To motivate a question about the nonnegativity of the Dirichlet energy of a weighted graph, we first discuss some background on curvatures and flows on complex manifolds.
Definition 1. Let
$(M^n, \omega )$
be a Hermitian manifold. The quadratic orthogonal bisectional curvature (from now on, QOBC) is the function
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu1.png?pub-status=live)
where the
$R_{\alpha \overline {\alpha } \gamma \overline {\gamma }}$
denote the components of the Chern connection of
$\omega $
with respect to the unitary frame
$\vartheta $
(a section of the unitary frame bundle
$\mathcal {F}_M$
).
This curvature first appeared implicitly in [Reference Bishop and Goldberg1] and is the Weitzenböck curvature operator (in the sense of [Reference Petersen12, Reference Petersen and Wink13, Reference Petersen and Wink14]) acting on real
$(1,1)$
-forms. (See [Reference Broder3] for alternative descriptions of the QOBC.) From [Reference Li, Wu and Zheng9], the QOBC is strictly weaker than the orthogonal bisectional curvature
$\text {HBC}_{\omega }^{\perp }$
(the restriction of the holomorphic bisectional curvature
$\text {HBC}_{\omega }$
to pairs of orthogonal
$(1,0)$
-tangent vectors). From [Reference Gu and Zhang8], the Kähler–Ricci flow on a compact Kähler manifold with
$\text {HBC}_{\omega }^{\perp } \geq 0$
converges to a Kähler metric
$\text {HBC}_{\omega } \geq 0$
. Hence, Mok’s extension [Reference Mok10] of the solution of the Frankel conjecture [Reference Mori11, Reference Siu and Yau16] shows that all compact Kähler manifolds with
$\text {HBC}_{\omega }^{\perp } \geq 0$
are biholomorphic to a product of Hermitian symmetric spaces (of rank
$\geq 2$
) and projective spaces. In particular, although
$\text {HBC}_{\omega }^{\perp }$
is an algebraically weaker curvature notion than
$\text {HBC}_{\omega }$
, the positivity of
$\text {HBC}_{\omega }^{\perp }$
does not generate new examples.
Wu et al. [Reference Wu, Yau and Zheng17] showed that if
$(M, \omega )$
is a compact Kähler manifold with
$\text {QOBC}_{\omega } \geq 0$
, then every non-Kähler class on the boundary of the Kähler cone affords a semi-positive representative (which is certainly not true in general; see [Reference Demailly, Peternell and Schneider7]). Further, Chau and Tam [Reference Chau and Tam4] showed that all harmonic
$(1,1)$
-forms are parallel on compact Kähler manifolds with
$\text {QOBC}_{\omega } \geq 0$
.
In [Reference Broder2], I showed that the real bisectional curvature of Yang and Zheng [Reference Yang and Zheng18] was best understood as a Rayleigh quotient. Continuing this program, we find that the QOBC is best understood as the Dirichlet energy of a certain weighted graph. In particular, we realise the difference of the Hodge and metric Laplacians, acting on
$(1,1)$
-forms, as the Dirichlet energy of a weighted graph.
To analyse the QOBC, we recall the following definition. Let G be a finite weighted graph, with vertices
$V(G) = \{ x_1, \ldots , x_n \}$
, and weighting specified by its adjacency matrix
$A \in \mathbb {R}^{n \times n}$
. The Dirichlet energy for a weighted graph is defined by
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu2.png?pub-status=live)
where
$f : V(G) \to \mathbb {R}$
is a function defined on the vertices of G.
To understand when the QOBC is nonnegative, we can ask the following (equally natural) question.
Question 2. Given a finite weighted graph
$(G, A)$
, where
$A \in \mathbb {R}^{n \times n}$
is a real matrix, what conditions on A are necessary or sufficient for the inequality
$\mathcal {E}(f) \geq 0$
to hold for all
$f : V(G) \to \mathbb {R}$
?
The main theorem of this note gives an answer to this problem. To this end, let us recall some terminology arising from distance geometry.
Definition 3. Let
$A = (A_{ij}) \in \mathbb {R}^{n \times n}$
be a real symmetric matrix. We say that A is a Euclidean distance matrix if there is a vector
$x =(x_1, \ldots , x_n) \in \mathbb {R}^n$
such that
$A_{ij} = (x_i - x_j)^2$
for each
$i,j=1,\ldots , n$
.
The set of all
$n \times n$
Euclidean distance matrices (EDMs) forms a convex cone that we denote by
$\mathbb {EDM}^n$
. Recall that the Frobenius inner product of two matrices
$A,B \in \mathbb {R}^{n \times n}$
is defined by
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu3.png?pub-status=live)
This dual pairing allows us to define the dual EDM cone
$(\mathbb {EDM}^n)^{\ast }$
.
Definition 4. The dual EDM cone
$(\mathbb {EDM}^n)^{\ast }$
is given by
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu4.png?pub-status=live)
Theorem 5. Let
$(G,A)$
be a weighted finite graph. Then the Dirichlet energy
$\mathcal {E}$
is nonnegative if and only if A lies in the dual EDM cone.
Proof.
If
$V(G) = \{ x_1, \ldots , x_n \}$
is the vertex set of some graph, then we may construct a Euclidean distance matrix
$B(f)$
from a graph function
$f : V(G) \to \mathbb {R}$
by setting
$B(f)_{ij} = (f(x_i) - f(x_j))^2$
. In particular, since
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu5.png?pub-status=live)
we see that the Dirichlet energy
$\mathcal {E}$
of a weighted graph
$(G, A)$
is nonnegative if and only if
$\text {tr}(AB) \geq 0$
for all Euclidean distance matrices
$B \in \mathbb {EDM}^n$
.
It is natural to ask what is the relation (if any) between the EDM cone (and its dual) and the
$\mathbb {PSD}^n$
cone, that is, the cone of (symmetric) positive semi-definite matrices. Dattorro [Reference Dattorro5] has shown that
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu6.png?pub-status=live)
Here,
$\mathbb {S}_{\text {H}}^n$
denotes the space of symmetric
$n \times n$
hollow matrices, that is, symmetric matrices with no nonzero entries on the diagonal, and
$\mathbb {S}_{\text {C}}^n$
denotes the geometric centring subspace
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu7.png?pub-status=live)
where
$\mathbf {e} = (1, \ldots , 1)^{t}$
. It is more natural to refer to
$\mathbb {S}_{\text {C}}^{n}$
as the annihilator of
$\mathbf {e} = (1, \ldots , 1)^{t} \in \mathbb {R}^{n}$
. The orthogonal complement of
$\mathbb {S}_{\text {C}}^{n}$
is then
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu8.png?pub-status=live)
In particular, standard properties of cones (see [Reference Dattorro6, page 434]) give the next proposition.
Proposition 6. Let
$(G,A)$
be a weighted finite graph. The Dirichlet energy
$\mathcal {E}$
is nonnegative if and only if A lies in
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu9.png?pub-status=live)
where
$\mathbb {D}^n$
is the cone of diagonal matrices.
Remark 7. As discussed in [Reference Broder3], let us introduce the nonstandard terminology of Perron weights. The well-known Schoenberg criterion [Reference Schoenberg15] states that a symmetric hollow matrix
$\Sigma $
is a Euclidean distance matrix if and only if it is negative semi-definite on the hyperplane
$H = \{ x \in \mathbb {R} : x^{t} \mathbf {e} = 0 \}$
, where
$\mathbf {e} = (1, \ldots , 1)^t$
. The Perron–Frobenius theorem asserts that the largest eigenvalue (the Perron root) of the EDM
$\Sigma $
is positive and occurs with eigenvector in the nonnegative orthant
$\mathbb {R}_{\geq 0}^n$
. Therefore, if
$\delta _1 \geq \delta _2 \geq \cdots \geq \delta _n$
denote the eigenvalues of a nontrivial Euclidean distance matrix
$\Sigma $
(that is, a Euclidean distance matrix with
$\delta _1> 0$
), then
$\delta _1> 0$
and
$\delta _2, \ldots , \delta _n \leq 0$
.
Definition 8. The Perron weights
$r_2, \ldots , r_n$
of an
$n \times n$
Euclidean distance matrix
$\Sigma $
, with eigenvalues
$\delta _1 \ge \delta _2 \ge \cdots \ge \delta _n$
, are the ratios
$r_k := -\delta _k/\delta _1$
.
With this terminology in mind, appealing to the eigenvalue characterisation of the dual EDM cone given in [Reference Broder3] yields the following corollary.
Corollary 9. Let
$A \in \mathbb {R}^{n \times n}$
be a real symmetric matrix with eigenvalues
$\lambda _1 \geq \lambda _2 \geq \cdots \geq \lambda _n$
. Then
$A \in (\mathbb {EDM}^n)^{\ast }$
if and only if, for every Euclidean distance matrix
$\Sigma $
, the Perron weights
$r_2, \ldots , r_k$
of
$\Sigma $
satisfy
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20230412142831535-0993:S0004972721001015:S0004972721001015_eqnu10.png?pub-status=live)
Remark 10. Let us note that the Perron weights of a Euclidean distance matrix always satisfy
$0 \le r_2 \le r_3 \le \cdots \le r_n \le 1$
.
Acknowledgement
The author would like to thank the referee for valuable suggestions and comments.