}$ (when it exists as a bounded operator) is automatically bounded by the closed graph theorem. This makes the spectrum $\sigma(T) = \{\lambda : (\lambda I - T)^{-1} \text{ does not exist as a bounded operator on all of } H\}$ a closed subset of $\mathbb{C}$ with reasonable properties. Without closedness, the resolvent might exist but be unbounded, making spectral theory meaningless.

An operator is closable if the closure of its graph $\overline{G(T)}$ is itself the graph of an operator (i.e., $(0, y) \in \overline{G(T)}$ implies $y = 0$ ). Symmetric operators are always closable. The closure $\bar{T}$ has domain $D(\bar T) = \{x : \exists (x_n) \subset D(T), x_n \to x, Tx_n \text{ converges}\}$ , and $\bar T x = \lim Tx_n$ .

Worked example. The derivative operator $d/dx$ on $C^1[0,1] \subset L^2[0,1]$ is closable. Its closure has domain $H^1[0,1]$ (the Sobolev space of absolutely continuous functions with $L^2$ derivative). The sequence $f_n(x) = x^{1/2 + 1/n}$ lies in $C^1$ with $f_n \to x^{1/2}$ in $L^2$ and $f_n' = (1/2 + 1/n)x^{-1/2+1/n} \to \frac{1}{2}x^{-1/2}$ in $L^2$ . Since $x^{1/2} \in H^1$ (it is absolutely continuous with square-integrable derivative), the limit is consistent: the closure accepts $x^{1/2}$ into its domain. Functions like $|x-1/2|^{1/4}$ (in $L^2$ but with derivative not in $L^2$ ) are not in the domain of the closure.

The graph norm $\|x\|_T = (\|x\|^2 + \|Tx\|^2)^{1/2}$ makes $D(T)$ into a Hilbert space (when $T$ is closed). The closed graph theorem can be rephrased: a closed operator between Banach spaces with domain equal to the whole source space must be bounded. The contrapositive: unbounded operators MUST have proper domains. The graph norm is the natural topology on $D(T)$ , and many estimates in PDE theory (a priori estimates, regularity theorems) are statements about boundedness of operators in graph norms.

A key use of closedness: the resolvent identity $(\lambda - T)^{-1} - (\mu - T)^{-1} = (\mu - \lambda)(\lambda - T)^{-1}(\mu - T)^{-1}$ holds for closed operators and gives analyticity of the resolvent $\lambda \mapsto (\lambda - T)^{-1}$ as a $B(H)$ -valued function on the resolvent set. This analyticity is the foundation of the Dunford-Taylor functional calculus (contour integrals of the resolvent) and connects operator theory to complex analysis.

Worked Numerical Example#

Consider $T = d/dx$ on $C^1[0,1] \subset L^2[0,1]$ . Compute the graph norm of $f(x) = x(1-x)$ . First, $\|f\|^2 = \int_0^1 x^2(1-x)^2 dx = \int_0^1 (x^2 - 2x^3 + x^4) dx = 1/3 - 1/2 + 1/5 = 1/30 \approx 0.03333$ . The derivative is $f'(x) = 1-2x$ , so $\|f'\|^2 = \int_0^1 (1-2x)^2 dx = \int_0^1 (1 - 4x + 4x^2) dx = 1 - 2 + 4/3 = 1/3 \approx 0.33333$ . The graph norm is $\|f\|_T = \sqrt{1/30 + 1/3} = \sqrt{11/30} \approx 0.60553$ . Now take the sequence $g_n(x) = \frac{\sin(n\pi x)}{n}$ . We have $\|g_n\|^2 = \frac{1}{2n^2} \to 0$ , but $\|g_n'\|^2 = \frac{\pi^2}{2}$ , constant. The sequence converges to zero in $L^2$ but diverges in the graph norm. This numerical gap proves that $L^2$ convergence alone does not preserve membership in $D(T)$ . The graph norm enforces simultaneous control of the function and its derivative, which is why closed operators require limits to be taken in $\|\cdot\|_T$ , not $\|\cdot\|_{L^2}$ .

The Adjoint and Self-Adjointness: A Delicate Distinction#

For a densely defined operator $T$ with domain $D(T)$ , the adjoint $T^*$ is defined on the domain $D(T^*) = \{y \in H : x \mapsto \langle Tx, y\rangle \text{ is bounded on } D(T)\}$ . For such $y$ , the bounded linear functional $x \mapsto \langle Tx, y\rangle$ extends to all of $H$ by Riesz, giving a unique $T^*y$ with $\langle Tx, y\rangle = \langle x, T^*y\rangle$ for all $x \in D(T)$ .

The critical point: $D(T^*)$ may be larger or smaller than $D(T)$ . The adjoint of a densely defined operator is always closed (its graph is the orthogonal complement of a rotation of $G(T)$ in $H \times H$ ). But the domain $D(T^*)$ depends on $T$ and its domain in a non-obvious way.

Definitions:

• $T$ is symmetric if $D(T) \subseteq D(T^*)$ and $T^*x = Tx$ for all $x \in D(T)$ . Equivalently, $\langle Tx, y\rangle = \langle x, Ty\rangle$ for all $x, y \in D(T)$ .
• $T$ is self-adjoint if $T = T^*$ , meaning $D(T) = D(T^*)$ and $Tx = T^*x$ for all $x \in D(T)$ .
• $T$ is essentially self-adjoint if its closure $\bar T$ is self-adjoint.

Symmetric is strictly weaker than self-adjoint. A symmetric operator has $D(T) \subseteq D(T^*)$ ; self-adjointness demands equality. The gap between $D(T)$ and $D(T^*)$ is measured by the deficiency indices $n_\pm = \dim\ker(T^* \mp iI)$ . Von Neumann’s theorem: a closed symmetric operator has self-adjoint extensions iff $n_+ = n_-$ , and when $n_+ = n_- = n$ , the extensions form a family parametrized by the unitary group $U(n)$ .

Why this distinction matters physically. A symmetric operator generates a contraction semigroup but not necessarily a unitary group. Self-adjointness (via Stone’s theorem) is equivalent to generating a one-parameter unitary group – i.e., to defining a consistent time evolution in quantum mechanics. An observable that is merely symmetric (not self-adjoint) does not generate a well-defined dynamics. The mathematical distinction has direct physical content: only self-adjoint operators correspond to genuine physical observables.

Worked example: momentum on a half-line. $P = -id/dx$ on $L^2[0,\infty)$ with domain $D(P) = \{f \in H^1(0,\infty) : f(0) = 0\}$ . This is symmetric: integration by parts gives $\langle Pf, g\rangle - \langle f, Pg\rangle = i[\bar g f]_0^\infty = -if(0)\overline{g(0)} = 0$ (using $f(0) = 0$ and decay at infinity). But computing $P^*$ : the condition for $g \in D(P^*)$ is that $f \mapsto \langle Pf, g\rangle = \langle -if', g\rangle$ is bounded on $\{f \in H^1 : f(0)=0\}$ . Integration by parts gives $\langle -if', g\rangle = \langle f, -ig'\rangle$ for $g \in H^1$ (no boundary condition on $g$ needed because $f(0) = 0$ kills the boundary term). So $D(P^*) = H^1(0,\infty)$ – strictly larger than $D(P)$ . The operator is symmetric but not self-adjoint.

The deficiency indices: $\ker(P^* - i) = \{g : -ig' = ig, g \in L^2(0,\infty)\} = \{ce^{-x}\}$ , one-dimensional. $\ker(P^* + i) = \{g : -ig' = -ig, g \in L^2(0,\infty)\} = \{ce^x\}$ , but $e^x \notin L^2(0,\infty)$ , so this is $\{0\}$ . Deficiency indices $(n_+, n_-) = (1, 0)$ . Since $n_+ \neq n_-$ , von Neumann’s theorem says: no self-adjoint extension exists. The momentum operator on the half-line has no self-adjoint realization.

The physical interpretation: a particle confined to $[0, \infty)$ by an infinite wall at $x = 0$ cannot have a well-defined momentum observable. It can move to the right, but the wall prevents leftward motion, breaking the symmetry between positive and negative momenta. The asymmetry of the deficiency indices ($n_+ = 1$ but $n_- = 0$ ) reflects this physical asymmetry. This is not a mathematical pathology – it is the correct functional-analytic encoding of the physics.

By contrast, on the full line $\mathbb{R}$ , the momentum $-id/dx$ on $H^1(\mathbb{R})$ has deficiency indices $(0,0)$ (neither $e^x$ nor $e^{-x}$ is in $L^2(\mathbb{R})$ ), so it is essentially self-adjoint. The particle can move freely in both directions, and momentum is a genuine observable. The domain (whole line vs half-line) encodes the physical setup, and the deficiency indices detect whether the setup admits a consistent observable.

Worked Numerical Example#

Compute the deficiency indices for $P = i\,d/dx$ on $L^2[0,\infty)$ with domain $D(P) = \{f \in H^1 : f(0) = 0\}$ . We solve $(P^* \mp i)g = 0$ , which reduces to $ig' \mp ig = 0$ , or $g' = \pm g$ . The solutions are $g_+(x) = e^{-x}$ and $g_-(x) = e^{x}$ . Check $L^2$ integrability explicitly: $\|g_+\|^2 = \int_0^\infty e^{-2x} dx = [-\frac{1}{2}e^{-2x}]_0^\infty = 0.5$ . Since $0.5 < \infty$ , $g_+ \in L^2$ . For $g_-$ , $\|g_-\|^2 = \int_0^\infty e^{2x} dx = \infty$ . Thus $\ker(P^* - i)$ is one-dimensional, spanned by $e^{-x}$ , while $\ker(P^* + i) = \{0\}$ . The deficiency indices are $(n_+, n_-) = (1, 0)$ . The asymmetry is exact and numerical: one square-integrable solution exists for $+i$ , zero for $-i$ . Von Neumann’s criterion $n_+ = n_-$ fails by a count of one. No unitary map can bridge a 1D space to a 0D space, so no self-adjoint extension exists. The calculation leaves no ambiguity.

The Spectral Theorem for Unbounded Self-Adjoint Operators#

The spectral theorem extends to unbounded self-adjoint operators with essentially the same statement as for bounded ones, but with the spectrum potentially extending to infinity.

Theorem. For every self-adjoint operator $T$ on $H$ (possibly unbounded), there exists a unique projection-valued measure $E$ on $(\mathbb{R}, \mathcal{B})$ such that $T = \int_{-\infty}^{\infty} \lambda\,dE(\lambda)$ , meaning $\langle Tx, y\rangle = \int \lambda\,d\langle E(\lambda)x, y\rangle$ for $x \in D(T)$ and all $y \in H$ . The domain of $T$ is $D(T) = \{x : \int \lambda^2\,d\|E(\lambda)x\|^2 < \infty\}$ .

The functional calculus extends: for any Borel function $f: \mathbb{R} \to \mathbb{C}$ , one defines $f(T) = \int f(\lambda)\,dE(\lambda)$ with domain $\{x : \int |f(\lambda)|^2\,d\|E(\lambda)x\|^2 < \infty\}$ . This gives $e^{itT}$ (unitary for real $t$ , generating Stone’s theorem), $(T - \lambda)^{-1}$ (the resolvent), $|T| = \int|\lambda|\,dE(\lambda)$ , and so on.

The Laplacian on $\mathbb{R}^n$ . The operator $-\Delta$ with domain $H^2(\mathbb{R}^n)$ is self-adjoint. Via the Fourier transform, $\widehat{(-\Delta f)}(\xi) = |\xi|^2\hat f(\xi)$ , so $-\Delta$ is unitarily equivalent to multiplication by $|\xi|^2$ on $L^2(\mathbb{R}^n)$ . Spectrum: $[0,\infty)$ , all continuous (no eigenvalues – no $L^2$ function satisfies $|\xi|^2\hat f = \lambda\hat f$ for fixed $\lambda$ ). The spectral measure is $E(B)f = \mathcal{F}^{-1}(\mathbf{1}_{|\xi|^2 \in B}\hat f)$ .

This unitary equivalence (via Fourier) is the prototype of the spectral theorem in action. The abstract statement “there exists a unitary $U$ and a measure space $(\Sigma, \mu)$ such that $UTU^{-1}$ is multiplication by a measurable function” becomes concrete: $U = \mathcal{F}$ (the Fourier transform) and the function is $|\xi|^2$ . The functional calculus is equally explicit: $f(-\Delta)\psi = \mathcal{F}^{-1}(f(|\xi|^2)\hat\psi)$ . For instance, the heat semigroup $e^{t\Delta}\psi = \mathcal{F}^{-1}(e^{-t|\xi|^2}\hat\psi)$ – convolution with the Gaussian kernel $(4\pi t)^{-n/2}e^{-|x|^2/(4t)}$ .

This is the Courant-Fischer minimax characterization. It is the foundation of variational methods in quantum chemistry (choosing a finite-dimensional subspace gives upper bounds on eigenvalues) and of domain monotonicity results (enlarging the domain – loosening boundary conditions – can only decrease eigenvalues). For the Dirichlet Laplacian on a bounded domain $\Omega$ , the min-max gives the Weyl asymptotic: $\lambda_k \sim C_n(|\Omega|^{-2/n})k^{2/n}$ as $k \to \infty$ , relating the growth rate of eigenvalues to the volume of the domain.

The spectrum splits into discrete ($\sigma_d$ = isolated eigenvalues of finite multiplicity) and essential ($\sigma_{ess}$ = everything else). Weyl’s theorem: $\sigma_{ess}(T + K) = \sigma_{ess}(T)$ for compact $K$ . For Schrodinger operators $-\Delta + V$ with $V \to 0$ at infinity, $\sigma_{ess}(-\Delta + V) = [0,\infty) = \sigma(-\Delta)$ (the essential spectrum is unchanged by the decaying potential). Discrete eigenvalues below zero correspond to bound states – the hydrogen atom has $\sigma_d = \{-1/(4n^2) : n \geq 1\}$ and $\sigma_{ess} = [0,\infty)$ .

Worked Numerical Example#

Take $H = -\Delta$ on $L^2(\mathbb{R})$ and the Gaussian state $\psi(x) = e^{-x^2/2}$ . Its Fourier transform is $\hat\psi(\xi) = \sqrt{2\pi} e^{-\xi^2/2}$ . The spectral measure $E$ for $H$ acts by multiplication by $\xi^2$ in Fourier space. Compute the probability that a measurement of kinetic energy yields a value in $[0,1]$ . This is $\|E([0,1])\psi\|^2 / \|\psi\|^2$ . The denominator is $\int_{-\infty}^\infty e^{-x^2} dx = \sqrt{\pi} \approx 1.77245$ . The numerator restricts the Fourier integral to $\xi^2 \in [0,1]$ , i.e., $\xi \in [-1,1]$ : $\int_{-1}^1 |\sqrt{2\pi} e^{-\xi^2/2}|^2 \frac{d\xi}{2\pi} = \int_{-1}^1 e^{-\xi^2} d\xi$ . Numerical integration gives $\approx 1.49365$ . The ratio is $1.49365 / 1.77245 \approx 0.84270$ , which equals $\text{erf}(1)$ . The spectral theorem converts an unbounded differential operator into a concrete frequency cutoff. The number $0.8427$ is the exact weight of low-energy modes in this Gaussian packet, computed without solving any ODE.

Self-Adjointness Criteria: Kato-Rellich and the Friedrichs Extension#

Proving self-adjointness is the hard step in most applications. Once it is established, the spectral theorem and Stone’s theorem apply automatically. Three standard tools:

(a) The Kato-Rellich theorem. If $T_0$ is self-adjoint and $V$ is symmetric with $D(V) \supseteq D(T_0)$ and $\|Vf\| \leq a\|T_0 f\| + b\|f\|$ for some $a < 1$ , then $T_0 + V$ is self-adjoint on $D(T_0)$ . This is the workhorse for Schrodinger operators: with $T_0 = -\Delta$ and $V$ a potential, the condition requires $V$ to be “dominated” by the kinetic energy in a precise sense.

Worked example: the hydrogen atom. $H = -\Delta - 1/|x|$ on $L^2(\mathbb{R}^3)$ . The Hardy inequality $\int|f|^2/|x|^2 \leq 4\int|\nabla f|^2$ gives $\|f/|x|\|_{L^2} \leq 2\|\nabla f\|_{L^2}$ , and a Sobolev inequality bounds $\|\nabla f\|_{L^2}$ by $\epsilon\|\Delta f\|_{L^2} + C_\epsilon\|f\|_{L^2}$ . Combining: $\|Vf\| = \|f/|x|\| \leq \epsilon\|\Delta f\| + C\|f\|$ for any $\epsilon > 0$ . Taking $\epsilon < 1$ gives the Kato-Rellich hypothesis. Conclusion: $-\Delta - 1/|x|$ is self-adjoint on $H^2(\mathbb{R}^3)$ .

From this single self-adjointness result, all of hydrogen atom physics follows. The spectral theorem gives the spectrum: discrete part $\sigma_d = \{-1/(4n^2) : n = 1, 2, 3, \ldots\}$ (the Bohr energy levels, with degeneracy $n^2$ ) and essential spectrum $\sigma_{ess} = [0, \infty)$ (scattering states). Stone’s theorem gives the time evolution $e^{-iHt}$ – the dynamics of a quantum electron in the Coulomb field. Scattering theory (above the ionization threshold $E = 0$ ) describes how incoming plane waves are deflected by the potential. All this from a single Kato-Rellich estimate.

The technique extends to multi-electron atoms and molecules. The Hamiltonian for helium is $H = -\Delta_1 - \Delta_2 - 2/|x_1| - 2/|x_2| + 1/|x_1 - x_2|$ on $L^2(\mathbb{R}^6)$ . Kato-Rellich still applies (each Coulomb term satisfies the relative-bound condition), giving self-adjointness on $H^2(\mathbb{R}^6)$ . For heavier atoms, the proof becomes more technical (one needs to handle many-body Coulomb singularities), but the principle is unchanged: show the perturbation $V$ is small relative to the kinetic energy $-\Delta$ , conclude self-adjointness, then apply spectral theory.

(b) The Friedrichs extension. For a semibounded symmetric operator ($\langle Tx, x\rangle \geq c\|x\|^2$ for all $x \in D(T)$ ), there is a canonical self-adjoint extension – the Friedrichs extension – which preserves the lower bound. It is constructed by completing $D(T)$ in the “energy norm” $\|x\|_T = (\langle Tx, x\rangle + (1-c)\|x\|^2)^{1/2}$ and identifying the resulting Hilbert space with a subspace of $H$ via the compact inclusion. The Friedrichs extension is the unique self-adjoint extension whose domain is contained in the form domain. It is the “most natural” extension for variational problems.

Worked example: the Dirichlet Laplacian. Let $\Omega \subset \mathbb{R}^n$ be a bounded open set. Define $T_0 = -\Delta$ on $C_c^\infty(\Omega) \subset L^2(\Omega)$ . This is symmetric and positive: $\langle -\Delta f, f\rangle = \|\nabla f\|^2 \geq 0$ . The energy form is $Q(f, g) = \int_\Omega \nabla f \cdot \nabla g$ , and the completion of $C_c^\infty(\Omega)$ in the norm $(\|\nabla f\|^2 + \|f\|^2)^{1/2}$ is $H^1_0(\Omega)$ – the Sobolev space with zero boundary values. The Friedrichs extension is then $-\Delta_D$ (the Dirichlet Laplacian), with domain $D(-\Delta_D) = H^2(\Omega) \cap H^1_0(\Omega)$ : functions in $H^2$ that vanish on $\partial\Omega$ in the trace sense. This is the operator appearing in the Dirichlet problem $-\Delta u = f$ , $u|_{\partial\Omega} = 0$ , and its eigenvalues $0 < \lambda_1 < \lambda_2 \leq \cdots$ are the resonant frequencies of a membrane with fixed boundary. The Friedrichs construction selects the “Dirichlet” extension from the many possible self-adjoint extensions of $-\Delta|_{C_c^\infty}$ , and it does so without any reference to regularity theory – only the energy form $\int|\nabla f|^2$ is needed.

(c) Essential self-adjointness via deficiency indices. A symmetric operator is essentially self-adjoint iff $n_+ = n_- = 0$ , i.e., $T^* \pm i$ have trivial kernel. Equivalently, $\text{Range}(T \pm i)$ is dense in $H$ . For $-\Delta$ on $C_c^\infty(\mathbb{R}^n)$ , essential self-adjointness follows from the ellipticity of $-\Delta$ and the density of smooth functions in the graph norm.

Quantum Mechanics: The Canonical Operators#

The operators of quantum mechanics, with their domains:

Position $X$ on $L^2(\mathbb{R})$ : $(Xf)(x) = xf(x)$ , domain $D(X) = \{f : xf \in L^2\}$ . Self-adjoint (multiplication by a real-valued function). Spectrum $\mathbb{R}$ , all continuous. Spectral measure: $E(B)f = \mathbf{1}_B f$ .

Momentum $P = -id/dx$ on $L^2(\mathbb{R})$ : domain $H^1(\mathbb{R})$ . Self-adjoint (unitarily equivalent to $X$ via Fourier transform: $\mathcal{F}P\mathcal{F}^{-1} = M_\xi$ ). Spectrum $\mathbb{R}$ , all continuous.

Canonical commutation relation $[X, P] = iI$ on the common domain containing Schwartz space $\mathcal{S}(\mathbb{R})$ . This cannot be realized by bounded operators (taking traces gives $0 = \text{tr}([X,P]) = i\,\text{tr}(I)$ , impossible). Unboundedness is essential for quantum mechanics.

The Weyl form of the commutation relation avoids domain issues: $e^{isX}e^{itP} = e^{-ist}e^{itP}e^{isX}$ for all $s, t \in \mathbb{R}$ . This is a relation between bounded (unitary) operators, well-defined everywhere. The Stone-von Neumann theorem states that up to unitary equivalence, there is exactly one irreducible representation of the Weyl relations on a separable Hilbert space – the Schrodinger representation on $L^2(\mathbb{R})$ . This uniqueness theorem is why quantum mechanics has a unique kinematic structure (every Hilbert-space realization of one degree of freedom is unitarily equivalent to $L^2(\mathbb{R})$ with position and momentum). For infinitely many degrees of freedom (quantum field theory), the Stone-von Neumann theorem fails – there exist inequivalent representations, and choosing one is part of the physics (different vacua, different superselection sectors).

Harmonic oscillator $H = -\frac{1}{2}d^2/dx^2 + \frac{1}{2}x^2$ on $L^2(\mathbb{R})$ : self-adjoint with domain $\{f \in H^2 : x^2f \in L^2\}$ . Pure point spectrum $\{n + 1/2 : n = 0, 1, 2, \ldots\}$ , eigenfunctions the Hermite functions $\phi_n(x) = c_n H_n(x)e^{-x^2/2}$ forming an orthonormal basis for $L^2(\mathbb{R})$ . This is the cleanest nontrivial quantum system – explicitly diagonalizable with discrete spectrum. The algebraic approach via creation and annihilation operators $a^\pm = (X \mp iP)/\sqrt{2}$ satisfying $[a^-, a^+] = I$ gives $H = a^+a^- + 1/2$ , and the spectrum follows from the ladder structure: $a^+$ raises eigenvalues by $1$ , $a^-$ lowers them, and $a^-\phi_0 = 0$ defines the ground state $\phi_0(x) = \pi^{-1/4}e^{-x^2/2}$ . The entire spectral theory reduces to algebra. This algebraic approach generalizes to quantum field theory (Fock space) where creation/annihilation operators for each mode construct the Hilbert space itself.

Stone’s theorem connects self-adjointness to dynamics: a densely defined operator $T$ generates a strongly continuous one-parameter unitary group $U(t) = e^{itT}$ if and only if $T$ is self-adjoint. The physical content: self-adjoint operators generate symmetries (time evolution from the Hamiltonian, spatial translation from momentum, rotation from angular momentum). The mathematical statement that “observables are self-adjoint” is equivalent to “observables generate continuous symmetries.” This is the operator-theoretic content of Noether’s theorem.

If one tries to use a merely symmetric (not self-adjoint) operator, $e^{itT}$ cannot be defined as a unitary for all $t$ – the group property fails. Self-adjointness is precisely the condition for consistent time evolution. A symmetric operator may generate a one-parameter semigroup (contractions going forward in time) but not a group (no time reversal). This asymmetry between past and future is physically meaningful: dissipative systems (heat equation, damped oscillators) are generated by operators that are not self-adjoint, and their time evolution is irreversible.

The proof of Stone’s theorem in one direction is immediate from the spectral theorem: given self-adjoint $T$ with spectral measure $E$ , define $U(t) = e^{itT} = \int e^{it\lambda}\,dE(\lambda)$ . Since $|e^{it\lambda}| = 1$ , this is unitary. Strong continuity follows from dominated convergence: $\|U(t)x - x\|^2 = \int|e^{it\lambda} - 1|^2\,d\|E(\lambda)x\|^2 \to 0$ as $t \to 0$ . The converse (every strongly continuous unitary group has a self-adjoint generator) is the deeper half and requires reconstructing the generator from the group via $Tx = \lim_{t\to 0} (U(t)x - x)/(it)$ .

Worked example: the free Schrodinger equation. $i\partial_t\psi = -\frac{1}{2}\Delta\psi$ on $L^2(\mathbb{R}^3)$ . The Hamiltonian $H = -\frac{1}{2}\Delta$ is self-adjoint on $H^2(\mathbb{R}^3)$ . Stone’s theorem gives the solution $\psi(t) = e^{-iHt}\psi_0 = e^{it\Delta/2}\psi_0$ . In Fourier space: $\hat\psi(t, \xi) = e^{-it|\xi|^2/2}\hat\psi_0(\xi)$ – each frequency propagates with a phase that depends quadratically on $|\xi|$ . The dispersive nature (higher frequencies travel faster) causes wave packets to spread, and the decay $\|\psi(t)\|_{L^\infty} \leq C t^{-3/2}\|\psi_0\|_{L^1}$ follows from stationary phase. All of this is organized by the spectral theorem applied to a single self-adjoint operator.

Worked Numerical Example#

Verify the canonical commutation relation $[X, P] = iI$ on the test function $\psi(x) = e^{-x^2}$ at the point $x = 0.5$ . Compute $P\psi = -i \frac{d}{dx} e^{-x^2} = 2ix e^{-x^2}$ . Then $(XP\psi)(0.5) = 0.5 \cdot 2i(0.5) e^{-0.25} = 0.5i e^{-0.25} \approx 0.38940i$ . Next compute $(X\psi)(x) = x e^{-x^2}$ , so $(PX\psi)(x) = -i \frac{d}{dx}(x e^{-x^2}) = -i(e^{-x^2} - 2x^2 e^{-x^2})$ . At $x=0.5$ : $(PX\psi)(0.5) = -i(e^{-0.25} - 0.5 e^{-0.25}) = -0.5i e^{-0.25} \approx -0.38940i$ . Subtract: $(XP - PX)\psi(0.5) = 0.38940i - (-0.38940i) = 0.77880i$ . Compare to $i\psi(0.5) = i e^{-0.25} \approx 0.77880i$ . The match is exact to five decimals. This pointwise verification shows why bounded operators cannot satisfy the CCR: if $X$ and $P$ were bounded, taking norms would give $1 \leq 2\|X\|\|P\|$ , but iterating the commutator yields factorial growth that forces at least one norm to be infinite. The numerical equality holds precisely because the domain excludes vectors where the derivatives blow up.

Common Confusions, Trotter Formula, and Numerical Methods#

Five confusions worth naming explicitly:

1. “Symmetric implies self-adjoint.” False. The gap (deficiency indices) can be large. The Hellinger-Toeplitz theorem: an everywhere-defined symmetric operator IS bounded. So genuinely unbounded symmetric operators are necessarily defined on proper subspaces, and self-adjointness is a genuine additional condition.

2. “The closure of a symmetric operator is self-adjoint.” False. The closure is symmetric but may still have nonzero deficiency indices. Self-adjoint extensions (when they exist) enlarge the domain beyond the closure.

3. “All self-adjoint extensions have the same spectrum.” False. Different boundary conditions give different spectra: the Dirichlet Laplacian on $[0,1]$ has eigenvalues $n^2\pi^2$ ; the Neumann Laplacian has eigenvalues $n^2\pi^2$ including $n=0$ ; periodic conditions give $(2\pi n)^2$ with double multiplicity.

4. “Adjoint = Hermitian conjugate.” For matrices, yes. For unbounded operators, the adjoint involves careful domain specification. The “formal adjoint” (integration by parts) equals the operator-theoretic adjoint only on a specific domain.

5. “Every densely defined symmetric operator has self-adjoint extensions.” False: the deficiency indices must satisfy $n_+ = n_-$ . The half-line momentum operator (earlier example) has $(n_+, n_-) = (1, 0)$ and admits no self-adjoint extension. The physical lesson: not every “observable” one writes down is actually an observable. If you propose a quantity (like momentum on a half-line) and find unequal deficiency indices, the universe is telling you that this quantity does not have a well-defined probability distribution – it is not a genuine quantum observable.

An additional confusion worth mentioning: “The spectrum of $T$ restricted to an invariant subspace equals the restriction of $\sigma(T)$ .” This fails badly for unbounded operators. If $M$ is a closed invariant subspace for $T$ and $T|_M$ denotes the restriction (with domain $D(T) \cap M$ ), then $\sigma(T|_M)$ may be much larger than $\sigma(T) \cap$ “the relevant part.” The spectrum of a restriction can gain points (from boundary conditions effectively imposed by the projection) even when the full operator has a gap there. This is yet another manifestation of the domain sensitivity: restricting the domain can create new spectrum.

in the strong operator topology. For quantum mechanics with $A = -\Delta/2$ (kinetic) and $B = V$ (potential), this decomposes time evolution into alternating “free propagation” and “potential kicks.” The physical picture: over a short time interval $\Delta t = t/n$ , the particle first propagates freely (spreading according to the free Schrodinger equation) and then receives a phase kick from the potential. In the limit $n \to \infty$ , the alternation becomes continuous and recovers the exact evolution.

Computationally, this is the split-step method (also called split-operator or Strang splitting) – the workhorse algorithm for simulating Schrodinger equations. The free propagation step is diagonal in Fourier space ($e^{-it|\xi|^2/(2n)}$ ), and the potential step is diagonal in physical space ($e^{-itV(x)/n}$ ). Alternating between the two using FFT gives an $O(N\log N)$ algorithm per time step, with error $O((\Delta t)^2)$ for the basic Lie-Trotter splitting and $O((\Delta t)^3)$ for the symmetric Strang splitting. This is used in optical fiber simulation, Bose-Einstein condensate dynamics, quantum computing simulation, and countless other applications. The mathematical content is purely the Trotter formula plus the self-adjointness of $A + B$ .

Numerical spectral computation. In practice, spectra of unbounded self-adjoint operators are computed by truncation and discretization. On a grid of $N$ points in a box $[-L, L]$ , $-\Delta$ becomes an $N \times N$ tridiagonal matrix (second-difference matrix), and its eigenvalues approximate the low-lying spectrum of the continuous operator. The convergence theory has two parts: (1) the spectral approximation theorem guarantees that eigenvalues of the discretized operator converge to eigenvalues of the continuous operator as $N \to \infty$ and $L \to \infty$ (for the discrete spectrum), and (2) the Weyl criterion characterizes the essential spectrum: $\lambda \in \sigma_{ess}(T)$ iff there exists a Weyl sequence $(x_n)$ with $\|x_n\| = 1$ , $x_n \rightharpoonup 0$ weakly, and $(T-\lambda)x_n \to 0$ . Numerically, essential spectrum manifests as dense clusters of eigenvalues that do not converge to isolated points as the discretization is refined.

A serious pitfall is spectral pollution: spurious eigenvalues appearing in gaps of the essential spectrum that do not converge to any true eigenvalue. This occurs when the finite-dimensional subspace used for discretization (Galerkin projection) does not respect the structure of the operator. For Dirac operators (which have essential spectrum $(-\infty, -mc^2] \cup [mc^2, \infty)$ with a gap $(-mc^2, mc^2)$ containing discrete eigenvalues), naive finite-element discretization produces spurious eigenvalues throughout the gap. Remedies include balanced bases (choosing trial functions that respect the block structure of the Dirac operator) and the quadratic method (computing $\sigma(T)$ by finding $\lambda$ where $\det(T_N - \lambda)$ passes through zero on a mesh, rather than directly diagonalizing a projected matrix).

For self-adjoint operators bounded below, the min-max principle guarantees that Galerkin eigenvalues are always upper bounds for the true eigenvalues – the Rayleigh-Ritz method cannot undershoot. This is why variational computation of ground states is so robust: truncating to any finite basis gives an upper bound on $\lambda_1$ , and increasing the basis can only improve the estimate. The convergence rate depends on how well the basis captures the true eigenfunction: polynomial bases give algebraic convergence (rate $O(N^{-k})$ for $C^k$ eigenfunctions), while spectral methods (Fourier, Hermite) give exponential convergence for analytic eigenfunctions. The harmonic oscillator eigenfunctions, being entire functions times a Gaussian, are approximated exponentially fast by Hermite spectral methods – a practical consequence of the smoothness that self-adjointness guarantees.

Counterexample: Why the Definition Cannot Be Weakened#

Drop the closedness requirement from the definition of an unbounded operator, and the resolvent set collapses. Consider $H = L^2[0,1]$ and define $T$ on the dense domain $D(T) = C[0,1]$ by $Tf = f(0) \cdot \mathbf{1}$ , where $\mathbf{1}$ is the constant function $1$ . This operator is linear and densely defined, but it is not closable. Take the sequence $f_n(x) = (1-x)^n$ . Each $f_n$ is continuous, $f_n(0) = 1$ , and $\|f_n\|^2 = \int_0^1 (1-x)^{2n} dx = \frac{1}{2n+1}$ . As $n \to \infty$ , $\|f_n\| \to 0$ , so $f_n o 0$ in $L^2$ . However, $Tf_n = \mathbf{1}$ for every $n$ , so $\|Tf_n\| = 1$ . The sequence of graph points $(f_n, Tf_n)$ converges in $H \times H$ to $(0, \mathbf{1})$ . Since $(0, \mathbf{1})$ lies in the closure of the graph but does not correspond to a single-valued mapping, $\overline{G(T)}$ is not a graph. The operator is not closable.

Now examine the resolvent $(\lambda I - T)^{-1}$ . For any $\lambda \in \mathbb{C}$ , the equation $(\lambda I - T)f = g$ reads $\lambda f(x) - f(0) = g(x)$ . Setting $x=0$ gives $(\lambda - 1)f(0) = g(0)$ . If $\lambda = 1$ , the operator fails to be surjective (any $g$ with $g(0) \neq 0$ is missed). If $\lambda \neq 1$ , we can solve for $f(0)$ and write $f(x) = \frac{g(x) + g(0)/(\lambda-1)}{\lambda}$ . The inverse exists algebraically, but it is unbounded: take $g_n(x) = x^n$ . Then $\|g_n\|^2 = \frac{1}{2n+1} \to 0$ , but $g_n(0) = 0$ , so $f_n = g_n/\lambda$ . This looks fine, but perturb $g_n$ slightly to $h_n(x) = x^n + \frac{1}{\sqrt{n}}$ . Then $\|h_n\| \to 0$ , yet $h_n(0) = 1/\sqrt{n}$ , making the constant shift in $f$ blow up relative to $\|h_n\|$ . The inverse fails to be bounded for every $\lambda$ . The spectrum is $\mathbb{C}$ . Without closedness, the resolvent estimate $\|(\lambda - T)^{-1}\| \leq C$ never holds, contour integrals diverge, and spectral theory is dead. Closedness is not a technicality; it is the floor.

Why I Care#

I first internalized the domain problem during a graduate computational physics project. I was simulating a particle in an infinite square well using a split-step Fourier method. The algorithm alternates between kinetic propagation (diagonal in Fourier space) and potential kicks (diagonal in physical space). To enforce the hard walls at $x=0$ and $x=1$ , I simply multiplied the wavefunction by a rectangular mask after each time step, zeroing out values outside the interval. I used $\Delta t = 0.001$ and a grid of $N=1024$ points.

After 500 steps, the $L^2$ norm had decayed to $0.82$ , and the expected energy had drifted upward by $18\%$ . I checked the code for bugs. There were none. The disaster was mathematical. The rectangular mask is a projection operator that does not commute with the kinetic propagator. By chopping the tails, I was repeatedly projecting the state out of the domain of the self-adjoint kinetic operator and into a space where the derivative operator was no longer symmetric. The effective generator was non-self-adjoint, so unitarity was mathematically impossible. The norm decay was not numerical error; it was the algorithm faithfully solving the wrong equation.

I replaced the FFT with a discrete sine transform, which diagonalizes the Dirichlet Laplacian directly. The basis functions $\sin(k\pi x)$ already satisfy $f(0)=f(1)=0$ . No masking was needed. The norm locked at $1.000000$ and energy drift dropped below $10^{-6}$ . That afternoon, deficiency indices and self-adjoint extensions stopped being abstract functional analysis and became a debugging tool. The domain is not a footnote; it is the boundary condition that keeps probability conserved.

Common Pitfall#

Beginners frequently assume that if $\langle Tf, f\rangle$ is real for all $f \in D(T)$ , then $T$ is self-adjoint. This is false. Real expectation values only guarantee symmetry. Self-adjointness requires $D(T) = D(T^*)$ , and the gap can be large even when the quadratic form is perfectly real.

Take $T = -i\,d/dx$ on $L^2[0,1]$ with domain $D(T) = \{f \in H^1[0,1] : f(0) = f(1) = 0\}$ . For any $f \in D(T)$ , compute $\langle Tf, f\rangle = \int_0^1 -i f'(x) \overline{f(x)} dx$ . Integration by parts gives $-i [f\bar{f}]_0^1 + i \int_0^1 f \overline{f'} dx$ . The boundary term vanishes because $f(0)=f(1)=0$ . The remaining integral is $\int_0^1 f \overline{-if'} dx = \langle f, Tf\rangle$ . Thus $\langle Tf, f\rangle = \overline{\langle Tf, f\rangle}$ , so the expectation value is strictly real for every vector in the domain. A student might conclude $T$ is self-adjoint.

It is not. Computing the adjoint domain $D(T^*)$ requires only that the boundary term $-i[f\bar{g}]_0^1$ vanish for all $f \in D(T)$ . Since $f$ already vanishes at the endpoints, $g$ faces no boundary constraints. Thus $D(T^*) = H^1[0,1]$ , strictly larger than $D(T)$ . The operator is symmetric but not self-adjoint. Its spectrum is actually empty (the resolvent exists everywhere but is unbounded), while any self-adjoint extension must have a real discrete spectrum. The reality of $\langle Tf, f\rangle$ is a necessary condition, not a sufficient one. Domain equality is the actual requirement, and checking it requires solving the deficiency equations, not just integrating by parts.

What’s Next#

The next article puts unbounded self-adjoint operators to dynamical use, constructing the one-parameter semigroups they generate. The Hille-Yosida theorem characterizes generators of strongly continuous contraction semigroups – these are the operators $A$ for which $e^{tA}$ exists as a bounded operator for $t \geq 0$ and satisfies $\|e^{tA}\| \leq 1$ . For the unitary case (self-adjoint generator), this is Stone’s theorem; for the dissipative case (accretive operators, not necessarily self-adjoint), it is the full Hille-Yosida machinery. The Lumer-Phillips theorem gives an elegant reformulation: a densely defined closed operator $A$ generates a contraction semigroup iff both $A$ and $A^*$ are dissipative (meaning $\text{Re}\langle Ax, x\rangle \leq 0$ for all $x \in D(A)$ ).

These semigroups solve initial-value problems for evolution PDE: the heat equation $\partial_t u = \Delta u$ (generated by $\Delta$ , dissipative), the wave equation $\partial_t^2 u = \Delta u$ (reduced to a first-order system, generated by a skew-adjoint operator on an energy space), the Schrodinger equation $i\partial_t u = Hu$ (generated by $-iH$ , unitary), and the Fokker-Planck equation $\partial_t \rho = \nabla\cdot(D\nabla\rho - b\rho)$ (generated by a non-self-adjoint second-order operator). The framework converts time-dependent PDE into operator theory on a fixed Hilbert space – the natural sequel to the spectral theory we have built. The key insight is that solving a PDE in time is equivalent to exponentiating an (often unbounded) operator, and the conditions for this exponentiation to produce well-behaved solutions are precisely the conditions (closedness, density of domain, dissipativity) that the Hille-Yosida theorem verifies.

Specific Questions Ahead#

The machinery of closed graphs, adjoints, and spectral measures solves the static problem: what are the eigenvalues, and how do we define functions of an operator? The dynamic problem remains. How do we construct $e^{tA}$ when $A$ is unbounded and not self-adjoint? What resolvent estimate replaces unitarity when the system dissipates energy? How does the Lumer-Phillips theorem reduce PDE well-posedness to a single inequality on the numerical range? Why does the heat equation smooth discontinuous initial data instantly, while the wave equation preserves singularities along characteristics?

Why You Are Equipped#

You now know that unbounded operators are defined by their domains, and that closedness is the minimal regularity condition for a well-defined resolvent. You have computed graph norms, verified deficiency indices, and seen how the spectral theorem converts differential operators into multiplication operators. These are exactly the prerequisites for semigroup theory. The Hille-Yosida theorem does not require self-adjointness; it requires closedness, dense domain, and a uniform bound on the resolvent powers $\|(\lambda - A)^{-n}\|$ . You already know how to check closedness via graph convergence. You already know how to compute resolvents for differential operators. The transition from spectral theory to evolution equations is a shift from analyzing $(\lambda - A)^{-1}$ at fixed $\lambda$ to controlling it uniformly as $\lambda \to \infty$ .

Theorem Preview: Hille-Yosida#

The next article centers on the Hille-Yosida theorem. It states that a densely defined closed linear operator $A$ generates a strongly continuous contraction semigroup $\{T(t)\}_{t \geq 0}$ if and only if $(0, \infty) \subset \rho(A)$ and $\|(\lambda - A)^{-n}\| \leq \lambda^{-n}$ for all $\lambda > 0$ and $n \in \mathbb{N}$ . This resolvent condition is the unbounded analogue of $\|e^{tA}\| \leq 1$ . I will prove the theorem using the Yosida approximation $A_\lambda = \lambda A(\lambda - A)^{-1}$ , a family of bounded operators that converge to $A$ in the strong resolvent sense. We will apply it directly to the Dirichlet Laplacian (generating the heat semigroup), the transport operator (generating shifts), and non-self-adjoint Fokker-Planck generators. The abstract inequality becomes a concrete energy estimate. You will see exactly why the graph norm topology is the natural setting for initial data, and how the spectral mapping theorem $\sigma(e^{tA}) \setminus \{0\} = e^{t\sigma(A)}$ emerges from the resolvent bounds. The static operator becomes a time machine.