Neural-Path/Notes
45 min

Bounded Linear Operators: Adjoints, Spectrum & Compact Operators

Bounded linear operators between Hilbert spaces generalize matrices to infinite dimensions — their spectrum characterizes invertibility and dynamical behavior, while compact operators form a tractable subclass with a complete spectral theory that mirrors finite-dimensional eigendecomposition.

Concepts

Gram matrix K(i,j) = k(xᵢ, xⱼ) for N=8 equally-spaced points on [0,1]. Color encodes kernel value: low high.

γ:medium
x1x2x3x4x5x6x7x8x1x2x3x4x5x6x7x81.001.001.001.001.001.001.001.00
RBF kernel with γ=2: K(x,x') = exp(-γ‖x-x'‖²). Large γ → narrow kernel → nearly diagonal Gram matrix (local similarity only).

A matrix acting on Rn\mathbb{R}^n can be diagonalized (if symmetric) or decomposed by singular values. Bounded linear operators generalize this to infinite-dimensional function spaces, where the question of when an operator is "diagonalizable" — and what its eigenvalues look like — becomes substantially more subtle. The spectrum of an operator replaces the eigenvalue list, but in infinite dimensions it can contain points that are neither eigenvalues nor their absence: a continuous spectrum where the operator is injective but not surjective, and a residual spectrum where the range is not even dense.

Bounded Linear Operators

A map T:XYT: X \to Y between normed spaces is linear if T(αx+βy)=αTx+βTyT(\alpha x + \beta y) = \alpha Tx + \beta Ty. It is bounded if

T=supx1TxY<\|T\| = \sup_{\|x\| \leq 1} \|Tx\|_Y < \infty

Boundedness is equivalent to continuity for linear maps. The space B(X,Y)B(X, Y) of bounded linear operators is itself a Banach space under this operator norm, and B(X,X)B(X, X) is a Banach algebra under composition.

Adjoint. For TB(H,K)T \in B(H, K) between Hilbert spaces, the adjoint T:KHT^*: K \to H is the unique operator satisfying Tx,yK=x,TyH\langle Tx, y \rangle_K = \langle x, T^*y \rangle_H for all xHx \in H, yKy \in K. Self-adjoint operators satisfy T=TT = T^* and are the infinite-dimensional analogue of symmetric matrices.

The Spectrum

For TB(X)T \in B(X) (a bounded operator on a Banach space) and λC\lambda \in \mathbb{C}, the operator TλIT - \lambda I may fail to be invertible for three distinct reasons:

  • Point spectrum σp(T)\sigma_p(T): TλIT - \lambda I is not injective — λ\lambda is an eigenvalue
  • Continuous spectrum σc(T)\sigma_c(T): TλIT - \lambda I is injective with dense range but not surjective (no inverse)
  • Residual spectrum σr(T)\sigma_r(T): TλIT - \lambda I is injective but range is not dense

The spectrum σ(T)=σpσcσr\sigma(T) = \sigma_p \cup \sigma_c \cup \sigma_r is always closed in C\mathbb{C} and non-empty. The resolvent set ρ(T)=Cσ(T)\rho(T) = \mathbb{C} \setminus \sigma(T) is where (TλI)1(T - \lambda I)^{-1} exists and is bounded.

For self-adjoint operators on Hilbert spaces, σ(T)R\sigma(T) \subset \mathbb{R} and σr(T)=\sigma_r(T) = \emptyset — the spectrum consists only of eigenvalues and continuous spectrum.

The three-part decomposition of the spectrum is forced by the definition of invertibility: an operator fails to be invertible if it is not injective (point spectrum), not surjective (residual or continuous spectrum depending on density of range). In finite dimensions, injective implies surjective for square maps, so this distinction collapses to eigenvalues vs. not. In infinite dimensions, injectivity and surjectivity decouple, creating the richer spectral structure — and the spectral type determines the dynamics of the associated differential operator.

Spectral Theorem for Compact Self-Adjoint Operators

An operator TB(H)T \in B(H) is compact if it maps bounded sets to precompact sets (sets whose closure is compact). Equivalently, if xnxx_n \rightharpoonup x weakly, then TxnTxTx_n \to Tx in norm. Compact operators can be approximated by finite-rank operators.

Spectral theorem for compact self-adjoint operators. Let T:HHT: H \to H be a compact self-adjoint operator on a Hilbert space. Then:

  1. Every nonzero λσ(T)\lambda \in \sigma(T) is an eigenvalue with finite-dimensional eigenspace
  2. Eigenvalues are real and form a sequence λ1,λ2,\lambda_1, \lambda_2, \ldots with λ1λ20|\lambda_1| \geq |\lambda_2| \geq \cdots \to 0
  3. The eigenvectors {ek}\{e_k\} corresponding to distinct eigenvalues are orthogonal and form a complete ONB for ran(T)\overline{\mathrm{ran}(T)}
  4. The spectral expansion holds: Tx=kλkx,ekekTx = \sum_k \lambda_k \langle x, e_k \rangle e_k

The only possible accumulation point of eigenvalues is 00. If HH is infinite-dimensional, 00 must lie in the spectrum (but may not be an eigenvalue).

Hilbert-Schmidt Operators

An operator T:L2(X)L2(X)T: L^2(X) \to L^2(X) defined by an integral kernel k:X×XRk: X \times X \to \mathbb{R},

Tf(x)=Xk(x,y)f(y)dyTf(x) = \int_X k(x, y)\,f(y)\,dy

is Hilbert-Schmidt if kL22= ⁣k(x,y)2dxdy<\|k\|_{L^2}^2 = \int\!\int |k(x,y)|^2\,dx\,dy < \infty. Every Hilbert-Schmidt operator is compact.

Singular Value Decomposition for Operators

For a compact operator T:HKT: H \to K, the operator TT:HHT^*T: H \to H is compact and self-adjoint with non-negative eigenvalues. The singular values are σk=λk(TT)\sigma_k = \sqrt{\lambda_k(T^*T)}, arranged in decreasing order σ1σ20\sigma_1 \geq \sigma_2 \geq \cdots \to 0. The operator SVD is

Tf=kσkf,vkukTf = \sum_k \sigma_k \langle f, v_k \rangle u_k

where {vk}\{v_k\} is an ONB for ran(T)\overline{\mathrm{ran}(T^*)} (right singular vectors) and {uk}\{u_k\} is an ONB for ran(T)\overline{\mathrm{ran}(T)} (left singular vectors). Truncating this expansion at rr terms gives the best rank-rr approximation to TT.

Fredholm alternative. For a compact operator TT and λ0\lambda \neq 0, either (TλI)(T - \lambda I) is invertible (unique solution for every right-hand side), or the equation (TλI)x=0(T - \lambda I)x = 0 has a nonzero solution (eigenvalue). There is no third option — this is the exact analogue of the Fredholm alternative for matrices.

Worked Example

The Integral Operator Is Hilbert-Schmidt and Compact

Let k:[0,1]2Rk: [0,1]^2 \to \mathbb{R} be square-integrable, and define Tf(x)=01k(x,y)f(y)dyTf(x) = \int_0^1 k(x,y)f(y)\,dy on L2([0,1])L^2([0,1]).

Boundedness: By Cauchy-Schwarz, Tf(x)201k(x,y)2dyf2|Tf(x)|^2 \leq \int_0^1 |k(x,y)|^2\,dy \cdot \|f\|^2. Integrating over xx:

Tf2kL22f2\|Tf\|^2 \leq \|k\|_{L^2}^2 \|f\|^2

so TkL2\|T\| \leq \|k\|_{L^2}. Thus TT is bounded and Hilbert-Schmidt.

Compactness: Write k(x,y)=j,lcjlej(x)el(y)k(x,y) = \sum_{j,l} c_{jl}\,e_j(x)e_l(y) in any ONB {ej}\{e_j\} of L2L^2. Truncating to NN terms gives a finite-rank operator TNT_N, and TTNkkNL20\|T - T_N\| \leq \|k - k_N\|_{L^2} \to 0 as NN \to \infty. Since finite-rank operators are compact and the compact operators form a closed set under norm limits, TT is compact.

Eigenvalues via Separation of Variables (Separable Kernel)

For a separable kernel k(x,y)=ϕ(x)ψ(y)k(x,y) = \phi(x)\psi(y), the operator Tf(x)=ϕ(x)01ψ(y)f(y)dy=ϕ(x)f,ψTf(x) = \phi(x)\int_0^1 \psi(y)f(y)\,dy = \phi(x)\langle f, \overline{\psi}\rangle is rank-1. Its only nonzero eigenvalue is λ=ϕ,ψ=01ϕ(y)ψ(y)dy\lambda = \langle \phi, \overline{\psi}\rangle = \int_0^1 \phi(y)\psi(y)\,dy with eigenvector ϕ/ϕ\phi/\|\phi\|. The finite-rank SVD truncation applied to a general kk extends this idea: the rr-term truncation is the sum of rr rank-1 operators, each from a pair of singular vectors.

Finite-Rank SVD Approximation

If TT is Hilbert-Schmidt with singular values σ1σ2\sigma_1 \geq \sigma_2 \geq \cdots, the best rank-rr approximation in operator norm is Tr=k=1rσk,vkukT_r = \sum_{k=1}^r \sigma_k \langle \cdot, v_k\rangle u_k, with approximation error TTr=σr+1\|T - T_r\| = \sigma_{r+1}. This is the operator analogue of matrix truncated SVD and justifies low-rank kernel approximations (Nyström method) in machine learning.

Connections

Where Your Intuition Breaks

The spectral theorem for compact self-adjoint operators looks like eigendecomposition, and in a sense it is — but the zero eigenvalue plays a special role that has no finite-dimensional analogue. In infinite dimensions, a compact operator TT must have 00 in its spectrum even if 00 is not an eigenvalue (Tx=0Tx = 0 has only the trivial solution). This is because a compact operator cannot be invertible on an infinite-dimensional space: if it were, the unit ball would be compact (image of a bounded set under a bounded invertible map), but Riesz's theorem says the unit ball in an infinite-dimensional normed space is never compact. The "invisibility" of 00 in the point spectrum does not mean TT is invertible — it means the failure of invertibility is more subtle than the existence of a kernel.

💡Intuition

The spectral theorem for compact self-adjoint operators is the infinite-dimensional generalization of eigendecomposition. Just as a symmetric matrix can be diagonalized in an ONB of eigenvectors, a compact self-adjoint operator on a Hilbert space expands in an ONB of eigenfunctions. The key difference: infinite-dimensional operators may have eigenvalue 00 with infinite-dimensional eigenspace, and the spectrum may have a non-trivial continuous part for non-compact operators. For compact operators, this complication disappears — the spectrum is purely discrete (plus possibly 00).

💡Intuition

The operator SVD is the conceptual foundation of principal component analysis, latent semantic analysis, and neural network weight matrix analysis. When a weight matrix WRm×nW \in \mathbb{R}^{m \times n} is viewed as a finite-rank operator from Rn\mathbb{R}^n to Rm\mathbb{R}^m, its singular values measure the "gain" in each principal direction. For kernel operators, the singular values σk\sigma_k decay at a rate determined by the smoothness of kk (smoother kernels have faster-decaying σk\sigma_k), which directly controls approximation quality via the Nyström method.

⚠️Warning

The spectral theorem fails for non-self-adjoint operators in ways that have no finite-dimensional analogue. A bounded operator on an infinite-dimensional Hilbert space may have an empty point spectrum (no eigenvalues at all), yet a non-trivial continuous spectrum. The left-shift operator on 2\ell^2, for example, has σp=\sigma_p = \emptyset but σ(T)\sigma(T) equals the closed unit disk. For some pathological operators (e.g., the Volterra operator Vf(x)=0xf(t)dtVf(x) = \int_0^x f(t)\,dt), every λ0\lambda \neq 0 lies in the resolvent and σ(V)={0}\sigma(V) = \{0\} — not an eigenvalue, just a spectral point. Non-self-adjoint spectral theory is significantly more complex and is an active research area.

Enjoying these notes?

Get new lessons delivered to your inbox. No spam.