Let
$\mathcal{F}$
be a Riemannian foliation on a closed manifold M. Fix any Riemannian metric on the leaves that is smooth on M. Let V be any Riemannian vector bundle with a flat Riemannian
$\mathcal{F}$
-partial connection. Let
$p=\dim\mathcal{F}$
,
$q=\operatorname{codim}\mathcal{F}$
and
$k=\operatorname{rank} V$
. Let
$(\Omega(\mathcal{F},V),d_\mathcal{F})$
be the leafwise de Rham complex with coefficients in V. Its reduced cohomology is denoted
${\mathcal H}(\mathcal{F},V)$
; this is the maximal Hausdorff quotient of its cohomology,
$H(\mathcal{F},V)/\overline0$
, using the topology on
$H(\mathcal{F},V)$
induced by the
$C^\infty$
topology on
$\Omega(\mathcal{F},V)$
. Let
$\delta_\mathcal{F}$
be the operator on
$\Omega(\mathcal{F},V)$
defined by the de Rham coderivative on the leaves. Then
$\Delta_\mathcal{F}:=d_\mathcal{F}\delta_\mathcal{F}+\delta_\mathcal{F} d_\mathcal{F}$
is the leafwise Laplacian on
$\Omega(\mathcal{F},V)$
.
Since M is closed, the Hilbertian space
$L^2\Omega(\mathcal{F},V)$
of
$L^2$
leafwise differential forms with coefficients in V is well defined. The operators
$d_\mathcal{F}$
,
$\delta_\mathcal{F}$
and
$\Delta_\mathcal{F}$
have closed extensions in
$L^2\Omega(\mathcal{F},V)$
, denoted in the same way. Moreover, the leafwise heat equation defines the leafwise heat operator
$e^{-t\Delta_\mathcal{F}}$
(
$t\gt 0)$
on
$\Omega(\mathcal{F},V)$
[Reference RoeRoe87], which is bounded on
$L^2\Omega(\mathcal{F},V)$
, where it is strongly convergent to a projection
$\Pi:L^2\Omega(\mathcal{F},V)\to\mathrm{ker}\,\Delta_\mathcal{F}$
; thus the notation
$\Pi=e^{-\infty\Delta_\mathcal{F}}$
may be used.
A Hilbert-space scalar product on
$L^2\Omega(\mathcal{F},V)$
can be defined in a standard way by using the product of the leafwise Riemannian density with any holonomy-invariant smooth non-vanishing transverse density (e.g., one induced by a bundle-like metric) and the Riemannian structure on V. Then
$\delta_\mathcal{F}$
is the adjoint of
$d_\mathcal{F}$
in
$L^2\Omega(\mathcal{F},V)$
, and therefore
$\Delta_\mathcal{F}$
is self-adjoint,
$e^{-t\Delta_\mathcal{F}}$
on
$L^2\Omega(\mathcal{F},V)$
can be given by the spectral theorem, and
$\Pi$
is an orthogonal projection of
$L^2\Omega(\mathcal{F},V)$
.
In [Reference Álvarez López and KordyukovÁLK01], Corollary C states that
$\Pi$
also preserves
$\Omega(\mathcal{F},V)$
, yielding a leafwise Hodge decomposition of
$\Omega(\mathcal{F},V)$
(involving
$\Delta_\mathcal{F}$
,
$d_\mathcal{F}$
and
$\delta_\mathcal{F}$
), and a corresponding Hodge isomorphism
${\mathcal H}(\mathcal{F},V)\cong\mathrm{ker}\,\Delta_\mathcal{F}$
(induced by
$\Pi$
). However, S. Goette provided us with the counterexample explained in Example 1 below.
Let us recall our intended proof in [Reference Álvarez López and KordyukovÁLK01, Section 4] to explain our error and the solution. Since any metric on the leaves that is smooth on M can be extended to a bundle-like metric on M, Corollary C follows from Theorem B if V is a trivial Riemannian vector bundle with a trivial
$\mathcal{F}$
-partial connection.
In the general case, we considered the principal O(k)-bundle of orthonormal frames of V,
$\pi:F\to M$
, whose total space is a closed manifold. The flat metric
$\mathcal{F}$
-partial connection on V can be understood as a completely integrable O(k)-invariant vector subbundle
$H\subset TF$
so that
$\pi_\ast:H_f\to T_{\pi(f)}\mathcal{F}$
is an isomorphism for every
$f\in F$
. Let
$\widehat{\mathcal{F}}$
be the corresponding foliation on F. As in Molino’s theory [Reference MolinoMol88],
$\pi^\ast V$
is a trivial Riemannian vector bundle, and the lift of the flat
$\mathcal{F}$
-partial connection on V is the trivial flat
$\widehat{\mathcal{F}}$
-partial connection on
$\pi^\ast V$
. Thus, if
$\widehat{\mathcal{F}}$
were also a Riemannian foliation, Corollary C would hold for
$(\Omega(\widehat{\mathcal{F}},\pi^\ast V), d_{\widehat{\mathcal{F}}})$
. Here, we could use any metric on the leaves of
$\widehat{\mathcal{F}}$
that is smooth on F, and the corresponding operators
$\delta_{\widehat{\mathcal{F}}}$
and
$\Delta_{\widehat{\mathcal{F}}}$
.
Consider the injective homomorphism
$\pi^\ast:(\Omega(\mathcal{F},V),d_\mathcal{F})\to(\Omega(\widehat{\mathcal{F}},\pi^\ast V),d_{\widehat{\mathcal{F}}})$
. Taking the lift of the given metric on the leaves of
$\mathcal{F}$
to the leaves of
$\widehat{\mathcal{F}}$
, the operators
$\delta_{\widehat{\mathcal{F}}}$
and
$\Delta_{\widehat{\mathcal{F}}}$
also correspond to
$\delta_\mathcal{F}$
and
$\Delta_\mathcal{F}$
by
$\pi^*$
. Then Corollary C for
$(\Omega(\mathcal{F},V), d_\mathcal{F})$
would follow from the case of
$(\Omega(\widehat{\mathcal{F}},\pi^\ast V), d_{\widehat{\mathcal{F}}})$
.
We wrongly thought that
$\widehat{\mathcal{F}}$
is always a Riemannian foliation, as in Molino’s theory, but this condition has to be assumed as a hypothesis. So the correct statement of Corollary C is the following.
Corollary C.
With the above notation, assume that
$\widehat{\mathcal{F}}$
is a Riemannian foliation. Then
$\Pi$
defines a continuous operator on
$\Omega(\mathcal{F},V)$
, we have the leafwise Hodge decomposition
and
$(t,\alpha)\mapsto e^{-t\Delta_\mathcal{F}}\alpha$
defines a continuous map
$[0,\infty]\times\Omega(\mathcal{F},V)\to\Omega(\mathcal{F},V)$
. Thus
${\mathcal H}(\mathcal{F},V)$
can be canonically identified with
$\mathrm{ker}\,\Delta_\mathcal{F}$
, and, if
$\mathcal{F}$
is oriented, the leafwise Hodge star operator on
$\mathrm{ker}\,\Delta_\mathcal{F}$
induces an isomorphism
${\mathcal H}^v(\mathcal{F},V)\cong{\mathcal H}^{p-v}(\mathcal{F},V^\ast)$
.
If V is the normal bundle
$\nu=TM/T\mathcal{F}$
with the Bott
$\mathcal{F}$
-partial connection and the Riemannian structure induced by any bundle-like metric, then F is the principal O(q)-bundle
$\widehat M$
of orthonormal frames of
$\nu$
. In this case,
$\widehat{\mathcal{F}}$
is a Riemannian foliation by Molino’s theory. It easily follows that Corollary C is also true when V is a vector bundle associated to
$\widehat M$
and any orthogonal representation of O(q), with the induced Riemannian structure and flat
$\mathcal{F}$
-partial connection. For instance, this applies to the vector bundle
$S^2(\nu)$
of symmetric bilinear forms on the fibers of
$\nu$
, used in Section 5 of [Reference Álvarez López and KordyukovÁLK01]. Thus, the change in Corollary C does not affect the rest of the paper [Reference Álvarez López and KordyukovÁLK01].
Example 1 (S. Goette). Consider
$\mathbb{S}^1\equiv\mathbb{R}/2\pi\mathbb{Z}$
, and let
$[x]\in \mathbb{S}^1$
denote the class of every
$x\in\mathbb{R}$
. Take the 2-torus
$M = \mathbb{S}^1\times \mathbb{S}^1$
with the foliation
$\mathcal{F}$
whose leaves are the fibers of the first factor projection to
$\mathbb{S}^1$
. It is the quotient of the foliation
$\widetilde{\mathcal{F}}$
on
$\widetilde M :=\mathbb{R}^2$
whose leaves are the fibers of the first factor projection to
$\mathbb{R}$
. Let V be the complex line bundle over M that is the quotient of the trivial complex line bundle
$\widetilde V := \widetilde M\times\mathbb{C}$
over
$\widetilde M$
by the action of
$\mathbb{Z}^2$
defined by the transformations
where
$i = \sqrt{-1}$
and
$m,n \in\mathbb{Z}$
. Consider also the trivial Hermitean structure on
$\widetilde V$
. Then the trivial flat
$\widetilde{\mathcal{F}}$
-partial connection
$\widetilde{\nabla}$
on
$\widetilde V$
is a Hermitean connection, and projects to a flat
$\mathcal{F}$
-partial Hermitean connection
$\nabla$
on V. Via the canonical identity
$\mathbb{C} \equiv \mathbb{R}^2$
,
$\widetilde V$
and V become oriented real vector bundles of rank 2, with trivial Riemannian structures so that
$\widetilde{\nabla}$
and
$\nabla$
are flat
$\mathcal{F}$
-partial Riemannian connections. Then
$\Omega^0(\mathcal{F}, V) = C^\infty(M; V)$
(the space of smooth sections of V), which can be considered as the subspace of
$C^\infty(\widetilde M; \widetilde V) \equiv C^\infty(\widetilde M, \mathbb{C})$
consisting of the smooth functions
$f: \widetilde M \to \mathbb{C}$
satisfying the equivariance condition
Such a function f satisfies
$d_{\widetilde{\mathcal{F}}}(f)=0$
if and only if it depends only on x; say
$f = f(x)$
. Then Equation (1) means that
If
$\sin(x)$
is irrational, then the set
$\{\,e^{2\pi m \sin(x) i}\mid m\in\mathbb{Z}\,\}$
is dense in the complex unit circle
$\{\,z\in\mathbb{C}\mid|z|=1\,\}$
. The corresponding section
$s \in C^\infty(M; V)$
is therefore invariant by multiplication by complex numbers of modulus 1 on the leaves
$\{[x]\} \times \mathbb{S}^1$
when
$\sin(x)$
is irrational, so s vanishes on those leaves, whose union is dense in M, and therefore s vanishes on M. This shows that
$\mathrm{ker}\, d_\mathcal{F} = 0$
in
$\Omega^0(\mathcal{F}, V)$
.
On the other hand, the function
$f(x,y) = e^{y \sin(x) i}$
satisfies Equation (1) and is equal to 1 on
$2\pi \mathbb{Z} \times \mathbb{R}$
, so it defines a section
$s\in C^\infty(M;V)$
satisfying
$\Delta_\mathcal{F} s = 0$
on the leaf
$L_0 = \{ [0] \} \times \mathbb{S}^1$
. Then
$s_t := e^{-t\Delta_\mathcal{F}}s$
remains independent of
$t \gt 0$
on
$L_0$
, and therefore the restriction to
$2\pi \mathbb{Z} \times \mathbb{R}$
of its lift to
$\widetilde M$
remains equal to f for all t. If the limit of
$s_t$
as
$t \to \infty$
were defined in
$C^\infty(M;V)$
, it would be in
$\mathrm{ker}\, d_\mathcal{F}$
and different from zero on
$L_0$
, a contradiction.
In this case, the principal O(2)-bundle F of orthonormal frames of V over M is the quotient of the bundle
$\widetilde F := \widetilde M\times \{\,z\in\mathbb{C}\mid|z|=1\,\}$
over
$\widetilde M$
by the action of
$\mathbb{Z}^2$
defined by the transformations
$T_{m,n}$
(
$m,n \in\mathbb{Z}$
). The foliation
$\widehat{\mathcal{F}}$
on F is the quotient of the foliation
$\widetilde{\mathcal{F}}$
on
$\widetilde F$
whose leaves are the fibers of the first factor projection to
$\mathbb{R}$
. The foliation
$\widehat{\mathcal{F}}$
is not Riemannian because its leaf closures do not define a stratified structure of F (see Section 5.4 of [Reference MolinoMol88]).
We would like to take this opportunity to correct the following error on page 140 in [Reference Álvarez López and KordyukovÁLK01]. For
$a \gt 1$
, the value of the integral
$\int_0^{t/a}(1/{\sqrt{(t-s)s}})\,{ds}$
is
$\pi/2+\arcsin(({2-a})/{a})$
instead of
$\pi+\arctan(({2-a})/{a})$
. This correction gives the estimate needed there.
Acknowledgements
We thank S. Goette for providing us with the example.
Funding statement
The authors are partially supported by the grants PID2020-114474GB-I00 (AEI/FEDER, UE) and ED431C 2019/10 (Xunta de Galicia, FEDER).