Open MathsAll courses

The Weyl-plane action and its continuous part

Translations and modulations on $L^2(\mathbb R)$ form a projective representation of the phase plane. Scalar projective multipliers disappear under conjugation, producing an honest action on all bounded operators. The same Weyl relations compute its integrability, Connes spectrum, centralizer, and invariant von Neumann subalgebras.

For a first reading, follow the Weyl relations and orthogonality in (Y1)–(Y13), then the continuous part in (Y28)–(Y37). The centralizer calculation (Y14)–(Y17) is a short matrix-unit application. The longer subgroup correspondence (Y18)–(Y27) can be read separately after the trace-class convolution, Fourier localization and synthesis arguments.

The norm-continuous part contains the identity and every Weyl unitary, so it strictly contains the compact operators. We characterize it by $L^1$ convolution and explain the discrepancy with the source exercise in the final sections.

The earlier inputs are the concrete predual construction, scalar integration and Fourier analysis, smooth filters, smooth Fourier cutoffs and inversion, normal integrated actions, the concrete Weyl commutant calculation, and the action and Connes spectra. The trace-class realization and Euclidean subgroup synthesis are proved below.

Let $\mathcal H=L^2(\mathbb R)$, with Lebesgue measure, and define

$$ (U_s\xi)(r)=\xi(r+s), \qquad (V_t\xi)(r)=e^{irt}\xi(r), \qquad W(s,t)=U_sV_t. \tag{Y1} $$

We identify the dual of $\mathbb R^2$ by $\chi_{(p,q)}(s,t)=e^{i(sp+tq)}$ and use Lebesgue measure $ds\,dt$.

Original lesson exposition: CC0-1.0 to the extent of rights held. Cited works retain their own terms.

The projective Weyl pair gives an honest plane action

Direct calculation gives

$$ U_sV_t=e^{ist}V_tU_s \tag{Y2} $$

and therefore

$$ W(s,t)W(a,b)=e^{-iat}W(s+a,t+b). \tag{Y3} $$

Thus $W$ is a strongly continuous projective unitary representation with multiplier $m((s,t),(a,b))=e^{-iat}$. Define

$$ \alpha_{(s,t)}=\operatorname{Ad}W(s,t) \quad\text{on }B(\mathcal H). \tag{Y4} $$

The two multipliers in a product of inner automorphisms cancel, so

$$ \alpha_z\alpha_w=\alpha_{z+w} \qquad(z,w\in\mathbb R^2). \tag{Y5} $$

Translations are strongly continuous on $L^2(\mathbb R)$ by FF-1; modulation continuity follows by dominated convergence, since $|e^{irt}-1|^2|\xi(r)|^2\le4|\xi(r)|^2$. Splitting a product difference proves joint strong continuity of $W(s,t)$. The adjoints are strongly continuous too, by the group formulas. For fixed $x$ and $\xi$, split \[ (W(z)xW(z)^*-W(z_0)xW(z_0)^*)\xi =W(z)x(W(z)^*-W(z_0)^*)\xi +(W(z)-W(z_0))xW(z_0)^*\xi . \] Both terms tend to zero; replacing $x$ by $x^*$ proves strong-star continuity. Every normal functional is a summable vector series by CP-04. The uniform bound $2\|x\|$ controls its tail, so finite vector tests prove point-ultraweak continuity on the whole predual. Thus (Y4) is a continuous action despite the multiplier in (Y3).

The commutator phase will be used repeatedly. If $z=(s,t)$ and $w=(a,b)$, then

$$ \alpha_z(W(w)) =e^{i\omega(z,w)}W(w), \qquad \omega(z,w)=sb-at. \tag{Y6} $$

The alternating form $\omega$ is nondegenerate.

Phase-space orthogonality proves integrability

For $\xi,\eta\in\mathcal H$, Fubini and scalar Plancherel give the exact orthogonality identity

$$ \int_{\mathbb R^2} |\langle\eta,W(s,t)\xi\rangle|^2\,ds\,dt =2\pi\|\xi\|^2\|\eta\|^2. \tag{Y7} $$

For fixed $s$, the product $h_s(u)=\xi(u)\overline{\eta(u-s)}$ is integrable by Cauchy–Schwarz. Tonelli gives \[ \int_{\mathbb R}\int_{\mathbb R}|h_s(u)|^2\,du\,ds =\|\xi\|^2\|\eta\|^2, \] so $h_s\in L^1\cap L^2$ for almost every $s$. After the change $u=r+s$, the coefficient in (Y7) is $\overline{\int h_s(u)e^{itu}\,du}$. Its modulus is therefore the modulus of the ordinary Fourier transform of $h_s$. The full $L^1\cap L^2$ compatibility and Plancherel theorem of FF-2 therefore give the factor $2\pi$ for almost every $s$. A second application of Tonelli proves (Y7) for all $\xi,\eta\in L^2$, with no boundedness or smoothness hypothesis on the vectors.

Write $\Theta_{\xi,\xi}$ for the positive rank-one operator $\zeta\mapsto\langle\zeta,\xi\rangle\xi$, using the course convention that the inner product is linear in the first variable. For every $\eta$, (Y7) says

$$ \int_{\mathbb R^2} \langle\alpha_z(\Theta_{\xi,\xi})\eta,\eta\rangle\,dz =2\pi\|\xi\|^2\|\eta\|^2. \tag{Y8} $$

The increasing net of compact positive orbit integrals therefore has the bounded ultraweak supremum

$$ \int_{\mathbb R^2}\alpha_z(\Theta_{\xi,\xi})\,dz =2\pi\|\xi\|^2 1. \tag{Y9} $$

Positive finite-rank operators are finite positive sums of such rank-one operators, so they all belong to the bounded positive integrability domain. For every $x\ge0$, the operators $x^{1/2}P_Fx^{1/2}$, where $P_F$ ranges over the projections onto finite-dimensional subspaces, are positive finite-rank operators bounded by $x$ and increase strongly to $x$. Their values under every positive normal functional increase to its value at $x$, and all normal-functional values converge, by the vector-series description. Thus the positive bounded-integral domain is order dense, as well as having ultraweakly dense linear span. The positive action integral is semifinite:

$$ \boxed{\alpha\text{ is an integrable action of }\mathbb R^2.} \tag{Y10} $$

No operator-norm integral over the whole plane is asserted in (Y8)–(Y9).

The Connes spectrum is the whole dual plane

An operator fixed by every $\alpha_z$ commutes with every translation and every modulation. Commutation with all $V_t$ makes it a multiplication operator, and commutation with all $U_s$ makes its multiplier translation invariant. The multiplier and translation-invariant-class steps, including almost-everywhere equivalence, are the full ND Weyl proof. Thus

$$ B(\mathcal H)^\alpha=\mathbb C1. \tag{Y11} $$

Formula (Y6) says that $W(a,b)$ is an eigenunitary for the dual frequency

$$ (p,q)=(b,-a). \tag{Y12} $$

As $(a,b)$ ranges over the plane, so does $(p,q)$. If an integrated kernel annihilates the entire action, applying it to $W(a,b)$ makes its Fourier transform vanish at every such frequency. Consequently the hull defining the action spectrum contains every $(p,q)$; equivalently each displayed eigenunitary contributes its own frequency. Thus the point spectrum and action spectrum both fill $\mathbb R^2$. The only nonzero projection in the fixed algebra (Y11) is $1$, so the intersection defining the Connes spectrum has only the unreduced action as a term. Consequently

$$ \boxed{\Gamma(\alpha)=\operatorname{Sp}(\alpha)=\mathbb R^2.} \tag{Y13} $$

The same calculation proves that every eigenspace is one dimensional. If $x$ has the character in (Y12), then $W(a,b)^*x$ is fixed by $\alpha$ and therefore scalar by (Y11).

Every commuting automorphism is a Weyl translation

Let $\beta\in\operatorname{Aut}(B(\mathcal H))$ commute with every $\alpha_z$. A star automorphism of a von Neumann algebra is normal: it and its inverse preserve order, hence every existing least upper bound of a bounded increasing net. The equivalence with ultraweak normality is the automatic normality proof. Here is its inner implementation in this setting. Choose an orthonormal basis $(\xi_i)$ and its matrix units $e_{ij}$, with one distinguished index $0$. The image $\beta(e_{00})$ is a minimal projection of $B(\mathcal H)$ and therefore has rank one: a projection of larger rank has a nontrivial rank-one subprojection. Choose a unit vector $\eta_0$ in its range and set $\eta_i=\beta(e_{i0})\eta_0$. The matrix-unit relations make these vectors orthonormal. Normality of $\beta$ gives $\sum_i\beta(e_{ii})=1$, so they form a basis. The unitary $T\xi_i=\eta_i$ implements $\beta$ on every matrix unit. Finite matrix compressions converge strongly on the bounded unit ball, and hence ultraweakly; normality extends the identity to every operator. Thus $\beta=\operatorname{Ad}T$. Commutation of the two inner automorphisms implies

$$ TW(z)T^*W(z)^*\in\mathbb T1. \tag{Y14} $$

Hence there is a scalar function $c$ with

$$ TW(z)T^*=c(z)W(z). \tag{Y15} $$

Strong continuity makes $c$ continuous. Applying (Y15) to the product law (Y3) shows $c(z+w)=c(z)c(w)$, so

$$ c(s,t)=e^{i(ps+qt)} \tag{Y16} $$

for unique $p,q\in\mathbb R$. Here the elementary character assertion can be seen without a duality theorem. For a continuous character $d:\mathbb R\to\mathbb T$, choose a neighborhood of zero on which $d(t)=e^{i\phi(t)}$ with continuous $|\phi(t)|<\pi/4$ and $\phi(0)=0$. The character law makes $\phi(s+t)-\phi(s)-\phi(t)$ an integer multiple of $2\pi$; when all three arguments are in that neighborhood its absolute value is less than $3\pi/4$, hence it is zero. Fix a positive $a$ in a smaller interval. Subdivision and repeated local addition show $\phi(ra)=r\phi(a)$ for rational $r$ with $ra$ in that smaller interval. Continuity extends this to real $r$, so $\phi(t)=pt$ there. Writing any $t$ as a sufficiently large integer multiple of a small argument gives $d(t)=e^{ipt}$ everywhere. Apply this to $c(s,0)$ and $c(0,t)$ and use $c(s,t)=c(s,0)c(0,t)$. Equality of two slopes at every sufficiently small real argument forces them equal. By (Y6), conjugation by $W(q,-p)$ has exactly the same effect (Y16) on every $W(s,t)$. Therefore $W(q,-p)^*T$ commutes with the irreducible Weyl family and is scalar. We obtain

$$ \boxed{ \{\beta\in\operatorname{Aut}(B(\mathcal H)): \beta\alpha_z=\alpha_z\beta\ \forall z\} =\alpha(\mathbb R^2).} \tag{Y17} $$

Faithfulness follows at the same time: nondegeneracy of $\omega$ makes $\alpha_z=\operatorname{id}$ possible only for $z=0$.

Closed subgroups classify invariant von Neumann subalgebras

For $E\subseteq\mathbb R^2$, define its symplectic annihilator by

$$ E^{\perp_\omega} =\{z\in\mathbb R^2:e^{i\omega(z,e)}=1\text{ for every }e\in E\}. \tag{Y18} $$

An eigenoperator calculation alone does not prove this classification. We supply the extra argument below: trace-class convolutions give a Gaussian correspondence with classical ideals, and closed subgroups of the Euclidean plane are synthesis sets.

Trace-class operators from the concrete predual

Let $E_{\mathcal H}$ be the completed projective tensor space constructed in CP-01–04. On algebraic tensors define \[ \iota(\xi\otimes\overline\eta)=\Theta_{\xi,\eta}, \qquad \Theta_{\xi,\eta}\zeta=\langle\zeta,\eta\rangle\xi . \tag{TC1} \] The bound $\|\Theta_{\xi,\eta}\|=\|\xi\|\|\eta\|$ makes $\iota$ contractive into $B(\mathcal H)$, so it extends to the completion. This extension is injective. Indeed, for a finite-rank orthogonal projection $P$, the tensor map $C_P=P\otimes\overline P$ is contractive and $\iota(C_Pu)=P\iota(u)P$. As $P$ increases over all finite-dimensional subspaces, $C_Pu\to u$ in the projective norm: this is immediate for a finite tensor sum and follows for the completion by contraction and approximation. The range of $C_P$ is the finite-dimensional tensor space $P\mathcal H\otimes\overline{P\mathcal H}$, and its rank-one matrix units are linearly independent as operators. Thus $\iota(u)=0$ makes every $C_Pu=0$, and the limit makes $u=0$.

Define $\mathcal T(\mathcal H)=\iota(E_{\mathcal H})$ and $\|S\|_1=\|\iota^{-1}S\|$. This is a Banach space and finite-rank operators are dense. CP-03 gives its complete dual: \[ \langle S,x\rangle=\operatorname{Tr}(Sx),\qquad \operatorname{Tr}(\Theta_{\xi,\eta}x)=\langle x\xi,\eta\rangle, \qquad \mathcal T(\mathcal H)^*=B(\mathcal H). \tag{TC2} \] Here the trace is defined by pairing with $1$. It agrees with the ordinary finite matrix trace on finite-rank operators. In particular $|\operatorname{Tr}(Sx)|\le\|S\|_1\|x\|$ and $\|\Theta_{\xi,\eta}\|_1=\|\xi\|\|\eta\|$ by CP-02.

For $a,b\in B(\mathcal H)$, $a\Theta_{\xi,\eta}b=\Theta_{a\xi,b^*\eta}$. Taking the infimum over finite tensor representations, then passing to the completion, proves $\|aSb\|_1\le\|a\|\|S\|_1\|b\|$. Finite-rank calculation gives $\operatorname{Tr}(Sx)=\operatorname{Tr}(xS)$; both sides are trace-norm continuous, so this holds for every $S\in\mathcal T$. Conjugation by a unitary is a trace-norm isometry, since its inverse is also contractive. CP-04 represents every $S$ by an absolutely summable series of rank-one operators. The equality of the induced dual topology with the ultraweak topology and the annihilator identity for every ultraweakly closed linear subspace are CP-04–05. Thus \[ N=(N_\perp)^\perp,\qquad N_\perp=\{S:\operatorname{Tr}(Sx)=0\text{ for every }x\in N\}. \tag{TC3} \]

For completeness, the Hilbert–Schmidt completion used below also embeds faithfully in operators. On finite-rank operators put $\|S\|_{\mathrm{HS}}^2=\sum_j\|Se_j\|^2$ for an orthonormal basis. Parseval on the finite-dimensional range makes this basis independent and gives its Hilbert inner product. Rank-one operators identify this space with the algebraic Hilbert tensor product $\mathcal H\otimes\overline{\mathcal H}$. The inequality $\|S\|\le\|S\|_{\mathrm{HS}}$ gives a bounded operator associated to every vector of the completion. If that operator is zero, its inner products with all rank-one tensors vanish, because these are its matrix coefficients; density forces the Hilbert tensor vector to be zero. Moreover the triangle inequality and $\|\Theta_{\xi,\eta}\|_{\mathrm{HS}}=\|\xi\|\|\eta\|$ give $\|S\|_{\mathrm{HS}}\le\|S\|_1$ first on finite sums and then on all of $\mathcal T$. Hence trace-norm convergence implies Hilbert–Schmidt convergence with the same resulting operator.

These arguments supply precisely the trace-class, normal-functional and Hilbert–Schmidt facts needed here. No singular-value expansion is required. Rank-one operators with Schwartz vectors are trace-norm dense: approximate each vector in a finite sum in $L^2$ and use the rank-one norm bound.

The scalar Fourier and ideal conventions

The earlier FF-1–2, smooth-filter proof and RF-1 give Gaussian integration, Fourier inversion, scalar Plancherel, smooth cutoffs and the required $L^1$ compatibility. In two variables apply the one-variable transforms successively; integrable majorants justify Fubini on Schwartz functions, and the $L^2$ isometries extend the resulting identities by density. Repeated differentiation under the integral and integration by parts show that Fourier transform and inverse preserve the Schwartz seminorms. In particular, multiplying a Schwartz transform by a smooth cutoff equal to one on expanding balls converges in all Schwartz seminorms, as do the inverse transforms. Schwartz density in $L^1$ and $L^2$ follows by approximating by compact continuous functions and convolving with smooth compact approximate identities; uniform convergence on a common compact set gives the claimed norms. These are also the scalar approximation constructions of the linked Fourier lessons.

The vector integrals use FF-3. For $m=1,2$, write $A(\mathbb R^m)=\{Fa:a\in L^1(\mathbb R^m)\}$ with the positive Fourier convention below. Its norm is $\|Fa\|_A=\|a\|_1$. Fourier injectivity makes the norm well-defined, and convolution makes it a Banach algebra. For a closed set $E$ put \[ I(E)=\{f\in A:f|_E=0\},\qquad j(E)=\overline{\{f\in A:\operatorname{supp}f\text{ is compact and disjoint from }E\}}^{\,\|\cdot\|_A}. \tag{TC4} \] For a closed ideal $J$, its hull is $h(J)=\{p:f(p)=0\text{ for every }f\in J\}$. These are the ideals in the local division and synthesis arguments below; no synthesis theorem is assumed.

Euclidean local division. Use the ordinary positive-sign transform $Fa(p)=\int a(t)e^{it\cdot p}\,dt$ on $\mathbb R^m$, $m=1,2$, and the norm $\|Fa\|_A=\|a\|_1$. Smooth compact functions are in this algebra by Schwartz inversion. Its compact Fourier functions are dense: multiplying $Fa$ by $\eta(p/R)$, where $\eta\in C_c^\infty$ is one near zero, convolves $a$ with a bounded $L^1$ approximate identity. Translation continuity proves convergence in $L^1$. Fourier injectivity follows from the preceding Plancherel argument, and vanishing at infinity follows by $L^1$ approximation from Schwartz kernels.

Let $f=Fa$ and $f(p_0)=c\ne0$. Choose $\chi\in C_c^\infty$, equal to one near zero, and write $\chi_\varepsilon(p)=\chi((p-p_0)/\varepsilon)$. If $b_\varepsilon$ is the inverse transform of $\chi(p/\varepsilon)$, then $b_\varepsilon(t)=\varepsilon^m b(\varepsilon t)$. Its $L^1$ norm is constant, and $\|\tau_y b_\varepsilon-b_\varepsilon\|_1\to0$ for every fixed $y$. The inverse kernel of $(f-c)\chi_\varepsilon$ is a modulation of

$$ \int a(y)e^{iy\cdot p_0} (\tau_y b_\varepsilon-b_\varepsilon)\,dy . \tag{Y18a} $$

This is an $L^1$ Bochner integral. Dominated convergence, with bound $2\|b\|_1|a(y)|$, makes its norm tend to zero. Choose $\varepsilon$ with $\|(f-c)\chi_\varepsilon\|_A<|c|$. In the unitized Fourier algebra, $d=c1+(f-c)\chi_\varepsilon$ has the inverse given by its convergent Neumann series. On a neighborhood of $p_0$, $d=f$. Multiply $d^{-1}$ by a smooth compact cutoff that is one on a smaller neighborhood. The result lies in $A$ and is a local inverse of $f$ there.

For a compact set on which $f$ has no zeros, choose finitely many such neighborhoods and smooth cutoffs $c_i$ supported inside them and equal to one on a smaller covering. The functions $q_1=c_1$, $q_i=c_i\prod_{k<i}(1-c_k)$ sum to one near that compact set. Multiplying each local inverse by $q_i$ and summing proves the compact local inverse assertion. It also proves the local ideal inclusion used below: if $J$ is a closed ideal, a compact Fourier function supported outside its common zero set has a finite decomposition $h=\sum_i hq_i$ with $hq_i=f_i(v_i hq_i)\in J$, where $f_i\in J$ is nonzero on the supporting neighborhood and $v_i$ is its local inverse. Taking norm closure gives $j(h(J))\subset J\subset I(h(J))$. This supplies the Euclidean instance directly; no local inverse theorem for an arbitrary LCA group is needed here.

Weyl synthesis lemma. Let $N\subseteq B(\mathcal H)$ be a unital von Neumann algebra satisfying $\alpha_z(N)=N$ for every $z$. Then

$$ L_N=\{w\in\mathbb R^2:W(w)\in N\} \tag{Y19} $$

is a closed subgroup and

$$ N=\{W(w):w\in L_N\}''. \tag{Y20} $$

Moreover, for every closed subgroup $L$,

$$ \{W(w):w\in L\}' =\{W(z):z\in L^{\perp_\omega}\}''. \tag{Y21} $$

For trace-class $S$, its Fourier–Wigner transform

$$ \mathcal W_S(w)=\operatorname{Tr}(S W(w)^*) \tag{Y22} $$

is continuous, vanishes at infinity, and separates trace-class operators. Here is the norm argument needed for the last assertion. Polarizing (Y7) in both vector variables and summing the resulting rank-one pairings gives, for every finite-rank $S$,

$$ \int_{\mathbb R^2}|\mathcal W_S(w)|^2\,dw =2\pi\|S\|_{\mathrm{HS}}^2. \tag{Y22a} $$

Take finite-rank $S_n$ converging to an arbitrary trace-class $S$ in trace norm. They also converge in Hilbert--Schmidt norm, so (Y22a) gives an $L^2$ limit of their transforms. Meanwhile $\|\mathcal W_{S_n}-\mathcal W_S\|_\infty\leq\|S_n-S\|_1$. On each compact set this uniform convergence is also $L^2$ convergence. Hence the $L^2$ limit is $\mathcal W_S$ almost everywhere, and (Y22a) holds for $S$. In particular, $\mathcal W_S=0$ implies $S=0$. Injectivity is proved by this norm identity, rather than inferred from injectivity on the finite-rank subspace alone.

For continuity and vanishing at infinity, approximate $S$ in trace norm by finite-rank operators formed from Schwartz vectors. Their transforms are Schwartz functions: explicitly,

$$ \mathcal W_{\Theta_{\xi,\eta}}(s,t) =\int e^{-irt}\xi(r-s)\overline{\eta(r)}\,dr. \tag{Y22b} $$

This is a partial Fourier transform of a Schwartz function of $(r,s)$. The same uniform bound makes $\mathcal W_S$ a uniform limit in $C_0$. Conjugating $S$ by $W(z)$ multiplies (Y22) by the character $w\mapsto e^{i\omega(z,w)}$. Consequently $N_\perp$ is invariant under conjugation by Weyl operators. The common zero set of its transforms is exactly $L_N$: by the preannihilator separation theorem, $\mathcal W_S(w)=0$ for every $S\in N_\perp$ if and only if $W(w)^*\in N$, equivalently $W(w)\in N$. The multiplication and adjoint formulas (Y3) make $L_N$ a subgroup. Strong continuity of $W$ and weak-star closedness of $N$ make it closed.

Trace-class convolution. For this proof only, use the normalized measure $d\mu(z)=ds\,dt/(2\pi)$ and symmetric Weyl operators

$$ D(s,t)=e^{-ist/2}W(s,t),\qquad Q_S(w)=\operatorname{Tr}(S D(w)^*),\qquad (J\xi)(r)=\xi(-r). \tag{Y22c} $$

Thus $D(z)^*=D(-z)$, $JD(z)J=D(-z)$, and $Q_S(a,b)=e^{iab/2}\mathcal W_S(a,b)$. Neither the zero sets nor the generated algebras change. For $f\in L^1(d\mu)$ and trace-class $T,S$, set

$$ \begin{aligned} f*T&=\int f(z)\alpha_z(T)\,d\mu(z),\\ (T*S)(z)&=\operatorname{Tr}\bigl(T\alpha_z(JSJ)\bigr),\\ F_\omega f(w)&=\int f(z)e^{i\omega(z,w)}\,d\mu(z). \end{aligned} \tag{Y22d} $$

We also write $T*f=f*T$. The first integral converges in trace norm: rank-one approximation and strong continuity of $D$ make trace-class orbits norm continuous. It satisfies $\|f*T\|_1\leq\|f\|_{L^1}\|T\|_1$. For two rank-one operators, the integrand defining $T*S$ is a product of two Weyl coefficients. Cauchy--Schwarz and (Y7), with the normalized measure, bound its $L^1$ norm by the product of the four vector norms. Apply this bound to finite sums of rank-one operators and take the infimum over both projective tensor representations. Trace-norm completion and completeness of $L^1$ then give

$$ \|T*S\|_{L^1}\leq\|T\|_1\|S\|_1,\qquad F_\omega(T*S)=Q_TQ_S,\qquad Q_{f*T}=(F_\omega f)Q_T. \tag{Y22e} $$

Here is a direct sign and normalization check for the middle identity. For finite-rank Schwartz operators with kernels $K_T,K_S$, the convolution is

$$ (T*S)(s,t)=\int_{\mathbb R^2} K_T(r,u)e^{it(u-r)}K_S(-u-s,-r-s)\,dr\,du. \tag{Y22f} $$

For fixed $s$, write (Y22f) as $\int H_s(v)e^{itv}\,dv$, where $H_s(v)=\int K_T(r,r+v)K_S(-r-v-s,-r-s)\,dr$. This is a Schwartz function of $(s,v)$. Taking $F_\omega$ at $(a,b)$, one-dimensional Fourier inversion evaluates $H_s$ at $a$; this justifies $u=r+a$ without exchanging a nonintegrable oscillatory integrand by Fubini. Put $v=-r-s$ in the remaining integral. The result is the product

$$ \left(\int K_T(r,r+a)e^{-ib(r+a/2)}\,dr\right) \left(\int K_S(v-a,v)e^{-ib(v-a/2)}\,dv\right) =Q_T(a,b)Q_S(a,b). \tag{Y22g} $$

All the functions in this calculation are Schwartz, so inversion is valid before the trace-norm approximation. The norm bound in (Y22e) and the uniform bound for $Q$ extend the identity to arbitrary trace-class operators. The last identity in (Y22e) follows directly by pairing the trace-norm integral with $D(w)^*$.

Classical Fourier injectivity and (Y22a) now prove associativity of these convolutions: both parenthesizations of three trace-class factors have transform $Q_TQ_SQ_R$. The same verification covers mixed classical and trace-class factors. Also $T*S=S*T$; cyclicity of trace and $JD(z)J=D(-z)$ give this directly.

A Gaussian recovers the trace-class subspace. Let $g_0(r)=\pi^{-1/4}e^{-r^2/2}$ and $G=\Theta_{g_0,g_0}$. Completing the square in (Y22b) gives

$$ Q_G(w)=e^{-|w|^2/4},\qquad F_\omega(G*G)(w)=e^{-|w|^2/2}. \tag{Y22h} $$

The classical ideal $(G*G)*L^1$ is dense in $L^1$. To prove this specific assertion, take a Schwartz function $h$ whose Fourier transform is smooth and compactly supported. Dividing that transform by the nowhere-zero Gaussian in (Y22h) gives another smooth compactly supported function; its inverse transform $b$ is Schwartz, and $h=(G*G)*b$. Such $h$ are dense in $L^1$: first approximate by Schwartz functions, then multiply their Schwartz transforms by smooth cutoffs equal to one on expanding balls. The inverse transforms converge in the Schwartz topology, hence in $L^1$. No unrestricted spectral-synthesis or Wiener-density assertion is used here.

Let $\mathcal M$ be a norm-closed conjugation-invariant trace-class subspace, and define

$$ \mathcal I_{\mathcal M} =\overline{\operatorname{span}\{T*S: T\text{ trace class},\ S\in\mathcal M\}}^{L^1}. \tag{Y22i} $$

Covariance makes this a closed translation-invariant classical subspace. It is an ideal: convolution by a classical kernel is a Bochner integral of translates, hence preserves it. Likewise $f*S\in\mathcal M$ for $S\in\mathcal M$. Associativity gives $T*(R*S)=(T*R)*S\in\mathcal M$, so

$$ \overline{\operatorname{span} \{T*f:T\text{ trace class},\ f\in\mathcal I_{\mathcal M}\}}^{\|\cdot\|_1} =\mathcal M. \tag{Y22j} $$

For the reverse inclusion, approximate a classical approximate identity in $L^1$ by kernels $(G*G)*b$. Acting on $S\in\mathcal M$, these kernels approach $S$ in trace norm. Their actions equal $G*(G*(b*S))$, whose inner classical factor is in $\mathcal I_{\mathcal M}$. This proves the missing inclusion in (Y22j). Formula (Y22e) and the nonvanishing $Q_G$ also show that the common Fourier zero set of $\mathcal I_{\mathcal M}$ is exactly the common zero set of the $Q_S$ for $S\in\mathcal M$.

Classical closed-subgroup synthesis. We prove $j(L)=I(L)$ in the Euclidean Fourier algebra for every closed subgroup $L\leq\mathbb R^2$, using the ideals of lesson 17. The symplectic Fourier transform differs from the ordinary one by an invertible linear change of frequency coordinates, which preserves this assertion.

First a point is a synthesis set in $\mathbb R^m$; the case $m=0$ has zero vanishing ideal and is immediate. Choose $\chi\in C_c^\infty(\mathbb R^m)$ equal to one near zero, and let $b_\varepsilon$ be the inverse transform of $\chi(\,\cdot/\varepsilon)$. Scaling gives uniformly bounded $L^1$ norms and $\|\tau_yb_\varepsilon-b_\varepsilon\|_1\to0$ for every fixed $y$. For a mean-zero $a\in L^1$, its convolution is the actual $L^1$ Bochner integral

$$ a*b_\varepsilon =\int a(y)(\tau_yb_\varepsilon-b_\varepsilon)\,dy \longrightarrow0\quad\text{in }L^1. \tag{Y22k} $$

The convergence follows by dominated convergence and the uniform norm bound. Consequently $(Fa)(1-\chi(\,\cdot/\varepsilon))$ approaches $Fa$ in Fourier-algebra norm and vanishes near zero. Multiplication by smooth expanding Fourier cutoffs then makes each approximant compactly supported without losing that vanishing. Modulation shifts the same proof to any point.

Every affine plane $V\times\{\lambda\}$ in $\mathbb R^d\times\mathbb R^m$ is therefore a synthesis set. Indeed, restriction of a Fourier transform to that plane is the Fourier transform of the bounded map

$$ r_\lambda a(u)=\int a(u,v)e^{iv\cdot\lambda}\,dv. \tag{Y22l} $$

Choose $\rho\in L^1(\mathbb R^m)$ with $\int\rho(v)e^{iv\cdot\lambda}\,dv=1$. The map $a\mapsto(r_\lambda a)\otimes\rho$ is a bounded projection. Approximate an element of $\ker r_\lambda$ by finite separated sums in $L^1(\mathbb R^{d+m})$, then apply one minus this projection. The resulting sums have transverse factors of zero $\lambda$-evaluation. The point-synthesis proof approximates each such factor by a compact Fourier function vanishing near $\lambda$; compact Fourier approximation of the other factors completes the affine-plane synthesis proof. Fourier injectivity identifies $\ker r_\lambda$ with the full vanishing ideal of the plane. The cases $d=0$ and $m=0$ are included, with the latter full-plane vanishing ideal equal to zero.

To handle $L$, choose a vector subspace $V\subset L$ of maximal dimension. Orthogonal projection gives $L=V+\Lambda$, where $\Lambda=L\cap V^\perp$ is closed and discrete. For if nonzero $\lambda_n\in\Lambda$ tended to zero, a subsequence of their unit directions would converge to a unit vector $v\in V^\perp$. For every real $t$, suitable integer multiples of $\lambda_n$ converge to $tv$, so closedness would put the line $\mathbb Rv$ in $L$, contradicting maximality.

Write the frequency coordinates as $(u,v)\in V\times V^\perp$. Every smooth compact cutoff in just one coordinate block acts as a bounded Fourier-algebra multiplier, because it convolves the inverse kernel in that block with an $L^1$ kernel. Thus the expressions $1\otimes\eta_R$ and $1\otimes u_\lambda$ below do not require the constant function to belong to the full Fourier algebra. For $f\in I(L)$, multiply in the transverse variable by smooth compact cutoffs $\eta_R(v)=\eta(v/R)$ with $\eta=1$ near zero. Their inverse kernels are an approximate identity in that variable, so $f_R=(1\otimes\eta_R)f\to f$ in Fourier-algebra norm. Only finitely many points of $\Lambda$ meet $\operatorname{supp}\eta_R$.

Choose disjoint smooth transverse cutoffs $u_\lambda$, equal to one near these finitely many points and supported in neighborhoods containing no other lattice points. Each $(1\otimes u_\lambda)f_R$ vanishes on $V\times\{\lambda\}$. Approximate it by the affine-plane result, then multiply the approximants by a slightly larger cutoff equal to one on $\operatorname{supp}u_\lambda$ and still avoiding all other lattice points. They are compact Fourier functions whose supports miss the whole of $L$. Hence each localized piece is in $j(L)$. The remainder $(1-\sum_\lambda1\otimes u_\lambda)f_R$ has compact transverse support disjoint from $\Lambda$. Expanding compact cutoffs in the $V$ variable approximate it in norm by compact Fourier functions with supports disjoint from $L$, so it too belongs to $j(L)$. Thus $f_R\in j(L)$, and closedness gives $f\in j(L)$. This proves subgroup synthesis.

Finish the Weyl synthesis proof. Take $\mathcal M=N_\perp$. Its classical ideal has common zero set $L_N$. The Euclidean local ideal inclusion just proved, together with the preceding subgroup proof and the definitions (TC4), implies that its Fourier image is exactly $I(L_N)$. If a trace-class $S$ has $Q_S|_{L_N}=0$, then $G*S\in\mathcal I_{\mathcal M}$ by (Y22e). Approximate a classical approximate identity by $(G*G)*b$ once again. Its action on $S$ equals $G*(b*(G*S))$, which lies in $\mathcal M$ by (Y22j). Taking the trace-norm limit puts $S$ in $\mathcal M$. The opposite inclusion is the definition of the common zero set. Since $Q_S$ and $\mathcal W_S$ have identical zeros, we obtain

$$ N_\perp =\{S\in\mathcal T(\mathcal H): \mathcal W_S(w)=0\text{ for every }w\in L_N\}. \tag{Y23} $$

This Gaussian argument supplies the exact correspondence and subgroup synthesis needed in Werner's Corollary 4.4(1), (4), (5), printed page 1409. The trace-class convolution identities correspond to Propositions 3.2 and 3.4, printed pages 1405–1407. Its proof uses classical Wiener approximation; the proof here uses explicit Gaussian division on compact smooth Fourier supports. The Euclidean subgroup argument above supplies the synthesis step directly.

Taking annihilators in (Y23) gives (Y20). To prove (Y21), set $C=\{W(w):w\in L\}'$. The character formula (Y6) makes $C$ $\alpha$-invariant. It also shows that $W(z)\in C$ holds exactly when $z\in L^{\perp_\omega}$. Applying (Y20) to $C$ proves (Y21). The same synthesis argument applies to this commutant because it is a unital ultraweakly closed algebra invariant under every $\alpha_z$.

Euclidean double annihilators. The remaining subgroup identity can also be proved here. For the decomposition $L=V+\Lambda$ obtained above, put $E=\operatorname{span}_{\mathbb R}\Lambda\subset V^\perp$. The discrete subgroup $\Lambda$ has a lattice basis in $E$. We give the argument in the dimensions at issue, rather than importing an arbitrary-group duality theorem.

If $\Lambda\ne\{0\}$, choose a shortest nonzero $\lambda_1\in\Lambda$. It exists because a closed discrete subset has finite intersection with every compact set. On the line $\mathbb R\lambda_1$, subtracting an integer multiple of $\lambda_1$ leaves a multiple with coefficient in $[0,1)$; minimal length makes that remainder zero. Thus $\Lambda\cap\mathbb R\lambda_1=\mathbb Z\lambda_1$. This settles $\dim E=1$.

If $\dim E=2$, project $\Lambda$ onto the line perpendicular to $\lambda_1$ within $E$. Zero is isolated in the projected subgroup. Otherwise choose nonzero projected values tending to zero, and subtract integer multiples of $\lambda_1$ from their lifts so that their coordinates along $\lambda_1$ lie in $[0,\|\lambda_1\|)$. The reduced lifts lie in a fixed compact rectangle. Its intersection with $\Lambda$ is finite, which is incompatible with those nonzero projected values tending to zero. A nonzero subgroup of $\mathbb R$ with zero isolated is $d\mathbb Z$ for a least positive $d$: distinct elements are separated by a fixed positive distance, so there are only finitely many in any bounded interval; division by the least positive element leaves zero remainder. Choose $\lambda_2\in\Lambda$ with projection $d$. For every $\lambda\in\Lambda$, subtract an integer multiple of $\lambda_2$ to remove that projection, then an integer multiple of $\lambda_1$. This proves $\Lambda=\mathbb Z\lambda_1+\mathbb Z\lambda_2$. The case $E=\{0\}$ requires no basis.

Let $x\notin L$ and write $x=x_V+x_{V^\perp}$. If $x_{V^\perp}\notin E$, its nonzero component perpendicular to $E$ gives a vector $p\in V^\perp\cap E^\perp$ with $p\cdot x=\pi$. This $p$ pairs to zero with $L$. If $x_{V^\perp}\in E$, express it in the lattice basis. Some coefficient $a_j$ is not an integer. The dual vector $p_j\in E$ satisfying $p_j\cdot\lambda_i=2\pi\delta_{ji}$ pairs with every member of $L$ in $2\pi\mathbb Z$, while $p_j\cdot x=2\pi a_j\notin2\pi\mathbb Z$. Thus in every case an ordinary Euclidean character separates $x$ from $L$.

For the symplectic convention (Y6), set $J_0(s,t)=(-t,s)$, a rotation of the plane distinct from the parity operator $J$ in (Y22c). Then $\omega(z,w)=J_0z\cdot w$. If $p$ is a separating vector just found, $w=J_0^{-1}p$ belongs to $L^{\perp_\omega}$ and $\omega(x,w)=-p\cdot x\notin2\pi\mathbb Z$. Alternation gives the opposite inclusion for elements of $L$. We have therefore proved

$$ (L^{\perp_\omega})^{\perp_\omega}=L \quad\text{for every closed }L\le\mathbb R^2. \tag{Y23a} $$

Now let $H\le\mathbb R^2$ be closed. By (Y21),

$$ B(\mathcal H)^{\alpha|_H} =\{W(h):h\in H\}' =\{W(w):w\in H^{\perp_\omega}\}''. \tag{Y24} $$

Conversely, if $N$ is invariant, put $H=L_N^{\perp_\omega}$. The locally proved identity (Y23a) and (Y20) give

$$ N=\{W(w):w\in L_N\}'' =B(\mathcal H)^{\alpha|_H}. \tag{Y25} $$

The subgroup is recovered exactly as

$$ H=\{z:\alpha_z(x)=x\text{ for every }x\in N\}, \tag{Y26} $$

because (Y6) and $L_N^{\perp_\omega}=H$ test every Weyl generator of $N$. Thus

$$ \boxed{ H\longmapsto B(\mathcal H)^{\alpha|_H} \text{ is an order-reversing bijection}} \tag{Y27} $$

from closed subgroups of $\mathbb R^2$ to unital $\alpha$-invariant von Neumann subalgebras.

The printed compact-operator equality is false

Let

$$ B(\mathcal H)_c =\{x\in B(\mathcal H): \|\alpha_z(x)-x\|\longrightarrow0\text{ as }z\to0\}. \tag{Y28} $$

The proposed equality $B(\mathcal H)_c=K(\mathcal H)$ cannot hold. First,

$$ 1\in B(\mathcal H)_c\setminus K(\mathcal H). \tag{Y29} $$

More structurally, (Y6) gives

$$ \|\alpha_z(W(w))-W(w)\| =|e^{i\omega(z,w)}-1|\longrightarrow0. \tag{Y30} $$

Every $W(w)$ is therefore in the continuous part, while a unitary on the infinite-dimensional space $L^2(\mathbb R)$ is not compact. These examples disprove the equality in Takesaki's Exercise XI.2.5(e). The strict inclusion and convolution characterization below describe the continuous part.

Convolution exactly characterizes the continuous part

For $f\in L^1(\mathbb R^2)$ and $x\in B(\mathcal H)$, define the ultraweak integral

$$ \alpha_f(x)=\int_{\mathbb R^2}f(z)\alpha_z(x)\,dz. \tag{Y31} $$

Normal-functional testing gives $\|\alpha_f(x)\|\le\|f\|_1\|x\|$. If $(\tau_hf)(z)=f(z-h)$, invariance of Haar measure gives

$$ \alpha_h(\alpha_f(x))=\alpha_{\tau_hf}(x), \qquad \|\alpha_h(\alpha_f(x))-\alpha_f(x)\| \le\|\tau_hf-f\|_1\|x\|. \tag{Y32} $$

Translation is norm continuous on $L^1$, so every convolution in (Y31) belongs to $B(\mathcal H)_c$. Conversely, let $(k_i)$ be a nonnegative $C_c(\mathbb R^2)$ approximate identity with integral one and supports shrinking to $0$. For $x\in B(\mathcal H)_c$,

$$ \|\alpha_{k_i}(x)-x\| \le\int k_i(z)\|\alpha_z(x)-x\|\,dz \longrightarrow0. \tag{Y33} $$

We obtain the characterization

$$ \boxed{ B(\mathcal H)_c =\operatorname{cl}_{\|\cdot\|} \{\alpha_f(x):f\in C_c(\mathbb R^2),\ x\in B(\mathcal H)\}.} \tag{Y34} $$

Finally, the valid compact-operator inclusion follows directly. For rank-one operators,

$$ \alpha_z(\Theta_{\xi,\eta}) =\Theta_{W(z)\xi,W(z)\eta} \tag{Y35} $$

and

$$ \|\Theta_{\xi,\eta}-\Theta_{\xi',\eta'}\| \le \|\xi-\xi'\|\,\|\eta\| +\|\xi'\|\,\|\eta-\eta'\|. \tag{Y36} $$

Strong continuity of $W$ makes every finite-rank orbit norm continuous. Automorphisms are isometric, so norm closure gives

$$ \boxed{K(\mathcal H)\subsetneq B(\mathcal H)_c.} \tag{Y37} $$

Problem. Why does point-strong continuity of $z\mapsto W(z)$ not make $z\mapsto\alpha_z(x)$ norm continuous for every $x\in B(\mathcal H)$?

Solution. Strong convergence of unitaries is uniform on each fixed finite-dimensional range, which proves (Y35) for finite-rank operators, but it is not uniform on the unit ball of an infinite-dimensional Hilbert space. For an explicit example, let $P$ be multiplication by $1_{[0,\infty)}$. Modulations commute with $P$, whereas $\alpha_{(s,0)}(P)$ is multiplication by $1_{[-s,\infty)}$. For every $s\ne0$ the two indicator functions differ by absolute value $1$ on an interval of positive measure, so $\|\alpha_{(s,0)}(P)-P\|=1$, although $U_s\to1$ strongly as $s\to0$. The $L^1$ regularization in (Y31)–(Y34) identifies exactly the operators for which norm continuity does hold. $\square$

The source locator is Takesaki, Theory of Operator Algebras II, Exercise XI.2.5, on printed page 351. Equations (Y1)–(Y17) prove clauses (a)–(c); (Y18)–(Y27), including the Gaussian correspondence and Euclidean subgroup proof (Y22c)–(Y22l), derive clause (d). Equations (Y28)–(Y30) diagnose clause (e) as printed, while (Y31)–(Y37) give the exact continuous-part replacement and retain the valid compact-operator inclusion. The trace-class facts used in the subgroup proof are constructed in (TC1)–(TC3), and its Fourier ideal conventions are (TC4).