Bounded, Compact, and Integral Operators

For a linear map between normed spaces, continuity is equivalent to one uniform estimate,

$$\|Tx\|_Y\leq C\|x\|_X.$$

The least possible $C$ is the operator norm. Compactness is stronger: a compact operator sends every bounded sequence to a sequence having a norm-convergent subsequence. An integral formula does not by itself imply either property. A square-integrable kernel does give a particularly useful answer: it defines a Hilbert–Schmidt operator, hence a compact bounded operator.

On an infinite-dimensional separable Hilbert space, the resulting hierarchy is strict:

$$\{\text{finite-rank operators}\} \subsetneq \mathcal S_2(H) \subsetneq \mathcal K(H) \subsetneq \mathcal B(H).$$

Here $\mathcal S_2(H)$ denotes the Hilbert–Schmidt operators, $\mathcal K(H)$ the compact operators, and $\mathcal B(H)$ the bounded operators. Compactness alone does not imply self-adjointness, normality, an orthonormal eigenbasis, or Hilbert–Schmidt summability. Those distinctions matter already for a regulated free-field covariance: it is compact at finite volume, but its Hilbert–Schmidt property depends on dimension, and the corresponding infinite-volume multiplier is not compact.

Required background. Banach and Hilbert Spaces, Completion, and Riesz Representation supplies completeness, orthonormal expansions, and Riesz representation.

Bounded and compact operator classes

Let $X$ and $Y$ be normed spaces over $\mathbb K\in\{\mathbb R,\mathbb C\}$, and let $H_1,H_2$ be Hilbert spaces. Spectral statements below are made over $\mathbb C$ unless explicitly specialized. The site convention makes the bra slot conjugate-linear and the ket slot linear:

$$\langle \alpha\eta|\xi\rangle = \alpha^*\langle\eta|\xi\rangle, \qquad \langle\eta|\alpha\xi\rangle = \alpha\langle\eta|\xi\rangle.$$

The symbol $T^\dagger$ denotes the Hilbert-space adjoint; $X^*$ continues to denote the continuous Banach dual. Every bounded operator on this page is defined on its whole source space. An operator such as a derivative, a Hamiltonian, or $-\Delta+m^2$ is normally defined only on a proper dense domain. Its closure and adjoint require domain data and belong to Unbounded Operators, Domains, Closure, and Adjoints.

The later page on Spectra, Resolvents, and Functional Calculus develops general spectral theory. The present page states only the compact operator consequences needed to distinguish discrete modes from continuous spectrum. Summability, traces, and determinants continue at Trace Ideals and Fredholm Determinants.

Bounded linear maps are exactly the continuous ones

A linear map $T:X\to Y$ is bounded when there is a finite constant $C\geq0$ such that

$$\|Tx\|_Y\leq C\|x\|_X \qquad \text{for every }x\in X.$$

This does not say that the image of the whole, unbounded space $X$ is a bounded subset of $Y$. It says that $T$ has a uniform linear growth bound. The operator norm is

$$\begin{aligned} \|T\|_{\mathrm{op}} &= \sup_{\|x\|_X\leq1}\|Tx\|_Y\\ &= \sup_{x\neq0} \frac{\|Tx\|_Y}{\|x\|_X}. \end{aligned}$$

Thus $\|Tx\|_Y\leq\|T\|_{\mathrm{op}}\|x\|_X$, and the norm is the smallest constant with this property.

For a linear map, the following conditions are equivalent:

1. $T$ is continuous on $X$;
2. $T$ is continuous at $0$;
3. $T$ is bounded on the closed unit ball; and
4. $\|T\|_{\mathrm{op}}<\infty$.

Only continuity at zero needs to be checked because

$$\|Tx-Tx_0\|_Y=\|T(x-x_0)\|_Y.$$

Conversely, if continuity at zero gives $\|Tx\|_Y<1$ whenever $\|x\|_X<\delta$, then scaling any nonzero $x$ to have norm less than $\delta$ yields a bound of the form $\|Tx\|_Y\leq C\|x\|_X$. Linearity turns a local continuity statement into a global norm estimate.

The bounded operators from $X$ to $Y$ form the normed space $\mathcal B(X,Y)$. If $Y$ is Banach, then $\mathcal B(X,Y)$ is Banach: an operator-norm Cauchy sequence $(T_n)$ makes $(T_nx)$ Cauchy in $Y$ for each $x$, and the pointwise limit $Tx=\lim_nT_nx$ is linear and bounded. The same uniform estimate then gives $\|T_n-T\|_{\mathrm{op}}\to0$.

Composition is compatible with the norm. If $T\in\mathcal B(X,Y)$ and $S\in\mathcal B(Y,Z)$, then

$$\|ST\|_{\mathrm{op}} \leq \|S\|_{\mathrm{op}}\|T\|_{\mathrm{op}}.$$

This inequality is often more important than an exact norm. It lets one control a long construction by controlling its factors.

A rank-one operator

For $u\in H_2$ and $v\in H_1$, define

$$T=|u\rangle\langle v|, \qquad T\xi=u\langle v|\xi\rangle.$$

Cauchy–Schwarz gives

$$\|T\xi\| \leq \|u\|\,\|v\|\,\|\xi\|.$$

When $u,v\neq0$, equality is attained at $\xi=v/\|v\|$, so

$$\big\||u\rangle\langle v|\big\|_{\mathrm{op}} = \|u\|\,\|v\|.$$

Finite sums of such maps have finite-dimensional range. They will be the basic approximants for compact and integral operators.

Values on basis vectors are not enough

Suppose $\{e_n\}_{n\geq1}$ is the standard orthonormal basis of $\ell^2$. The rule

$$Ae_n=ne_n$$

is perfectly finite on every basis vector, but $\|Ae_n\|=n$ has no uniform bound. It therefore does not define a bounded operator on all of $\ell^2$. It defines an unbounded operator only after one specifies a suitable domain, for example

$$\mathcal D(A) = \left\{ x\in\ell^2: \sum_{n=1}^{\infty}n^2|x_n|^2<\infty \right\}.$$

Checking finitely many modes, or checking every mode separately without a uniform estimate, cannot establish boundedness.

The bounded Hilbert adjoint

Let $T\in\mathcal B(H_1,H_2)$. For fixed $\eta\in H_2$, the map

$$\xi\longmapsto\langle\eta|T\xi\rangle$$

is a bounded linear functional on $H_1$. Riesz representation therefore gives a unique vector $T^\dagger\eta\in H_1$ satisfying

$$\boxed{ \langle\eta|T\xi\rangle_{H_2} = \langle T^\dagger\eta|\xi\rangle_{H_1} } \qquad \text{for all }\eta\in H_2,\ \xi\in H_1.$$

The dependence on $\eta$ is linear. Indeed, both the bra in the first expression and the representing-vector slot in the second are conjugate-linear, so the two conjugations cancel. Standard consequences are

$$\begin{aligned} \|T^\dagger\|_{\mathrm{op}}&=\|T\|_{\mathrm{op}},\\ (ST)^\dagger&=T^\dagger S^\dagger,\\ (T^\dagger)^\dagger&=T. \end{aligned}$$

For the rank-one map above,

$$\big(|u\rangle\langle v|\big)^\dagger = |v\rangle\langle u|.$$

This construction uses that $T$ is bounded and defined everywhere. The adjoint of a proper-domain operator has a domain determined by a boundedness condition on a pairing; it cannot be obtained by silently reusing the bounded formula.

Compactness controls bounded sequences

Let $B_X=\{x\in X:\|x\|_X\leq1\}$. A bounded linear map $K:X\to Y$ is compact when $K(B_X)$ has compact closure in $Y$. Equivalently, $K$ is compact when every bounded sequence $(x_n)$ in $X$ has a subsequence $(x_{n_j})$ for which $(Kx_{n_j})$ converges in norm.

The word “compact” therefore describes the image of bounded sets, not the size of $\|K\|_{\mathrm{op}}$. Multiplying a nonzero compact operator by a large scalar keeps it compact, while multiplying a noncompact operator by a small nonzero scalar does not make it compact.

Several permanence properties follow directly:

• every finite-rank bounded operator is compact;
• if $K:X\to Y$ is compact and $A:Y\to Z$ and $B:W\to X$ are bounded, then $AKB:W\to Z$ is compact;
• an operator-norm limit of compact operators is compact when the codomain is Banach; and
• for Hilbert spaces, $K$ is compact exactly when $K^\dagger$ is compact.

The last statement is the Hilbert-space form of Schauder’s theorem. The closedness statement is worth distinguishing from pointwise, or strong, operator convergence.

The identity and shift are bounded but not compact

On an infinite-dimensional Hilbert space, choose an orthonormal sequence $(e_n)$. The sequence lies in the unit ball, but

$$\|e_n-e_m\|=\sqrt2 \qquad(n\neq m).$$

It has no norm-convergent subsequence. Consequently the identity is not compact.

The unilateral shift $S:\ell^2\to\ell^2$,

$$Se_n=e_{n+1},$$

has $\|S\|_{\mathrm{op}}=1$, but $(Se_n)$ is again orthonormal. It too is not compact. Boundedness gives norm control; it does not force subsequence convergence.

Finite-rank approximation and its boundary

On Hilbert spaces, compact operators are exactly the operator-norm limits of finite-rank operators. One direction follows because finite-rank maps are compact and compact maps are operator-norm closed. In the other direction, compactness lets a finite-dimensional projection approximate the compact closure of the image of the unit ball uniformly.

This equivalence must not be transferred without qualification to arbitrary Banach spaces. Norm limits of finite-rank maps are always compact, but some Banach spaces lack the approximation property, and compact maps there need not admit the corresponding finite-rank norm approximation.

Even on Hilbert space, strong convergence is too weak. Let $P_N$ project $\ell^2$ onto $\operatorname{span}\{e_1,\ldots,e_N\}$. For every fixed $x\in\ell^2$,

$$P_Nx\longrightarrow x,$$

but

$$\|I-P_N\|_{\mathrm{op}}=1$$

for every $N$. Thus the identity is a strong limit of finite-rank compact operators without being compact.

The boundedness and compactness results in this section follow Kehle 2025, §§2.1 and 6.2, PDF and Teschl 2014, §§0.5 and 6.2, PDF.

Diagonal operators separate the classes

Let $a=(a_n)$ be a scalar sequence and initially define

$$D_ae_n=a_ne_n$$

on the finite sequences in $\ell^2$. Then

$$\|D_ax\|_2^2 = \sum_{n=1}^{\infty}|a_n|^2|x_n|^2.$$

It follows that

$$D_a\in\mathcal B(\ell^2) \quad\Longleftrightarrow\quad a\in\ell^\infty, \qquad \|D_a\|_{\mathrm{op}}=\sup_n|a_n|.$$

If $a_n\to0$, the truncations

$$D_a^{(N)}e_n = \begin{cases} a_ne_n,&n\leq N,\\ 0,&n>N \end{cases}$$

have finite rank and satisfy

$$\|D_a-D_a^{(N)}\|_{\mathrm{op}} = \sup_{n>N}|a_n| \longrightarrow0.$$

Hence $D_a$ is compact. Conversely, if $a_n$ does not tend to zero, there are an $\varepsilon>0$ and distinct indices $n_j$ with $|a_{n_j}|\geq\varepsilon$. The vectors $D_ae_{n_j}$ are mutually orthogonal and separated by at least $\sqrt2\,\varepsilon$, so they have no convergent subsequence. Therefore

$$\boxed{ D_a\text{ is compact} \quad\Longleftrightarrow\quad a_n\to0. }$$

The examples $a_n=1$, $a_n=n^{-1/2}$, and $a_n=n^{-1}$ will distinguish bounded, compact, and Hilbert–Schmidt behavior below. Notice also that $a_n\to0$ need not make $D_a$ finite rank.

Integral kernels need hypotheses

Let $(M,\mu)$ and $(N,\nu)$ be $\sigma$-finite measure spaces whose $L^2$ spaces are separable, as they are for the standard Borel measure spaces used in the applications below. Use output-first notation:

$$(K_kf)(y) = \int_M k(y,x)f(x)\,\mathrm d\mu(x), \qquad y\in N.$$

The displayed integral must exist in an appropriate almost-everywhere sense, and its output must belong to the claimed target space. Measurability, integrability, and the use of Tonelli or Fubini are hypotheses, not consequences of writing a kernel symbol.

The Schur test gives boundedness

One useful sufficient criterion is the unweighted Schur test. Suppose $k:N\times M\to\mathbb C$ is measurable and

$$\begin{aligned} \mathop{\operatorname{ess\,sup}}_{y\in N} \int_M|k(y,x)|\,\mathrm d\mu(x)&\leq A,\\ \mathop{\operatorname{ess\,sup}}_{x\in M} \int_N|k(y,x)|\,\mathrm d\nu(y)&\leq B \end{aligned}$$

for finite $A,B$. Then the integral operator, first defined on a suitable dense class, extends uniquely to a bounded map

$$K_k:L^2(M,\mu)\longrightarrow L^2(N,\nu)$$

with

$$\|K_k\|_{\mathrm{op}}\leq\sqrt{AB}.$$

Indeed, Cauchy–Schwarz with respect to the measure $|k(y,x)|\,\mathrm d\mu(x)$ gives

$$|(K_kf)(y)|^2 \leq \left(\int_M|k(y,x)|\,\mathrm d\mu(x)\right) \left(\int_M|k(y,x)||f(x)|^2\,\mathrm d\mu(x)\right).$$

Integrating over $y$, applying the first bound, and then Tonelli and the second bound yields

$$\|K_kf\|_2^2\leq AB\|f\|_2^2.$$

The Schur bounds establish boundedness, not compactness. For example, if $h\in L^1(\mathbb R^s)$ is nonzero, then

$$(Kf)(y) = \int_{\mathbb R^s}h(y-x)f(x)\,\mathrm d^sx$$

is bounded on $L^2(\mathbb R^s)$ with $\|K\|_{\mathrm{op}}\leq\|h\|_1$. It is not compact. Choose an input $f$ for which $h*f\neq0$ and translate $f$ by vectors tending mutually far apart. The outputs are the corresponding translates of the nonzero $L^2$ function $h*f$ and have no norm-convergent subsequence. The kernel is also not square-integrable on $\mathbb R^s\times\mathbb R^s$ when $h\neq0$.

A square-integrable kernel is compact

Now suppose

$$k\in L^2(N\times M,\nu\otimes\mu).$$

For almost every $y$, Cauchy–Schwarz in $x$ gives

$$|(K_kf)(y)|^2 \leq \left(\int_M|k(y,x)|^2\,\mathrm d\mu(x)\right) \|f\|_2^2.$$

Tonelli then gives the operator-norm estimate

$$\boxed{ \|K_kf\|_{L^2(N)} \leq \|k\|_{L^2(N\times M)} \|f\|_{L^2(M)}. }$$

Thus $K_k$ has a unique bounded extension from any dense class on which the integral was first defined.

There is more structure. Finite sums of product functions are dense in the product $L^2$ space, so choose

$$k_r(y,x) = \sum_{j=1}^{m_r}a_{rj}(y)b_{rj}(x)^*$$

with $\|k-k_r\|_2\to0$. The associated operator is

$$K_{k_r}f = \sum_{j=1}^{m_r}a_{rj}\langle b_{rj}|f\rangle,$$

which has finite-dimensional range. The preceding estimate gives

$$\|K_k-K_{k_r}\|_{\mathrm{op}} \leq \|k-k_r\|_2 \longrightarrow0.$$

Therefore $K_k$ is compact. This proof is constructive: approximating the kernel in $L^2$ produces finite-rank operator approximations in operator norm.

The Hilbert–Schmidt norm

For separable Hilbert spaces, a bounded operator $T:H_1\to H_2$ is Hilbert–Schmidt when, for one and hence every orthonormal basis $(e_n)$ of $H_1$,

$$\|T\|_{\mathrm{HS}}^2 := \sum_n\|Te_n\|_{H_2}^2 <\infty.$$

Parseval’s identity shows that this sum is independent of the chosen basis. It also gives

$$\|T\|_{\mathrm{op}}\leq\|T\|_{\mathrm{HS}}.$$

If $P_N$ projects onto the first $N$ basis vectors, then $TP_N$ has finite rank and

$$\|T-TP_N\|_{\mathrm{op}}^2 \leq \sum_{n>N}\|Te_n\|^2 \longrightarrow0.$$

Every Hilbert–Schmidt operator is consequently compact.

For the integral operator above, Parseval and Tonelli give the exact Hilbert–Schmidt identity

$$\boxed{ \|K_k\|_{\mathrm{HS}} = \|k\|_{L^2(N\times M)}. }$$

This equality is stronger than, and must not be confused with, the earlier operator-norm inequality

$$\|K_k\|_{\mathrm{op}}\leq\|k\|_2.$$

The diagonal operator is Hilbert–Schmidt exactly when

$$\|D_a\|_{\mathrm{HS}}^2=\sum_n|a_n|^2<\infty.$$

It follows that

• $D_1=I$ is bounded but not compact;
• $D_{1/\sqrt n}$ is compact but not Hilbert–Schmidt; and
• $D_{1/n}$ is Hilbert–Schmidt and compact but has infinite rank.

These examples prove all three inclusions in the opening hierarchy are strict. Whether an operator has a summable trace is a separate, stronger question deferred to the trace-ideal page.

The adjoint kernel and symmetry

For $f\in L^2(M)$ and $g\in L^2(N)$, Fubini gives

$$\begin{aligned} \langle g|K_kf\rangle_{L^2(N)} &= \int_N\int_M g(y)^*k(y,x)f(x)\, \mathrm d\mu(x)\mathrm d\nu(y)\\ &= \langle K_k^\dagger g|f\rangle_{L^2(M)}, \end{aligned}$$

where

$$(K_k^\dagger g)(x) = \int_N k(y,x)^*g(y)\,\mathrm d\nu(y).$$

Thus the adjoint kernel, with its arguments reversed, is

$$k^\dagger(x,y)=k(y,x)^*.$$

When $M=N$ with the same measure, the almost-everywhere condition

$$k(y,x)=k(x,y)^*$$

makes $K_k$ self-adjoint. Mere symmetry without complex conjugation is not the correct condition for a complex Hilbert space.

Square-integrability is sufficient, not necessary. The identity on $L^2(\mathbb R^s)$ has the formal distributional kernel $\delta^{(s)}(y-x)$, but this is not an $L^2$ function and the identity is not compact. Conversely, merely knowing that $k$ is pointwise bounded on an infinite-measure product does not define a bounded $L^2$ operator. The constant kernel $k(y,x)=1$ on $\mathbb R^s\times\mathbb R^s$ already fails: for many $L^2$ inputs the integral is undefined, and when it is a nonzero constant the output is not in $L^2(\mathbb R^s)$.

The Schur criterion and the square-integrable-kernel and Hilbert–Schmidt statements follow Teschl 2014, §§0.6 and 6.3, PDF, with the kernel order translated to the site’s inner-product convention.

A compact integral operator need not have eigenvectors

The Volterra operator on $L^2([0,1])$ is

$$(Vf)(x)=\int_0^x f(y)\,\mathrm dy.$$

Its kernel is

$$v(x,y)=\mathbf 1_{\{0\leq y\leq x\leq1\}},$$

and

$$\|v\|_2^2 = \int_0^1\int_0^x\mathrm dy\,\mathrm dx = \frac12.$$

Therefore $V$ is Hilbert–Schmidt and compact. Its adjoint is

$$(V^\dagger g)(y)=\int_y^1g(x)\,\mathrm dx,$$

so $V$ is not self-adjoint.

It has no eigenvectors. If $Vf=\lambda f$ with $\lambda\neq0$, then $Vf$ is absolutely continuous, hence so is $f$, and differentiation gives

$$\lambda f'(x)=f(x) \quad\text{almost everywhere}, \qquad f(0)=0.$$

The only solution is $f=0$. For $\lambda=0$, the identity $Vf=0$ also implies $f=0$ almost everywhere. Thus compactness alone does not even guarantee one eigenvector, much less an orthonormal eigenbasis.

Exactly what compactness says about spectrum

For a compact operator $K$ on an infinite-dimensional complex Hilbert space, the Riesz–Schauder conclusions needed here are:

• every nonzero point of $\sigma(K)$ is an eigenvalue;
• each nonzero eigenspace is finite-dimensional;
• the nonzero spectrum is finite or countable; and
• its only possible accumulation point is $0$.

The Fredholm alternative gives the first statement, while compactness makes the unit ball in a nonzero eigenspace compact and therefore forces that eigenspace to be finite-dimensional. The accumulation claim also has a direct compactness proof. If there were distinct eigenvalues $|\lambda_n|\geq\varepsilon>0$, let $M_n$ be the span of eigenvectors for $\lambda_1,\ldots,\lambda_n$. Riesz’s lemma gives $y_n\in M_n$ with

$$\|y_n\|=1, \qquad \operatorname{dist}(y_n,M_{n-1})\geq\frac12.$$

The spaces $M_n$ are invariant, and

$$(K-\lambda_n I)y_n\in M_{n-1}.$$

Consequently the bounded sequence $y_n/\lambda_n$ has images under $K$ separated by at least $1/2$, contradicting compactness. Thus there are only finitely many distinct eigenvalues outside every disk $|\lambda|<\varepsilon$. Countability and accumulation only at zero follow.

The point $0$ belongs to the spectrum in infinite dimension because a compact operator cannot have a bounded inverse: otherwise $I=K^{-1}K$ would be compact. Yet $0$ need not be an eigenvalue. The injective compact diagonal operator $D_{1/n}$ and the Volterra operator both show this.

More geometry requires more hypotheses. If $K=K^\dagger$ is compact and self-adjoint on a Hilbert space, then its nonzero eigenvalues are real, their eigenspaces for distinct eigenvalues are orthogonal, and

$$H = \ker K \mathbin{\widehat\oplus} \overline{\operatorname{span}} \{u_j:Ku_j=\lambda_ju_j,\ \lambda_j\neq0\}.$$

After adjoining an orthonormal basis of $\ker K$, one obtains an orthonormal eigenbasis, and

$$Kx = \sum_j\lambda_j u_j\langle u_j|x\rangle, \qquad \lambda_j\longrightarrow0$$

when there are infinitely many nonzero eigenvalues. A compact normal operator has an analogous orthonormal eigenvector expansion, but a general compact non-normal operator need not be diagonalizable, as the Volterra example demonstrates.

The general nonzero-spectrum statement and compact self-adjoint theorem are treated in Kehle 2025, §§6.3–6.4, PDF and Teschl 2014, §6.2, PDF. General spectra, continuous spectrum, resolvents, spectral measures, and functional calculus require the later spectral page.

The hypotheses and conclusions can be summarized as follows:

Hypothesis on an everywhere-defined map	Licensed conclusion
bounded linear	continuous, with uniform norm control
compact	bounded sequences have image subsequences converging in norm
compact endomorphism on a complex Hilbert space	nonzero spectrum consists of eigenvalues with finite-dimensional eigenspaces, accumulating only at zero
compact and normal on a Hilbert space	orthonormal eigenvector expansion
compact and self-adjoint	the eigenvalues in that expansion are real
square-integrable kernel on $L^2$	Hilbert–Schmidt, hence compact

No row licenses a conclusion from a weaker hypothesis appearing above it.

Finite-volume free covariance

Compactify all $d$ Euclidean directions and consider the complex Hilbert space

$$H_L=L^2(\mathbb T_L^d), \qquad \mathbb T_L^d=(\mathbb R/L\mathbb Z)^d,$$

and the orthonormal Fourier modes

$$e_n(x) = L^{-d/2} \exp\left(\frac{2\pi i\,n\cdot x}{L}\right), \qquad n\in\mathbb Z^d.$$

For $m>0$, define the finite-volume Euclidean free covariance directly by its bounded Fourier action:

$$C_{m,L}e_n = c_ne_n, \qquad c_n = \frac{1}{ \left(\frac{2\pi}{L}\right)^2|n|^2+m^2 }.$$

The notation $C_{m,L}=(-\Delta+m^2)^{-1}$ is useful, but the displayed multiplier is the definition needed here; the domain theory of the unbounded operator $-\Delta+m^2$ is deferred.

The diagonal tests give immediately

$$\|C_{m,L}\|_{\mathrm{op}}=\frac1{m^2}, \qquad C_{m,L}=C_{m,L}^\dagger\geq0.$$

Because $c_n\to0$ as $|n|\to\infty$, $C_{m,L}$ is compact in every finite Euclidean dimension $d$. Let $P_N$ project onto the modes with $|n|\leq N$. Then

$$C_{m,L}^{(N)} = P_NC_{m,L}P_N$$

has finite rank and

$$\begin{aligned} \|C_{m,L}-C_{m,L}^{(N)}\|_{\mathrm{op}} &= \sup_{|n|>N} \frac1{ \left(\frac{2\pi}{L}\right)^2|n|^2+m^2 }\\ &\leq \frac1{ \left(\frac{2\pi N}{L}\right)^2+m^2 } \longrightarrow0. \end{aligned}$$

This is a controlled norm estimate for a discrete momentum cutoff, not merely convergence mode by mode.

This example is also the promised compact-resolvent application. If

$$A_{m,L}=-\Delta+m^2$$

is supplied with its standard periodic Sobolev domain $\mathcal D(A_{m,L})=H^2(\mathbb T_L^d)$, then it is positive self-adjoint, $0$ lies in its resolvent set, and

$$(A_{m,L}-0I)^{-1}=C_{m,L}$$

is compact. Thus $A_{m,L}$ has compact resolvent. The multiplier calculation establishes the inverse and compactness used here; the general domain and resolvent theory remains with the later unbounded-operator and spectral pages.

Compact does not always mean Hilbert–Schmidt

The covariance is Hilbert–Schmidt exactly when

$$\|C_{m,L}\|_{\mathrm{HS}}^2 = \sum_{n\in\mathbb Z^d} \frac1{ \left[ \left(\frac{2\pi}{L}\right)^2|n|^2+m^2 \right]^2 } <\infty.$$

At large $|n|$, the summand behaves like $|n|^{-4}$, while a shell of radius $R$ contains order $R^{d-1}$ lattice points. The series therefore converges exactly when

$$d<4.$$

For $d\geq4$, $C_{m,L}$ remains compact but is not Hilbert–Schmidt. Formally its periodic kernel is

$$G_{m,L}(x,y) = \frac1{L^d} \sum_{n\in\mathbb Z^d} \frac{ \exp\left(\frac{2\pi i\,n\cdot(x-y)}{L}\right) }{ \left(\frac{2\pi}{L}\right)^2|n|^2+m^2 }.$$

For $d<4$ this series represents an $L^2$ kernel in the corresponding Fourier sense. For $d\geq4$ it does not represent an $L^2$ kernel, even though the bounded compact operator remains well defined by its multiplier. The failure is ultraviolet, not a failure of compactness.

The assumption $m>0$ also matters. At $m=0$, the zero mode has a vanishing denominator. One must remove that mode, impose a zero-mean condition, add an infrared regulator, or otherwise specify a different problem before claiming a bounded covariance.

Infinite volume removes compactness

On $L^2(\mathbb R^d)$, the analogous massive covariance becomes, after the unitary Fourier transform, multiplication by

$$c(p)=\frac1{|p|^2+m^2}.$$

It is bounded with norm $m^{-2}$, but it is not compact. Choose a momentum ball on which $c(p)\geq\varepsilon>0$, divide it into infinitely many pairwise disjoint measurable sets $E_j$ of positive finite measure, and set

$$g_j = \frac{\mathbf 1_{E_j}}{|E_j|^{1/2}}.$$

The $g_j$ are orthonormal, while the vectors $cg_j$ have disjoint supports and norms at least $\varepsilon$. Hence $(cg_j)$ has no norm-convergent subsequence.

Finite volume has replaced continuous momentum by a discrete lattice whose covariance eigenvalues tend to zero. Infinite volume restores infinitely many mutually separated wavepackets in any momentum region of nonzero measure. This mathematical comparison does not say that finite volume makes every bounded operator compact, nor does it establish the spectrum of an interacting QFT.

The free scalar propagator motivating this Euclidean covariance is reviewed in Schwartz 2014, §6.2. The passage here to a positive Euclidean multiplier and a periodic finite-volume regulator is explicit and limited to this operator test. The developed physical treatment of discrete spectral weights, continua, positivity, and two-point normalization belongs to Spectral Decomposition of Two-Point Functions, with Schwartz 2014, §24.2.1 as the application source. The site’s $(+---)$ convention governs the Lorentzian discussion there; the Euclidean operator above has no Lorentzian signature choice.

What the operator tests establish

The principal question has three different answers:

• linear continuity is exactly a finite operator norm;
• compactness is boundedness plus norm-subsequence control on every bounded sequence; and
• an integral kernel defines a bounded or compact map only after measurable size conditions such as the Schur bounds or square-integrability have been verified.

Finite-rank approximations converge in operator norm for compact operators on Hilbert space, but strong convergence alone is insufficient. A square-integrable kernel is Hilbert–Schmidt and compact, but compact operators need not be Hilbert–Schmidt or have ordinary function kernels. Finally, for a compact Hilbert-space operator, normality or self-adjointness licenses an orthonormal eigenvector expansion; compactness by itself licenses only the Riesz–Schauder structure away from zero.

For QFT-facing approximations, this yields a practical sequence of checks: state the Hilbert space and domain, prove the operator norm estimate, decide compactness, test any stronger summability separately, and identify which conclusion changes when a regulator or finite volume is removed.

Common pitfalls

Bounded does not mean finite-dimensional or compact. The identity and the unilateral shift have operator norm one on $\ell^2$, but neither is compact.

Modewise finiteness is not a uniform bound. The rule $Ae_n=ne_n$ is finite on every basis vector and still unbounded. The supremum over the unit ball is the relevant test.

Strong finite-rank approximation does not prove compactness. The projections $P_N$ converge strongly to the identity while staying operator norm distance one from it.

A kernel formula is not an operator theorem. The integral must be measurable, exist almost everywhere, and produce a vector in the announced target space. Schur and $L^2$ estimates are useful sufficient conditions, not necessary conditions.

Compact does not imply Hilbert–Schmidt. The operator $D_{1/\sqrt n}$ and the finite-volume covariance in dimension $d\geq4$ are compact without having a finite Hilbert–Schmidt norm.

Compact does not imply an orthonormal eigenbasis. Normality is the missing geometric hypothesis. The compact Volterra operator has no eigenvectors at all.

A formal diagonal value is not automatically a trace. An $L^2$ kernel is an almost-everywhere equivalence class and need not have a canonical diagonal representative. On a non-atomic continuum space, the diagonal has product measure zero, so changing values there does not even change the $L^2$ kernel. Trace formulas require stronger hypotheses developed on Trace Ideals and Fredholm Determinants.

Finite-volume conclusions do not automatically survive infinite volume. The massive covariance is compact on a torus and noncompact on $\mathbb R^d$. State which topology and regulator are being removed before claiming spectral convergence.

Exercises

Compute a rank-one norm and adjoint

Let $T=|u\rangle\langle v|:H_1\to H_2$. Prove its norm formula, compute $T^\dagger$, and determine when $T$ is self-adjoint in the case $H_1=H_2$.

Solution

Cauchy–Schwarz gives

$$\|T\xi\| = \|u\|\,|\langle v|\xi\rangle| \leq \|u\|\,\|v\|\,\|\xi\|.$$

If $u,v\neq0$, equality holds for $\xi=v/\|v\|$, so $\|T\|_{\mathrm{op}}=\|u\|\,\|v\|$. The zero cases satisfy the same formula. Moreover,

$$\langle\eta|T\xi\rangle = \langle\eta|u\rangle\langle v|\xi\rangle.$$

The vector $v\langle u|\eta\rangle$ represents this functional because the first slot is conjugate-linear. Hence

$$T^\dagger\eta = v\langle u|\eta\rangle, \qquad T^\dagger=|v\rangle\langle u|.$$

The map is self-adjoint exactly when $|u\rangle\langle v|=|v\rangle\langle u|$. Apart from the zero operator, this holds precisely when $u=\alpha v$ for a real scalar $\alpha\neq0$; equivalently, the rank-one map is a real multiple of $|v\rangle\langle v|$.

Classify diagonal operators

For each sequence below, classify $D_ae_n=a_ne_n$ as bounded, compact, and Hilbert–Schmidt:

1. $a_n=(-1)^n$;
2. $a_n=n^{-1/2}$;
3. $a_n=n^{-1}$; and
4. $a_n=n$.

Solution

Use

$$\begin{aligned} D_a\text{ bounded} &\Longleftrightarrow \sup_n|a_n|<\infty,\\ D_a\text{ compact} &\Longleftrightarrow a_n\to0,\\ D_a\text{ Hilbert–Schmidt} &\Longleftrightarrow \sum_n|a_n|^2<\infty. \end{aligned}$$

Thus:

1. $D_{(-1)^n}$ is bounded but not compact or Hilbert–Schmidt.
2. $D_{n^{-1/2}}$ is bounded and compact, but not Hilbert–Schmidt because $\sum_n1/n$ diverges.
3. $D_{n^{-1}}$ is Hilbert–Schmidt, hence compact and bounded.
4. The rule $D_ne_n=ne_n$ is unbounded and is not an everywhere-defined bounded, compact, or Hilbert–Schmidt operator on $\ell^2$.

Test an integral operator and its adjoint

For the Volterra kernel

$$v(x,y)=\mathbf 1_{\{0\leq y\leq x\leq1\}},$$

compute the Hilbert–Schmidt norm of $V$, derive $V^\dagger$, and decide whether $V$ is self-adjoint.

Solution

The kernel occupies a triangle of area $1/2$, so

$$\|V\|_{\mathrm{HS}}^2 = \int_0^1\int_0^1|v(x,y)|^2 \,\mathrm dy\,\mathrm dx = \frac12.$$

Hence $\|V\|_{\mathrm{HS}}=1/\sqrt2$ and $V$ is compact. Reversing the arguments and conjugating gives

$$v^\dagger(y,x)=v(x,y),$$

or, in operator form,

$$(V^\dagger g)(y)=\int_y^1g(x)\,\mathrm dx.$$

This differs from $\int_0^y g(x)\,\mathrm dx$, so $V$ is not self-adjoint.

Compare finite and infinite volume

For the covariance $C_{m,L}$:

1. derive the momentum-cutoff operator-norm error;
2. determine in which Euclidean dimensions it is Hilbert–Schmidt;
3. explain what fails at $m=0$; and
4. prove that the infinite-volume multiplier is not compact.

Solution

Because the Fourier basis diagonalizes both $C_{m,L}$ and $P_N$,

$$\|C_{m,L}-P_NC_{m,L}P_N\|_{\mathrm{op}} = \sup_{|n|>N} \frac1{ \left(\frac{2\pi}{L}\right)^2|n|^2+m^2 } \leq \frac1{ \left(\frac{2\pi N}{L}\right)^2+m^2 }.$$

The Hilbert–Schmidt norm squared is the sum of the squared eigenvalues. Its large-$|n|$ terms behave as $|n|^{-4}$, so lattice counting gives convergence exactly for $d<4$.

At $m=0$, the $n=0$ coefficient diverges. Removing the zero mode or adding an infrared regulator is additional data, not an automatic step.

In infinite volume, work in momentum space and choose infinitely many disjoint positive-measure sets $E_j$ inside a ball on which $(|p|^2+m^2)^{-1}\geq\varepsilon$. The normalized indicators of the $E_j$ are orthonormal, and their multiplied images have disjoint supports and norm at least $\varepsilon$. The images therefore have no norm-convergent subsequence, so the multiplier is not compact.

References

• Christoph Kehle, Introduction to Functional Analysis, PDF, lecture notes for MIT 18.102, Spring 2025, especially §§2.1, 5.4, and 6.1–6.4. These sections develop bounded maps, bounded adjoints, compactness, finite-rank approximation, the Fredholm alternative, and the compact self-adjoint spectral theorem. It uses the same conjugate-linear-first-slot convention as this page.
• Matthew D. Schwartz, Quantum Field Theory and the Standard Model, Cambridge University Press, 2014, especially §§6.2 and 24.2.1. These sections derive the free scalar propagator and the spectral decomposition of scalar two-point functions. The page makes the Euclidean and finite-volume regulator explicit and hands the developed physical interpretation to Foundations.
• Gerald Teschl, Mathematical Methods in Quantum Mechanics: With Applications to Schrödinger Operators, PDF, second edition, Graduate Studies in Mathematics 157, American Mathematical Society, 2014, especially §§0.5–0.6 and 6.2–6.3. These sections establish bounded operators, compact self-adjoint spectral structure, the singular-value canonical form, square-integrable kernels, and Hilbert–Schmidt operators. Its inner-product convention agrees with the site convention. The author’s second-edition errata, PDF, updated March 18, 2026, was checked for the cited sections.