Convex Cones

Rolf Schneider

Universität Freiburg

Summer School “New Perspectives on Convex Geometry”

CIEM, Castro Urdiales, September 3–7, 2018

1/1
A closed convex cone in Rd , briefly a cone, is a closed, convex
6 C ⊆ Rd with λx ∈ C for x ∈ C and λ ≥ 0.
set ∅ =

2/1
A closed convex cone in Rd , briefly a cone, is a closed, convex
6 C ⊆ Rd with λx ∈ C for x ∈ C and λ ≥ 0.
set ∅ =

We consider the following question:

Let C, D be cones. Let D undergo a uniform random rotation Θ.

What is the probability that C and ΘD have a nontrivial

intersection, that is, C ∩ ΘD 6= {o}?

2/1
3/1
The answer is immediate if C and D happen to be subspaces:

(
0 if dim C + dim D ≤ d,
P(C ∩ ΘD 6= {o}) =
1 if dim C + dim D > d.

Thus, the probability depends only on the sum of the

dimensions of C and D.

4/1
The answer is immediate if C and D happen to be subspaces:

(
0 if dim C + dim D ≤ d,
P(C ∩ ΘD 6= {o}) =
1 if dim C + dim D > d.

Thus, the probability depends only on the sum of the

dimensions of C and D.

The great surprise is that for cones one can define a number,
the statistical dimension, such that a similar result holds
approximately.

4/1
The answer is immediate if C and D happen to be subspaces:

(
0 if dim C + dim D ≤ d,
P(C ∩ ΘD 6= {o}) =
1 if dim C + dim D > d.

Thus, the probability depends only on the sum of the

dimensions of C and D.

The great surprise is that for cones one can define a number,
the statistical dimension, such that a similar result holds
approximately.

Where does this come from?

4/1
The answer is immediate if C and D happen to be subspaces:

(
0 if dim C + dim D ≤ d,
P(C ∩ ΘD 6= {o}) =
1 if dim C + dim D > d.

Thus, the probability depends only on the sum of the

dimensions of C and D.

The great surprise is that for cones one can define a number,
the statistical dimension, such that a similar result holds
approximately.

Where does this come from?

The use of convex optimization for signal demixing under a

certain random model.

4/1
Michael B. McCoy: A geometric analysis of convex demixing. PhD
Thesis, California Institute of Technology, 2013.

Quotation:

“Our framework includes a random orientation model for the

constituent signals that ensures the structures are incoherent.
This work introduces a summary parameter, the statistical
dimension, that reflects the intrinsic complexity of a signal.
. . . demixing succeeds with high probability when the sum of
the complexities is less than the ambient dimension; otherwise,
it fails with high probability.
The fact that a phase transition between success and failure
occurs in demixing is a consequence of a new inequality in
conic integral geometry. Roughly speaking, this inequality
asserts that a convex cone behaves like a subspace whose
dimension is equal to the statistical dimension of the cone.”

5/1
This was continued and expanded in
D. Amelunxen, M. Lotz, M.B. McCoy, J.A. Tropp: Living on the edge:
phase transitions in convex programs with random data. Information
and Inference 3 (2014), 224–294.
(Winner of the “Information and Inference Best Paper Prize”
2015).

6/1
This was continued and expanded in
D. Amelunxen, M. Lotz, M.B. McCoy, J.A. Tropp: Living on the edge:
phase transitions in convex programs with random data. Information
and Inference 3 (2014), 224–294.
(Winner of the “Information and Inference Best Paper Prize”
2015).

Quotation: “This paper introduces a summary parameter, called

the statistical dimension, that canonically extends the
dimension of a linear subspace to the class of convex cones.
The main technical result demonstrates that the sequence of
intrinsic volumes of a convex cone concentrates sharply around
the statistical dimension.
This fact leads to accurate bounds on the probability that a
randomly rotated cone shares a ray with a fixed cone.”

6/1
 uk(c)

d(c)
Flc' 3' conctnratioh of conic in§insic.volumes' This plor displays thc conic inuinsic volumes,t(c) ofa circularcone c c Rtlt
withangler/6'Thedisributionccrnced'*te§rharplyaroundthe§iati§ticaldimensionä(C)^c32.5.SeeSection3:4lorfurther
discussi,on of this example.

Ao*eIr,,.n^xe"r, L rtt, I't* CoY, TvoPf L üA1

7/1
Applications described in that paper: Phase transitions in
random convex optimization problems

• compressed sensing; `1 minimization for identifying a sparse

vector from random linear measurements
• regularized linear inverse problems with random
measurements
• demixing problems under a random incoherence model
• cone programs with random affine constraints

8/1
The core from the viewpoint of convex geometry
C, D are cones, not both subspaces, and Θ is a uniform
random rotation.

We ask for the probability

P(C ∩ ΘD 6= {o}).

9/1
The core from the viewpoint of convex geometry
C, D are cones, not both subspaces, and Θ is a uniform
random rotation.

We ask for the probability

P(C ∩ ΘD 6= {o}).
In principle, the answer is known for a long time:

Spherical integral geometry (Santaló 1976, Glasauer 1995),

translated into the conical setting, yields the following.

9/1
The core from the viewpoint of convex geometry
C, D are cones, not both subspaces, and Θ is a uniform
random rotation.

We ask for the probability

P(C ∩ ΘD 6= {o}).
In principle, the answer is known for a long time:

Spherical integral geometry (Santaló 1976, Glasauer 1995),

translated into the conical setting, yields the following.

First, one needs to define the conic intrinsic volumes

v1 (C), . . . , vd (C) of a closed convex cone C ⊆ Rd .

Second, the conic kinematic formula provides the expectation

d
X
Evk (C ∩ ΘD) = vi (C)vd+k−i (D)
i=k
for k = 1, . . . , d. 9/1
Third, a version of the spherical Gauss–Bonnet theorem says
that
b d−1
2
c
X
2 v2k+1 (C) = 1 if C is not a subspace.
k=0

Since C ∩ ΘD is, with probability one, either {o} (in which case
vk (C ∩ ΘD) = 0 for k ≥ 1) or not a subspace, this implies that

b d−1
2
c
1{C ∩ ΘD 6= {o}} = 2
X
v2k+1 (C ∩ ΘD) almost surely
k=0

and hence
b d−1
2
c d
X X
P(C ∩ ΘD 6= {o}) = 2 vi (C)vd+2k+1−i (D).
k=0 i=2k+1

10 / 1
This is the promised explicit answer.

11 / 1
This is the promised explicit answer.

However, it is useless for most applications, since one cannot

compute the conic intrinsic volumes.

11 / 1
This is the promised explicit answer.

However, it is useless for most applications, since one cannot

compute the conic intrinsic volumes.

Fortunately, one can prove concentration of the conic intrinsic

volumes around the statistical dimension, and this allows
valuable conclusions.

11 / 1
This is the promised explicit answer.

However, it is useless for most applications, since one cannot

compute the conic intrinsic volumes.

Fortunately, one can prove concentration of the conic intrinsic

volumes around the statistical dimension, and this allows
valuable conclusions.

————————————————————————————

The following gives an introduction to the conic intrinsic

volumes, the conic kinematic formula, and the concentration
result.

11 / 1
Convex cones

C d denotes the set of closed convex cones in Rd .

Let C ∈ C d .
The dual or polar cone of C is

C ◦ := {x ∈ R d : hx, yi ≤ 0 for all y ∈ C}.

We have C ◦◦ = C.

For each x ∈ Rd , there is a unique point ΠC (x) ∈ C such that

kx − ΠC (x)k = min{kx − yk : y ∈ C}.

This defines the metric projection or nearest-point map ΠC of C.

12 / 1
T.c,r)

fl",c4

13 / 1
Lemma. (Moreau decomposition)
For all x ∈ Rd ,
x = ΠC (x) + ΠC ◦ (x),
where
hΠC (x), ΠC ◦ (x)i = 0.

14 / 1
Lemma. (Moreau decomposition)
For all x ∈ Rd ,
x = ΠC (x) + ΠC ◦ (x),
where
hΠC (x), ΠC ◦ (x)i = 0.

Proof.

W.l.o.g., x ∈ Rd \ (C ∪ C ◦ ) (otherwise trivial).

u := x − ΠC (x) is an outer normal vector of a supporting

hyperplane H of C through ΠC (x). We have o ∈ H.

⇒ u ∈ C◦

14 / 1
o ∈ H ⇒ ΠC (x) ⊥ u

The hyperplane H 0 ⊥ ΠC (x) through o supports C ◦ .

⇒ ΠC ◦ (x), the point in C ◦ nearest to x, is also the point in H 0

nearest to x.

⇒ ΠC ◦ (x) = u

The points x, ΠC (x), ΠC ◦ (x), o are the vertices of a rectangle.

⇒ assertion

15 / 1
Conic intrinsic volumes of polyhedral cones

First, we consider only polyhedral cones.

For these, the quickest definition of the conic intrinsic volumes
requires the k -skeleton
[
skelk (C) := relint F ,
F ∈Fk (C)

where Fk denotes the set of k -faces of C.

Let g be a standard Gaussian random vector in Rd .

Definition. For k = 0, . . . , d, the k th conic intrinsic volume of C

is defined by

vk (C) := P(ΠC (g) ∈ skelk (C)).

16 / 1
Some explanations:
The distribution of g is the standard Gaussian measure,
Z
1 1 2
γd (A) = √ d e− 2 kxk dx.
2π A

The Gaussian measure on a subspace L is denoted by γL .

The relation between the spherical Lebesgue measure σd−1 of

a Borel set A ⊆ Sd−1 (unit sphere) and the Gaussian measure
of the spanned cone A∨ := {λa : a ∈ A, λ ≥ 0} is given by

σd−1 (A)
γd (A∨ ) = .
σd−1 (Sd−1 )

(Hence, ‘angles’ in the following are spherical volumes.)

17 / 1
We define the internal angle of C at o by

β(o, C) := γhCi (C),

where hCi is the linear hull of C, and the external angle of C at

its face F by
γ(F , C) := γhF i⊥ (N(C, F )),
where N(C, F ) is the normal cone of C at F .

18 / 1
We define the internal angle of C at o by

β(o, C) := γhCi (C),

where hCi is the linear hull of C, and the external angle of C at

its face F by
γ(F , C) := γhF i⊥ (N(C, F )),
where N(C, F ) is the normal cone of C at F .

Recall that
vk (C) := P(ΠC (g) ∈ skelk (C)).

For a face F ∈ Fk (C) we have

ΠC (x) ∈ relint F ⇔ x ∈ (relint F ) + N(C, F ).

18 / 1
 19 / 1
IL
LJ
\)
z
Therefore,

P(ΠC (g) ∈ relint F ) = γd (F + N(C, F )) = β(o, F )γ(F , C).

The latter holds since the Gauss measure of a direct sum is the
product of the Gauss measures of the summands.

We conclude that

vk (C) = P(ΠC (g) ∈ skelk (C))

X
= γd (F + N(C, F ))
F ∈Fk (C)
X
= β(o, F )γ(F , C).
F ∈Fk (C)

20 / 1
Relations between conic intrinsic volumes
A duality relation

Let F ∈ Fk (C), then G := N(C, F ) ∈ Fd−k (C ◦ ) and

N(C ◦ , G) = F , hence F + N(C, F ) = G + N(C ◦ , G).
It follows that

vk (C) = vd−k (C ◦ ) for k = 0, . . . , d.

21 / 1
Relations between conic intrinsic volumes
A duality relation

Let F ∈ Fk (C), then G := N(C, F ) ∈ Fd−k (C ◦ ) and

N(C ◦ , G) = F , hence F + N(C, F ) = G + N(C ◦ , G).
It follows that

vk (C) = vd−k (C ◦ ) for k = 0, . . . , d.

A trivial relation

From the definition

vk (C) := P(ΠC (g) ∈ skelk (C)),

it follow immediately that

v0 + · · · + vd = 1.
21 / 1
A nontrivial relation

If C is not a subspace, then

v0 (C) − v1 (C) + v2 (C) − · · · + (−1)d vd (C) = 0.

This follows from the identity

(−1)dim F 1F −N(C,F ) (x) = 0

X
for x ∈ Rd \ U,
F ∈F (C)

where U is the union of all faces of dimension < d − 1 of all

cones F − N(C, F ), F ∈ F(C).

The identity is due to P. McMullen (sketched 1975, proved

1981, reproduced in Schneider and Weil, Stochastic and
Integral Geometry, 2008).

22 / 1
A side remark

The identity

(−1)dim F 1F −N(C,F ) (x) = 0

X
(∗)
F ∈F (C)

was recently proved for all x ∈ Rd , without an exceptional set.

More generally, the following identity for convex polyhedra (not

necessarily bounded) holds.

Let P be a convex polyhedron,

• F a face of P,
• A(F , P) the angle cone of P at F ,
• N(P, F ) the normal cone of P at F .

23 / 1
Theorem 1. Let E 6= P be a face of P. Then

(−1)dim F 1A(E,F )−N(P,F ) = 0.

X

E⊆F ∈F (P)

The proof, in

R. Schneider, Combinatorial identities for polyhedral cones. Algebra i

Analiz 29 (2017), 279–295

uses the generalized Brianchon–Gram–Sommerville relation

(on the level of indicator functions) and employs the incidence
algebra of the face lattice of P.

If P is a cone (not a subspace) and E is the lineality space of P,

then A(E, F ) = F , hence Theorem 1 gives (*).

24 / 1
The identity

(−1)dim F 1F −N(C,F ) = 0
X
(∗)
F ∈F (C)

was extended to all polyhedra by Hug and Kabluchko 2018.

(End of the side remark)

25 / 1
Back to the identity

v0 (C) − v1 (C) + v2 (C) − · · · + (−1)d vd (C) = 0.

Above, we had excluded subspaces. If C is a subspace, then

v0 (C) − v1 (C) + v2 (C) − · · · + (−1)d vd (C) = (−1)dim C .

For a formula comprising both cases, we need the Euler

characteristic.

26 / 1
More generally, for a closed convex set K ⊂ Rd , define

type(K ) := (k, ),

where k is the largest dimension of an affine subspace

contained in K , and  = 1 or 0 according to whether the
line-free kernel of K is bounded or not.

Theorem and Definition. There is a unique real valuation χ on

U(CC d ) (the set of all finite unions of closed convex sets), the
Euler characteristic, with

χ(K ) = (−1)k  if type(K ) = (k, ), for K ∈ CC d .

Hadwiger’ (1955) elementary existence proof extends.

27 / 1
With this definition, we have for any polyhedral cone C
d
X
(−1)k vk (C) = χ(C).
k=0

Together with
d
X
vk (C) = 1,
k=0
this yields
b d−1
2
c
X
2 v2k+1 (C) = 1 − χ(C).
k=0

This is important for our initial question, since almost surely

b d−1
2
c 
X 1 if C ∩ ΘD =
6 {o},
2 v2k+1 (C ∩ ΘD) =
0 if C ∩ ΘD = {o}.
k=0

28 / 1
Remark. The relation
b d−1
2
c
X
2 v2k+1 (C) = 1 − χ(C)
k=0

is a version of the spherical Gauss–Bonnet theorem.

For smooth submanifolds of the sphere, it appears in Santaló
1955, 1976.

29 / 1
Remark. The relation
b d−1
2
c
X
2 v2k+1 (C) = 1 − χ(C)
k=0

is a version of the spherical Gauss–Bonnet theorem.

For smooth submanifolds of the sphere, it appears in Santaló
1955, 1976.

————————————————————————————

Up to now, we have only considered polyhedral cones.

We use a Steiner formula to extend the conic intrinsic volumes

to general closed convex cones.

29 / 1
Steiner formulas

M.B. McCoy, J.A.Tropp, From Steiner formulas for cones to

concentration of intrinsic volumes. Discrete Comput. Geom. 51
(2014), 926–963:

Theorem 2. (Master Steiner formula) For a polyhedral cone C

and a measurable function f : R2+ → R+ define
 
ϕf (C) := Ef kΠC (g)k2 , kΠC ◦ (g)k2 .

Then
d
X
ϕf (C) = Ik (f ) · vk (C),
k=0

where
Ik (f ) = ϕf (Lk ), Lk ∈ G(d, k ).

30 / 1
Explicitly,
Z ∞Z ∞
ωk ωd−k 1 2 +s 2 )
Ik (f ) = √ d
f (r 2 , s2 )e− 2 (r r k−1 sd−k−1 ds dr
2π 0 0

for k = 1, . . . , d − 1 and
Z ∞
ωd 1 2
I0 (f ) = √ d f (0, s2 )e− 2 s sd−1 ds,
2π 0
Z ∞
ωd 1 2
Id (f ) = √ d f (r 2 , 0)e− 2 r r d−1 dr .
2π 0

31 / 1
Essentials of the proof:

ϕf (C)
 
= Ef kΠC (g)k2 , kΠC ◦ (g)k2

d h   i
E f kΠC (g)k2 , kΠC ◦ (g)k2 1relint F (ΠC (g) .
X X
=
k=0 F ∈Fk (C)

In the following, we use the Moreau decomposition.

Then we use the fact that the Gauss measure of a direct sum is
the product of the Gauss measures of the summands.

32 / 1
For F ∈ Fk (C),
h   i
E f kΠC (g)k2 , kΠC ◦ (g)k2 1relint F (ΠC (g)
Z
1   1 2
=√ d
f kΠC (z)k2 , kΠC ◦ (z)k2 e− 2 kzk λd (dz)
2π F ⊕N(C,F )
Z Z
1 1 2 +kyk2 )
=√ d
f (kxk2 , ky k2 )e− 2 (kxk λd−k (dy)λk (dx).
2π F N(C,F )

The rest is computation.

33 / 1
For F ∈ Fk (C),
h   i
E f kΠC (g)k2 , kΠC ◦ (g)k2 1relint F (ΠC (g)
Z
1   1 2
=√ d
f kΠC (z)k2 , kΠC ◦ (z)k2 e− 2 kzk λd (dz)
2π F ⊕N(C,F )
Z Z
1 1 2 +kyk2 )
=√ d
f (kxk2 , ky k2 )e− 2 (kxk λd−k (dy)λk (dx).
2π F N(C,F )

The rest is computation.

Specializations of the Master Steiner formula yield ‘usual’

Steiner formulas, for the volume of parallel sets.

But for cones, we replace ‘volume’ by Gauss measure, and

‘parallel sets’ can be interpreted differently.
33 / 1
34 / 1
(a) The Gaussian Steiner formula

Here we consider the usual parallel set of C at distance λ ≥ 0,

{x ∈ Rd : dist(x, C) ≤ λ} = C + λB d .

The choice
1 if b ≤ λ2 ,

f (a, b) =
0 otherwise
in the Master Steiner formula yields

35 / 1
Corollary 1. For a polyhedral cone C, the Gaussian measure
of the parallel set C + λB d is given by
d
X
γd (C + λB d ) = fk (λ) · vk (C)
k=0

with Z λ
ω 1 2
fk (λ) = √ d−k
d−k
e− 2 s sd−k−1 ds
2π 0

for k = 0, . . . , d − 1 and fd ≡ 1.

This allows an approach to the conic intrinsic volumes that

parallels the approach to the classical intrinsic volumes of
convex bodies.

36 / 1
(b) The spherical Steiner formula
(in a conic interpretation)

Let da (x, y) denote the angle between the vectors x, y ∈ Rd ,

 
x y
da (x, y) = arccos , , x, y 6= o.
kxk ky k

The angular distance of x ∈ Rd \ C ◦ from a cone C 6= {o} is

da (x, C) = min{da (x, y) : y ∈ C \ {o}}

kΠC (x)k
= arccos .
kxk

The angular parallel set of C 6= {o} at distance λ ≥ 0 is

Cλa = {x ∈ Rd \ C ◦ : da (x, C) ≤ λ}.

37 / 1
The choice
1 if a ≤ b tan2 λ,

f (a, b) =
0 otherwise
in the Master Steiner formula yields:

Corollary 2. For a polyhedral cone C, the Gaussian measure of

the angular parallel set Cλa at distance 0 ≤ λ < π/2 is given by

d
X
γd (Cλa ) = gk (λ) · vk (C)
k=0

with Z λ
ω ω
gk (λ) = k d−k cosk−1 ϕ sind−k−1 ϕ dϕ
ωd 0
for k = 1, . . . , d − 1 and gd ≡ 1.

38 / 1
McCoy and Tropp (2014) have elegant reformulations of the
special Steiner formulas:

Corollary 3. For λ ≥ 0,
d
X
P(dist 2 (g, C) ≤ λ) = P(Xd−k ≤ λ) · vk (C),
k =0

where Xd−k is a random variable following the chi-square

distribution with d − k degrees of freedom.

Corollary 4. If u is a uniform random vector in the sphere Sd−1 ,

then, for λ ∈ [0, 1],
d
X
P(dist 2 (u, C) ≤ λ) = P(Bd−k,k ≤ λ) · vk (C),
k=0

where Bd−k,k is a random variable following the beta

distribution with parameters j/2 and k /2.
39 / 1
We can now extend all this by continuity.

Let C∗d denote the set of closed convex cones 6= {o} in Rd .

On C∗d , we define the angular Hausdorff metric by

δa (C, D) := min{ε ≥ 0 : C ⊆ Dεa , D ⊆ Cεa }.

With respect to this metric, polarity is a local isometry:

If C, D ∈ C∗d are cones 6= Rd with δa (C, D) < π/2, then

δa (C ◦ , D ◦ ) = δa (C, D)

(Glasauer 1995).

40 / 1
For 0 ≤ λ < π/2, let

µλ (C) := γd (Cλa )

denote the Gaussian measure of the angular parallel set of C at

angular distance λ.

If Ci , C ∈ C∗d and Ci → C in the angular Hausdorff metric, then

µλ ((Ci )aλ ) → µλ (Cλa ).

Further, any cone C ∈ C∗d can be approximated arbitrarily

closely by polyhedral cones, with respect to the angular
Hausdorff metric.

With this, the following theorem can be proved.

41 / 1
Theorem 3.
To every cone C ∈ C∗d , there exist nonnegative numbers
v0 (C), . . . , vd (C) such that, for every λ with 0 ≤ λ < π/2, the
Gaussian measure of the angular parallel set Cλa is given by

d
X
γd (Cλa ) = gk (λ) · vk (C)
k=1

with Z λ
ω ω
gk (λ) = k d−k cosk−1 ϕ sind−k−1 ϕ dϕ
ωd 0
for k = 1, . . . , d − 1 and gd ≡ 1. Further, vk ({o}) = δk,0 .
Each mapping vk : C d → R is a continuous valuation.

vk (C) is called the kth conic intrinsic volume of the cone C.

42 / 1
Let us briefly resume where we are.

Our goal is to compute and estimate the probability

P(C ∩ ΘD 6= {o}),

where C, D are given convex cones (not both subspaces) and

Θ is a uniform random rotation.

By now, we know how to express the indicator function of the

crucial event C ∩ ΘD 6= {o} in terms of conic intrinsic volumes
of C ∩ ΘD, namely (a.s.)

b d−1
2
c 
X 1 if C ∩ ΘD =
6 {o},
2 v2k+1 (C ∩ ΘD) =
0 if C ∩ ΘD = {o}.
k=0

43 / 1
Recall that Θ is a uniform random rotation, that is, a random
variable with values in the rotation group SOd , with distribution
given by the Haar probability measure ν on SOd .

Hence, we have to compute

Z
Evk (C ∩ ΘD) = vk (C ∩ ϑD) ν(dϑ).
SOd

The computation of such integrals is a task of integral geometry.

The relevant kinematic formulas are well known in Euclidean
space, and moderately well known in spherical space.

Spherical integral geometry, from the conic viewpoint, received

new interest when its applicability in convex programming was
discovered.

44 / 1
Theorem 4. (Conic kinematic formula)
If C, D ∈ C d are closed convex cones, then
Z d
X
vk (C ∩ ϑD) ν(dϑ) = vi (C)vd+k−i (D)
SOd i=k

for k = 1, . . . , d, and
Z d X
X d−i
v0 (C ∩ ϑD) ν(dϑ) = vi (C)vj (D).
SOd i=0 j=0

The second part of the theorem follows from the first, since
v0 + · · · + vd = 1.

45 / 1
Duality at work

Duality immediately yields the following:

Theorem 5.
If C, D ∈ C d are closed convex cones, then
Z d−k
X
vd−k (C + ϑD) ν(dϑ) = vi (C)vd−k−i (D)
SOd i=0

for k = 1, . . . , d, and
Z d
X d
X
vd (C + ϑD) ν(dϑ) = vi (C)vj (D).
SOd i=0 j=d−r

46 / 1
In the following, we sketch a new proof, which is due to

D. Amelunxen, M. Lotz, Intrinsic volumes of polyhedral cones: a

combinatorial perspective. Discrete Comput. Geom. 58 (2017),
371–409.

We restrict ourselves to polyhedral cones (the general case can

be obtained by approximation) and to the principal ideas.

These consist in clever integrations over suitable subgroups of

the rotation group and applying (of course) Fubini’s theorem.

First we prove two special subcases of the two theorems above

(which are used in the general proof).

47 / 1
Lemma 1. Let C, D ⊂ Rd be polyhedral cones, where

dim C = i, dim D = j, i + j = d + k > d.

Then Z
vk (C ∩ ϑD) ν(dϑ) = vi (C)vj (D).
SOd

For the proof, recall that we use hCi for the linear hull of the
cone C.

We also use SOhCi for the rotation group of the subspace hCi
and νhCi for its Haar probability measure.

48 / 1
For ν-almost all ϑ ∈ SOd , we have dim(hCi ∩ ϑhDi) = k.
Therefore,
Z
vk (C ∩ ϑD) νdϑ)
SOd
Z
= γhCi∩ϑhDi (C ∩ ϑD) ν(dϑ)
SOd
Z Z
= 1C (x)1ϑD (x)γhCi∩ϑhDi (dx) ν(dϑ).
SOd hCi∩ϑhDi

Here we replace ϑ by ϑρ with ρ ∈ SOhDi , which does not

change the outer integral (and observe that ρhDi = hDi).

Hence, we can integrate over all ρ with respect to νhDi .

This gives

49 / 1
 Z
vk (C ∩ ϑD) νdϑ)
SOd
Z Z Z
= 1C (x)1D (ρ−1 ϑ−1 x)
SOhDi SOd hCi∩ϑhDi

×γhCi∩ϑhDi (dx) ν(dϑ) νhDi (dρ)

Z Z "Z #
= 1C (x) 1D (ρ ϑ x) νhDi (dρ)
−1 −1
SOd hCi∩ϑhDi SOhDi

×γhCi∩ϑhDi (dx) ν(dϑ).

By properties of the Haar measure, the integral in brackets is

equal to
γhDi (D) = vj (D).

50 / 1
Thus, we obtain:

Z
vk (C ∩ ϑD) νdϑ)
SOd
Z Z
= vj (D) 1C (x) γhCi∩ϑhDi (dx) ν(dϑ).
SOd hCi∩ϑhDi

Now we play the same trick again.

The outer integral does not change if we replace ϑ by σϑ with

σ ∈ SOhCi .

Integrating over all σ, we get

51 / 1
 Z
vk (C ∩ ϑD) νdϑ)
SOd
Z Z Z
= vj (D) 1C (x)
SOhCi SOd σ(hCi∩ϑhDi)

×γσ(hCi∩ϑhDi) (dx) ν(dϑ) νhCi (dσ)

Z Z "Z #
= vj (D) 1C (σx) νhCi (dσ)
SOd hCi∩ϑhDi SOhCi

×γhCi∩ϑhDi (dx) ν(dϑ)

= vj (D)vi (C),

since γhCi∩ϑhDi and ν are probability measures.

This proves Lemma 1.

52 / 1
Lemma 2. Let C, D ⊂ Rd be polyhedral cones, where

dim C = i, dim D = j, i + j = d − k < d.

Then Z
vd−k (C + ϑD) ν(dϑ) = vi (C)vj (D).
SOd

For ν-almost all ϑ we have dim(hCi + ϑhDi) = i + j = d − k.

Therefore, using the product property of the Gaussian measure,
Z
vd−k (C + ϑD) = 1C+ϑD (y )γhCi+ϑhDi (dy)
hCi+ϑhDi
Z
= 1C+ϑD+(hCi+ϑhDi)⊥ (x) γd (dx).
Rd

53 / 1
For ν-almost all ϑ, there is the unique ϑ-decomposition

x = xC,ϑ + xD,ϑ + xϑ

with

xC,ϑ ∈ hCi, xD,ϑ ∈ ϑhDi, xϑ ∈ (hCi + ϑhDi)⊥ .

Apply ρ ∈ SOhCi , to get

ρx = ρxC,ϑ + ρxD,ϑ + ρxϑ

with

ρxC,ϑ ∈ hCi, ρxD,ϑ ∈ ρϑhDi, ρxϑ ∈ (hCi + ρϑhDi)⊥ ,

which is the ρϑ-decomposition of ρx.

54 / 1
 55 / 1
-1
v4a 6
§
X
+
v\)
4
X /r
t\
l,,\
I t,,
t\
ltt
t\
t\
J \
Thus, we obtain
Z
vd−k (C + ϑD) ν(dϑ)
SOd

as rewritten above
Z Z
= 1C+ϑD+(hCi+ϑhDi)⊥ (x) γd (dx) ν(dϑ)
SOd Rd

apply the ϑ-decomposition

Z Z
= 1C (xC,ϑ )1ϑD (xD,ϑ ) γd (dx) ν(dϑ)
SOd Rd

replace ϑ by ρϑ and x by ρx with ρ ∈ SOhCi

Z Z
= 1C ((ρx)C,ρϑ )1ρϑD ((ρx)D,ρϑ ) γd (dx) ν(dϑ)
SOd Rd

(ρϑ)-decomposition of ρx = ρ(ϑ-decomposition of x)

56 / 1
 Z Z
= 1C (ρxC,ϑ )1ρϑD (ρxD,ϑ ) γd (dx) ν(dϑ)
SOd Rd

integrate over all ρ

Z Z Z
= 1C (ρxC,ϑ )1ϑD (xD,ϑ ) γd (dx) ν(dϑ) }νhCi (dρ)
SOhCi SOd Rd

apply Fubini
Z Z "Z #
= 1C (ρxC,ϑ )νhCi (dρ) 1D (ϑ−1 xD,ϑ ) γd (dx) ν(dϑ)
SOd Rd SOhCi

integral in brackets = vi (C)

Z Z
= vi (C) 1D (ϑ−1 xD,ϑ ) γd (dx) ν(dϑ)
SOd Rd

replace ϑ by ϑσ −1 with σ ∈ SOhDi , integrate over all σ

= vi (C)vj (D). This proves Lemma 2.

57 / 1
Aiming at the proof of the conic kinematic formula for
polyhedral cones, we recall that
X
vk (C) = β(o, F )γ(F , C)
F ∈Fk (C)
X
= vk (F )vd−k (N(C, F )).
F ∈Fk (C)

Here we introduce the abbreviation

ϕF (C) := vk (F )vd−k (N(C, F )) if dim F = k.

Hence, we have to integrate

X
vk (C ∩ ϑD) = ϕJ (C ∩ ϑD).
J∈Fk (C∩ϑD)

Thus, we need to consider the k -faces of C ∩ ϑD.

58 / 1
Definition. The cones F , G intersect transversally, written
F t G, if

dim(F ∩ G) = dim F + dim G − d and relint F ∩ relint G 6= ∅.

Lemma. Let C, D be polyhedral cones. For ν-almost all

ϑ ∈ SOd , each k face J of C ∩ ϑD is of the form

J = F ∩ ϑG

with F ∈ Fi (C), G ∈ Fj (D), where i + j = d + k and F t ϑG.

59 / 1
Therefore,
Z
vk (C ∩ ϑD) ν(dϑ)
SOd
X Z
ϕF ∩ϑD (C ∩ ϑD)1{F t ϑG} ν(dϑ).
X X
=
i+j=k+d F ∈Fi (C) G∈Fj (D) SOd

We have to show that

Z
ϕF ∩ϑD (C ∩ ϑD)1{F t ϑG} ν(dϑ) = ϕF (C)ϕG (D)
SOd

for dim F + dim G = d + k.

60 / 1
For ν-almost all ϑ,

ϕF ∩ϑD (C ∩ ϑD) = vk (F ∩ ϑG)vd−k (N(C, F ) + ϑN(D, G)),

thus we have to prove that

Z
vk (F ∩ ϑG)vd−k (N(C, F ) + ϑN(D, G)) ν(dϑ)
SOd

= ϕF (C)ϕG (D)

for dim F + dim G = d + k.

Lemmas 1 and 2 treated corresponding integrals where one of

the factors in the integrand is missing.

61 / 1
The proof:
Z
vk (F ∩ ϑG)vd−k (N(C, F ) + ϑN(D, G)) ν(dϑ)
SOd

replace ϑ by ρϑ with ρ ∈ SOhF i and integrate over all ρ

Z Z
= vk (F ∩ ρϑG)vd−k (N(C, F ) + ρϑN(D, G)) ν(dϑ)
SOhF i SOd

×νhF i (dρ)

Fubini; N(C, F ) = ρN(C, F ); SOd -invariance of vd−k

Z "Z #
= vk (F ∩ ρϑG) νhF i (dρ)
SOd SOhF i

×vd−k (N(C, F ) + ϑN(D, G)) ν(dϑ)

62 / 1
 apply Lemma 1 in hF i to the integral in brackets
Z
= vi (F ) vk (ϑG ∩ hF i)vd−k (N(C, F ) + ϑN(D, G)) ν(dϑ)
SOd

use SOd -invariance of vk

Z
= vi (F ) vk (G ∩ ϑ−1 hF i)vd−k (N(C, F ) + ϑN(D, G)) ν(dϑ)
SOd

replace ϑ by ϑσ with σ ∈ SOhGi and integrate over all σ

Z Z
= vi (F ) vk (σG ∩ ϑ−1 hF i)vd−k (N(C, F ) + ϑN(D, G))
SOhGi SOd

×ν(dϑ) νhGi (dσ)

Fubini

63 / 1
 Z "Z #
= vi (F ) vk (σG ∩ ϑ−1 hF i) νhGi (dσ)
SOd SOhGi

×vd−k (N(C, F ) + ϑN(D, G)) ν(dϑ)

apply Lemma 1 in hGi to the integral in brackets

Z
= vi (F )vj (G) vd−k (N(C, F ) + ϑN(D, G)) ν(dϑ)
SOd

apply Lemma 2

= vi (F )vj (G)vd−k (N(C, F ))vd−j (N(D, G)

= ϕF (C)ϕG (D).

This finishes the proof of the conic kinematic formula for

polyhedral cones.
Approximation extends this to general convex cones.
64 / 1
A side remark

Much of this can be localized.

For η ⊂ Rd × Rd , let

η + := {(λx, µy) : (x, y ) ∈ η, λ, µ > 0}.

The biconic σ-algebra is defined by

b d × Rd ) := {η ∈ B(Rd × Rd ) : η + = η}.
B(R
b d × Rd ), let
For a polyhedral cone C and for η ∈ B(R

Ωk (C, η) := P (ΠC (g) ∈ skel (C), (ΠC (g), ΠC ◦ (g)) ∈ η) .

Then Ωk (C, ·) is a measure, the kth conic support measure of

C, and Ωk (C, Rd × Rd ) = vk (C).

65 / 1
Much of what we said about conic intrinsic volumes extends to
conic support measures.

• Explicit representation:
X Z Z
Ωk (C, η) = 1η (x, y ) γhF i⊥ (dy ) γhF i (dx).
F ∈Fk F N(C,F )

• Duality:

Ωk (C, η) = Ωd−k (C ◦ , η∗), η∗ := {(y , x) : (x, y) ∈ η}.

• Od -equivariance:

Ωk (ϑC, ϑη) = Ωk (C, η), ϑ ∈ Od .

66 / 1
• Master Steiner formula: Set
h   i
ϕf (C, η) := E f kΠC (g)k2 , kΠC ◦ (g)k2 1η (ΠC (g), ΠC ◦ (g)) .

Then
d
X
ϕf (C, η) = Ik (f ) · Ωk (C, η).
k=0
• Conic kinematic formula: Define the conic curvature
measures by

Φk (C, A) := Ωk (C, A × Rd ), b d ).
A ∈ B(R
b d ),
Then, for A, B ∈ B(R
Z d
X
Φk (C ∩ ϑD, A ∩ ϑB) ν(dϑ) = Φi (C, A)Φd+k−i (D, B).
SOd i=k

By approximation and weak continuity, an extension to general

closed convex cones is possible.
67 / 1
Some references:

R. Schneider: Intersection probabilities and kinematic formulas for

polyhedral cones. Acta Math. Hungar. 155 (2018), 3–24.

R. Schneider: Conic support measures. arXiv:1807.03614v1

(end of side remark)

68 / 1
Some references:

R. Schneider: Intersection probabilities and kinematic formulas for

polyhedral cones. Acta Math. Hungar. 155 (2018), 3–24.

R. Schneider: Conic support measures. arXiv:1807.03614v1

(end of side remark)

————————————————————————————
The following is based on:

D. Amelunxen, M. Lotz, M.B. McCoy, J.A. Tropp: Living on the edge:

phase transitions in convex programs with random data. Information
and Inference 3 (2014), 224–294.

M.B. McCoy, J.A. Tropp, From Steiner formulas for cones to

concentration of intrinsic volumes. Discrete Comput. Geom. 51
(2014), 926–963:

68 / 1
Back to the initial question! It now has the following answer:

P (C ∩ ΘD 6= {o}) = E1 {C ∩ ΘD 6= {o}}
b d−1
2
c d
X X
= 2 vj (C)vd+2k+1−j (D)
k=0 j=2k+1

d−1
X d
X
= (1 + (−1)i+1 ) vj (C)vd+i−j (D).
i=1 j=i

An explicit computation of the conic intrinsic volumes is only

possible in very special cases.
One example: If C is the nonnegative orthant in Rd , then
 
−d d
vk (C) = 2 .
k

69 / 1
Concentration of the conic intrinsic volumes

The dimension of a linear subspace L plays a particular role:


1 if dim L = k ,
vk (L) =
0 otherwise.
If also C is a subspace, then

0 if dim C + dim L ≤ d,
P(C ∩ ΘL 6= {o}) =
1 if dim C + dim L > d.

The statistical dimension δ(C) of a general closed convex cone

C has the property that the conic intrinsic volumes of C
concentrate near δ(C), therefore:
P(C ∩ ΘL 6= {o}) is close to 0 if δ(C) + dim L is smaller than d,
P(C ∩ ΘL 6= {o}) is close to 1 if δ(C) + dim L is larger than d.

70 / 1
 uk(c)

d(c)
Flc' 3' conctnratioh of conic in§insic.volumes' This plor displays thc conic inuinsic volumes,t(c) ofa circularcone c c Rtlt
withangler/6'Thedisributionccrnced'*te§rharplyaroundthe§iati§ticaldimensionä(C)^c32.5.SeeSection3:4lorfurther
discussi,on of this example.

Ao*eIr,,.n^xe"r, L rtt, I't* CoY, TvoPf L üA1

71 / 1
The statistical dimension
Recall that the conic intrinsic volumes of a cone C satisfy

vk (C) ≥ 0, v0 (C) + · · · + vd (C) = 1.

Definition. The intrinsic volume random variable VC of a cone

C ∈ C d is defined as a random variable with values in
{0, . . . , d} and with distribution

P(VC = k) = vk (C), k = 0, . . . , d.

By duality,

P(VC ◦ = k) = vk (C ◦ ) = vd−k (C) = P(VC = d−k) = P(d−VC = k ),

hence
VC ◦ ∼ d − VC
(where ∼ means equality in distribution).
72 / 1
Definition and Theorem. The statistical dimension of the cone
C ∈ C d is the number
d
X
δ(C) := EVC = kvk (C) = EkΠC (g)k2 .
k=0

Proof of the last equality:

Z
2
EkΠC (g)k = kΠC (g)k2 γd (dx)
Rd
Z ∞  
= γd {x ∈ Rd : kΠc (x)k2 > t} dt
0
Z ∞  
= P kΠC (g)k2 > t dt.
0

Let C be polyhedral. By the Gaussian Steiner formula in the

version of Corollary 1,
73 / 1
 P(kΠC (g)k2 > t) = P(dist2 (g, C ◦ ) > t)

= 1 − P(dist2 (g, C ◦ ) ≤ t)
d
X
= 1− P(Xd−k ≤ t)vk (C ◦ )
k =0

d
X
= 1− [1 − P(Xk > t)] vk (C)
k=0

d
X
= P(Xk > t)vk (C).
k=0
R∞
Since 0 P(Xk > t) dt = EXk = k, the assertion follows (then
approximation).

74 / 1
Properties of the statistical dimension:

• intrinsic

• rotation-invariant, continuous valuation

• increasing under set inclusion

• δ(C) + δ(C ◦ ) = d

• δ(L) = dim L for a subspace L

• δ(C) = 12 d if C is self-dual (i.e., C ◦ = −C)

• δ(C × D) = δ(C) + δ(D)

The last relation will be explained later.

75 / 1
A variance estimate
By the same argument that above yielded
d
X
EkΠC (g)k2 = (EXk ) vk (C),
k=0

we can prove that

d 
X 
EkΠC (g)k4 = EXk2 vk (C).
k=0

Since EXk2 = k 2 + 2k , we obtain

EkΠC (g)k4 = EV2C + 2δ(C).

This yields an expression for the variance of the intrinsic

volume random variable VC :
76 / 1
 Var VC = EV2C − (EVC )2

= EkΠC (g)k4 − 2δ(C) − δ(C)2

= EkΠC (g)k4 − (EkΠC (g)k2 )2 − 2δ(C)

= Var(kΠC (g)k2 ) − 2δ(C).

We apply the Gaussian Poincaré inequality

Var f (g) ≤ E(k∇f (g)k2 )

to the function f (x) = kΠC (x)k2 and use that

∇kΠC (x)k2 = 2ΠC (x).

The result is that

Var VC ≤ 2δ(C).

77 / 1
Possibly an improvement: Since VC ◦ ∼ d − VC , we have

Var VC = Var VC ◦

and hence
Var VC ≤ 2(δ(C) ∧ δ(C ◦ )),
where a ∧ b := min{a, b}.

Tschebyscheff’s inequality yields a first concentration result:

  Var V 2
P |VC − δ(C)| ≥ λ δ(C) ≤ 2 C ≤ 2 .
p
λ δ(C) λ

The following gives sharper concentration.

78 / 1
Theorem 6. For a closed convex cone C, define

ω(C)2 := δ(C) ∧ δ(C ◦ )

and
−λ2 /4
 
pC (λ) := exp for λ ≥ 0.
ω(C)2 + λ/3
Then
P(|VC − δ(C)| ≥ λ) ≤ pC (λ)
for λ ≥ 0.

79 / 1
Theorem 6. For a closed convex cone C, define

ω(C)2 := δ(C) ∧ δ(C ◦ )

and
−λ2 /4
 
pC (λ) := exp for λ ≥ 0.
ω(C)2 + λ/3
Then
P(|VC − δ(C)| ≥ λ) ≤ pC (λ)
for λ ≥ 0.

Before saying a few words about the proof, we indicate how this
leads to estimates of

P(C ∩ ΘD 6= {o}).

We consider first the special case where D is a subspace L.

79 / 1
Theorem 7. Let C ∈ C d , and let L be a subspace. Then the
following holds for λ ≥ 0:

δ(C) + dim L ≤ d − λ ⇒ P(C ∩ ΘL 6= {o}) ≤ pC (λ),

δ(C) + dim L ≥ d + λ ⇒ P(C ∩ ΘL 6= {o}) ≥ 1 − pC (λ).

80 / 1
Theorem 7. Let C ∈ C d , and let L be a subspace. Then the
following holds for λ ≥ 0:

δ(C) + dim L ≤ d − λ ⇒ P(C ∩ ΘL 6= {o}) ≤ pC (λ),

δ(C) + dim L ≥ d + λ ⇒ P(C ∩ ΘL 6= {o}) ≥ 1 − pC (λ).

Proof. We write dim L = d − m, then

δ(C) + dim L ≤ d − λ ⇔ m ≥ δ(C) + λ,

and similarly in the second case.

We recall the general formula (C not a subspace)

b d−1
2
c d
X X
P (C ∩ ΘD 6= {o}) = 2 vj (C)vd+2k+1−j (D)
k=0 j=2k+1

and observe that now vk (D) = δk,d−m (Kronecker symbol).

80 / 1
This gives
P(C ∩ ΘL 6= {o}) = 2hm+1 (C)
with the half-tail function

hm (C) := vm (C) + vm+2 (C) + . . .

Define also the tail function

tm (C) := vm (C) + vm+1 (C) + . . .

Then, with a subspace E ⊃ L of dimension d − m + 1,

2hm+1 (C) = P(C ∩ ΘL 6= {o}) ≤ P(C ∩ ΘE 6= {o}) = 2hm (C)

and hence 2hm+1 (C) ≤ hm+1 (C) + hm (C) = tm (C), thus

P(C ∩ ΘL 6= {o}) ≤ tm (C).

81 / 1
Since, by the definition of VC ,
X
P(VC − δ(C) ≥ λ) = vk (C)
k≥δ(C)+λ

assumption m ≥ δ(C) + λ
X
≥ vk (C) = tm (C),
k≥m

we get

P(C ∩ ΘL 6= {o}) ≤ tm (C) ≤ P(VC − δ(C) ≥ λ) ≤ pC (λ)

by Theorem 6.

This is the first assertion of Theorem 7.

The second assertion is obtained similarly.

82 / 1
Some remarks about the crucial

Theorem 6. For a closed convex cone C, define

ω(C)2 := δ(C) ∧ δ(C ◦ )

and
−λ2 /4
 
pC (λ) := exp for λ ≥ 0.
ω(C)2 + λ/3
Then
P(VC − δ(C) ≥ λ) ≤ pC (λ),
P(VC − δ(C) ≤ −λ) ≤ pC (λ)
for λ ≥ 0.

83 / 1
McCoy and Tropp derive it from the stronger

Theorem 8.
Let C ∈ C d . Define the function ψ by

ψ(u) := (u + 1) log(u + 1) − u for u ≥ −1,

and ψ(u) = ∞ for u < −1. Then, for all λ ≥ 0,

P {VC − δ(C) ≥ λ}
     
1 λ ◦ −λ
≤ exp − max δ(C)ψ , δ(C )ψ
2 δ(C) δ(C ◦ )

and

P {VC − δ(C) ≤ −λ}

     
1 −λ ◦ λ
≤ exp − max δ(C)ψ , δ(C )ψ .
2 δ(C) δ(C ◦ )

84 / 1
Their proof first expresses the moment generating function of
the intrinsic volume random variable as
1 
EeηVC = EeξkΠC (g)k with ξ =
2
1 − e−2η .
2
The proof uses the Master Steiner formula and the moment
generating functions of chi-square random variables.

The main ingredient of the subsequent estimations is the

Gaussian logarithmic Sobolev inequality.

The rest is analysis.

85 / 1
After the special case C, L, where L is a subspace, it remains to
consider the case of two general cones C, D.

It is convenient to consider their product C × D.

Convention. For two Euclidean spaces Rd1 , Rd2 , their

Cartesian product Rd1 × Rd2 is always equipped with the scalar
product given by

h(x, y), (x 0 , y 0 )i := hx, x 0 i1 +hy , y 0 i2 , (x, y), (x 0 , y 0 ) ∈ Rd1 ×Rd2 ,

where h· , ·ir denotes the scalar product in Rdr , r = 1, 2.

Then C × D can be considered as a direct orthogonal sum

C ⊕ D.

86 / 1
Some remarks about C × D

First remark:

ΠC×D (x, y) = (ΠC (x), ΠD (y)) for (x, y) ∈ Rd × Rd .

The proof makes use of the Moreau decomposition.

Second remark:

δ(C × D) = δ(C) + δ(D).

Proof:

Assume R2d = L1 ⊕ L2 , dim Lr = d, C ⊂ L1 , D ⊂ L2 ;

identify C × D with C ⊕ D.

87 / 1
 Z
2
δ(C ⊕ D) = EkΠC⊕D (g)k = kΠC⊕D (x)k2 γ2d (dx)
R2d
Z Z  
= kΠC (y)k2 + kΠD (z)k2 γL2 (dz) γL1 (dy))
L1 L2

= δ(C) + δ(D).

Third remark:
X
vk (C × D) = vi (C)vj (D).
i+j=k

88 / 1
To prove X
vk (C × D) = vi (C)vj (D),
i+j=k

we can assume that C, D ⊂ Rd × Rd ; C ⊂ Rd × {o};

D ⊂ {o} × R d and C × D is the direct orthogonal sum C ⊕ D.

Each face F ∈ Fk (C ⊕ D) is of the form F = Fi ⊕ Gj with

Fi ∈ Fi (C), Gj ∈ Fj (D), i + j = k .

We have

N(C ⊕ D, Fi ⊕ Gj ) = N(C, Fi ) ⊕ N(D, Gj ).

The assertion then follows after computing the internal and

external angles.

89 / 1
Theorem 9. For closed convex cones C, D, define

σ(C, D)2 := (δ(C) ∧ δ(C ◦ )) + (δ(D) + δ(D ◦ ))

and
−λ2 /4
 
pC,D (λ) := exp for λ ≥ 0.
σ(C, D)2 + λ/3
Then
P(|VC×D − δ(C × D)| ≥ λ) ≤ pC,D (λ)
for λ ≥ 0.

The proof uses the previous estimates, and in addition that

X
vk (C × D) = vi (C)vj (D).
i+j=k

90 / 1
With this, we can write (if C, D are not both subspaces)

b d−1
2
c
X X
P(C ∩ ΘD 6= {o}) = 2 vi (C)vj (D)
k=0 i+j=d+2k+1

b d−1
2
c
X
= 2 vd+2k+1 (C × D)
k=0

= 2hd+1 (C × D).

This is convenient for the proof of the following theorem.

91 / 1
Theorem 10. Let C, D ∈ C d be cones. Then, for λ ≥ 0,

δ(C) + δ(D) ≤ d − λ ⇒ P(C ∩ ΘD 6= {o}) ≤ pC,D (λ),

δ(C) + δ(D) ≥ d + λ ⇒ P(C ∩ ΘD 6= {o}) ≥ 1 − pC,D (λ).

92 / 1
Theorem 10. Let C, D ∈ C d be cones. Then, for λ ≥ 0,

δ(C) + δ(D) ≤ d − λ ⇒ P(C ∩ ΘD 6= {o}) ≤ pC,D (λ),

δ(C) + δ(D) ≥ d + λ ⇒ P(C ∩ ΘD 6= {o}) ≥ 1 − pC,D (λ).

Proof.

We can assume that D, C are not both subspaces. Then, as

shown above,

P(C ∩ ΘD 6= {o}) = 2hd+1 (C × D).

As shown before,

2hd+1 (C × D) ≤ td (C × D).

92 / 1
The assumption δ(C) + δ(D) ≤ d − λ gives

d ≥ δ(C) + δ(D) + λ = δ(C × D) + λ

and hence

P(VC×D − δ(C × D) ≥ λ)
X X
= vk (C × D) ≥ vk (C × D) = td (C × D).
k≥δ(C×D)+λ k≥d

Thus,

P(C ∩ ΘD 6= {o}) ≤ P(VC×D − δ(C × D) ≥ λ) ≤ pC,D (λ),

by Theorem 9.

The second estimate is obtained similarly.

93 / 1