Spectral action and heat kernel trace for Ricci flat

manifolds from stochastic flow over second

quantized L2-differential forms.
Sita Gakkhar∗, Matilde Marcolli,
January 2, 2024

Abstract
arXiv:2401.00643v1 [math-ph] 1 Jan 2024

A quantum stochastic differential equation (qsde) on Fock space over L2 differential 1-forms is given from
the small “time” flow of which the trace of the connection Laplacian heat kernel for the spinor endomorphism
bundle can be computed over any compact Ricci-flat Riemannian manifold. The existence of the stochastic
flow is established by adapting the construction from [14]. When the manifold supports a parallel spinor –
Ricci-flatness is a required integrability condition for parallel spinors, the trace of Dirac Laplacian heat kernel
of the spinor bundle can be recovered. For 4-manifolds, this corresponds to the spectral action, and realizes
Einstein-Hilbert action as a stochastic flow.

1 Introduction
This article attempts to formalize the idea that Einstein-Hilbert action for a Riemannian manifold can be viewed
as arising from random fluctuations acting on the spinor bundle. The model of random fluctuations is provided by
a stochastic flow generated by the Dirac Laplacian. The connection between Dirac Laplacian, noncommutative
geometry and gravity is well established (see [15] and references therein for an account), the new contribution
here is the probabilistic perspective.
This realization of spectral action as a stochastic flow is suggested by explicit computations of spectral action
which yield Brownian bridge integrals along with the observation that boson Fock space can be viewed as a
Wiener space, with a preliminary exploration of the perspective put forward in [9] where covariant quantum
diffusions on almost-commutative spectral triples are considered. This article treats canonical spectral triples,
computing explicitly the structure matrix for the diffusion generated by the endomorphism connection Laplacian
acting on the algebra of functions over any compact Riemannian manifold. In absence of a natural action with
respect to which the generator is covariant, the standard constructions using Picard iterates (see Sinha and
Goswami [14], and also [1]) care adapted to show the existence of the solution. From this flow, when there exists
a parallel spinor, spectral action can be evaluated.
Some remarks on notation. By Riemannian (M, g), we mean a Riemannian manifold M with metric g. The
connection on the tangent bundle of M , T M , is always the Levi-Civita connection unless specified otherwise.
When clear from context, the same symbol is used of the connection ∇ on a Hermitian or Riemannian bundle
E and the dual connection on dual bundle E ∗ . After fixing a local orthonormal frame about any p ∈ M ,
(Xi )i∈dim M , ∇Xi will be used interchangeably with ∇i . For local coordinates (xi ) about any p ∈ M , ∂i will
∂
denote the coordinate vector fields ∂x i
. [n] is the set {i ∈ N, i ≤ n} where N, with convention that 0 6∈ N.
[nP: m] denotes the set {n, n + 1 . . . m}. The linear span we mean finite linear span denoted by LinSpan(V) :=
{ i∈[k] αi ai : αi ∈ K, ai ∈ V} where K = R, C is clear from context; if K = R then the subscript is dropped.
Throughout Γ(H) denotes the symmetric (boson) Fock space over the any space H, while E(H) denotes the
exponential vectors given by E(v) = ⊕∞ n=0 (n! )
−1/2 ⊗n
v for v ∈ H. From section 3 onwards, we make use of
Einstein notation. All calculation prior to section 3 are using at the center of normal coordinates where the
metric is identity, for clarity, the explicit summations are used.

1.1 Organization and overview

In the remainder of this section, we introduce the background on stochastic flows and spectral action, and delve
into motivating ideas. In section 2, the structure matrix for connection Laplacian is computed following the
standard prescription. The noise space turns out to be L2 differential forms and the flow lives on the Fock space
∗ Sitanshu Gakkhar, email. The first author would like to thank Debashish Goswami for helpful discussions on qsde’s of Evans-

Hudson type, and also Branimir Ćaćić for discussions on Dirac operators

1
over the differential forms. The section concludes by writing the structure matrix and the qsde for the stochastic
flow in the coordinate free quantum stochastic calculus notation; the relevant material from quantum stochastic
differential equations an appendix reviewing quantum stochastic differential equations is included at the end
(appendix B). Necessary bounds for controlling growth of Laplacian powers are established in section 3. The
existence of flows for the derived qsde is established in section 4 by providing estimates that can be plugged into
the standard theory.

1.2 Spectral action and stochastic flows

The canonical spectral triple for Riemannian (M, g) which carries a spin structure is the data (C ∞ (M ), L2 (S), DM ),
L2 (S) being the Hilbert space of square integrable sections of a spinor bundle S → M , and DM the Dirac operator
associated to the lift of Levi-Civita connection to the spinor bundle[16, pg 67]. We will take S to be any spinor
bundle associated to T M and DM to be the Atiyah-Singer operator D / (see [11, ex II.5.9]), then we have that
D/ 2 = △ + κ/4 where △ is the connection Laplacian for the connection on S and κ is the scalar curvature[11,
Thm II.8.8].
The bosonic spectral action is the linear funtional SbM := Tr f (DM /Λ) for a choice of test function f which
2
we take as e−x and Λ a cutoff parameter[7, § 5.1, 16, § 7.1], so that the parameter t of the Dirac heat semigroup
2
e−tDM will be taken to satisfy t = Λ−2 . From the asymptotic expansion limΛ→∞ SbM for a Riemannian spin
4-manifold M , the Einstein-Hilbert action, SEH , can be recovered[7, § 5.3, 16, § 8.3].
2
Since DM , DM are self-adjoint on L2 (S), suppose (φi,j ) is an basis of orthonormal eigensections with eigen-
values λi,j for DM where j ∈ [ni ] runs over the multiplicity ni of eigenvalue λ2i,j := λ2i Define the state
2 2
P P qP
Φz = 1/N (z) i:λi ≤z j∈[ni ] φji with normalization N (z) := i:λi ≤z ni then

2 X X 2 2 2
lim hΦz , e−tDM Φz iN (z) = hφi,j , e−tDM φi,j i = lim hΦz , e−tDM Φz i = Tr e−tDM
z→∞ z→∞
i:λi ≤z j∈ni

2
Therefore, the bosonic spectral action can be approximated by expectation of small time expectation of e−t(DM )
in state Φz for z large:
2 2
SEH ≈ √lim Tr e−tDM = √lim lim hΦz , e−tDM Φz iN (z)
t→0 t→0 z→∞
2
This motivates the interest in evaluating hΦz , e −tDM
Φz i. The approach we take is that of a quantum stochastic
2
dilation associated to the heat semigroup, e−tDM , which is a quantum dynamical semigroup by [9].
The issue is complicated by the fact that for the existence of Evans-Husdon flow, the generator must anni-
hilate identity, that is, the semigroup must be conservative. The Dirac laplacian acting on endomorphisms by
composition does not satisfy this. We instead have to work with the endomorphism connection and the associated
endomorphism connection laplacian and endomorphism Dirac laplacian and then extract the spinor bundle Dirac
/ 2 = △, the semigroup of
laplacian from it (see [9] more detailed discussion on this). Because of Ricci flatness, D
End
interest will be e−t △ for the endomorphism connection laplacian △End.
Remark 1.1. If the spinor bundle could be replaced by the Clifford bundle, then the associated Dirac laplacian
does generate a flow of Evans-Hudson type and the endomorphism trick is not needed.
A quantum stochastic dilation of the Evans-Hudson type1 for the quantum dynamical semigroup e−tL on the
*
C -algebra A := C(M ) ⊂ End(L2 (S)) is a family of ∗-homomorphisms jt : A → A′′ ⊗ B(Γ(R+ , k0 )) where k0 is
a Hilbert space, called the noise (or multiplicity) Hilbert space that is constructed from the generator −L, such
that the following holds
• There exists an ultra-weak dense ∗-subalgebra A0 ⊂ A such that the map-valued process Jt : A ⊗
E(L2 (R+ , k0 )) → A′′ ⊗ Γ(L2 (R+ , k0 )) with Jt (x ⊗ E(f ))u := jt (x)(uE(f )) for x ∈ A, u ∈ H := L2 (S),
f ∈ L2 (R+ , k0 ), satisfying the qsde:

dJt = Jt ◦ (aδ (dt) + a†δ (dt) + Λσ (dt) + IL (dt)), J0 = 1 (1)

on A0 ⊗E(L2 (R+ , k0 )) where δ : A0 → A⊗k0 , σ : A0 → A⊗B(k0 ) are linear maps, called the structure maps
2
for the qsde, derived from the generator L for e−tDM , A0 ⊂ Dom(L), and aδ , a†δ , Λσ , IL the fundamental
processes with respect to which stochastic integral is defined.
• For all u, v ∈ H, x ∈ A,

hvE(0), jt (x)E(0)i = hv, e−tL ui (2)

1 see appendix B for a brief review of quantum stochastic processes and integrals, and [14, 13] for detailed standard theory

2
 Therefore, hΦz , e−tL Φz i is realized as operator algebraic expectation of jt with respect vacuum state |Φz E(0)i ∈
L (S) ⊗ Γ(L2 (R+ , k0 )), where jt is obtained from the flow Jt for the Evans-Hudson qsde (equation 1). Schemes
2

for solving for Evans-Hudson qsde, for example, using Picard iterates, provide a way to algorithmically construct
the flow.

1.3 Dirac Laplacian, endomorphism connection and parallel spinors

To start we note the following sign conventions of the Laplacians. Primarily the signs are fixed so the Laplace-
Beltrami operator has non-negative spectrum, and signs on all other Laplacians cascade from there. On Rieman-
nian (M, g), M compact, without boundary, Trg denotes the trace of a covariant tensor taken after identifying
with a contravariant tensor via the metric g, Trg (h) := g ij hij . Note that trace on contravariant tensor, e.g.
vector fields, is simply the sum. For X ∈ Γ(T M ), div(X) = Tr(∇X). The Laplace-Beltrami operator is taken
as the operator with non-negative spectrum, that is, −div(∇) = − Tr(∇·,· ), where ∇·,· is the second invariant
derivative ∇2V,W := ∇V ∇W − ∇∇V W : Γ(E) → Γ(E). The connection Laplacian is ∇∗ ∇ where ∇∗ is adjoint of
the connection ∇ : Γ(E) → Γ(E) ⊗ T ∗ M . Equivalently, ∇∗ ∇ = − Tr(∇·,· ). Further, △ = −g ij ∇i ∇j .
Let ∇ be any connection on the vector bundle E → M , (M, g) a compact Riemannian P manifold. The
connection Laplacian at p ∈ M , △ = ∇∗ ∇ in local coordinates (ei ) is given by △ = −( i ∇i ∇i − ∇∇i ei ). To
evaluate △φ at any pP∈ M and φ ∈ Γ(E), we will use Riemann normal coordinates centered at p so ∇i ej vanish,
yielding △φ(p) = − i ∇i ∇i φ(p).
The endomorphism connection ∇End on the bundle End(E) = E ⊗ E ∗ associated to a connection ∇ on the
Hermitian vector bundle E over the Riemannian manifold M is such that for X ∈ T M, ∇End e X,
= ∇X ⊗ 1 + 1 ⊗ ∇
X
e ∗
where ∇ is the dual connection on E . The endomorphism Laplacian is defined as usual: at p ∈ M in Riemann
normal coordinates centered at p (denoting ∇e ∗ ∇,
e ∇e by △, ∇ again),
!
X X
△End = − ∇End
i ∇End
i =− △⊗1+2 ∇i ⊗ ∇i + 1 ⊗ △
i i

Note that as E ⊗ E ∗ is balanced over C(M ), the action of C(M ) on End(E) can be written as f · 1End =
P ∗
i (f · hi ) ⊗ hi ; this convention is used for all computation with Laplacian expressed in this tensor form. It’s
also very useful to note that in any local coordinates , ∇End acts by commutator: if over chart UP , the connection
has potential A, ∇ = d + A, then for a local orthonormal frame (µi ) and dual frame (µj ), ∇End ij σji µi ⊗ µj =
P i j
P j
ij (dσj )µi ⊗ µ + jk [σA − Aσ]jk µk ⊗ µ . In particular, since 1End is given by the identity matrix locally,
End
it follows (see [9]) that ∇ (1End ) = 0. This implies that again in normal Riemann coordinates centered at p
yields that for any f ∈ C ∞ (M ),
X X X
△End (f 1End )(p) = − ∇End
i ∇End
i (f · 1End )(p) = ∇End
i (∂i f 1End ) = − ∂i ∂i f · 1End (p) (3)
i i i

Proposition 1.2. For f ∈ C ∞ (M ), △End(E) (f · 1End ) = △M (f ) · 1End .

Proof. Let Γkij be the Christoffel symbols for Levi-Civita connection, then in local coordinates about x ∈ M ,
P P
△M (f ) = − ij g ij (x)(∂i ∂j − k Γkij ∂k )f (see, for instance, [2, pg 66]) and for the endomorphism Laplacian,
X X
△End (f · 1) = − g ij (x)(∇End
i ∇End
j − Γkij ∇End
k )(f · 1End )
ij k
 
X X
= − g ij (x)(∂i ∂j − Γkij ∂k )f  · 1End = △M (f )1
ij k

where we used ∇End (1End ) = 0, ∇End

X (f ) · 1End = X(f ) · 1End .

Now for any constant c, Tr(e−t(△+c) ) is just e−ct Tr(e−t △ ). If there exists a parallel section φ for ∇ (equiva-
lently for the connection on the dual bundle), then with Hφ = {s ⊗ φ : s ∈ Γ(E)}, and appropriate normalization
on φ,
End
Tr(e−t △ ) = Tr|Hφ (e−t △ ) (4)
End
The parallel section, therefore, allows one to extract the heat kernel trace Tr(e−t △ ) from the Tr(e−t △ ).
In the setting of the canonical spectral triple, we specialize to the spinor bundle. The existence of a par-
allel spinor constrains the holonomy of Levi-Civita connection of the manifold. In particular, this implies for
Riemannian (M, g) that the Ricci tensor vanishes[8, § 6.3], therefore, D2 = △.

3
Remark 1.3. A remark on the Lorentzian and Kahler analogs: the existence of a parallel spinor constrains
the holonomy[8], so is a strict condition. But the analysis is expected to work for Dirac operators coming from
SpinC . The extension to the complex setting, especially Kahler manifolds, should also follow easily, and these
provide an interesting class of examples. In the Lorentzian setting, there’s a richer supply of parallel spinors,
however, the essential difficulty there is that the spectrum of the Dirac operator is no longer discrete and the
regularity requirements become unclear. The spectral action principle for Lorentzian scattering space obtained by
[6] suggests the obvious question of a probabilistic interpretation in the Lorentzian setting as well. Since C(M ) is a
End
commutative algebra, so positivity is equivalent to complete positivity, therefore, by same ideas, e−t △ ⊂ C(M )
is a completely positive semigroup, and the question is reasonable.
Remark 1.4. A remark on in absence of a parallel spinor: The requirement of the parallel section can be
worked around sometimes, for example, S 1 × S 3 where S 1 carries the flat connection. By [9], there is a quantum
stochastic flow for the Laplacian for the endomorphism connection associated to the homogeneous H-connection
on the spinor bundle over S 3 . Since S 3 is also symmetric the homogeneous connection and the Levi-Civita
connection agree. However, there are no parallel spinors on S 3 , but there do exist Killing spinors with Killing
constant ±1/2. In such a setting, it’s possible to use the homogeneous space construction for quantum diffusion on
the endomorphism bundle to get at the spectral action and the heat kernel, the idea being to modify the connection
to make a Killing spinor parallel. Let △′± be the connection Laplacians, with k = ±1/2, for the connections ∇′± ,
acting by ∇′X φ = ∇X φ ± kX · φ where · is the Clifford action, and ∇ the Levi-Civita connection, then on S n ,
the Wietzenbock identity (D ± k)2 = △′± + 41 (n − 1)2 holds (this is a calculation, that works more generally than
2 2
S n ). Using that Killing spinors are parallel for △± , Tr(e−t(D±k) ) = e−t/4(n−1) Tr(e−t △± ) can be computed. It
just needs to be checked that the flow exists for △± . For S 1 × S 3 examples, the Dirac operator and its square
can be explicitly computed by specializing the Dirac operator for Robertson-Walker metrics to a constant warping
function[4].

2 The structure maps for the Laplacian generated flow

End End
From [9], the heat semigroup e−t △ is a quantum dynamical semigroup on End(E) ≡ E ⊗E ∗ with e−t △ (1) =
1 for all t. We will work with the semigroup living on C(M ) ⊂ End(E). To derive the qsde associated to the
heat semigroup, we start by computing the flow the structure matrix for the associated Evans-Hudson flow
following the standard prescription (see [14]): first we compute the kernel for the generator L = △End on the
A∞ = C ∞ (M ) acting on End(E) defined by KL : X × X → B(E ⊗ E ∗ ) for X := A∞ × A∞ , where KL for any
given any L : X → X is defined by

X × X ∋ ((f1 , f2 ), (g1 , g2 )) → L(f1∗ f2∗ g2 g1 ) + f1∗ L(f2∗ g2 )g1 − L(f1∗ f2∗ g2 )g1 − f1∗ L(f2∗ g2 g1 ) ∈ C(M )
P
For a basis of eigensections (hi ) of △, write i hi ⊗ h∗i = 1 ∈ End(E). As ∇X φ = df (X)φ + f ∇X (φ), for
fm , gm ∈ C(M ), the contribution of ∇i ⊗ ∇i term to the kernel ((f1 , f2 ), (g1 , g2 )) → L(f1∗ f2∗ g2 g1 ) + f1∗ L(f2∗ g2 )g1 −
L(f1∗ f2∗ g2 )g1 − f1∗ L(f2∗ g2 g1 ) vanishes because fm , gn are acting by multiplication on the ∇X φ term and they all
commute, while for the term df (X)φ, d(f1∗ f2∗ g2 g1 ) + f1∗ d(f2∗ g2 )g1 − f1∗ d(f2∗ g2 g1 ) − d(f1∗ f2∗ g2 )g1 vanishes as well
since

(f1∗ d(f2∗ g1 g2 ) + (df1∗ )f2∗ g1 g2 ) + f1∗ d(f2∗ g1 )g2 − f1∗ d(f2∗ g2 g1 ) − d(f1∗ f2∗ g2 )g1 (5)
= (df1∗ )f2∗ g1 g2 + f1∗ d(f2∗ g1 )g2 − ((df1∗ )(f2∗ g2 )g1 + f1∗ d(f2∗ g2 )g1 ) = 0
P
By commutativity of C(M ), 1 ⊗ △ also has no contribution, since by convention the f · 1End = i (f · hi ) ⊗ h∗i .
The only contribution to the kernel comes from the piece △ ⊗ 1. Suppose φ is an eigensection with eigenvalue λ.
Then with f ∈ C ∞ (M ), we have
X X
△f φ = − ∇ek ∇ek f φ = − ∇ek (df (ek )φ + f ∇ek φ)
k k
X
X

=− d2 f (ek , ek )φ + 2df (ek )∇ek φ + ∇ek ∇ek φ = − d2 f (ek , ek )Φ + 2df (ek )∇ek Φ + f λφ
k k

Again by commutativity of P C(M ), λf φ term does not contribute to the kernel, while a the same calculation as
equation 5 establishes
P that k 2df (ek )∇ek φ contributes
Pzero. The only contribution
P to the kernel comes from
the term − k d2 f (ek , ek )Φ. Computing the kernel for k d2 f (ek , ek ) gives −2 ( k da1 (ek )db1 (ek )) a2 b2 .

2.1 The Kolmogorov decomposition

Following Goswami-Sinha construction of the flow generator, we compute the Kolmogorov decomposition for the
kernel for L = − △End /2 (to avoid factors of −2), the data we need will be derived from the structure theory of

4
this kernel,
!
X
KL ((a1 , a2 ), (b1 , b2 )) = da1 (ek )db1 (ek ) a2 b2 (6)
k

As in [14, Thm 2.2.7], the Kolmogorov decomposition is taken to be the reproducing kernel Hilbert space

RL = LinSpan{KL (·, b)u : b ∈ C ∞ (M ), u ∈ E}

with map V : C ∞ (M ) → B(E, RL ) for x ∈ C ∞ (M ) given by

V (x) : E → RL , E ∋ u → KL (·, x)u ∈ RL

By definition, KL (·, x)u is total in RL making the decomposition minimal.

P
Remark 2.1. Comparing
P with equation 6 the term a2 b2 k da1P(ek )db1 (ek ) in local coordinates can be interpret
as the form a2 k da1 (ek )dek evaluated on the vector field b2 k db1 (ek ) ∂e∂k . So each KL (·, b)u is valued in
Γ(T M ) ⊗ L2 (M, E) and for each x = (a1 , a2 ) ∈ C ∞ (M ) × C ∞ (M ), V (x) is a 1-form acts by contracting with
Γ(T M ) component of KL (·, b)u.

2.2 The structure matrix

The structure matrix is the map Θ : A0 → B(H ⊗ (C ⊕ k0 )) for a dense subalgebra A0 ⊂ A, and Hilbert spaces
H, k0 (alternatively, a map Θ : A0 → A0 ⊗ B(C ⊕ k0 ), A0 ⊂ B(H)) given by
 
L(x) δ † (x)
Θ(x) = (7)
δ(x) σ(x)

where L : A0 → A0 is a ∗-linear map such that L(xy) − L(x)y − xL(y) = δ † (x)δ(y), δ : A0 → A0 ⊗ k0 is a

π-derivation, Θ(1) = 0. Additionally, σ(x) = π(x) − x ⊗ 1. The maps △, δ, σ satisfy these conditions iff the first
order Ito product formula holds[1, lemma 2.2].
For the structure matrix L is the semigroup generator − △End. Following [14], δ, σ are extracted from the
minimal Kolmogorov decomposition: the decomposition (RL , V ), induces the following maps on A∞ := C ∞ (M ) ⊂
C(M ) = A,

ρ : A∞ → B(RL ), ρ(x)(V (·, b)u) = V (·, xb)u

α : A∞ → B(E, RL ), α(x) = V (x, 1)

With remark 2.1 and equation

P 6 in mind, ρ(f ), f ∈ C ∞ (M ), is multiplication by f on RL while α(f ) acts by
contraction with 1-form k df (ek )dek . The representation ρ is the identity map: C ∞ (M ) is interpreted as acting
by multiplication on RL , and α is a derivation.
The construction of the structure maps proceeds as in [14, Thm 6.6.1]. To start define the Hilbert A-module
E = {α(x)y : x, y ∈ A∞ } where the closure is with respect to operator norm for B(HE , RL ). A has right action
on E by multiplication (where the norm density of A∞ ⊂ A is utilized) and the A-valued inner product is
E × E ∋ (a, b) → ha, bi = a∗ b.
Now E can be identified α(f )g ∈ E can be identified with g ⊗ df ∈ A ⊗C* W ⊂ A ⊗C* L2 (Ω1 (M )) where
W = {df : f ∈ C ∞ (M )} (note M is compact) is a closed subspace of L2 (Ω1 (M )) = W ⊕ W ⊥ (the A valued
innerproduct is pointwise tensor contraction). However, A ⊗C* L2 (Ω1 (M )) = L2 (Ω1 (M )) since A = C(M ) the
tensor product is C(M )-balanced: a ⊗ ω ∈ A ⊗C* L2 (Ω1 (M )) is simply 1 ⊗ aω.
This choice of the Hilbert space L2 (Ω1 (M )) simplifies the application of Kasparov stabilization theorem[14,
Thm 4.1.10]. Now the Hilbert-C* -module A ⊗C* L2 (Ω1 (M )) = A ⊗ L2 (Ω1 (M )) = L2 (Ω1 (M )), and by Kasparov
stabilization theorem, there’s a unitary map t′ : E ⊕ A ⊗ L2 (Ω1 (M )) → A ⊗ L2 (Ω1 (M )). In this case it can be
explicitly computed, but it turns out not to matter.
The t′ yields a unitary embedding t : E → A ⊗C* L2 (Ω1 (M )), t = t′ |E . Note that since the C* -module
A ⊗C* L2 (Ω1 (M )) has a basis, one does not need Kasparov’s stabilization theorem and an abstract embedding,
everything can be done explicitly.
Define δ(x) = t(α(x)). Now ρ induces a left action ρ̂ on E, ρ̂(x)(α(y)) = (α(xy) − α(x)y). But as α is a
ρ-derivation, ρ̂(x) = xα(y), so ρ̂(x) is multiplication by x and is again identity representation of A acting by
multiplication on sections on the bundle. Set π(x) = tρ̂(x)t∗ , again π = 1 (so the explicit form of t does not
come into play). Therefore, with equation 7 in reference,
• The multiplicity space k0 = L2 (Ω1 (M ))
• δ : A → A ⊗ k0 , δ(f ) = 1 ⊗ df .

5
 • The requirement
P that δ(x)∗ δ(y) = L(x∗ y) − x∗ L(y) − L(x)∗ y forces δ † (f ) := δ(f ∗ )∗ to act by δ(f )† (1 ⊗ ω) =
hdf, ωi = i df (ei )ω(ei ) ∈ C(M ) (using f = f ∗ ); so δ(f )† (a ⊗ ω) = ahdf, ωi works.
• σ = 0 since σ(x) = π(x) − x ⊗ 1k0 since π : A0 → A0 ⊗ B(k0 ) is identity
End
• L is the generator − △End /2 for the semigroup e−t △ /2
.
Therefore, the structure matrix for the Laplacian generated flow is the map
 
− △End(f · 1)/2 δ † (f )
A0 ∋ f → Θ(f ) = ∈ B(H ⊗ (C ⊕ k0 )) (8)
1 ⊗ df 0

3 Growth bounds and continuity

Let (M, g) be a Ricci-flat Riemannian manifold with Levi-Civita connection ∇. Since action of △End on C(M )
can be identified with the Laplace-Beltrami operator △M ≡ △ (and the lift, the rough Laplacian, to the tensor
bundle), we will work with △. Let φi be the an eigenfunction of the Laplace-Beltrami operator, △ = −g ab ∇a ∇b ,
with eigenvalue λ2i , △φi = λ2i φi . Define F as the generated by finite products and sums of φi ’s,
 
Y
F = LinSpan  φi : k ∈ N
i∈[k]
Q
F is norm-dense inside C(M ), and F1 := { i∈[k] φi : k = 1} ⊂ F forms a basis for C(M ). Set dF = {df : f ∈
F1 } ⊂ k0 as a norm-dense subspace, (k0 )∞ . The constructions from Sinha and Goswami [14], Belton and Wills
[1] both proceed by controlling the growth of the flow generator on a dense algebra, a role here played by F .
Note we are only interested in bound on the △, we can work with either △ = ±g ab ∇a ∇b to avoid tracking signs
if needed.

3.1 Growth of Laplacian iterates

For x ∈ M fix a small neighborhood of x, V , with U ⊂ V open, such that Ū ⊂ V , with coordinates (xi ) and
(∂xi ) ≡ ∂i the coordinate vector fields. For any multi-index α we denote ∂ α = ∂αk . . . ∂α1 , and same for ∇α with
∇i ≡ ∇∂i . Ricci-flatness implies for all f ∈ C ∞ (M ) (see, for instance, [5, lemma 1.36]),

[△, ∇i ](f ) = Ricki ∇k (f ) = 0 (9)

m ...
On (p, q)-tensors s, s′ there’s a natural innerproduct by contraction with g ij , gij , hs, s′ i = g i1 j1 . . . gm1 n1 s′ i1 ...
1
snj11...
...
.
The Levi-Civita connection has a lift to the tensor bundle and an associated connection Laplacian, both also de-
noted ∇, △. Denote by ∇k u the k th -covariant derivative and define the point-wise length with the innerproduct[10,
§ 2.2.1]:

ℓ(∇k u)2 = g i1 j1 . . . g ik jk (∇k u)i1 ...ik (∇k u)j1 ...jk = h∇k u, ∇k ui (10)

Lemma 3.1. For all f ∈ C ∞ (M ),

[△, ∇i ]∇jk ∇jk−1 . . . ∇j1 (f ) = 0
Proof. Using

(∇k u)i1 ...ik = ∇i1 ∇i2 . . . ∇ik (u) (11)

and that g ij commutes with △, ∇ since ∇ is metric compatible,

′ ′ ′
ℓ([△, ∇i ]∇jk ∇jk−1 . . . ∇j1 (f ))2 = g ii g j1 j1 . . . g jk jk ([△, ∇i ]∇jk . . . ∇j1 (f ))([△, ∇′i ]∇jk′ . . . ∇j1′ (f ))
′ ′ ′
= g ii ([△, ∇i ][△, ∇i′ ])(g j1 j1 . . . g jk jk ∇jk . . . ∇j1 (f )∇jk′ . . . ∇j1′ (f )) (12)
′ ′
Since h := g j1 j1 . . . g jk jk ∇jk . . . ∇j1 (f )∇jk′ . . . ∇j1′ (f ) ∈ C ∞ , from equation 9,
′
ℓ([△, ∇i ]∇jk ∇jk−1 . . . ∇j1 (f ))2 = g ii ([△, ∇i ]([△, ∇i′ ])(h) = 0.

Remark 3.2. Similar bounds can be obtained for Einstein manifolds where the Ricci tensor is proportional to
the metric. But since the analysis here needs a parallel spinor, the Evans-Hudson qsde over Einstein manifolds
are not considered.

6
 √
Proposition 3.3. For all f ∈ C ∞ (M ), k△(f )k ≤ dim M ℓ(∇2 (f )) and △k (f ) ≤ (dim M )k/2 ℓ(∇k (f ))
Proof. Using ℓ(gab )2 = g ac g bd gab gcd = g cd gcd = dim M
√
△f = g ab ∇a ∇b f = g ca g bd gcd ∇a ∇b f = hgcd , ∇a ∇b f i ≤ ℓ(gcd )ℓ(∇a ∇b f ) ≤ dim M ℓ(∇a ∇b f )

where we used Cauchy-Schwarz. The bound (dim M )k/2 ℓ(∇k (f )) follows identically using ℓ(gi1 j1 . . . gik jk )2 =
(dim M )k .
The growth of Laplacian and its powers is clear on φi ’s, since △k (φi ) = λ2k i φi . Controlling the growth on
products of φi ’s will require control over
q
h∇k φi , ∇φkj i(x) ≤ k∇k φi (x)k2 k∇k φj (x)k = ℓ(∇k φi (x))ℓ(∇k φj (x)) (13)

(using eq 10 with k = 1 and Cauchy-Schwartz inequality). To see this note how Laplacian and its iterated powers
act on products of φi ’s.
Proposition 3.4. [Laplacian on products] For φj , φi ,
i △(φi φj ) = φi △(φj ) + φi △(φj ) + 2h∇φi , ∇φj i
ii For any k,

△h∇k φi , ∇k φj i = h∇k △φi , ∇k φj i + h∇k φi , ∇k △φj i + 2h∇k+1 φi , ∇k+1 φj i (14)

Proof. These are straightforward computations, see appendix A for details.

Therefore, it’s sufficient to bound ∇k (φj ) ∞
. For φj with λj ≥ 1, the bound follows from [3, lemma 2.7]:
Theorem 3.5 (Bin [3]). With λ ≥ 1,
X
ℓ(∇k φj (x))2 ≤ CM,x λ2k+dim M−1 (15)
λj ∈[λ,λ+1]

The following corollary is immediate since M is compact and CM = (λj + 1)dim M−1 supx∈M CM,x and
Kj = (λj + 1)2 can be used.
Corollary 3.6. For M compact, for any φj with λj ≥ 1, ℓ(∇k φj (x))2 ≤ CM Kjk .

One expects that ∇k φj L2 (M) should decay to zero for j with λ2j < 1, and since φj ’s are smooth this is
enough to establish a uniform bound. However, this will need to be leveraged locally and the boundary for the
local chart will need to be taken into account. Recall the integration on parts formula for tensor fields[12, pg 50,
149] when M does have a boundary,
Z Z Z
h∇F, GidVg = hF ⊗ N ♭ , GidVĝ − hF, div(G)idVg (16)
M ∂M M

where ĝ is the induced metric on ∂M, dVg , dVĝ the associated volume forms, ·♭ the musical isomorphism, N the
outward unit normal at ∂M , and F, G tensor fields, div(G) = Trg (∇G), the trace being over the last two indices.
Note if G = ∇H then, −div(G) = △(H).
Proposition 3.7. For u = φj , with j such that λ2j < 1, ∇k u ∞
≤ ∇k−1 u ∞

Proof. Suppose for some x ∈ M , h∇k u(x), ∇k u(x)i − h∇k−1 h(x), ∇k−1 u(x)i > 0. Then since u is smooth, there
exists an open neighborhood U of x such that on U , h∇k u, ∇k ui − h∇k−1 u, ∇k−1 ui > 0. Now let ψ be such that
supp(ψ) ⊂ U is compact, ψ ≥ 0 on U and ψ > 0 on open V ⊂ U , then
Z Z
ψh∇k u, ∇k ui − ψh∇k−1 u, ∇k−1 uidVg = hψ∇k u, ∇k ui − ψh∇k−1 u, ∇k−1 uidVg > 0 (17)
M U

Now hψ∇k u, ∇k ui = h∇(ψ∇k−1 u), ∇k ui − h∇ψ · ∇k−1 u, ∇k ui, and for the first term
Z Z Z
h∇(ψ∇k−1 u), ∇k uidVg = h·, ·idVĝ + hψ∇k−1 u, −div(∇k u)idVg
U ∂U U

7
 R
where ∂U h·, ·idVĝ = 0 since ψ = 0 on ∂U , while −div(∇k u) = △∇k−1 u = ∇k−1 △u using M is Ricci flat.
Therefore, we have
Z Z Z
ψh∇k u, ∇k ui = ψh∇k−1 u, ∇k−1 △ui − h∇ψ · ∇k−1 u, ∇k uidVg
M U U
Z Z
2 k−1 k−1
=λ ψh∇ u, ∇ ui − h∇ψ · ∇k−1 u, ∇k uidVg
U supp(∇ψ)

This yields
Z
0< ψh∇k u, ∇k ui − ψh∇k−1 u, ∇k−1 uidVg (18)
U
Z Z Z
= λ2 ψh∇k−1 u, ∇k−1 ui − h∇ψ · ∇k−1 u, ∇k uidVg − ψh∇k−1 u, ∇k−1 uidVg
U supp(∇ψ) U
Z Z
2
= ψ(λ2 − 1) ∇k−1 u dVg − h∇ψ · ∇k−1 u, ∇k uidVg (19)
U supp(∇ψ)
R
Define the linear functional ω(ψ) := supp(∇ψ)
h∇ψ · ∇k−1 u, ∇k uidVg . Note


h∇ψ · ∇k−1 u, ∇k ui = g i1 j1 ∇i1 ψ g i2 j2 . . . g ik jk (∇k−1 u)i2 ...ik (∇k u)j1 j2 ...jk


= g i1 j1 ∇i1 ψ g i2 j2 . . . g ik jk (∇i2 . . . ∇ik u)(∇j1 ∇j2 . . . ∇jk u)
= 12 g i1 j1 ∇i1 ψ∇j1 (ℓ(∇k−1 u, ∇k−1 u)2 ) = 21 h∇ψ, ∇ℓ(∇k−1 u, ∇k−1 u)2 i (20)

By showing that there exists a ψ that makes ω(ψ) ≥ 0, since (λ2 − 1) < 0, from equation 19 it will follow that
equation 18 cannot hold.
Assume that U is small enough to be covered by a geodesic normal coordinates, and consider polar coordinates
on U centered at x. Define τsc on U for c, s ∈ R>0 such that τs (x) = c and then decays linearly in radially
outwards direction with slope −s to 0 at ∂BR (x) with c, s such that supp(τsc ) ⊂ U , R depending on c, s. Then
τsc is continuous, piecewise continuously differentiable, with compact support in U , so weakly-differentiable, and
∇τsc = −s1BR (x) (there’s enough slack to work with mollified versions of τ ’s, but weak-differentiablility suffices
for simplicity). If for some τsc , ω(τsc ) ≥ 0 then that ψ = τsc is the required ψ.
If not, then ω(τsc ) < 0 for all c small enough to have support in U . By rescaling wlog assume c = s = 1, and
set τ1 := τ11 ( otherwise the constants are messy). For such τ1 , define τ1′ such that τ1′ (x) = 0, and τ1′ increases
linearly to 1 at ∂B1 (x), and outside of B1 (x), τ1′ = 0. Then ω(τ1′ ) = −ω(τ1 ) = δ > 0 since ∇τ1′ = −∇τ1 on
supp(∇τ1′ ) = supp(∇τ1 ). It remains to make τ1′ continuous without changing ω(τ1′ ) too much. For this set
′′
τ1,r = τ1′ on B1 (x), τ1,r
′′
= 0 on B1+r (x)c , and on B1+r (x)c \ B1 (x), τ1,r
′′
decays linearly to 0 on ∂B1+r (x). Finally,
′′
since for all r > 0 small enough, τ1,r is piecewise continuous, continuously differentiable and compactly supported
in U , it remains to check for any ǫ > 0 there exists rǫ > 0 such that for all r < rǫ , ω(τs′ ) − ω(τs,r ′′
) ≤ ǫ. Note
that
Z
2 ω(τ1′ ) − ω(τ1,r
′′
) = hτs′′ , −div(∇(ℓ(∇k−1 (u))2 ))idVg (21)
B1+r (x)(x)\B1 (x)

using equation 20 and so ω(τ1′ ) − ω(τ1,r

′′
) ≤ ǫ for small r because −div(∇(ℓ(∇k−1 (u))2 )) = △(ℓ(∇k−1 (u))2 ),
′′
τs,r being continuous (with M compact) are bounded.
The idea above generalizes to all φi ’s giving a simple proof for growth bounds on covariant derivatives from
[3] in the Ricci flat setting.
Corollary 3.8. For u = φj , with λ2 := λ2j , ∇k u ∞
≤ λ2 ∇k−1 u ∞

Proof. Assume not, then on some open U ⊂ M , for all x ∈ U , for some c > 1,

h∇k u(x), ∇k u(x)i − cλ2 h∇k−1 h(x), ∇k−1 u(x)i > 0

and as in proposition 3.7 for some ψ ≥ 0 compactly supported in U, ψ > 0 on an open set,
Z Z
hψ∇k u(x), ∇k u(x)idVg − cλ2 ψh∇k−1 h(x), ∇k−1 u(x)idVg > 0 (22)
U U
Z Z Z
k k k−1 k
with hψ∇ u(x), ∇ u(x)idVg = h∇(ψ∇ u(x)), ∇ u(x)idVg − h∇ψ · ∇k−1 u, ∇k uidVg
U U U

8
Therefore, equation 22 becomes
Z Z Z
λ2 hψ∇k−1 u(x), ∇k−1 u(x)idVg − h∇ψ · ∇k−1 u, ∇k uidVg − cλ2 ψh∇k−1 h(x), ∇k−1 u(x)idVg
U U U
Z Z
2 k−1 k−1
= λ (1 − c) hψ∇ u(x), ∇ u(x)idVg − h∇ψ · ∇k−1 u, ∇k uidVg > 0
U U

But choosing ψ as in 3.7, since λ2 (1 − c) < 0, the last inequality cannot hold.
Therefore, ∇k φi ∞ is uniformly bounded for all φi . Collecting this with propositions 3.3 and 3.4 along with
corollary 3.6, yields easily implies the following bounds.
Theorem 3.9. For all k ∈ N, φ ∈ F ,
i There exists constants Cφ , Kφ such that △k (φ) ≤ Cφ Kφk

ii For any k, N ∈ N, the map △k ⊗ 1MatN : MatN (F ) → MatN (C ∞ (M )), △k ⊗ 1MatN (ξ) ≤ Cξ Kξk for
ξ = [ξij ] ∈ MatN (F ). The same holds for (△ ⊗ 1MatN )k .
Proof. Part i) follows from estimates on △k , how △k acts on products of φi ’s and ∇ being a derivation. For part
ii), by growth bounds on △k , △k (ξij ) ≤ Cij Mijk for some constants Cij , Mij . so
k 
 
X X 
△k ⊗ 1MatN (ξ) = [△k (ξij )] ≤ max  Cij ·  Mij   . (23)
i
j j

Note that theorem 3.9, ii) is needed to guarantee that the map-valued process Jt evaluated on exponential vectors
is pointwise bounded.

4 Picard iterates for the Laplacian generated flow

We now specialize the map-valued Evans-Hudson qsde’s to the Laplacian flow generator, Θ: from equation 8,
 
− △(f )/2 δ † (f )
A∞ ⊃ F ∋ f → Θ(f ) = ∈ B(H ⊗ (C ⊕ k0 )) (24)
1 ⊗ df 0

where k0 = L2 (Ω1 (M )), the noise space, and H = L2 (S), the square integrable spinors on which C(M ) acts. Recall
(k0 )∞ := dF = {df : f ∈ F } ⊂ L2 (Ω1 (M ). k = L2 (R+ , k0 ) For f ∈ C ⊕ k0 . For f ∈ k0 , set fˆ := 1 ⊕ f ∈ C ⊕ k0
and identify f with 0 ⊕ f ∈ C ⊕ k0
Towards establishing the existence of the quantum flow of Evans-Hudson type under the much weaker reg-
ularity assumption and without the Frechet space scaffolding from [14, Thm 8.1.38]; the usual ideas from the
literature can be leveraged, however, crucial estimates need to be made from ground up which we establish next.
Only a sketch of known results into which these estimates are plugged are included how they made fit into the
scheme. A crucial point is an extension of the squareroot trick, which is needed since in general the algebra
F , the algebra consisting of finite products and sums of eigenfunctions of △, is not closed under squareroots
since derivatives of squareroots grow factorially while on F the covariant derivatives satisfy growth bounds of
type ℓ(∇k f ) ≤ Cf Mfk . For background on map-valued qsdes, we refer to the included appendix B. To start,
recall the following estimates for map-valued processes, aδ , a†δ , IL being the fundamental processes (appendix
equations 40,41):
Estimate 4.1. [14, Thm 5.4.7, 8.1.37] For a map-valued integrable process Ys ,
Z t 2 Z t
2
Ys ◦ (aδ + IL )(ds)(x ⊗ E(f ))u ≤ et kYs (L(x) + hδ(x∗ ), f (s)i) ⊗ E(f ))uk ds (25)
0 0
Z t 2 Z t 2

Ys ◦ (a†δ )(ds)(x ⊗ E(f ))u ≤ et f
Ys (δ(x)E(f ))u + kYs (hf (s), δ(x)i)E(f )uk2 ds (26)
0 0

There’s the following characterization for an integrable map-valued process generated by the structure maps
L = − △/2, δ from Θ through the Picard iteration scheme, the convergence of which will yields the solution to
the qsde needed.

9
Lemma 4.2. [14, lemma 8.1.37] Let V = {(dF -valued simple functions}, E(V) the exponential vectors, and
J (0) : F ⊗ E(V) → A ⊗ Γ(k0 ) be the identity map, then with J (0) = 1,
Z t
,J (n+1)
(t) = J (n) (s) ◦ (a†δ + aδ + IL )(ds), J (n+1) : F ⊗ E(V) → A ⊗ Γ(k0 ) (27)
0

each J n is an a map-valued integrable process (by definition linear, but not necessarily completely smooth),
Additionally, the following estimates hold ,
Z 2 Z 2
!
2 t t
(n+1)
Jt (x ⊗ E(f ))u ≤2 Js(n) ◦ (aδ + IL )(ds)(x ⊗ E(f ))u + Js(n) ◦ (a†δ )(ds)(x ⊗ E(f ))u (28)
0 0

Proof. The continuity requirements for existence of the integral are satisfied since for each fixed E(f ) and a, the
maps δ, △ are bounded. The inequalities follow from standard theory. The Laplacian flow generator satisfies
much weaker assumptions, therefore the resultant process is not completely smooth.
P (n)
The Picard iterates defined by SN (t) = n≤N Jt (x ⊗ E(f )) can be shown to converge on the exponential
vectors following same scheme as [14, Thm 8.1.38] after plugging in the following estimates which need to
be obtained differently as Θ has much less regularity. To motivate the estimates we sketch the convergence
arguments.
The first term in r.h.s.for equation 28, using the definition of map-valued integrals (see appendix B. 42) can
be recursively expanded using via estimate 4.1, inequality 25:
Z t 2
Z t Z s 2
(n−1)
et Js(n) (L(x) + hδ(x∗ ), f (s)i) ⊗ E(f ))u ds = et Js1 ◦ (aδ + IL )(ds1 )(x ⊗ E(f ))u ds
0 0 0

Since f is simple, all terms depending on f in above can be uniformly bound by a constant Bf , and recursively
applying inequality 25 to the r.h.s., till reaching J (0) = 1 yields:
2
Z t Z s Z s1
2
J (n+1) (t)(a ⊗ Ef ) ≤ (2et B)n kE(f )k . . . Φf(s)
ˆ (Φfˆ(s1 ) . . . Φfˆ(sn−1 ) (x)) dsds1 . . . dsn−1
0 0 0
(29)

where Φf (s) (x) := L(x) + hδ(x∗ ), f (s)i. Since f is simple, Range[f ] is finite: for each s ∈ [0, t], f (s) ∈ {dζi ∈
dF , i ∈ [r], r ∈ N} ≡ Range[f ]. This means one must control

Φξin (Φξin−1 (. . . Φξi1 (x))) (30)

where each ξik ∈ Range[f ]. Similarly, for the second term in r.h.s.for equation 28, from equation 26 (also, see
appendix equation 44), we have
Z t 2 Z t 2 2

Js(n) ◦ (a†δ )(ds)(x ⊗ E(f ))u ≤ et Jg
(n) (δ(x) ⊗ E(f ))u + J (n) (hf (s), δ(x)i)E(f )u ds
0 0

The kJ (n) (hf (s), δ(x)i)E(f )uk2 term is controlled exactly as equation 29 via a corresponding estimate on nested
Φ′f (s) (x) := hf (s), δ(x)i. The kJg (n) (δ(x) ⊗ E(f ))uk2 term lives on the Fock spaces it can be controlled by

same recursive expansion using Φ′′f (s) (x) := δ(x) ⊗ E(f (s)) since by definition (see appendix equation 43 and
remark B.1), integral with respect to a†δ is given by –
Z t  Z t 
J (n)
(s) ◦ (a†δ )(ds) (x ⊗ E(f ))u = J^
n (s)(δ(x) ⊗ E(f ))u ds
s
0 0
Z t Z s  
= J (n−1) (s) ◦ (a†δ )(ds1 ) (δ(x) ⊗ E(fs )) uds
0 0
Z t Z s 
= J^
(n−1) ((δ ⊗ 1)(δ(x) ⊗ E(f )))u ds ds
s 1
0 0

so one needs to bound both of the following

Φ′ξin (Φ′ξi (. . . Φ′ξi1 (x))) , kδ n (x)k (31)

n−1

10
Estimate 4.3. There exist constants, Cf,x , Mf,x such that

n
i) Φξin (Φξin−1 (. . . Φξi1 (x))) ≤ Cf,x Mf,x ii) Φ′ξin (Φ′ξin−1 (. . . Φ′ξi1 (x))) ≤ Cf,x Mf,x
n
iii) kδ m (x)k ≤ Cf,x Mf,x
n

That is, three terms (equations 30,31) satisfy bound of form CK n for C, K depending on the simple function
f and independent from n (note the constants for Φ, Φ′ , Φ′′ may be different, but taking the maximum, yields a
single pair that works).
Proof. i) Define an := Φξin (Φξin−1 (. . . Φξi1 (x))). Since each ξik ∈ dF , let ξik = dzk . Now Φzk (a′ ) = △′ a′ +
hda′ , dzk i = △′ a′ + h∇a′ , ∇zk i where △′ := − △/2. Therefore,
 
′ 1 in jn in jn
kan k = Φzn (an−1 ) = △ an−1 + h∇an−1 , ∇zn i = gin jn − ∇ ∇ an−1 + ∇ an−1 ∇ zn
2
  
1 in jn 1 in−1 jn−1 in−1 jn−1
= gin jn gin−1 jn−1 − ∇ ∇ − ∇ ∇ an−2 + ∇ an−2 ∇ zn−1
2 2
   
in 1 in−1 jn−1 in−1 jn−1 jn
+ gin jn gin−1 jn−1 ∇ − ∇ ∇ an−2 + ∇ an−2 ∇ zn−1 ∇ zn
2
Therefore, each recursive expansion of Φzk doubles the number of terms. After n recursive expansion there are 2n
terms, consisting of n copies of the metric, gin jn gin−1 jn−1 . . . g11 j1 , contracting against the covariant derivatives
acting on x, z1 . . . zn . Note that zi ’s appear in increasing order left to right by the recurence structure, and x is
at the left most, making appearance only at the last step. Let Pn+1 [I] be a partition of set I := {ik , jk : k ∈ [n]}
into n + 1 ordered sets Pn+1 = {Ai ⊂ I : i ∈ [n + 1]}, with some Ai ’s possibly empty. For A ∈ Pn+1 [I] with
∇A , A = [a1 , a2 . . . a|A| ] denote the operator ∇a1 ∇a2 . . . ∇a|A| ,and by ∇A the corresponding operator with indices
lowered. Then after n recursive expansions, each term corresponds to some partition Pn+1 [I] = {Ai : i ∈ [n + 1]}.
Writing the tensor contraction as an explicit innerproduct,

gin jn gin−1 jn−1 . . . g11 j1 ∇A1 x∇A2 z1 ∇A3 z2 . . . ∇An+1 zn = hgin jn gin−1 jn−1 , ∇A1 x∇A2 z1 ∇A3 z2 . . . ∇An+1 zn i

where with empty Ai meaning the term is missing. Applying Cauchy-Schwarz, and dropping the −1/2 factors
as we are only interested in an upper bound on each term,
q
hgin jn gin−1 jn−1 , ∇A1 x∇A2 z1 ∇A3 z2 . . . ∇An+1 zn i ≤ (dim M )n/2 h∇A1 x∇A2 z1 . . . ∇An+1 zn , ∇A1 x∇A2 z1 ∇An+1 zn i
q
≤ (dim M )n/2 h∇A1 x, ∇A1 xi . . . h∇An+1 zn , ∇An+1 zn i
≤ (dim M )n/2 ℓ(∇|A1 | x)ℓ(∇|A2 | )z1 . . . ℓ(∇|An+1 | zn )

For each φ ∈ {x, zn } ≡ Tx,f , there are constants Cφ , Mφ such that ℓ(∇k φ) ≤ Cφ Mφk . Wlog, assume each
Cφ = 1, by absorbing it into Mφ . NowPsince each dzn ∈ Range[f ], |Range[f ]|= r, the choice M = maxφ∈Tx,f Mφ
is independent of n, and using that i∈[n+1] |Ai |= 2n, yields

hgin jn gin−1 jn−1 , ∇A1 x∇A2 z1 ∇A3 z2 . . . ∇An+1 zn i ≤ ((dim M )1/2 M 2 )n

And finally accounting 2n terms with above inequality holding for each, yields the needed bound

kan k ≤ (2(dim M )1/2 M 2 )n

ii) This follows identically to part i).

iii) Since δ : A → A ⊗ k0 , x → 1 ⊗ dx, therefore δ 2 (x) = (δ ⊗ 1)(δ(x)) = (δ ⊗ 1)(1 ⊗ dx) = 0. So the bound is
simply
sup kE(fs )k kdxk
s∈[0,t]

With these estimates available, the following results are easily obtained exactly as in [14].
Theorem 4.4. Suppose a ∈ F , and f ∈ U where U is the set of simple function taking values in (k0 )F
∞ such that
ft ∈ {ξi : i ∈ [r]} for all t. Define
Z t
(0)
J = 1, J (n+1)
= J (n) (s) ◦ (a†δ + aδ + IL )(ds) (32)
0

where δ, L from the structure matrix Θ then

11
 P (n)
1. J(t) = Jt (a ⊗ E(f )) converges
n
Rt
2. J(t) = 1 + 0 J(s)(a†δ + aδ + IL )(ds) holds
3. For all a1 , a2 ∈ F , f1 , f2 simple and (k0 )F
∞ -valued, v1 , v2 ∈ H,

hJt (a1 ⊗ Ef1 )v1 , Jt (a2 ⊗ Ef2 )v2 i = hv1 ⊗ Ef1 , Jt (a∗1 a2 ⊗ Ef2 )v2 i
J(t)(1 ⊗ Ef1 )v1 = v1 Ef1

P (n) (n)
Proof. From the estimate 4.3, SN (t) = n≤N Jt (a ⊗ E(f )) is convergent since for each Jt (a ⊗ E(f )),

(n)
Jt (a ⊗ E(f )) ≤ Cf,x (2et BMf,x )n /n!

This yields part 1. Part 2. is by the definition of the stochastic integral, and 3. is standard theory as in [14,
Thm 8.1.38].
It remains to check that the integral extends from E(U) to Γ(k0 ). For this we need to extend the squareroot
trick using regularity of the flow generator.

4.1 The extended squareroot trick

It follows from quantum Ito formulae that for u, v ∈ H, h, f ∈ V ⊂ L2 (R+ , k0 ) ≡ k,

hJt (a ⊗ E(f ))u, Jt (b ⊗ E(h))vi = huE(f ), Jt (a∗ b ⊗ E(h))vi (33)

Jt (1 ⊗ E(f ))u = uE(f ) (34)

Define
(n)
jtn (a)(vEf ) := Jt (a ⊗ Ef )v
so jtn is unital with the factorization property (equation 33), and jtn (a) is a linear operator on a dense subspace
K := H ⊗ E(V) ⊂ H ⊗ Γ(k). For any v, f , jtn is bounded pointwise on F .
Proposition 4.5. For all a ∈ F there exists K such that

kΘ(a)k ≤ K kakW 2,∞

where k·kW 2,∞ is the Sobolev norm kakW 2,∞ = kak + k∇ak + ∇2 (a)
Proof. For a, ψ ∈ F , f ′ ∈ dF with f ′ = df, f ∈ F and w ∈ C,
    
− △(a)/2 h·, dai ψ⊗w −(△(a)/2)ψ ⊗ w + ψ ⊗ hdf, dai
= (35)
1 ⊗ da 0 ψ ⊗ df ψ ⊗ w ⊗ da

Therefore, the inequality kΘ(a)k ≤ 4 dim(M ) kakW 2,∞ follows, since

2 2 2 2 2 2 2
kΘ(a)(ψ ⊗ (w ⊕ df ))k ≤ k△(a)/2k kψ ⊗ wk + kdak kψ ⊗ df k + kdak kψ ⊗ wk

(n−1)
Now F ⊂ W k,p (M ) since F ⊂ C ∞ (M ), M compact for all p, k. Because kΘ(a)k ≤ K kakW 2,∞ , if Jt is
(n)
bounded for each t, then Jt is continuous on F wit respect to W 2,2 -norm topology. We will use this to show
(n−1) (n)
that if Jt is bounded, then jt is positive on K. Then using jtn is positive on K it will be checked that for
(0)
every a ∈ F , jtn (a) ∈ B(K) and that it extends from B(K) to B(H ⊗ Γ(k)). The base case is jt = 1 ∈ B(K)
n n
which is obviously positive. Then from jt : F → B(H ⊗ Γ(k)), it extends to jt : C(M ) → B(H ⊗ Γ(k)).
(n−1)
Lemma 4.6. Suppose Jt ∈ B(K), then jtn is a positive map on a ∈ F , a > 0.
√
Proof. Suppose a ∈ F is positive. We want to show jtn (a) is positive as well. If a ∈ F , then
√ √
hu, jtn (a)ui = hjtn ( a)u, jtn ( a)ui ≥ 0 (36)

for every u ∈ K,√hence jt (a) is positive.

So assume a 6∈ F where a is positive,
√ a(m) > 0, m ∈ M . Since LinSpan(F ) is dense in C(M ), for any
ǫ > 0, there exists f ∈ F such that k a − f k ≤ O(ǫ) meaning a − f 2 ≤ O(ǫ). Additionally f can be chosen
so a − f 2 W 2,∞ ≤ O(ǫ), so f 2 approximates a in Sobolev W 2,∞ -norm as well.

12
 √ √
To √see why this is possible note that since a > 0, a ∈ C ∞ (M ), therefore, a ∈ L2 (M ), additionally for all
k, ∇k ( a) ∈ L2 (M ), with
√ X X √ √
a= αi φi and λ2k k k
i αi = h∇ ( a), ∇ ( a)iL2 (M) < ∞
i i
√ Pn √
where αi = hφi , ai and Ricci-flatness was used. So the sequence ( i=1 αi φi )n∈N converging to a in L2 (M )
is a bounded in every W k,2 . For sufficiently
P large k, the embedding W k,2 (M ) ⊂ W 2,2 (M ) is compact by
Rellich–Kondrachov theorem, that is, (P ni=1 αi φi )n has a Cauchy, and so a convergent subsequence; wlog let this
n
subsequence be denoted by the same i=1 αi φi := an . For (ai ) to be convergent in W 2,2 (M ), kai kL2 (M) must
√ P
vanish, so the only possible limit is P a. Now suppose P the tail ∞ i=n αi φi does not vanish in W
2,∞
. This means for
k k
some xP ∈ M for some
P k ∈ {0, 1, 2}, h i αi ∇ φi , i α i ∇ φi i(x) > 0. But then by the following
√ argument shows
that h i αi ∇k φi , i αi ∇k φi iL2 (U) > 0 contradicting the convergence in W 2,2 (M ). So a can be approximated
arbitrarily well in W 2,∞ (M ).
P P P
ClaimP 4.7. Suppose
P a = i αi φi ∈ C ∞ (M ), then h∇k i αi φi , ∇k i αi φi i(x) > 0 for some x ∈ M implies
h∇k i αi φi , ∇k i αi φi iL2 (M) > 0. In particular, this holds for k = 0.

Proof. As before in proposition 3.7 by smoothness of a, this holds for all x ∈ U for some open set U , and with
ψ ≥ 0 compactly supported on U , ψ > 0 on an open V ⊂ U ,
Z Z X X
k k k k
sup ψ(x) · h∇ a, ∇ aiL2 (M) = sup ψ(x) · h∇ a, ∇ aidVg ≥ hψ∇k αi φi , ∇k αi φi idVg > 0
x∈U x∈U M U i i

Since ψ is compactly supported and bounded, the claim follows. The k = 0, ∇k = 1 specialization is identical.
Now define
Wa = {a} ∪ {f 2 : f ∈ F with a − f2 W 2,∞
≤ 1/n, n ∈ N}
(n−1)
then as kΘ(a′ )k ≤ K ka′ kW 2,∞ and Jt is bounded on K by hypothesis, the bound in lemma 4.2, implies norm
is continuous map with respect to k·kW 2,∞ -topology on Wa :

k·k : Wa → R, a′ → kjt (a′ )k

If jtn (a) is not positive, then there exists u ∈ K such that hu, jtn (a)ui < 0. Since norm is continuous, the
map a′ → hu, jtn (a′ )ui is also continuous on Wa : by Cauchy-Schwartz inequality, hu, jtn (a′ )ui ≤ kuk kjtn (a′ )uk ≤
2
kuk K ′ K ka′ kW 2,∞ where K ′ depends on u which we fixed and Jtn−1 . This continuity means hu, jtn (·)ui < 0
on some neighborhood containing a in Wa . However, for any neighborhood U of a in Wa , w ∈ U, w 6= a implies
w = f 2 , f ∈ F , so hu, jtn (f 2 )ui ≥ 0 by equation 36. Therefore, jtn (a) must be positive.
2
Lemma 4.8. If jtn is a positive map on positive a, a > 0, then kjtn (a)k ≤ kak
p
Proof. Let x ∈ F so (1 + ǫ) kxk − x ∈ F and positive for any ǫ > 0. Define Φǫ (x) := (1 + ǫ) kxk 1 − x ∈ C(M ).
Approximate Φǫ (x) from below by z ∈ F . Then jtn (Φǫ (x)2 − z 2 ) > 0 because Φǫ (x)2 − z 2 > 0 and jtn is positive.
This yields hθ, (jtn (Φǫ (x)2 ) − jtn (z 2 ))θi ≥ 0 and we have

hθ, jtn (Φǫ (x)2 )θi ≥ hθ, jtn (z 2 )θi ≥ 0

Now the usual squareroot trick takes over: since jtn is unital,
(n)
0 ≤ kjtn (z)θk2 = hθ, jtn (z 2 )θi ≤ hθ, jt ((1 + ǫ) kxk 1 − x)θi (37)
2
hθ, jtn (x)θi ≤ hθ, jtn ((1 + ǫ) kxk 1)θi ≤ (1 + ǫ) kxk hθ, jtn (1)θi = (1 + ǫ) kxk kθk (38)
2
Since ǫ was arbitrary, hθ, jtn (x)θi ≤ kxk kθk . Finally,

kjtn (x)θk2 = hjtn (x)θ, jtn (x)θi = hθ, jtn (x∗ x)θi ≤ kx∗ xk kθk2 = kxk2 kθk2 (39)
2
. So kjtn (x)k ≤ kxk , and the bound on kjtn k is uniform.
Now from density
P of F , V and K, each jtn extends from a map jtn : F → B(K) to P
jtn : A → B(H ⊗ Γ(k)).
Since SN (t) = n∈[N ] J n converges, so does S = limN →∞ SN , and therefore limn→∞ jtn is the needed flow.
Precisely, we have the following result:
Theorem 4.9. Following notation from theorem 4.4 from define jt (a)(v1 Ef1 ) := Jt (a ⊗ Ef1 )v1 , then

13
 1. jt : F → B(H ⊗ E(U)) is a unital ∗-homomorphism
2. jt extends to jt : F → B(H ⊗ Γ(k0 ))
3. jt extends to jt : A → B(H ⊗ Γ(k0 ))
Remark 4.10. A remark on Sinha and Goswami [14] Frechet structures: Proposition 4.5, along with the growth
bounds on ℓ(∇k φ) suggests that convergence of the stochastic integrals can approached via a variant of complete
smoothness regularity condition introduced by [14]. In absence of the group action, the Frechet space structure
on k0 has to be obtained by other methods – in this setting ∇k is the natural candidate for defining the Sobolev
norms on dF ⊂ k0 . Some work is required to strap that into the Frechet machinery, since [14] require norms to
come from derivations on A = C(M ), but covariant derivative is a derivation on the tensor bundle. Note that the
usual Sobolev norms k·kW k,∞ can be used to obtain estimates of the type needed: for instance for estimate 4.3,
n2
i), using that kabkW p,∞ ≤ kakW p,∞ kbkW p,∞ yields a slightly weaker bound of form Mf,x which is still enough to
get convergence.
Remark 4.11. A remark on Belton and Wills [1] construction: The polynomial bounded growth condition that
is leveraged for the Laplacian is similar to one obtained by [1]. However, this growth cannot be bound on the
entire Hilbert space as operator Θ(a) is only defined on C ∞ (M ), and it becomes necessary to work with a dense
subspace and utilize some mild continuity to push the necessary estimates through. The continuity is also needed
for the extension of the squareroot trick.

A Laplacian on products
Proposition A.1 (Laplacian on products). For φj , φi ,
i △(φi φj ) = φi △(φj ) + φi △(φj ) + 2h∇φi , ∇φj i
ii For any k,

△h∇k φi , ∇k φj i = h∇k △φi , ∇k φj i + h∇k φi , ∇k △φj i + 2h∇k+1 φi , ∇k+1 φj i

Proof. With △ = g ab ∇a ∇b , M Ricci flat, using g ij = g ji , ∇a (φi φj ) = φj ∇a (φi ) + ∇a (φj )φi yields:

△(φi φj ) = g ab ∇a ∇b (φi φj ) = g ab (∇a ∇b (φi φj )) = g ab (∇a (φi ∇b (φj ) + φj ∇b (φi )))

= f △(φj ) + φj △(φi ) + 2g ab ∇a (φi )∇b (φj ) = f △(φj ) + φj △(φi ) + 2h∇φi , ∇φj i

For △ = −g ab ∇a ∇b , we get −2h∇φi , ∇φj i instead for the mixed term.

To see ii), note that by i) and for k-tensors f, h with multi-indices a, b, g ab := g a1 b1 . . . g ak bk , we have

△hf, hi = △g ab fa hb = g ab △fa hb = g ab fa △hb + g ab (△fa )hb + 2g ab g cd ∇c fa ∇d hb

= h△f, hi + hf, △hi + 2h∇f, ∇hi

and then since[∇, △] = 0 as M is Ricci-flat, as needed

△h∇k φi , ∇k φj i = h∇k △φi , ∇k φj i + h∇k φi , ∇k △φj i + h∇k+1 φi , ∇k+1 φj i

B Appendix: Map-valued Evans-Hudson quantum sde’s

Suppose A0 is a dense ∗-subalgebra inside the C * -algebra A ⊂ B(H), k0 is the noise space with k := L2 (R+ , k0 ), kt :=
L2 ((0, t), k0 ), k t := L2 ((t, ∞), k0 ). Γ = Γs (k). Γt := Γ(L2 ((0, t), k0 ), Γt := Γ(L2 ((t, ∞), k0 ). Similarly, f ∈
k, f t , ft are projections onto k t , kt , ft := f 1[0,t] , f t : f 1[t,∞) . k̂0 := C ⊕ K0 .
We will only work with annihilation and creation processes since the conservation process in the Laplacian
generated processes is zero. For a map Lin(D0 , H ⊗ k0 ), D0 ⊂dense H, if R(u) := a ⊗ b, then for ∆ ⊂ (t, ∞), using
the mapping k0 ∋ b → b1∆ ∈ kt , Rt∆ is defined by

D0 ⊗ Γt ∋ u ⊗ ψ → Rt∆ (u ⊗ ψ) := a ⊗ ψ ⊗ (b1∆ ) ∈ H ⊗ Γt ⊗k t

The associated creation process creates k0 component of R on interval ∆:

a†R (∆) := a† (Rt∆ )

14
The corresponding annihilation process is defined by using k0 component of R to annihilate:
Z  
(D0 ⊗ Γt ) ⊗ Γt ∋ ut E(f t ) → aR (δ)(ut E(f t )) = hR, f (s)ids ut E(f t )
∆

where hR, f (s)i is viewed as an operator() on H.

The Hudson-Parthasarathy quantum stochastic calculus on Hilbert spaces are set in the the Schröedinger
picture of quantum dynamics. Suppose (Ht )t≥0 is a family of linear operators on H⊗Γ with {vft⊗n ψ t } ⊂ Dom(Ht )
for v ∈ D1 ⊂dense H, ft ∈ kt , ft simple, right continuous and valued in V ⊂dense k0 , ψ t ∈ Γt , with Ht = Ĥt ⊗ 1Γt
for some map Ĥt : {H ⊗ E(kt )} ⊃ Dom(Ĥt ) → H ⊗ Γ, such that sup−≤s≤t kHs (uE(f ))k ≤ krt uk for all t, where
′ ′
rt depending on t, f only is a closable
Pm map in Lin(D1 , H ) for some Hilbert space H depending only on f .
If Ht is simple, that is, Ht = i=0 Hti 1[ti ,ti+1 ) (t), 0 = t0 < . . . tm < tm+1 = ∞, then for M as one of the
fundamental processes2 aR , a†R , t1,
Z t m
X
Hs M (ds) = Hti M ([ti , ti+1 ) ∩ [0, t])
0 0

For general processes the integral is obtained as an appropriate limit.

The Heisenberg formalism is captured by operator valued processes: for an adapted, regular process, Y (t) :
A⊗Γ ⊃ Dom(Y (t)) → A⊗Γ, Yg (t) : A⊗k0 ⊗Γt ⊃ Dom(Yg (t)) → A⊗Γt ⊗k0 defined by Yg
(t) = (Y (t)⊗1k0 )Swap23
where Swap23 (a1 ⊗ a2 ⊗ a3 ) = a1 ⊗ a3 ⊗ a2 ,

] Swap23 Y (s)⊗1
Y (s) : A ⊗ k0 ⊗ E(ks ) −−−−−→ A ⊗ E(ks ) ⊗ k0 −−−−−→ A ⊗ Γ(ks ) ⊗ k0

The map-valued processes for any δ : A0 → A ⊗ k0 are defined as follows

X X X X †
aδ (∆)( xi ⊗ E(fi ))u := aδ(x∗i ) (∆)(uE(fi )), a†δ (∆)( xi ⊗ E(fi ))u := aδ (xi )(∆)(uE(fi )) (40)
i i i i
X X
IL (∆)( xi ⊗ E(fi ))u := |∆|(L(xi )u) ⊗ E(fi ) (41)
i i

and the map-valued integrals are defined by: with u ∈ H, f ∈ L4loc , x ∈ A,

Z t  Z t
Y (s) ◦ (aδ + IL )(ds) (x ⊗ E(f ))u = Y (s) ((L(x) + hδ(x∗ ), f (s)i ⊗ E(f )) u ds (42)
0 0
Z t  Z t 
Y (s) ◦ (a†δ )(ds) (x ⊗ E(f ))u = a†Ye ,x (ds) uE(f ) (43)
0 0

where

a†Ye ,x (s)(uE(fs )) := Y]
(s)(δ(x) ⊗ E(fs ))u (44)

Remark B.1. A standard technique involves using iterated integrals arising through Picard iteration, requiring
iteration of the map δ : A → k0 of the form δ(δ(x)) := (δ ⊗ 1)(δ(x)) : A → A ⊗ k0 ⊗ k0 , more generally,
δ m : A → A ⊗ k0⊗m . To this, Ye is extended as follows –

Swap23
Y (s)⊗1

Y]
(s) : A ⊗ ⊕m k0⊗m ⊗ E(ks ) −−−−−→ A ⊗ E(ks ) ⊗ ⊕m k0⊗m −−−−−→ A ⊗ Γ(ks ) ⊗ ⊕m k0⊗m

To make sense of the Pcreation process on k0⊗m , one just needs an identification k0⊗m → kt , a natural choice is
via a1 ⊗ a2 . . . am → i∈[0:m−1] ai 1i|∆|+∆ , where the tensor product leads to the interpretation as a partition
refinement.
Now assume L(A∞ ) ⊂ A∞ . With V0 = (k0 )∞ , define
• Vt = {V0 -valued simple functions in kt }
• V = {V0 -valued simple functions}

One defines a map-valued integrable process with respect to aδ , a†δ , IL as follows:

Definition B.2. An integrable completely smooth map-valued process is an adapted process (Y (s))s≥0 : A∞ ⊗
E(V) → A ⊗ Γ(k) such that:
2 The integral with respect to the conservation process is not treated here, but the treatment is analogous

15
 1. For each t ≥ 0, f ∈ V, Y (t)(a ⊗ E(f )) ∈ (A ⊗ Γ(k))∞ and the map following map is completely smooth:

A∞ ∋ a → Ωt,f (a) : Y (t)(a ⊗ E(f )) ∈ (A ⊗ Γ(k))∞ (45)

2. For every m ∈ N and fixed X ∈ E∞ m

:= (A ⊗ k̂0⊗m )∞ , f ∈ V, any separable Hilbert space H′ , the ampilation
e t,f := Ωt,f ⊗ 1H′ : (A ⊗ H )∞ → (A ⊗ Γ ⊗H′ )∞ is continuous.
Ω ′

3. For every fixed a ∈ A∞ , f ∈ V, Y] e t,f ,

(s) := Ω

Sa (s) : H ⊗ E(Vs ) ∋ uE(fs ) → Y]

(s)(δ(a ⊗ E(fs )))u ∈ H ⊗ Γs ⊗k0 (46)

is continuous
4. For every a ∈ A∞ , ξ ∈ V0 the map

s → Y (s)((L(a)) + hδ(a∗ ), ξi)) ⊗ E(f ) (47)

is strongly continuous.

References
[1] A. C. Belton and S. J. Wills. “An algebraic construction of quantum flows with unbounded generators”. In:
Annales de l’IHP Probabilités et statistiques. Vol. 51. 1. 2015, pp. 349–375.
[2] N. Berline, E. Getzler, and M. Vergne. Heat kernels and Dirac operators. Springer, 2003.
[3] X. Bin. “Derivatives of the spectral function and Sobolev norms of eigenfunctions on a closed Riemannian
manifold”. In: Annals of Global Analysis and Geometry 26 (2004), pp. 231–252.
[4] A. H. Chamseddine and A. Connes. “Spectral action for Robertson-Walker metrics”. In: Journal of High
Energy Physics 2012.10 (2012), pp. 1–30.
[5] B. Chow, P. Lu, and L. Ni. Hamilton’s Ricci flow. Vol. 77. American Mathematical Society, Science Press,
2023.
[6] N. V. Dang and M. Wrochna. “Complex powers of the wave operator and the spectral action on Lorentzian
scattering spaces”. In: arXiv preprint arXiv:2012.00712 (2020).
[7] A. Devastato, M. Kurkov, and F. Lizzi. “Spectral noncommutative geometry standard model and all that”.
In: International Journal of Modern Physics A 34.19 (2019), p. 1930010.
[8] J. Figueroa-O’Farrill. Spin Geometry. version:18.05.2017. url: http://empg.maths.ed.ac.uk/Activities/Spin.
[9] S. Gakkhar. “Quantum diffusion on almost commutative spectral triples and spinor bundles”. In: arXiv
preprint arXiv:2211.03319 (2022). url: https://arxiv.org/abs/2211.03319v2.
[10] E. Hebey. Nonlinear analysis on manifolds: Sobolev spaces and inequalities: Sobolev spaces and inequalities.
Vol. 5. American Mathematical Soc., 2000.
[11] H. B. Lawson and M.-L. Michelsohn. Spin Geometry (PMS-38), Volume 38. Princeton university press,
2016.
[12] J. M. Lee. Introduction to Riemannian manifolds. 2nd ed. Springer, 2018.
[13] K. Parthasarathy. An introduction to quantum stochastic calculus. Vol. 85. Springer Science & Business
Media, 1992.
[14] K. B. Sinha and D. Goswami. Quantum stochastic processes and noncommutative geometry. Vol. 169.
Cambridge University Press, 2007.
[15] J. Tolksdorf and T. Ackermann. “The generalized Lichnerowicz formula and analysis of Dirac operators.”
In: Journal für die reine und angewandte Mathematik 1996 (471 1996).
[16] W. D. Van Suijlekom. Noncommutative geometry and particle physics. Springer, 2015.

16