Spectral Theory of Dynamical Systems and the Halmos-von Neumann Theorem

Intuition Beginner

A dynamical rule is a nonlinear, possibly tangled thing: it bends and folds the space in ways that are hard to compare across two different systems. The spectral viewpoint replaces the rule with a tame linear stand-in. Instead of tracking where points go, track what happens to measurements — functions that read off a number at each point. The rule "compose with the map" sends one measurement to another, and this rewriting of measurements is a clean linear operation. Studying that operation, rather than the map itself, is the spectral idea: listen to the system instead of watching it.

Here is the everyday picture. Strike a bell and you hear not one tone but a chord — a bundle of pure frequencies, each at its own pitch. Two bells can be compared by comparing their chords. A measure-preserving rule has its own chord: special measurements that, under the rule, simply get multiplied by a fixed phase, like a pure tone that keeps its pitch and only shifts. These special measurements are its eigenfunctions, and their phases are its eigenvalues — the pitches in its chord. A rotation of a circle has a full ladder of these pitches. A thoroughly scrambling rule, by contrast, has no pure tones above the constant background — all hiss, no pitch.

The headline result is a perfect score for the pure-tone systems. If a rule's chord is all pitches and no hiss — what we call discrete spectrum — then the chord alone determines the rule completely: two such systems with the same set of pitches are the same system, merely relabelled. And every such system turns out to be a rotation on a circle-like space. So at the orderly end of dynamics, the chord is a perfect fingerprint.

The takeaway: the spectral view trades the nonlinear map for a linear operator on measurements, reads off its chord of pure tones, and discovers that for the most orderly systems — those built entirely from pure tones — the chord pins the system down exactly, identifying it as a rotation. The fingerprint is perfect there, and only there.

Visual Beginner

Picture each measure-preserving rule as a sound and draw its spectrum — a bar chart of which pure pitches it contains, plus a smear of pitchless hiss.

The top panel is a rotation: a clean ladder of pure tones, no hiss — discrete spectrum. The bottom is a Bernoulli shift: only the constant tone, the rest pure hiss — continuous spectrum. The middle mixes both. The side diagram is the punchline: for discrete-spectrum systems the chord is a perfect fingerprint, but matching hiss does not match the systems.

Worked example Beginner

We read the chord of a small rotation and check that its pure tones form a tidy pattern — they multiply together like clock arithmetic.

Step 1. The system. Take the circle as the numbers from $0$ up to $1$, wrapping, and the rule that adds a fixed step $a$ and wraps. A measurement is a function on the circle. The pure-tone measurements are the waves $w_k$ that wind around the circle $k$ times, for each whole number $k$.

Step 2. Apply the rule to a pure tone. Shifting the input by $a$ shifts a wave that winds $k$ times by a fixed phase: the wave $w_k$ comes back as itself multiplied by the number $c_k$, where $c_k$ is the point on the unit circle at angle $k$ times $a$ (in turns). So each $w_k$ is an eigenfunction, and its eigenvalue — its pitch — is $c_k$.

Step 3. The pitches multiply. Take the step $a=0.25$ (a quarter turn). Then the wave $w_1$ has pitch $c_1$ at a quarter turn, $w_2$ has pitch $c_2$ at a half turn, $w_3$ at three-quarters, and $w_4$ at a full turn, which is back to the start, pitch $1$. So the distinct pitches are at $0$, a quarter, a half, and three-quarters of a turn: exactly four of them, and the next wave repeats the cycle.

Step 4. The group pattern. List those four pitches as quarter-turn marks $0,1,2,3$ (in quarters). Combine the pitch of $w_2$ with the pitch of $w_3$ by multiplying the waves: $w_2$ times $w_3$ is $w_5$, whose pitch is $5$ quarter-turns $=1$ quarter-turn after a full loop, i.e. mark $1$. And $2+3=5=1$ in quarter-turn clock arithmetic. The pitches combine by adding their marks on a $4$-hour clock.

What this tells us: the pure tones of a rotation are not a random scatter — they close up into a tidy system where combining two tones (by multiplying the waves) adds their positions on a clock. That clock-arithmetic structure is what "the eigenvalues form a group" means, and it is the seed of the big theorem: knowing this group of pitches is enough to reconstruct the whole rotation.

Check your understanding Beginner

Formal definition Intermediate+

Throughout, $(X,\mathcal{B},\mu ,T)$ is an invertible measure-preserving system on a Lebesgue probability space in the sense of 38.04.02, with Koopman operator $U_T:L^2(\mu )\to L^2(\mu )$, $U_{T}f=f\circ T$. Since $T$ preserves $\mu$ and is invertible, $U_T$ is a unitary operator 02.11.08: $⟨U_{T}f,U_{T}g⟩=\int (f\circ T)\overline{(g\circ T)}d\mu =\int f\bar{g}d\mu =⟨f,g⟩$, and $U_T^{-1}=U_{T^{-1}}$. Write $L_0^2(\mu )=\{f:\int fd\mu =0\}$, the orthocomplement of the constants, which $U_T$ preserves.

Definition (spectral isomorphism). Two systems $(X,\mu ,T)$ and $(Y,\nu ,S)$ are spectrally isomorphic if their Koopman operators are unitarily equivalent: there is a unitary $W:L^2(\mu )\to L^2(\nu )$ with $W{U}_T=U_{S}W$. A measurable isomorphism is a measure-space isomorphism $\varphi :X\to Y$ (a.e.-defined bijection, both directions measurable, ${\varphi }_{\ast }\mu =\nu$) with $\varphi \circ T=S\circ \varphi$ a.e.; it induces the unitary $Wf=f\circ {\varphi }^{-1}$, which intertwines the Koopman operators and carries constants to constants and $L^{\infty }$ to $L^{\infty }$. Hence measurable isomorphism implies spectral isomorphism. The converse is the central question of this unit.

Definition (maximal spectral type; discrete and continuous spectrum). By the spectral theorem for the unitary $U_T$ 02.11.08 there is a projection-valued measure $E$ on the circle $\mathbb{T}$ with $U_T={\int }_{\mathbb{T}}zdE(z)$. For $f\in L^2$ the spectral measure ${\sigma }_f$ is the finite positive measure with $⟨U_T^{n}f,f⟩={\int }_{\mathbb{T}}{z}^{n}d{\sigma }_f(z)$ (Herglotz-Bochner). The maximal spectral type $\sigma$ is the equivalence class of measures ${\sigma }_f$ under mutual absolute continuity that dominates every ${\sigma }_g$; it decomposes uniquely as $\sigma ={\sigma }_{\mathrm{pp}}+{\sigma }_{\mathrm{c}}$ into its atomic (pure point) and continuous parts, inducing an orthogonal Koopman-von Neumann decomposition $$L^2(\mu )=H_{\mathrm{pp}}\oplus H_{\mathrm{c}},$$ where $H_{\mathrm{pp}}$ is the closed span of eigenfunctions of $U_T$ and $H_{\mathrm{c}}$ is the subspace on which $U_T$ has continuous spectral type. The system has discrete (or pure point) spectrum if $H_{\mathrm{pp}}=L^2(\mu )$, i.e. the eigenfunctions span $L^2$.

Definition (eigenvalue group). An eigenvalue of $U_T$ is $\lambda \in \mathbb{T}$ admitting $f\ne 0$ with $U_{T}f=\lambda f$. Write $e(T)\subseteq \mathbb{T}$ for the set of eigenvalues. If $T$ is ergodic, every eigenvalue is simple (its eigenspace is one-dimensional) and every eigenfunction has constant modulus a.e.: $U_{T}f=\lambda f$ gives $|f|\circ T=|f|$, so $|f|$ is invariant, hence a.e. constant by 38.04.02. The product and conjugate of eigenfunctions are eigenfunctions, so $e(T)$ is a subgroup of $\mathbb{T}$; under ergodicity it is moreover countable.

Canonical examples. (i) Rotation on a compact abelian group $G$: $Tx=x+a$ on $(G,\text{Haar})$ with $\overline{\{na\}}=G$ (so $T$ is ergodic). Every character $\gamma \in \hat{G}$ is an eigenfunction ($\gamma (x+a)=\gamma (a)\gamma (x)$, eigenvalue $\gamma (a)$); the characters form an orthonormal basis of $L^2$, so $T$ has discrete spectrum and $e(T)=\{\gamma (a):\gamma \in \hat{G}\}\cong \hat{G}$. (ii) Irrational rotation $x\mapsto x+\alpha$ on $\mathbb{T}$: discrete spectrum, $e(T)=\{e^{2\pi ik\alpha }:k\in \mathbb{Z}\}$, a dense cyclic subgroup. (iii) Bernoulli shift / hyperbolic toral automorphism: continuous spectrum on $L_0^2$ ($e(T)=\{1\}$), the opposite extreme — weak mixing of 38.05.01. (iv) Anzai skew product $T(x,y)=(x+\alpha ,y+x)$ on ${\mathbb{T}}^2$: mixed spectrum — discrete part from the base rotation, continuous part in the fibres.

Counterexamples to common slips Intermediate+

• Spectral isomorphism is weaker than measurable isomorphism in general. The Koopman operator forgets the multiplicative ($L^{\infty }$) structure and the point realisation. Two systems can have unitarily equivalent Koopman operators yet not be measurably isomorphic; entropy is the standard separator (Theorem in Advanced results).

• Discrete spectrum is the exception where the two notions coincide. Halmos-von Neumann is precisely the statement that within the discrete-spectrum class, spectral isomorphism upgrades to measurable isomorphism. Do not extrapolate this coincidence beyond pure point spectrum.

• Eigenvalues lie on the unit circle and form a group, not a vector space. Since $U_T$ is unitary, $|\lambda |=1$; the group law on $e(T)$ is multiplication in $\mathbb{T}$ (equivalently addition of angles), inherited from multiplying eigenfunctions. An "eigenvalue" off the unit circle signals an error.

• Ergodicity is needed for simplicity and the modulus-constant property. Without ergodicity an eigenvalue can have a multi-dimensional eigenspace and eigenfunctions of non-constant modulus; the Halmos-von Neumann machine runs on ergodic systems.

• The eigenvalue group must be a subgroup of $\mathbb{T}$ to be realisable. The Halmos-von Neumann representation theorem requires the prescribed countable subgroup $\Lambda \le \mathbb{T}$; the model group is $G=\hat{\Lambda }$ (with $\Lambda$ given the discrete topology), and the rotation element is dual to the inclusion $\Lambda \hookrightarrow \mathbb{T}$.

Key theorem with proof Intermediate+

Theorem (Halmos-von Neumann). Let $(X,\mathcal{B},\mu ,T)$ be an ergodic measure-preserving system on a Lebesgue probability space with discrete spectrum. Then:

1. (Group structure) The eigenvalue set $e(T)$ is a countable subgroup of $\mathbb{T}$, every eigenvalue is simple, and the eigenfunctions of unit modulus form an orthonormal basis of $L^2(\mu )$ that is a group under pointwise multiplication.
2. (Representation) $T$ is measurably isomorphic to the rotation $Rx=x+a$ on the compact abelian group $G=\widehat{e(T)}$ (the Pontryagin dual of $e(T)$ with the discrete topology) equipped with Haar measure, where $a\in G$ is the element $a(\lambda )=\lambda$ for $\lambda \in e(T)$.
3. (Classification) Two ergodic discrete-spectrum systems are measurably isomorphic iff they have the same eigenvalue group. Consequently, on the discrete-spectrum class, spectral isomorphism = measurable isomorphism.

Proof. (1) By ergodicity every eigenvalue is simple and every eigenfunction has constant modulus (the formal-definition argument). Normalise each eigenfunction to unit modulus; this fixes it up to a unit scalar. If $U_{T}f=\lambda f$ and $U_{T}g=\mu g$ with $|f|=|g|=1$, then $U_T(fg)=\lambda \mu fg$ and $U_T(\bar{f})=\bar{\lambda }\bar{f}$, so the unit-modulus eigenfunctions are closed under multiplication and conjugation; hence $e(T)$ is a subgroup of $\mathbb{T}$. Discrete spectrum says the eigenfunctions span $L^2$; choosing one unit-modulus eigenfunction ${\chi }_{\lambda }$ per $\lambda \in e(T)$ (with ${\chi }_1\equiv 1$, ${\chi }_{\lambda }{\chi }_{{\lambda }^{\prime }}={\chi }_{\lambda {\lambda }^{\prime }}$ after fixing the scalars consistently using a section of the group) gives an orthonormal basis indexed by the group $e(T)$, which is therefore countable since $L^2$ is separable.

(2) Let $\Lambda =e(T)$ with the discrete topology and $G=\hat{\Lambda }$ its Pontryagin dual, a compact abelian group with Haar measure $m$. By Pontryagin duality $\hat{G}=\Lambda$, so the characters of $G$ are exactly the elements $\lambda \in \Lambda$, viewed as functions $g\mapsto g(\lambda )$ on $G$; they form an orthonormal basis of $L^2(G,m)$. Define $W:L^2(G,m)\to L^2(X,\mu )$ on this basis by $W\lambda ={\chi }_{\lambda }$ and extend linearly; $W$ is unitary, carries the character $\lambda$ to the eigenfunction ${\chi }_{\lambda }$, and intertwines the rotation with $T$: the Koopman operator of $Rg=g+a$ acts on the character $\lambda$ by $\lambda (g+a)=\lambda (a)\lambda (g)=a(\lambda )\lambda (g)=\lambda \cdot \lambda (g)$ (since $a(\lambda )=\lambda$), i.e. $U_R\lambda =\lambda \cdot \lambda$, exactly as $U_T{\chi }_{\lambda }=\lambda {\chi }_{\lambda }$. So $W{U}_R=U_{T}W$.

It remains to upgrade this $L^2$ unitary to a genuine measurable isomorphism. The map $W$ sends the multiplicative group of characters $\{\lambda \}$ to the multiplicative group of unit-modulus eigenfunctions $\{{\chi }_{\lambda }\}$, preserving products: $W(\lambda {\lambda }^{\prime })={\chi }_{\lambda {\lambda }^{\prime }}={\chi }_{\lambda }{\chi }_{{\lambda }^{\prime }}=(W\lambda )(W{\lambda }^{\prime })$. A unitary preserving the multiplication of a generating algebra of bounded functions and sending $1$ to $1$ is an algebra isomorphism $L^{\infty }(G)\to L^{\infty }(X)$ between the function algebras (the characters generate $L^{\infty }(G)$ as a von Neumann algebra and their images generate $L^{\infty }(X)$, by discreteness of spectrum). By the von Neumann / Mackey point-realisation theorem for Lebesgue spaces, an isomorphism of the measure algebras intertwining the actions is induced by an a.e.-defined measure isomorphism $\varphi :X\to G$ with $\varphi \circ T=R\circ \varphi$. Thus $T\cong R$ measurably.

(3) If two ergodic discrete-spectrum systems are measurably isomorphic they are spectrally isomorphic, and unitary equivalence of $U_T,U_S$ carries eigenfunctions to eigenfunctions preserving eigenvalues, so $e(T)=e(S)$. Conversely, if $e(T)=e(S)=\Lambda$, part (2) gives $T\cong R\cong S$ through the same model rotation on $\hat{\Lambda }$, so $T\cong S$ measurably. Finally, spectral isomorphism of two discrete-spectrum systems forces $e(T)=e(S)$ (a unitary intertwiner matches point spectra with multiplicities, and here all multiplicities are one), whence measurable isomorphism by what was just shown; so the two notions coincide on this class. $\square$

Bridge. This theorem builds toward the global organisation of measure-preserving systems by their spectral type, and it appears again in the relative theory as the construction of the Kronecker factor — the maximal discrete-spectrum factor — that sits beneath any ergodic system. The foundational reason the classification is clean is that for discrete spectrum the eigenfunctions are simultaneously an orthonormal basis and a group, so the Hilbert-space datum and the algebraic datum coincide; this is exactly the Pontryagin duality $G=\hat{\Lambda }$ turning a subgroup of the circle into a compact group whose characters are the eigenfunctions. The central insight is that spectral isomorphism, which in general remembers only the unitary $U_T$ and forgets the pointwise multiplication, recovers that multiplication for free here because eigenfunctions multiply — putting these together, the lost $L^{\infty }$ structure is reconstructible from the eigenvalue group, which is why discrete spectrum is exactly the regime where spectral and measurable isomorphism merge. This is dual to the weak-mixing extreme of 38.05.01, where $H_{\mathrm{pp}}$ collapses to the constants and the spectral invariant retains almost nothing.

Exercises Intermediate+

Advanced results Master

Theorem 1 (Halmos-von Neumann classification). An ergodic measure-preserving system on a Lebesgue probability space has discrete spectrum iff it is measurably isomorphic to an ergodic rotation on a compact abelian group with Haar measure. Two such systems are measurably isomorphic iff their eigenvalue groups coincide, and a countable subgroup $\Lambda \le \mathbb{T}$ arises as $e(T)$ for some ergodic discrete-spectrum $T$ iff it is realised by the rotation by $a$ on $\hat{\Lambda }$ with $a(\lambda )=\lambda$. On the discrete-spectrum class the eigenvalue group is a complete invariant, and spectral isomorphism coincides with measurable isomorphism ^{[Halmos-von Neumann 1942]}.

Theorem 2 (the Kronecker factor). Every ergodic system $(X,\mu ,T)$ has a maximal factor with discrete spectrum, the Kronecker factor $(Z,m,R)$, with $Z$ a compact abelian group and $R$ a rotation; $L^2(Z)\subseteq L^2(X)$ is precisely $H_{\mathrm{pp}}$, the closed span of eigenfunctions, and $H_{\mathrm{c}}=L^2(Z{)}^{\perp }$ carries the weakly mixing part of the dynamics relative to $Z$. The system is weakly mixing iff its Kronecker factor is a single point, and has discrete spectrum iff $Z=X$. The Kronecker factor is the spectral-theoretic incarnation of the dichotomy of 38.05.01 and the base of the Furstenberg structure tower ^{[Cornfeld-Fomin-Sinai 1982]}.

Theorem 3 (spectral isomorphism is strictly coarser; the entropy obstruction). Spectral isomorphism does not imply measurable isomorphism. Two Bernoulli shifts of different entropy have countable Lebesgue spectrum on $L_0^2$ and are therefore spectrally isomorphic, yet Kolmogorov-Sinai entropy ($\log 2$ versus $\log 3$ for $B(\frac{1}{2},\frac{1}{2})$ and the uniform $3$-shift) is a measurable-isomorphism invariant that distinguishes them; by Ornstein's theorem entropy is in fact a complete invariant for Bernoulli shifts, none of which is detected by the unitary $U_T$. Hence spectral invariants determine neither entropy nor measurable-isomorphism type off the discrete-spectrum class ^{[Anzai 1951]}.

Theorem 4 (maximal spectral type and multiplicity; the unitary invariants). The complete unitary-equivalence invariant of $U_T$ is the pair (maximal spectral type $\sigma$, multiplicity function $M:\mathbb{T}\to \{1,2,\dots ,\infty \}$), by the Hahn-Hellinger multiplicity theory for normal operators. Discrete spectrum is the case $\sigma$ purely atomic with $M\equiv 1$ on the atoms (under ergodicity); weak mixing is $\sigma {|}_{L_0^2}$ continuous; Bernoulli/$K$-systems realise $\sigma {|}_{L_0^2}$ Lebesgue with $M\equiv \infty$ (countable Lebesgue spectrum). The fine question of which spectral types and multiplicities occur for measure-preserving transformations — e.g. Banach's problem of simple Lebesgue spectrum, and the realisation of finite multiplicities — remains a central open program of spectral theory ^{[Nadkarni 1998]}.

Theorem 5 (rigidity of discrete spectrum under factors and products). The discrete-spectrum class is closed under factors, inverse limits, and (ergodic) direct products: if $T,S$ are ergodic discrete-spectrum with eigenvalue groups ${\Lambda }_T,{\Lambda }_S$, then $T\times S$ (if ergodic, i.e. ${\Lambda }_T\cap {\Lambda }_S$ generates compatibly) has discrete spectrum with eigenvalue group ${\Lambda }_T{\Lambda }_S\le \mathbb{T}$, and any factor of a discrete-spectrum system has eigenvalue group a subgroup of the original. The eigenvalue group is thus a functor from the category of ergodic discrete-spectrum systems with factor maps to countable subgroups of $\mathbb{T}$ with inclusions, and Halmos-von Neumann is the statement that this functor is an equivalence onto its image ^{[Walters 1982]}.

Synthesis. The five results are one statement read at successive depths, and the foundational reason they cohere is that the Koopman operator splits $L^2$ into a pure-point part where dynamics is a group rotation and a continuous part where the unitary forgets everything finer than spectral type and multiplicity. This is the central insight that the eigenvalue group is simultaneously a Hilbert-space invariant and an algebra of bounded functions, so on $H_{\mathrm{pp}}$ the spectral datum is the measurable datum — putting these together, Halmos-von Neumann, the Kronecker factor, and the functoriality of Theorem 5 are three faces of the single fact that discrete spectrum is the regime of complete spectral classification. The entropy obstruction is exactly the complementary fact: on $H_{\mathrm{c}}$ the maximal spectral type and multiplicity are a strictly coarser invariant than measurable isomorphism, and entropy is the first measurable invariant the unitary cannot see, so the spectral and measurable theories generalise the mixing hierarchy of 38.05.01 into a precise statement of what unitary equivalence remembers. The bridge is that the same Herglotz-Bochner spectral measure that diagnosed weak mixing as continuous type now, at the opposite extreme of atomic type with simple multiplicity, reconstructs the entire system as a rotation — and this is dual to the Bernoulli world, where Lebesgue spectrum of infinite multiplicity is spectrally featureless yet measurably stratified by entropy, the foundational reason spectral theory and entropy theory are two complementary classification programs rather than one.

Full proof set Master

Proposition 1 (eigenfunctions of an ergodic system form a group; eigenvalues are simple). For ergodic $T$, the unit-modulus eigenfunctions of $U_T$ are closed under multiplication and conjugation, each eigenvalue is simple, and $e(T)\le \mathbb{T}$.

Proof. If $U_{T}f=\lambda f$ with $f\ne 0$ then $|f|\circ T=|f|$, so $|f|$ is invariant and a.e. equal to a constant $c>0$ by ergodicity 38.04.02; normalise $c=1$. For unit-modulus eigenfunctions $f,g$ with eigenvalues $\lambda ,\beta$, $U_T(fg)=\lambda \beta fg$ and $U_T\bar{f}=\bar{\lambda }\bar{f}$, both unit-modulus and nonzero, so $\lambda \beta ,\bar{\lambda }\in e(T)$ and $1\in e(T)$ via constants; hence $e(T)\le \mathbb{T}$. If $U_{T}f=U_{T}g=\lambda \cdot$ with $f,g$ unit-modulus eigenfunctions, then $g/f$ has $U_T(g/f)=g/f$, so $g/f$ is invariant, a.e. constant, and $g$ is a scalar multiple of $f$: the eigenspace is one-dimensional. $\square$

Proposition 2 (group rotations have discrete spectrum and prescribed eigenvalue group). For a compact abelian $G$ with Haar measure and an ergodic rotation $Rx=x+a$, $R$ has discrete spectrum with $e(R)=\{\gamma (a):\gamma \in \hat{G}\}\cong \hat{G}$.

Proof. Characters $\gamma \in \hat{G}$ satisfy $\gamma \circ R=\gamma (a)\gamma$ and form an orthonormal basis of $L^2(G)$ by Pontryagin/Peter-Weyl, so the eigenfunctions span $L^2$ and $R$ has discrete spectrum. Ergodicity (from $\overline{\{na\}}=G$, equivalently $\gamma (a)\ne 1$ for $\gamma \ne 1$) makes eigenvalues simple, so every eigenfunction is a scalar multiple of a character and $e(R)=\{\gamma (a)\}$. The map $\gamma \mapsto \gamma (a)$ is an injective homomorphism $\hat{G}\to \mathbb{T}$ (injective because $\gamma (a)=1$ with ergodicity forces $\gamma =1$), so $e(R)\cong \hat{G}$. $\square$

Proposition 3 (Halmos-von Neumann representation). An ergodic discrete-spectrum $T$ with $\Lambda =e(T)$ is measurably isomorphic to the rotation by $a$ ($a(\lambda )=\lambda$) on $G=\hat{\Lambda }$ with Haar measure.

Proof. Give $\Lambda$ the discrete topology; $G=\hat{\Lambda }$ is compact abelian and $\hat{G}=\Lambda$. Choose unit-modulus eigenfunctions ${\chi }_{\lambda }$ ($U_T{\chi }_{\lambda }=\lambda {\chi }_{\lambda }$) multiplicatively (${\chi }_1=1$, ${\chi }_{\lambda }{\chi }_{{\lambda }^{\prime }}={\chi }_{\lambda {\lambda }^{\prime }}$): the obstruction is a $\mathbb{T}$-valued $2$-cocycle on $\Lambda$ which is a coboundary since the eigenspaces are one-dimensional and we may rescale along a transversal, fixing the multipliers. By discrete spectrum $\{{\chi }_{\lambda }{\}}_{\lambda \in \Lambda }$ is an orthonormal basis of $L^2(\mu )$. Define $W:L^2(G)\to L^2(\mu )$ by $W\lambda ={\chi }_{\lambda }$ on characters; $W$ is unitary and $W{U}_R=U_{T}W$ because $U_R\lambda =a(\lambda )\lambda =\lambda \cdot \lambda$ matches $U_T{\chi }_{\lambda }=\lambda {\chi }_{\lambda }$. Multiplicativity $W(\lambda {\lambda }^{\prime })={\chi }_{\lambda {\lambda }^{\prime }}={\chi }_{\lambda }{\chi }_{{\lambda }^{\prime }}=(W\lambda )(W{\lambda }^{\prime })$ makes $W$ a $\ast$-isomorphism of the measure algebras $L^{\infty }(G)\to L^{\infty }(\mu )$ intertwining the actions; the von Neumann point-realisation theorem on Lebesgue spaces yields an a.e. measure isomorphism $\varphi$ with $\varphi \circ T=R\circ \varphi$. $\square$

Proposition 4 (eigenvalue group is a complete invariant on the discrete-spectrum class). Two ergodic discrete-spectrum systems are measurably isomorphic iff they have equal eigenvalue groups; equivalently, on this class spectral isomorphism = measurable isomorphism.

Proof. Measurable isomorphism implies spectral isomorphism, which preserves the point spectrum (with multiplicities), so $e(T)=e(S)$. Conversely $e(T)=e(S)=\Lambda$ gives, by Proposition 3, $T\cong R\cong S$ through the same model rotation on $\hat{\Lambda }$, hence $T\cong S$ measurably. Finally a spectral isomorphism of two discrete-spectrum systems matches their (simple) point spectra, forcing $e(T)=e(S)$, and the previous sentence then yields a measurable isomorphism; so the two notions coincide. $\square$

Proposition 5 (spectral isomorphism does not determine entropy). There exist spectrally isomorphic systems with different Kolmogorov-Sinai entropy; hence spectral isomorphism is strictly weaker than measurable isomorphism off the discrete-spectrum class.

Proof. The Bernoulli shifts $B(\frac{1}{2},\frac{1}{2})$ and the uniform $B(\frac{1}{3},\frac{1}{3},\frac{1}{3})$ each have $U_T$ with simple eigenvalue $1$ on constants and countable Lebesgue spectrum on $L_0^2$: $L_0^2\cong {⨁}_{j=1}^{\infty }{L}^2(\mathbb{T},\text{Leb})$ with $U_T$ acting as multiplication by $z$ on each summand (the past-future orthonormal wandering subspaces of the shift generate this decomposition). Two unitaries of countable Lebesgue spectrum are unitarily equivalent, so the shifts are spectrally isomorphic. Their entropies are $\log 2\ne \log 3$, and entropy is invariant under measurable isomorphism (Kolmogorov-Sinai), so they are not measurably isomorphic. $\square$

Connections Master

• The mixing hierarchy and weak mixing 38.05.01 is the opposite pole of this unit on the same axis: weak mixing is continuous spectral type on $L_0^2$ (the death of all eigenfunctions, $H_{\mathrm{pp}}=\mathbb{C}1$), whereas discrete spectrum is purely atomic type ($H_{\mathrm{pp}}=L^2$). The Koopman-von Neumann decomposition $L^2=H_{\mathrm{pp}}\oplus H_{\mathrm{c}}$ proved there is exactly the splitting whose pure-point part Halmos-von Neumann classifies as a group rotation and whose continuous part is the weak-mixing remainder.

• Ergodicity, unique ergodicity, and equidistribution 38.04.02 supplies the indispensable hypotheses: ergodicity is what makes eigenvalues simple and eigenfunctions of constant modulus, so the eigenfunctions form a group; and the uniquely ergodic irrational rotation studied there is the prototype discrete-spectrum system, its eigenvalue group $\{e^{2\pi ik\alpha }\}$ being the simplest infinite case of the Halmos-von Neumann invariant.

• Hilbert space and the spectral theorem 02.11.08 is the operator-theoretic engine: the Koopman operator is unitary, its projection-valued measure furnishes the maximal spectral type and multiplicity, and the Hahn-Hellinger theory provides the complete unitary invariant that this unit specialises to the dynamical setting, with discrete spectrum the atomic-simple extreme.

• The Kolmogorov-Sinai entropy and the generator theorem 38.06.02 provides the invariant that spectral theory cannot see: entropy separates the spectrally isomorphic Bernoulli shifts of Theorem 3, marking the precise boundary of the spectral method and motivating the entropy theory as the complementary classification program for the continuous-spectrum world.

• Oseledets multiplicative ergodic theorem and Lyapunov exponents 38.07.01 sit downstream in the smooth theory: Lyapunov exponents are further measurable invariants invisible to the Koopman operator, and the contrast between the unitary spectral data and the cocycle spectral data (Oseledets splitting) parallels the discrete-versus-continuous dichotomy organised here.

Historical & philosophical context Master

The spectral approach originates with Bernard Koopman, whose 1931 Proceedings of the National Academy of Sciences note ^{[Koopman 1931]} observed that composition with a measure-preserving transformation is a unitary operator on $L^2$, converting a nonlinear dynamical problem into linear operator theory and immediately enabling von Neumann's 1932 mean ergodic theorem. The decisive classification came in the second of the two 1942 Annals of Mathematics papers of Paul Halmos and John von Neumann, "Operator methods in classical mechanics II" ^{[Halmos-von Neumann 1942]}, which proved that an ergodic system with discrete spectrum is determined up to isomorphism by its eigenvalue group and is realised as a rotation on a compact abelian group — the first complete isomorphism classification of a natural class of dynamical systems, and a paradigm for the use of Pontryagin duality in ergodic theory.

The boundaries of the spectral method were charted over the following decades. Hidehiko Anzai's 1951 study of skew products on the torus ^{[Anzai 1951]} produced systems mixing discrete and continuous spectrum and early examples distinguishing spectral from metric invariants, foreshadowing the central limitation. That limitation became decisive with Kolmogorov and Sinai's introduction of entropy in 1958-59, a measurable-isomorphism invariant invisible to the Koopman operator, and with Donald Ornstein's 1970 theorem that entropy is a complete invariant for Bernoulli shifts, exhibiting spectrally isomorphic systems of different entropy. The maximal-spectral-type and multiplicity theory was systematised by Cornfeld, Fomin, and Sinai and later by Nadkarni, and the realisation problem for spectral multiplicities — including Banach's problem of simple Lebesgue spectrum — remains an active research frontier.

Bibliography Master

@article{HalmosvonNeumann1942,
  author  = {Halmos, Paul R. and von Neumann, John},
  title   = {Operator methods in classical mechanics, II},
  journal = {Annals of Mathematics},
  volume  = {43},
  number  = {2},
  year    = {1942},
  pages   = {332--350}
}

@article{Koopman1931,
  author  = {Koopman, Bernard O.},
  title   = {Hamiltonian systems and transformations in Hilbert space},
  journal = {Proceedings of the National Academy of Sciences},
  volume  = {17},
  number  = {5},
  year    = {1931},
  pages   = {315--318}
}

@article{Anzai1951,
  author  = {Anzai, Hidehiko},
  title   = {Ergodic skew product transformations on the torus},
  journal = {Osaka Mathematical Journal},
  volume  = {3},
  number  = {1},
  year    = {1951},
  pages   = {83--99}
}

@article{Ornstein1970,
  author  = {Ornstein, Donald S.},
  title   = {Bernoulli shifts with the same entropy are isomorphic},
  journal = {Advances in Mathematics},
  volume  = {4},
  number  = {3},
  year    = {1970},
  pages   = {337--352}
}

@book{Walters1982,
  author    = {Walters, Peter},
  title     = {An Introduction to Ergodic Theory},
  publisher = {Springer},
  series    = {Graduate Texts in Mathematics},
  volume    = {79},
  year      = {1982}
}

@book{CornfeldFominSinai1982,
  author    = {Cornfeld, Isaac P. and Fomin, Sergei V. and Sinai, Yakov G.},
  title     = {Ergodic Theory},
  publisher = {Springer},
  series    = {Grundlehren der mathematischen Wissenschaften},
  volume    = {245},
  year      = {1982}
}

@book{Nadkarni1998,
  author    = {Nadkarni, Mahendra G.},
  title     = {Spectral Theory of Dynamical Systems},
  publisher = {Birkh\"auser},
  series    = {Birkh\"auser Advanced Texts},
  year      = {1998}
}

@book{Petersen1983,
  author    = {Petersen, Karl},
  title     = {Ergodic Theory},
  publisher = {Cambridge University Press},
  year      = {1983}
}

@book{Glasner2003,
  author    = {Glasner, Eli},
  title     = {Ergodic Theory via Joinings},
  publisher = {American Mathematical Society},
  series    = {Mathematical Surveys and Monographs},
  volume    = {101},
  year      = {2003}
}