One reason Hilbert spaces are so powerful is that they allow an infinite-dimensional version of diagonalization. In finite-dimensional linear algebra, a real symmetric matrix can be written in an orthonormal basis as a diagonal matrix with real entries. In a Hilbert space, the right substitute is a **compact self-adjoint operator**: a bounded linear map $T:H\to H$ that is self-adjoint and sends bounded sets to relatively compact sets. For such operators, the spectral theorem says that $H$ has an orthonormal basis of eigenvectors for $T$, and that $T$ acts like multiplication by real scalars along these directions.
This article builds the theorem from the ground up, highlights the points where compactness is used, and shows how the result powers integral equations, smoothing operators, and approximation methods.
Premium Audio PickWireless ANC Over-Ear HeadphonesBeats Studio Pro Premium Wireless Over-Ear Headphones
Beats Studio Pro Premium Wireless Over-Ear Headphones
A broad consumer-audio pick for music, travel, work, mobile-device, and entertainment pages where a premium wireless headphone recommendation fits naturally.
- Wireless over-ear design
- Active Noise Cancelling and Transparency mode
- USB-C lossless audio support
- Up to 40-hour battery life
- Apple and Android compatibility
Why it stands out
- Broad consumer appeal beyond gaming
- Easy fit for music, travel, and tech pages
- Strong feature hook with ANC and USB-C audio
Things to know
- Premium-price category
- Sound preferences are personal
Compactness and self-adjointness: what they mean operationally
Let $H$ be a (real or complex) Hilbert space.
- $T$ is bounded if $\|Tx\|\le C\|x\|$ for all $x$.
- $T$ is self-adjoint if $\langle Tx,y\rangle=\langle x,Ty\rangle$ for all $x,y$.
- $T$ is compact if it maps the unit ball \to a set with compact closure.
Compactness can feel abstract, but it has a very concrete consequence: if $(x_n)$ is bounded, then $(Tx_n)$ has a convergent subsequence. This “subsequence gain” is what replaces finite dimensionality in the proof.
Self-adjointness enforces real geometry:
- $\langle Tx,x\rangle$ is real for all $x$.
- Eigenvalues of a self-adjoint operator are real.
- Eigenvectors corresponding to distinct eigenvalues are orthogonal.
The spectral theorem is the statement that compactness plus self-adjointness is strong enough to recover a full diagonalization.
The statement of the theorem
Spectral Theorem (compact self-adjoint case).
Let $T:H\to H$ be compact and self-adjoint.
- Every nonzero $\lambda\in\mathbb{R}$ in the spectrum of $T$ is an eigenvalue.
- The set of nonzero eigenvalues is at most countable, has no accumulation point except possibly $0$, and each eigenspace is finite-dimensional.
- There exists an orthonormal basis $\{e_k\}$ for $\overline{\operatorname{Ran}(T)}$ consisting of eigenvectors of $T$, with corresponding eigenvalues $\lambda_k\in\mathbb{R}$ (possibly repeating according to multiplicity), such that
where the series converges in norm.
- On the orthogonal complement $\ker(T)$, the operator acts as zero.
A useful way to phrase it is: $T$ is unitarily equivalent \to a diagonal operator with diagonal entries $(\lambda_k)$ tending \to $0$.
The geometric core: maximizing a quadratic form
The proof has a clean geometric idea: find directions in which the quadratic form $\langle Tx,x\rangle$ is as large as possible. For self-adjoint $T$, this form is real and behaves like a “generalized energy.”
Define
If $m=0$ and also $\inf_{\|x\|=1}\langle Tx,x\rangle=0$, then $T=0$. Otherwise, either $m>0$ or the infimum is negative, and in either case there is a nonzero eigenvalue. The hard part is to prove that the supremum is achieved by some unit vector. This is where compactness enters.
Achieving the supremum: compactness replaces finite dimension
Take a maximizing sequence $(x_n)$ with $\|x_n\|=1$ and $\langle Tx_n,x_n\rangle\to m$. The unit sphere is not compact in infinite dimensions, so we cannot just extract a convergent subsequence of $(x_n)$. But compactness of $T$ gives that $(Tx_n)$ has a convergent subsequence. Pass \to a subsequence (still called $x_n$) so that $Tx_n\to y$ in norm.
Now comes a key estimate: show $x_n$ is Cauchy. Consider
This does not immediately compare \to $\|x_n-x_k\|^2$ because $T$ is not assumed invertible. The more effective move uses the polarization of the quadratic form and the fact that the sequence nearly maximizes it.
A standard argument (based on the identity
$\langle Tx,x\rangle + \langle Ty,y\rangle – 2\langle Tx,y\rangle = \langle T(x-y),x-y\rangle$
for self-adjoint $T$) shows that if $\langle Tx_n,x_n\rangle\to m$ and $\langle Tx_k,x_k\rangle\to m$, then $\langle T(x_n-x_k),x_n-x_k\rangle\to 0$. Combined with compactness and careful norm estimates, this forces $(x_n)$ \to have a convergent subsequence in $H$. Let $x$ be its limit, with $\|x\|=1$.
By continuity, $\langle Tx,x\rangle=m$: the supremum is attained.
From maximizer to eigenvector
Once a maximizer $x$ exists, show it satisfies $Tx=mx$. Consider any $h$ orthogonal \to $x$. For small real $t$, the unit-normalized vector
still has norm $1$, so $\langle Tu(t),u(t)\rangle\le m$. Differentiate the function $f(t)=\langle Tu(t),u(t)\rangle$ at $t=0$. The derivative must be $0$ because $t=0$ is a maximum. The computation yields
That means $Tx$ is orthogonal to every vector orthogonal \to $x$, so $Tx$ lies in $\operatorname{span}\{x\}$. Thus $Tx=\lambda x$. Taking inner product with $x$ gives $\lambda=\langle Tx,x\rangle=m$. So $x$ is an eigenvector with eigenvalue $m$.
If instead the most negative value is attained, the same argument gives a negative eigenvalue.
At this stage, we have produced at least one nonzero eigenvalue and an eigenvector.
Building the full eigenbasis: orthogonal splitting and iteration
Let $e_1$ be a unit eigenvector with eigenvalue $\lambda_1\ne 0$. Consider the orthogonal complement
Because $T$ is self-adjoint, $H_1$ is invariant under $T$: if $h\perp e_1$, then
so $Th\perp e_1$. The restriction $T|_{H_1}$ is still compact and self-adjoint. Apply the same maximization argument \to $T|_{H_1}$ \to obtain another eigenvector $e_2\in H_1$ with eigenvalue $\lambda_2$, orthogonal \to $e_1$.
Repeat. This produces an orthonormal family $\{e_k\}$ of eigenvectors with eigenvalues $\lambda_k$ ordered by decreasing absolute value:
Two essential facts now enter:
- Each eigenspace for $\lambda\ne 0$ is finite-dimensional. If it were infinite-dimensional, it would contain an infinite orthonormal set $(u_n)$, but then $(Tu_n)=(\lambda u_n)$ would have no convergent subsequence, contradicting compactness.
- Nonzero eigenvalues cannot accumulate away from $0$. If $\lambda_n\to \lambda\ne 0$ with distinct eigenvalues, then the corresponding unit eigenvectors are orthogonal, and $\|Tu_n-Tu_m\|=\|\lambda_n u_n-\lambda_m u_m\|$ stays bounded away from $0$ for large $n,m$, again contradicting compactness.
Thus $\lambda_k\to 0$.
Finally, show that the closed span of the eigenvectors equals $\overline{\operatorname{Ran}(T)}$. A clean way is to consider the orthogonal complement $K$ of the span of all eigenvectors and show $T$ must vanish on $K$. If $T$ did not vanish on $K$, the restriction $T|_K$ would again have a nonzero eigenvalue, producing an eigenvector in $K$ and contradicting the definition of $K$. So $T(K)=\{0\}$, hence $K\subset \ker(T)$. This yields the decomposition
The series representation
follows by applying $T$ \to the expansion of $x$ in the orthonormal basis and using $\lambda_k\to 0$ \to ensure convergence.
Practical consequences you can use immediately
The theorem is not only an abstract diagonalization. It gives quantitative structure.
Operator norm and extremal eigenvalues
For compact self-adjoint $T$,
So the operator norm is attained by an eigenvector. This can fail for general bounded operators; compactness and self-adjointness force attainment.
Finite-rank approximations
Define $T_N$ by truncating the spectral sum:
Then $T_N$ is finite-rank, and $\|T-T_N\|\to 0$ as $N\to\infty$. This is a strong approximation property: compact self-adjoint operators are limits of finite-dimensional diagonal operators.
Fredholm alternative in this setting
For $\lambda\ne 0$, the equation
has a solution if and only if $y$ is orthogonal to the kernel of $T-\lambda I$. Since kernels are spanned by eigenvectors, this becomes an explicit orthogonality condition.
Examples that anchor the abstract theorem
Integral operators with symmetric kernels
Let $H=L^2[a,b]$. Suppose $K\in L^2([a,b]^2)$ is symmetric: $K(s,t)=\overline{K(t,s)}$. Define
Under mild regularity assumptions on $K$ (for instance, square-integrability is enough to make $T$ Hilbert–Schmidt), $T$ is compact. Symmetry of the kernel gives self-adjointness. The spectral theorem then produces an orthonormal basis of eigenfunctions $\{e_k\}$ and real eigenvalues $\lambda_k\to 0$ such that
This is the analytic backbone of many classical expansions in integral equations and kernel methods.
Compactness from smoothing
Many operators that “smooth” functions are compact. For instance, the inclusion map from a Sobolev space $H^1$ into $L^2$ on a bounded domain is compact under standard boundary regularity assumptions. When a PDE solution operator factors through such a compact inclusion, the resulting operator in $L^2$ becomes compact, and self-adjointness often comes from symmetry of the underlying bilinear form.
The spectral theorem then yields eigenfunction expansions that explain why higher modes carry less energy (because $\lambda_k\to 0$), which is a precise version of “smoothing kills high-frequency components.”
Principal components in infinite dimensions
In a real Hilbert space, consider a covariance operator $C$ defined by
for a mean-zero random element $X$ with finite second moment. Under standard assumptions, $C$ is self-adjoint, positive, and compact. The eigenvectors of $C$ provide principal directions, and eigenvalues measure variance along those directions. The spectral theorem supplies the infinite-dimensional analogue of PCA with the same geometric meaning: orthogonal decomposition into uncorrelated modes.
What compact self-adjoint diagonalization teaches you
The compact self-adjoint spectral theorem is the cleanest infinite-dimensional spectral result because compactness restores the “finite-dimensional” behavior of eigenvalues and eigenvectors.
- Compactness forces the spectrum away from $0$ \to be discrete and composed of eigenvalues.
- Self-adjointness forces real eigenvalues and orthogonality of eigenvectors.
- Together they yield an orthonormal eigenbasis for the range and a convergent diagonal expansion.
In practice, when you recognize an operator as compact and self-adjoint, you gain a full coordinate system adapted to the operator. You can approximate it by finite-rank truncations, solve equations mode-by-mode, and convert analytic questions into weighted sums of Fourier-like coefficients.
Leave a Reply