PROBABILISTIC SCHUBERT CALCULUS

arXiv:1612.06893v3 [math.AG] 19 Jan 2018

PETER BÜRGISSER AND ANTONIO LERARIO

Abstract. We initiate the study of average intersection theory in real Grassmannians. We

define the expected degree edeg G(k, n) of the real Grassmannian G(k, n) as the average number
of real k-planes meeting nontrivially k(n − k) random subspaces of Rn , all of dimension n − k,
where these subspaces are sampled uniformly and independently from G(n − k, n). We express
edeg G(k, n) in terms of the volume of an invariant convex body in the tangent space to the
Grassmanian, and prove that for fixed k ≥ 2 and n → ∞,
1
edeg G(k, n) = deg GC (k, n) 2 εk +o(1) ,
where deg GC (k, n) denotes the degree of the corresponding complex Grassmannian and εk is
monotonically decreasing with limk→∞ εk = 1. In the case of the Grassmannian of lines, we
prove the finer asymptotic
 2 n
8 π

edeg G(2, n + 1) = 5/2
√ 1 + O(n−1 ) .
3π n 4
The expected degree turns out to be the key quantity governing questions of the random
enumerative geometry of flats. We associate with a semialgebraic set X ⊆ RPn−1 of dimension
n − k − 1 its Chow hypersurface Z(X) ⊆ G(k, n), consisting of the k-planes A in Rn such
whose projectivization intersects X. Denoting N := k(n − k), we show that
N
Y |Xi |
E # (g1 Z(X1 ) ∩ · · · ∩ gN Z(XN )) = edeg G(k, n) · m|
,
i=1
|RP
where each Xi is of dimension m = n − k − 1, the expectation is taken with respect to
independent uniformly distributed g1 , . . . , gm ∈ O(n) and |Xi | denotes the m-dimensional
volume of Xi .

1. Introduction
1.1. Motivation. Classical enumerative geometry deals with questions like: “How many lines
intersect four lines in three-dimensional space in general position?” or more generally: “How
many lines intersect four curves of degrees d1 , . . . , d4 in three-space in general position?”
The answer to this type of questions is provided by a beautiful, sophisticated machinery,
which goes under the name of Schubert calculus. The problem is reduced to a computation in
the cohomology ring of the Grassmann manifold: the set of lines intersecting a given line (or
curve) represents a codimension one cycle (a Schubert variety) in the complex Grassmannian
GC (2, 4) of 2-dimensional vector subspaces in C4 (i.e., lines in CP3 ), and one has to count the
number of points of intersection of four generic copies of this cycle. When counting complex
solutions, the answer to the first question is 2; for the second the answer is 2 · d1 · · · d4 .
We refer to the survey [28] and any of the monographs [24, 17, 35, 14] for a treatment of
Schubert calculus.
Date: January 22, 2018.
2010 Mathematics Subject Classification. 14N15, 14Pxx, 52A22, 60D05.
Key words and phrases. enumerative geometry, real Grassmannians, Schubert calculus, integral geometry,
random polytopes.
First author partially supported by DFG grant BU 1371/2-2.
1
2 PETER BÜRGISSER AND ANTONIO LERARIO

Over the real numbers this approach fails—there is no generic number of real solutions,
which can already be seen in the basic problem of counting the real solutions to a polynomial
equation. Understanding even the possible outcomes for higher dimensional versions of the
problem becomes increasingly complicated. A main question considered in real enumerative
geometry is whether the number of complex solutions can be realized over the reals, in which
case one speaks of a fully real problem, see Sottile [47, 48]. Schubert calculus is known to be
fully real [46, 54]. Recently Finashin and Kharlamov [15] have identified a class of enumerative
problems for which one can get an analog of Schubert calculus over the reals, where solutions
are counted with signs (these are Schubert problems which are expressible in terms of what in
[15] is called Euler-Pontryagin ring).
The goal of our work is to study enumerative problems over the reals in a probabilistic way,
by answering questions like:

(1.1) “On average, how many real lines intersect four random lines in RP3 ?”.

This approach offers a broad point of view to the subject, by seeking typical properties of random
arrangements. As we will see, the theory we will present will blend ideas from (real) algebraic
geometry with integral geometry and the theory of random polytopes.
Classical Schubert calculus deals with the intersection of Schubert varieties in general position.
Our paper is a first attempt at developing such a theory over the reals from the probabilistic
point of view. By probabilistic Schubert calculus we understand the investigation of the expected
number of points of intersection of real Schubert varieties in random position. We confine our
study to special Schubert varieties, thus ignoring the more complicated flag conditions. For the
sake of simplicity, we also restrict our presentation to the codimension one case, i.e., intersection
of hypersurfaces in the real Grassmannian, even though our methods work in more generality, as
pointed out in different remarks throughout the text. It may be interesting to note that many
results in Schubert calculus, going back to Schubert and Pieri, were first shown in the special
cases of codimension one, for GC (2, n + 1), and for special Schubert varieties, before reaching
complete generality. In that sense, our paper fits in nicely with that tradition. We plan to treat
the case of higher codimension in a future publication.

1.2. Expected degree of real Grassmannians. Generalizing the question (1.1), the number
of lines meeting 2n − 2 generic projective subspaces of dimension n − 2 in CPn equals the degree
of the complex Grassmann manifold GC (2, n + 1), which is known to be the Catalan number
1 2n−2


n n−1 and asymptotically, when n → ∞, behaves as:

4n−1 −1


(1.2) deg GC (2, n + 1) = √ 1 + O(n ) .
n3/2 π
Similarly, the degree of GC (k, n) equals the number of k-planes in Cn meeting nontrivially k(n−k)
generic subspaces, all of dimension n − k.
For the study of the corresponding problem over the reals, we can introduce the following
probabilistic setup. Recall that for every (k, n) there is a unique probability distribution on the
real Grassmannian G(k, n), which is invariant under the action of O(n); we call this distribution
the uniform distribution. A random k-dimensional linear subspace of Rn is obtained by sampling
from this distribution.

Definition. We define the expected degree of G(k, n) as the average number edeg G(k, n) of
real k-planes meeting nontrivially k(n − k) many random, real, independent subspaces of Rn , all
of dimension n − k.
 PROBABILISTIC SCHUBERT CALCULUS 3

More specifically, we define the special Schubert variety Σ(k, n) ⊆ G(k, n) as the set of k-
planes in Rn intersecting a fixed linear subspace V of dimension n − k nontrivially. Then Σ(k, n)
is a codimension-one real algebraic subset of G(k, n), and the expected degree equals
edeg G(k, n) = E # (g1 Σ(k, n) ∩ · · · ∩ gN Σ(k, n)) ,
where N := k(n − k) = dim G(k, n), and the expectation is taken over independent uniformly
distributed g1 , . . . , gN ∈ O(n).
The expected number of lines meeting four random lines in RP3 equals edeg (2, 4) = 1.7262...
(the exact value is not known, see Proposition 6.7 for an explicit expression involving an iterated
integral). More generally, for the Grassmannian of lines, we prove the following asymptotic
(Theorem 6.8), which should be compared with its complex analogue (1.2).
Theorem. The expected degree of the Grassmannian of lines equals
8  π 2 n
1 + O(n−1 ) .


edeg G(2, n + 1) = 5/2
√
3π n 4
For the general case, we define√ρk := E kXk, where X ∈ Rk is a standard normal Gaussian
k-vector. (It is known that ρk ≤ k, which is asymptotically sharp; see (2.15).) We show the
following explicit upper bound
r N
|G(k, n)| π ρk
edeg G(k, n) ≤ · √ .
|RPk(n−k) | 2 k
Note that this specializes to edeg G(2, n) ≤ (π 2 /4)n−2 . Moreover, we prove that (in the logarith-
mic scale) this upper bound is asymptotically sharp, deducing the following (see Theorem 6.5).
Theorem. For fixed k, as n → ∞, the following asymptotic holds:
√
!
π Γ k+12
log edeg G(k, n) = kn log − O(log n).
Γ k2

A general, guiding theme of our work is to compare the number of complex solutions with
the expected number of real solutions. With this regard, we obtain:
1
Corollary. For fixed k ≥ 2, we have edeg G(k, n) = deg GC (k, n) 2 εk +o(1) for n → ∞, where εk
is monotonically decreasing with limk→∞ εk = 1. (See Corollary 6.6 for an explicit expression
for εk .)
This means that for large n, the expected degree of the real Grassmannian exceeds the square
root of the degree of the corresponding complex Grassmannian, and 12 (εk − 1) measures the
deviation in the exponent. If also k → ∞, we have an asymptotic square root law, edeg G(k, n) =
1
deg GC (k, n) 2 +o(1) , saying that the expected number of solutions is roughly the square root of
the number of complex solutions.

1.3. Random incidence geometry over the reals. The expected degree of Grassmannians
turns out to be the key quantity governing questions of random incidence geometry, as we
discuss now. Given semialgebraic subsets X1 , . . . , XN ⊂ RPn−1 of dimension n − k − 1, where
N = k(n − k), we will say that they are in random position if they are randomly translated by
elements g1 , . . . , gN sampled independently from the orthogonal group O(n) with the uniform
distribution. The problem of counting the number of k-planes whose projectivization intersects
X1 , . . . , XN in random position can be geometrically described as follows.
4 PETER BÜRGISSER AND ANTONIO LERARIO

We associate with a semialgebraic subset X ⊆ RPn−1 of dimension n − k − 1 the Chow

hypersurface
Z(X) := {k-planes in G(k, n) whose projectivization intersects X} ⊆ G(k, n),
which is the real analog of the associated Chow variety in complex algebraic geometry. The
special real Schubert variety Σ(k, n) is obtained by taking for X a linear space. In the following,
we denote by |X| the (n − k − 1)-dimensional volume of the set of smooth points of X.
A translation of X by an element g ∈ O(n) corresponds to a translation of the associated Chow
hypersurface Z(X) by the induced action of the same element on G(k, n). In the geometry of the
Grassmannian, the random incidence problem is equivalent to the computation of the average
number of intersection points of random translates of the sets Z(X1 ), . . . , Z(XN ).
The following theorem decouples the problem into the computation of the volume of X1 , . . . , XN
and the determination of the expected degree of the Grassmann manifold (see Theorem 4.10).
Note that over the complex numbers, the same decoupling result is a consequence of the ring
structure of the cohomology of the Grassmannian (compare §2.5 below).
Theorem. The average number of k-planes whose projectivization intersects semialgebraic sets
X1 , . . . , XN of dimension n − k − 1 in random position in RPn−1 equals
N
Y |Xi |
(1.3) E # (g1 Z(X1 ) ∩ · · · ∩ gN Z(XN )) = edeg G(k, n) · .
i=1
|RPn−k−1 |

Example 1.1 (Lines intersecting four random curves in three-space). Let us consider the problem
of counting the number of lines intersecting four random curves in RP3 . A possible model for
random algebraic varieties, which has attracted a lot of attention over the last years [7, 32, 33,
18, 19, 20, 21], is the so called Kostlan model, which we describe now for the case of space curves.
First let us recall that a random Kostlan polynomial f ∈ R[x0 , . . . , x3 ](d) is defined as
X
f (x) = ξα xα α3
0 · · · x3 ,
0

|α|=d

d!
where the ξα are independent centered gaussian variables with variance α0 !···α3!
. This model
is invariant under the action of O(4) by change of variables (there are no preferred points
or directions in projective space). A random Kostlan curve is defined as the zero set of two
independent Kostlan polynomials
C = {f (x) = g(x) = 0}.
1
It is well known [30] that the expectation of the length of C equals (deg f deg g) 2 |RP1 |. Using
this fact in the above theorem, applied to four random independent curves C1 , . . . , C4 , and
combined with the O(4)-invariance of the model, it is easy to show that the expected number of
lines intersecting the four curves equals
4
Y 1
E #{lines intersecting C1 , . . . , C4 } = edeg G(2, 4) · (deg fi deg gi ) 2 .
i=1

Observe that the randomness in this problem does not come from putting the curves in random
position, but rather from sampling them from a distribution which is invariant under random
translations: this allows to exchange the order of the expectations without changing the result.
(More generally a similar argument can be applied to other invariant models and in higher
dimensions.)
 PROBABILISTIC SCHUBERT CALCULUS 5

1.4. Iterated kinematic formula and volume of Schubert varieties. A key technical
ingredient of our work is a generalization of the kinematic formula for homogeneous spaces in
Howard [26] to multiple intersections (Theorem 7.2). This generalization is novel, but the proof
is technically inspired by [26], so that we have moved it to the Appendix 7.5. What one should
stress at this point is that there is no “exact” kinematic formula for the Grassmann manifold
G(k, n), meaning that the evaluation of kinematic integrals depends on the class of submanifolds
we consider. This is in sharp contrast with the case of spheres or projective spaces, and is
essentially due to the fact that the stabilizer O(k) × O(n − k) of the action is too small to enforce
transitivity at level of tangent spaces. In the context of enumerative geometry, however, one
can still produce an explicit formula, thanks to the observation that the orthogonal group acts
transitively on the tangent spaces of Chow hypersurfaces (and in particular Schubert varieties).
We use the resulting formula for the derivation of (1.3). In practice this requires the com-
putation of the volume of the associated Chow hypersurfaces (Theorem 3.19), as well as the
evaluation of an average angle between them (see §1.5 below).
By means of classical integral geometry in projective spaces, in Proposition 4.8 we prove that

|Z(X)| |X|
= .
|Σ(k, n)| |RPdim X |
We note this is analogous to the corresponding result over C; see (2.8). The computation of
the volume of the special Schubert variety Σ(k, n) requires more work, and constitutes a result
of independent interest (Theorem 4.2), due to its possible applications in numerical algebraic
geometry [8].
Theorem. The volume of the special Schubert variety Σ(k, n) satisfies:
|Σ(k, n)| |RPk−1 | |RPn−k−1 |
=π· · ,
|G(k, n)| |RPk | |RPn−k |
where the volume |G(k, n)| of the Grassmann manifold is given by (2.14).
This result shows a striking analogy with the corresponding result over C; see (2.7).
Remark 1.2. A known result in integral geometry, [40, Eq. (17.61)], gives the measure of the set
of projective (k − 1)-planes meeting a fixed spherically convex set Q in RPn−1 . A natural idea
would be to apply this result to the ǫ-neighborhood Qǫ in RPn−1 around a fixed (n− k − 1)-plane
and to consider the limit for ǫ → 0 (after some scaling). However, this argument is flawed since
neighborhoods Qǫ are not spherically convex!
Finally, let us mention that the integral geometry arguments can also be applied to the classical
setting over C. There, the situation is considerably simpler, since the scaling constant αC (k, m)
appearing in the corresponding integral geometry formula (Theorem 4.14) can be explicitly
determined. On the other hand, we do not have a closed formula for its real version α(k, m),
which is captured by the expected degree of Grassmannians. As a consequence, it turns out that
classical results of enumerative geometry over C can be obtained via the computation of volumes
and the evaluation of integrals. We illustrate this general observation by the computation of the
degree of GC (k, n) in the proof of Corollary 4.15.
Remark 1.3. It is interesting to observe that over the complex numbers the degree of GC (k, n) can
be obtained in two different ways: (1) intersecting the Grassmannian in the Plücker embedding
with N = k(n − k) generic hyperplanes; (2) intersecting N many generic copies of the Schubert
variety ΣC (k, n) inside GC (k, n) itself. The second method is equivalent to intersecting the image
of the Plücker embedding with N very nongeneric hyperplanes.
6 PETER BÜRGISSER AND ANTONIO LERARIO

By contrast, over the real numbers the two procedures give very different answers: (1) av-
eraging the intersection of the real Grassmannian in the Plücker embedding with N random
hyperplanes gives |G(k,n)|
|RPN | (by classical integral geometry); (2) averaging the intersection of N
random copies of Σ(k, n) produces an intrinsic answer and gives the expected degree.
1.5. Link to random convex bodies. The proof of our results on the expected degree em-
ploys an interesting and general connection between expected absolute random determinants
and random polytopes that are zonoids (Vitale [55]). This allows to express the average angle
between random Schubert varieties in terms of the volume of certain convex bodies, for which
we coined the name Segre zonoids C(k, m), see §5.1. These are zonoids of matrices in Rk×m that
are invariant under the action of the group O(k) × O(m). Via the singular value decomposition,
the Segre zonoid fibers over a zonoid D(k) of singular values in Rk (if k ≤ m) and its volume
can be studied this way. Via a variant of Laplace’s method [16] we prove that, asymptotically
for m → ∞ and up to a subexponential factor eo(m) , the volume of C(k, m) equals the volume of
the smallest ball including C(k, m). The volume of this ball is not the same, but related to what
one gets when carrying out the analogous argument for computing the degree of the complex
Grassmann manifold.

1.6. Related work. A square root law (in the form of an equality) was for the first time dis-
covered by Kostlan [30] and Shub and Smale [44], who found a beautiful result on the expected
number of real solutions to random polynomial systems. This work has strongly inspired ours.
Paul Breiding [5] has recently discovered results in the same spirit for the (expected) number
of (real) eigenvalues of tensors. Related to this work, the second author of the present paper,
together with S. Basu, E. Lundberg and C. Peterson have recently obtained results on counting
real lines on random real hypersurfaces [3], obtaining a similar square root law (already con-
jectured by the first author of the current paper). When the zero set of a random system of
polynomial equations is not zero dimensional, its size may be measured either in terms of volume
or Euler characteristic, which was investigated by the first author of the current paper in [7].
We refer to the papers [12] and [11] that investigate the average number of real solutions in
different contexts. More recently, the subject of random algebraic geometry, meant as the study
of topological properties of random real algebraic sets, has become very popular. Asymptotic
square root laws have been found for Betti numbers of Kostlan hypersurfaces by Gayet and
Welschinger [19, 20, 21] (and in general for invariant models by the second author of the current
paper, Fyodorov and Lundberg [18]); similar results for Betti numbers of intersection of random
quadrics have been obtained by the second author of the current paper and Lundberg [32, 33].

1.7. Acknowledgments. We are very grateful to Frank Sottile who originally suggested this
line of research. We thank Paul Breiding, Kathlén Kohn, Chris Peterson, and Bernd Sturmfels
for discussions. We also thank the anonymous referees for their comments.

2. Preliminaries
2.1. Real Grassmannians. Suppose that E is a Euclidean L∞ vector space of dimension n. We
k
have an inner product on the exterior algebra Λ(E) = k=0 Λ (E), which in coordinates is
described as follows. Let e1 , . . . , en be an orthonormal basis of E and put eI := ei1 ∧ · · · ∧ eik
for I = {i1 , . . . , ik } with i1 < · · · < ik . Then the eI form an orthonormal basis of Λk (E), where
I runs over all subsets of [n] with |I| = k. Hence, for aI ∈ R,
X X
k aI e I k 2 = |aI |2 .
U I
 PROBABILISTIC SCHUBERT CALCULUS 7

Pp P
Let vi = j=1 xij ej with xij ∈ R. Then v1 ∧ · · · ∧ vk = I det(XI )eI , where XI is the k × k
submatrix obtained from X = (xij ) by selecting the columns indexed by the numbers in I.
Hence
X
(2.1) kv1 ∧ · · · ∧ vk k2 = | det(XI )|2 = det(hvi , vj i)1≤i,j≤k ,
I

where the second equality is the well known Binet-Cauchy identity. (Clearly, the above quantity
Pk Pp
does not change if we substitute the vi by j=1 gij vj with (gij ) ∈ SLk .) If wi = j=1 yij ej
with yij ∈ R, we obtain for the scalar product between two simple k-vectors (compare [31])

hv1 ∧ · · · ∧ vk , w1 ∧ · · · ∧ wk i = det(hvi , wj i)1≤i,j≤k .

Let G (k, E) denote the set of oriented k-planes in E. Let L ∈ G+ (k, E) and v1 , . . . , vk be
+

an oriented orthonormal basis of E. Then v1 ∧ · · · ∧ vk is independent of the choice of the basis.

Via the (injective) Plücker embedding G+ (k, E) → Λk (E), L 7→ v1 ∧ · · · ∧ vk we can identify
G+ (k, E) with the set of norm-one, simple k-vectors in Λ(E). We note that the orthogonal
group O(E) acts isometrically on Λ(E) and restricts to the natural action on G+ (k, E) that,
for g ∈ O(E), assigns to an oriented k-plane L its image gL. By this identification, G+ (k, E)
is a smooth submanifold of Λk (E) and inherits an O(E)-invariant Riemannian metric from the
ambient euclidean space. We denote by |U | the volume of a measurable subset of G+ (k, E).
The real Grassmann manifold G(k, E) is defined as the set of k-planes in E. We can identify
G(k, E) with the image of G+ (k, E) under the quotient map q : S(Λ(E)) → P(Λ(E)) from the
unit sphere of Λ(E), which forgets the orientation. The Riemannian metric on G(k, E) is defined
as the one for which q : G+ (k, E) → G(k, E) is a local isometry. (Note that this map is a double
covering.)
The uniform probability measure on G(k, E) is defined by setting:
|V |
Prob(V ) := , for all measurable V ⊆ G(k, E).
|G(k, E)|
We just write G(k, n) := G(k, Rn ) and G+ (k, n) := G+ (k, Rn ) where E = Rn has the standard
inner product.
The Plücker embedding allows to view the tangent space TA G(k, E) at A ∈ G(k, n) as a
subspace of Λk (E). While this is sometimes useful for explicit calculations, it is often helpful to
take a more invariant viewpoint. We canonically have TA G(k, n) = Hom(A, A⊥ ), compare [25,
Pk
Lect. 16]. To α ∈ Hom(A, A⊥ ) there corresponds the tangent vector i=1 a1 ∧ · · · ∧ ai−1 ∧
α(ai ) ∧ ai+1 ∧ · · · ∧ ak , where A = a1 ∧ · · · ∧ ak . Moreover, if g∗ : G(k, E) → G(k, E) denotes
the action corresponding to g ∈ O(E), then its derivative DA g∗ : TA G(k, E) → Tg(A) G(k, E) at
A ∈ G(k, E) is the map Hom(A, A⊥ ) → Hom(g(A), g(A)⊥ ), α 7→ g ◦ α ◦ g −1 . It is easy to verify
that the norm of a tangent vector α ∈ Hom(A, A⊥ ) in the previously defined Riemannian metric
equals the Frobenius norm of α.
Remark 2.1. It is known that (see [31, Section 6.4]) that for (k, n) 6= (2, 4), there is a unique
(up to multiples) O(n)-invariant Riemannian metric on G+ (k, n); moreover for all k, n there is a
unique (up to multiples) O(n)-invariant volume form and consequently a unique O(n)-invariant
probability distribution on G+ (k, n).
2.2. Regular points of semialgebraic sets. We refer to [4] as the standard reference for real
algebraic geometry and recall here some not so well known notion from real algebraic geometry.
Let M be a smooth real algebraic variety, e.g., Rn or a real Grassmannian, and let S ⊆ M
be a semialgebraic subset. A point x ∈ S is called regular of dimension d if x has an open
8 PETER BÜRGISSER AND ANTONIO LERARIO

neighborhood U in M such that S ∩ U is a smooth submanifold of U of dimension d. The

dimension of S can be defined as the maximal dimension of regular points in S (which is well
defined since regular points always exist). The points which fail to be regular are called singular.
We denote the set of singular points by Sing(S) and say that S is smooth if Sing(S) = ∅.
Moreover, we write Reg(S) := S \ Sing(S). The following is stated in [10, §4.2] without proof,
refering to [34]. A proof can be found in [50].
Proposition 2.2. The set Sing(S) is semialgebraic and dim Sing(S) < dim S. In particular,
S 6= Sing(S).
In the sequel it will be convenient to say that generic points of a semialgebraic set S satisfy
a certain property if this property is satisfied by all points except in a semialgebraic subset of
positive codimension in S. (For example, in view of the previous proposition, generic points of
a semialgebraic set S are regular points.)
2.3. Complex Grassmannians. The complex Grassmann manifold GC (k, n) is defined as the
set of complex k-planes in Cn . It can be identified with its image under the (complex) Plücker
embedding ι : GC (k, n) → P(Λ(Cn )), that is defined similarly as over R in §2.1. We note that the
unitary group U (n) acts isometrically on Λ(Cn ) and restricts to the natural action on GC (k, n)
that, for g ∈ U (n), assigns to a complex k-plane L its image gL. Since the image of ι is Zariski
closed, we can view GC (k, n) as a complex projective variety, which inherits an U (n)-invariant
Hermitian metric from the ambient space. The real part of this Hermitian metric defines a
Riemannian metric on GC (k, n). We denote by |V | the volume of a measurable subset V of
GC (k, n). The degree of GC (k, n) as a projective variety is well known, see Corollary 4.15.
Let H ⊆ GC (k, n) be an irreducible algebraic hypersurface in GC (k, n). It is known that the
vanishing ideal of H in the homogeneous coordinate ring of GC (k, n) is generated by a single
equation in the Plücker coordinates; cf. [22, Chap. 3, Prop. 2.1]. We shall call its degree the
relative degree rdegH of H. This naming is justified by the following observation. The image ι(H)
of H under the Plücker embedding ι is obtained by intersecting ι(GC (k, n)) with an irreducible
hypersurface of degree rdegH and hence, by Bézout’s theorem,
(2.2) deg ι(H) = rdegH · deg GC (k, n).
For dealing with questions of incidence geometry, we introduce the Chow variety ZC (X) ⊆
GC (k, n) associated with an irreducible complex projective variety X ⊆ CPn−1 of dimension
n − k − 1. One defines ZC (X) as the set of k-dimensional linear subspaces A ⊆ Cn such that
their projectivization satisfies P(A) ∩ X 6= ∅. Is it known that Z(X) is an irreducible algebraic
hypersurface in GC (k, n); moreover rdegZC (X) = deg X; cf. [22, Chap. 3, §2.B]. If we choose X
to be a linear space, then ZC (X) is a special Schubert variety of codimension one that we denote
by ΣC (k, n) (notationally ignoring the dependence on X).
2.4. Degree and volume. A fundamental result in complex algebraic geometry due to Wirtinger
states that degree and volume of an irreducible projective variety X ⊆ CPn are linked via
(2.3) |X| = deg X · |CPm |, where m = dim X;
see [38, §5.C] or [43, Chap. VIII, §4.4]. This is a key observation, since the interpretation of the
degree as a volume ratio paves the way for arriving at analogous results over the reals.
Equation (2.3), combined with the fact that the Plücker embedding is isometric, implies
(2.4) |GC (k, n)| = deg GC (k, n) · |CPN |,
where N := dim GC (k, n) = k(n − k). If X ⊆ CPn−1 is an irreducible projective variety
of dimension n − k − 1, we obtain for the Chow hypersurface ZC (X) by the same reasoning,
 PROBABILISTIC SCHUBERT CALCULUS 9

using (2.2),
(2.5) |ZC (X)| = deg ι(ZC (X)) · |CPN −1 | = rdegZC (X) · deg GC (k, n) · |CPN −1 |.
In particular, taking for X a linear space, we get for the special Schubert variety:
(2.6) |ΣC (k, n))| = deg GC (k, n) · |CPN −1 |.
Dividing this equation by (2.4) we obtain for the volume of ΣC (k, n), cf. [8]:
|ΣC (k, n)| |CPN −1 | N |CPk−1 | |CPn−k−1 |
(2.7) = N
= =π· · .
|GC (k, n))| |CP | π |CPk | |CPn−k |
We remark that this formula can also by obtained in a less elegant way by calculus only (see
first version of [8]). In Theorem 4.2 we derive a real analogue of (2.7), based on the calculation
of the volume of the tube around the special Schubert variety Σ(k, n).
We can easily derive more conclusions from the above: dividing the equations (2.5) and (2.6),
we get:
|ZC (X)| |X|
(2.8) = rdegZC (X) = deg X = .
|ΣC (k, n))| |CPdim X |
In Proposition 4.8 we derive an analogous result over R.
2.5. Intersecting hypersurfaces in complex Grassmannians. For the sake of comparison,
we describe the classical intersection theory in the restricted setting of intersecting hypersurfaces
of GC (k, n).
Many results in enumerative geometry over C can be obtained from the following well known
consequence of Bézout’s theorem. In Theorem 4.10 we provide a real analogue of this result.
Theorem 2.3. Let H1 , . . . , HN be irreducible algebraic hypersurfaces in GC (k, n), where N :=
k(n − k). Then we have:
#(g1 H1 ∩ . . . ∩ gN HN ) = deg GC (k, n) · rdegH1 · . . . · rdegHN
for almost all (g1 , . . . , gN ) ∈ (GLn )N . In particular, if X1 , . . . , XN ⊆ Pn (C) are irreducible
projective subvarieties of dimension n − k − 1, then:
#(g1 ZC (X1 ) ∩ . . . ∩ gN ZC (XN )) = deg GC (k, n) · deg X1 · . . . · deg XN
for almost all (g1 , . . . , gN ) ∈ (GLn )N .
Proof. A general result by Kleiman [27] implies that g1 H1 , . . . , gN HN meet transversally, for
almost all g1 , . . . , gn . Recall the Plücker embedding ι : GC (k, n) ֒→ P(Λk Cn ). Note that ι(Hi ) is
obtained by intersecting ι(GC (k, n)) with an irreducible hypersurface Hi in P(Λk Cn ) of degree
rdegHi . We thus have ι(g1 H1 ) ∩ . . . ∩ ι(g1 HN ) = ι(GC (k, n)) ∩ g1 H1 ∩ . . . ∩ gN HN and the
assertion follows from Bézout’s theorem. 
2.6. Convex bodies. For the following see [41]. By a convex body K we understand a nonempty,
compact, convex subset of Rd with nonempty interior. The support function hK of K is defined
as:
for u ∈ Rd .

hK (u) := max hx, ui | x ∈ K
One calls {x ∈ Rd | hx, ui = hK (u)} the supporting hyperplane in the direction u ∈ Rd \ {0}. The
support function hK is a positively homogeneous, subadditive function that, by the hyperplane
separation theorem, characterizes K as follows:
x ∈ K ⇐⇒ ∀u ∈ Rd hx, ui ≤ hK (u).
10 PETER BÜRGISSER AND ANTONIO LERARIO

If 0 ∈ K one defines the radial function rK of K by:

for u ∈ Rd \ {0}.

rK (u) := max t ≥ 0 | tu ∈ K
The radius kKk of K is defined as the maximum of the function rK on the unit sphere S d−1 .
Lemma 2.4. Let K ⊆ Rd be a convex body containing the origin. Then the radial function rK
and the support function hK of K have the same maximum on S d−1 . Moreover, a direction
u ∈ S d−1 is maximizing for rK if and only u is maximizing for hK .
Proof. Let r denote the maximum of rK . Then K ⊆ B(0, r) and hence hK ≤ hB(0,r) , which
implies max hK ≤ max hB(0,r) = r. For the other direction let u ∈ S d−1 . Then hx, ui ≤ hK (u) for
all x ∈ K. In particular, setting x := rK (u)u, which lies in K, we obtain rK (u) = hrK (u)u, ui ≤
hK (u), and consequently max rK ≤ max hK . The claim about the maximizing directions follows
easily by tracing our argument. 

We recall also the following useful fact [41, Corollary 1.7.3].

Proposition 2.5. Let K ⊆ Rd be a convex body and u ∈ Rd \{0}. Then the support function hK
of K is differentiable at u if and only if the intersection of K with the supporting hyperplane
of K in the direction u contains only one point x. In this case x = (∇hK )(u). In particular, if
S denotes the set of nonzero points at which hK is smooth, we have ∇hK (S) ⊆ ∂K.
The set Kd of convex bodies in Rd can be turned into a metric space by means of the Hausdorff
metric. The Hausdorff distance of the sets K, L ∈ Kd is defined by:
δ(K, L) := min t ≥ 0 | K ⊆ L + tB d and L ⊆ K + tB d ,


where B d denotes the closed unit ball and + the Minkowski addition.
We introduce now special classes of convex bodies that will naturally arise in our work. A
zonotope is the Minkowski sum of finitely many line segments. A zonoid is obtained as the limit
(with respect to the Hausdorff metric) of a sequence of zonotopes. It is easy to see that closed
balls are zonoids.

2.7. The zonoid associated with a probability distribution. The following is from Vi-
tale [55]. A random convex body X is a Borel measurable map from a probability space to Kd .
Suppose that the radius kXk of X has a finite expectation. Then one can associate with X its
expectation E X ∈ Kd , which is characterized by its support function as follows:
(2.9) hE X (u) = E hX (u) for all u ∈ Rd .
Suppose now Y is a random vector taking values in Rd . We associate with Y the zonoid CY :=
E [0, Y ] in Rd , which is defined if E kY k < ∞. In the following we assume that Y and −Y have
the same distribution. By (2.9), the support function h of CY is given by:
1
(2.10) hCY (u) = E |hu, Y i| for u ∈ Rd .
2
Example 2.6. Suppose that Y ∈ Rd is standard Gaussian. The support function h1 of CY
satisfies by (2.10):
1 1 1
h1 (e1 ) = E |he1 , Y i| = E |Y1 | = √ .
2 2 2π
√
The function h1 is O(d)-invariant and hence√h1 (u) = kuk/ 2π. It follows that the associated
zonoid CY is the ball √12π B d with radius 1/ 2π and centered at the origin.
 PROBABILISTIC SCHUBERT CALCULUS 11

The random variable Ỹ := Y /kY k is uniformly distributed in the sphere S d−1 and Ỹ and kY k
are independent. Hence the support function h2 of CỸ satisfies:
1 |hu, Y i| 1
h2 (u) = E = h1 (u).
2 kY k ρd
Hence the zonoid associated with Ỹ equals the ball ρ1d √12π B d . Therefore, if we sample uniformly
in the sphere of radius ρd around the origin, the associated zonoid is the same as for the standard
normal distribution, namely √12π B d .

Let MY ∈ Rd×d be the random matrix whose columns are i.i.d. copies of Y . Our goal is to
analyze the expectation of the absolute value of the determinant of MY . The following result
due to Vitale [55] is crucial for our analysis in §5.
Theorem 2.7. If Y is a random vector taking values in Rd such that E kY k < ∞, then
E |MY | = d! |CY |.
More information on this can be found in Schneider’s book [41, §3.5], where CY is called the
zonoid whose generating measure is the distribution of Y . (This is under the assumption that
Y ∈ S d−1 .) More results in the spirit of Theorem 2.7 can be found in [41, §5.3].
2.8. Various volumes. In order to express the volume of orthogonal groups etc., we introduce
the multivariate Gamma function Γk (a), defined by:
Z
k+1
Γk (a) := e−trA (det A)a− 4 dA,
A>0

for ℜ(a) > 21 (k − 1), where the integral is over the cone of positive definite k × k real symmetric
matrices. This clearly generalizes the classical Gamma function: Γ1 (a) = Γ(a). It is known
that [37, Theorem 2.1.12]:
Γk (a) = π k(k−1)/4 Γ(a)Γ(a − 1/2) · · · Γ(a − (k − 1)/2).
The orthogonal group O(k) := {Q ∈ Rk×k | QT Q = 1} is a compact smooth submanifold
of Rk×k . Its tangent space at 1 is given by the space of skew-symmetric matrices T1 O(k) =
{A ∈ Rk×k | AT + A = 0}, on which we define the inner product hA, Bi := 21 tr(AT B) for
A, B ∈ T1 O(k). We extend this to a Riemannian metric on O(k) by requiring that the left-
multiplications are isometries and call the resulting metric the canonical one. (Note that this
differs by a factor 21 from the metric on O(k) induced by the Euclidean metric on Rk×k .) Then,
for v in the unit sphere S k−1 := {x ∈ Rk | kxk = 1}, the orbit map O(k) → S k−1 , Q 7→ Qv is a
Riemannian submersion.
Similarly, if G(k, n) is given the Riemannian metric defined above, it is easy to verify that
the quotient map O(n) → G(k, n) is a Riemannian submersion. We also define a Riemannian
metric on the Stiefel manifold S(k, m) := {Q ∈ Rm×k | QT Q = 1} for 1 ≤ k ≤ m such that
O(m) → S(k, m), Q = [q1 , . . . , qm ] 7→ [q1 , . . . , qk ] is a Riemannian submersion, where qi denotes
the ith column of Q. The following is well known [37, Thm. 2.1.15, Cor. 2.1.16]:
k2 +k k2 km
2k π 4 2k π 2 2k π 2
(2.11) |O(k)| = =
, |S(k, m)| =
.
Γ(k/2)Γ((k − 1)/2) · · · Γ(1/2) Γk 2 k
Γk m 2
k
In particular, we note that |S k−1 | = 2π 2 /Γ k2 .

The unitary group U (k) := {Q ∈ Ck×k | Q∗ Q = 1} is a compact smooth submanifold of

k×k
C . Its tangent space at 1 is given by the space of skew-hermitian matrices
T1 U (k) = {A ∈ Ck×k | A∗ + A = 0}.
12 PETER BÜRGISSER AND ANTONIO LERARIO

We define an inner product on T1 U (k) by setting for A = [aij ], B = [bij ]:

X 1X X X
(2.12) hA, Bi := ℑ(aii )ℑ(bii ) + aij b̄ij = ℑ(aii )ℑ(bii ) + aij b̄ij ,
i
2 i i<j
i6=j

where ℑ(z) denotes the imaginary part of z ∈ C. We extend this to a Riemannian metric
on U (k) by requiring that the left-multiplications are isometries and call the resulting metric
the canonical one. It is important to realize that this metric is essentially different from the
Riemannian metric on U (k) that is induced by the Euclidean metric of Ck×k . (The reason
is the contribution in (2.12) from the imaginary elements on the diagonal; in the analogous
situation of the orthogonal group, the Riemannian metrics differ by a constant factor only.) A
useful property of the canonical metric (whose induced volume form is proportional to the Haar
measure) is that for v in the unit sphere S(Ck ) of Ck , the orbit map U (k) → S(Ck ), Q 7→ Qv is
a Riemannian submersion. This implies |U (k)| = |U (k − 1)| · |S(Ck )| and we obtain from this:
k2 +k
2k π 2
(2.13) |U (k)| = Qk−1 .
i=1 i!
We also state the following well known formulas:
|O(n)| |U (n)|
(2.14) |G(k, n)| = , |GC (k, n)| = .
|O(k)| · |O(n − k)| |U (k)| · |U (n − k)|
2.9. Some useful estimates. We define ρk := E kXk, where X ∈ Rk is a standard normal
Gaussian k-vector. It is well known that:
√
2 Γ( k+12 )
(2.15) ρk = .
Γ( k2 )
E.g., ρ2 = π2 . For the following estimate see [9, Lemma 2.25]:
p

√ √
r
k
(2.16) · k ≤ ρk ≤ k.
k+1
√
Hence ρk is asymptotically equal to k.
Lemma 2.8. For fixed k and n → ∞ we have:
1 1
log |G(k, n)| = − kn log n + kn log(2eπ) + O(log n),
2 2
log |GC (k, n)| = −kn log n + kn log(eπ) + O(log n),
log deg GC (k, n) = kn log k + O(log n).
|O(n)|
Proof. By (2.14) we have |G(k, n)| = |O(k)|·|O(n−k)| . From (2.11) we get:
n
|O(n)| kn k−k2 Y 1
= 2k π 2 + 4 · .
|O(n − k)| Γ(j/2)
j=n−k+1

Consequently, using the asymptotic log Γ(x) = x log x − x + O(log x) as x → ∞, we get:

|O(n)| 1
log = kn log π − k log Γ(n/2) + O(log n).
|O(n − k)| 2
1 n n n
= kn log π − k log − + O(log n)
2 2 2 2
kn 1 1 1
= log π − kn log n + kn log 2 + kn + O(log n),
2 2 2 2
 PROBABILISTIC SCHUBERT CALCULUS 13

from which the assertion on log |G(k, n)| follows.

For |GC (k, n)| we argue similarly. Using (2.13), we get:
k−1
|U (n)| k−k2 Y 1
= 2k π kn+ 2 .
|U (n − k)| i=0
(n − k + i)!

and the assertion on log |GC (k, n)| follows.

The third assertion follows easily from the formula for deg GC (k, n) in Corollary 4.15. 

3. Integral geometry of real Grassmannians

3.1. Joint density of principal angles between random subspaces. Let A ≃ Rk and B ≃
Rℓ be two subspaces of Rn with corresponding orthonormal bases (a1 , . . . , ak ) and (b1 , . . . , bℓ ).
By slightly abusing notation we denote by A = [a1 , . . . , ak ] and B = [b1 , . . . , bℓ ] also the matrices
with the columns ai and bj , respectively. By the singular value decomposition theorem, there
exist U ∈ O(k) and V ∈ O(ℓ) such that:
 
D 0
(3.1) U T AT BV =
0 0
where D = diag(σ1 , . . . , σr ) and r = min{k, ℓ}; we adopt the convention 1 ≥ σ1 ≥ · · · ≥ σr ≥ 0.
We recall that the principal angles 0 ≤ θ1 ≤ · · · ≤ θr ≤ π/2 between A and B are defined by
σi = cos θi for i = 1, . . . , r. Clearly, they are independent of the choice of the orthonormal bases.
We note that m := dim(A ∩ B) equals the number of zero principal angles between A and B.
Lemma 3.1. There exist orthonormal bases (a1 , . . . , ak ) and (b1 , . . . , bℓ ) of A and B, respec-
tively, such that ai = bi for i ≤ m and hai , bj i = δij cos θi for all i and j.
Proof. Put C := A ∩ B. If C = 0, then the columns of the matrices AU and BV satisfy the
desired property. In the general case, we apply this argument to the orthogonal complement
of C in A and B. 

The following theorem is a generalization of [1, Equation (3)] and can be found in [2, Appendix
D.3]. We present a different and considerably shorter proof in Appendix 7.1.
Theorem 3.2. Let k ≤ ℓ and k + ℓ ≤ n. We fix an ℓ-dimensional subspace B ⊆ Rn and sample
A ∈ G(k, n) with respect to to the uniform distribution. Then the joint probability density of the
principal angles 0 ≤ θ1 ≤ · · · ≤ θk ≤ π/2 between A and B is given by
k
Y Y
(cos θj )l−k (sin θj )n−l−k (cos θi )2 − (cos θj )2 ,


(3.2) p(θ1 , . . . , θk ) = ck,l,n
j=1 i<j

where
k2
2k π 2 Γk n2


ck,l,n := k

l

n−l

.
Γk 2 Γk 2 Γk 2

Example 3.3. For a fixed 2-plane B ⊆ R4 , the joint density of the principal angles 0 ≤ θ1 ≤ θ2 ≤
π/2 between B and a uniform A ∈ G(2, 4) is given by p(θ1 , θ2 ) = 2((cos θ1 )2 − (cos θ2 )2 ) and we
can easily verify that
Z π/2 Z π/2
2 (cos θ1 )2 − (cos θ2 )2 dθ2 dθ1 = 1.


0 θ1
14 PETER BÜRGISSER AND ANTONIO LERARIO

More generally, the joint density p(θ1 , θ2 ) of the principal angles 0 ≤ θ1 ≤ θ2 ≤ π/2 between a
random 2-plane in G(2, n + 1) and a fixed subspace of Rn+1 of dimension n − 1 is given by
p(θ1 , θ2 ) = (n − 1)(n − 2)(cos θ1 )n−3 (cos θ2 )n−3 (cos θ1 )2 − (cos θ2 )2 .

3.2. Subvarieties of Grassmannians with transitive action on tangent spaces. The or-
thogonal group O(n) acts transitively and isometrically on G(k, n). For an element g ∈ O(n), let
g∗ : G(k, n) → G(k, n) denote the corresponding action and DA g∗ : TA G(k, n) → Tg∗ (A) G(k, n)
its derivative at A ∈ G(k, n). See §2.2 for the definition of generic points.
Definition 3.4. Let Z ⊆ G(k, n) be a semialgebraic set. We say that Z has transitive action
on its tangent spaces if, for generic regular points A1 , A2 of Z, there exists g ∈ O(n) such that
g∗ (A1 ) = A2 and DA1 g∗ (TA1 Z) = TA2 Z.
Example 3.5. Consider the subvariety Ω := {A ∈ G(k, n) | Rk−1 ⊆ A ⊆ Rk+1 }, which is
isomorphic to a real projective line. For determining the tangent space of Ω at A we can assume
w.l.og. that A = Rk . Then the tangent vectors α ∈ Hom(A, A⊥ ) to Ω are characterized by
α(Rk−1 ) = 0 and α(ek ) ∈ Rek+1 , so that TA Ω = R ek ⊗ ek+1 . This implies that Ω has transitive
action on its tangent spaces.
Tangent spaces of Grassmannians have a product structure: in fact we have a canonical
isometry (cf. §2.1):
TA G(k, n) ≃ Hom(A, A⊥ ) ≃ A ⊗ A⊥
and note that dim A = k and dim A⊥ = n − k.
We will focus on a special type of hypersurfaces, whose normal spaces exploit the product
structure of TA G(k, n).
Definition 3.6. A semialgebraic subset M ⊆ G(k, n) of codimension one is called a coisotropic
hypersurface, if, for generic points A ∈ M, the normal space NA M is spanned by a rank one
vector (see [22, §4.3] and [29]).
Lemma 3.7. A coisotropic hypersurface has transitive action on its tangent spaces.
Proof. Let A1 , A2 ∈ M be regular points. By assumption, there are unit length vectors ui ∈ A
and vi ∈ A⊥ for i = 1, 2 such that NA1 M = Ru1 ⊗ v1 and NA2 M = Ru2 ⊗ v2 . There is g ∈ O(n)
such that g(A1 ) = A2 , g(u1 ) = u2 , and g(v1 ) = v2 . Let g∗ : G(k, n) → G(k, n) denote the
multiplication with g. Then DA1 g∗ maps NA1 M to NA2 M, and hence it maps TA1 M to TA2 M
(compare §2.1). 

Remark 3.8. Proposition 4.6 below implies that codimension one Schubert varieties (and more
generally Chow hypersurfaces) have transitive actions on their tangent space. One can show
that the same is true for all Schubert varieties. This shows that the method of this paper in
principle extends to include random intersections of general Schubert varieties, as we plan to
discuss in a future work.
The integral geometry formula to be discussed involves a certain average scaling factor between
random subspaces, that we introduce next.

3.3. Relative position of real subspaces. The relative position of two subspaces V, W of a
Euclidean vector space E can be quantified by a volume like quantity that we define now. This
quantity is crucial in the study of integral geometry in homogeneous spaces (see [26]).
 PROBABILISTIC SCHUBERT CALCULUS 15

3.3.1. Relative position of two real subspaces. Suppose that E is a Euclidean vector space of
dimension n. We have an induced inner product on Λ(E), cf. §2.1. Let V, W be linear subspaces
of E of dimensions k, m, respectively, such that k + m ≤ n. We define the following quantity:
(3.3) σ(V, W ) := kv1 ∧ . . . ∧ vk ∧ w1 ∧ . . . ∧ wm k,
where v1 , . . . , vk and w1 , . . . , wm are orthonormal bases of V and W , respectively. (This is clearly
independent of the choice of the orthonormal bases.) Note that if k = m and oriented versions
of V and W are interpreted as simple k-vectors in Λk (E), then σ(V, W ) is the sine of the angle
between V and W (viewed as vectors in the Euclidean space Λk (E)). So the quantity σ(V, W )
in a sense measures the relative position of V, W . We can make this more precise by expressing
σ(V, W ) in terms of the principal angles between V and W .
Lemma 3.9. The following is true:
(1) σ(W, V ) = σ(V, W ).
(2) σ(V, W ) = sin θ1 · . . . · sin θk if k ≤ m and the θi are the principal angles between V
and W .
(3) 0 ≤ σ(V, W ) ≤ 1.
(4) We have σ(V, W ) = 0 iff V ∩ W 6= 0 and σ(V, W ) = 1 iff V and W are orthogonal.
Proof. The first assertion is obvious. For the second we use the well known fact (cf. [8, Corol-
lary 2.4]) that there is an orthonormal basis e1 , f1 , . . . , ek , fk , g1 , . . . , gn−2k of E such that V is
the span of e1 , . . . , ek and
W = span{e1 cos θ1 + f1 sin θ1 , . . . , ek cos θk + fk sin θk , g1 , . . . , gm−k }.
This implies the second claim. The remaining assertions follow immediately from this. 

3.3.2. Relative position of two complex subspaces. We briefly discuss how to extend the previous
discussion to complex spaces. Suppose that F is a hermitian vector space with n = dimC F and
V, W ⊆ F are C-subspaces of C-dimensions k and m, respectively. We note that an analogue
of Lemma 3.9 holds. As in (3.3) we define σC (V, W ) := kv1 ∧ . . . ∧ vk ∧ w1 ∧ . . . ∧ wm k where
v1 , . . . , vk and w1 , . . . , wm are orthonormal bases of V and W , respectively. On the other hand,
we can view F as a Euclidean vector space of dimension 2n with respect to the induced inner
product hv, wiR := ℜhv, wi. Then v1 , iv1 , . . . , vk , ivk is an orthonormal basis of V , seen as an
R-vector space (note that v and iv are orthogonal for all v ∈ F ). We conclude that

σC (V, W ) = σ(V, W )2 .

3.3.3. Average relative position of two subspaces in a tensor product. We assume now that E is
a product E = E1 ⊗ E2 of Euclidean vector spaces Ei such that hx ⊗ y, x′ ⊗ y ′ i = hx, x′ ihy, y ′ i for
x, x′ ∈ E1 , y, y ′ ∈ E2 . The product of orthogonal groups K := O(E1 ) × O(E2 ) acts isometrically
on E. Clearly, K is a compact group with a uniform invariant measure.
If V and W are given linear subspaces of E, we measure their average relative position by
moving V and W with uniformly random transformations k1 , k2 ∈ K and taking the average
value of σ(k1 V, k2 W ). Of course, it suffices to move one of the subspaces.
Definition 3.10. Let V and W be linear subspaces of E. The average scaling factor of V and W
is defined as σ̄(V, W ) := E k∈K σ(kV, W ), taken over a uniform k ∈ K = O(E1 ) × O(E2 ).
From the definition it is clear that σ̄(V, W ) only depends on the K-orbits of V and W ,
respectively.
16 PETER BÜRGISSER AND ANTONIO LERARIO

Remark 3.11. In many cases, V and W factor themselves into subspaces as V = V1 ⊗ V2 and
W = W1 ⊗ W2 . In this case, it is clear that σ̄(V, W ) only depends on the dimensions of the
involved spaces Vi , Wi , Ei , i = 1, 2.
3.4. Intersecting random semialgebraic subsets of Grassmannians. We now provide
tailor made formulations of general results in Howard [26] in the situation of real Grassmannians.
3.4.1. Intersecting two random semialgebraic sets. We begin with a general observation. Let A ∈
G(k, n) be the subspace spanned by the first k standard basis vectors, A := e1 ∧. . .∧ek ∈ G(k, n),
and note that the tangent space E := TA G(k, n) ≃ A ⊗ A⊥ can be seen as a tensor product of
vector spaces. The stabilizer subgroup of A in O(n) can be identified with K := O(A) × O(A⊥ )
and acts on TA G(k, n). Recall that DA g∗ : TA G(k, n) → TA G(k, n) denotes the derivative of the
action g∗ : G(k, n) → G(k, n) induced by g ∈ O(n). Moreover, NA M ⊆ TA G(k, n) denotes the
normal subspace of a semialgebraic subset M ⊆ G(k, n) at a regular point A.
Lemma 3.12. Suppose that M is a semialgebraic subset of G(k, n) with transitive action on
its tangent spaces. Let Bi ∈ M be generic regular points of M and gi ∈ O(n) such that Bi is
mapped to A, for i = 1, 2. Then DB1 g1 ∗ (NB1 M) and DB2 g2 ∗ (NB1 M) are subspace of TA G(k, n)
that lie in the same K-orbit.
Proof. There is h ∈ O(n) such that hB1 = B2 and DB1 h∗ maps TB1 M to TB2 M, since M
has transitive action on its tangent spaces. We have g1 = kg2 h, where k := g1 (g2 h)−1 ∈ K.
Moreover,
DB1 g1 ∗ (TB1 M) = DA k∗ (DB2 g2 ∗ (DB1 h∗ (TB1 M))) = DA k∗ (DB2 g2 ∗ (TB2 M)).
Therefore, DA k∗ , which can be seen as an element of K, maps DB2 g2 ∗ (NB2 M) to DB1 g1 ∗ (NB1 M).

We can now define an important quantity that enters the integral geometry formula to be
discussed.
Definition 3.13. Let M and N be semialgebraic subsets of G(k, n) with transitive action on
their tangent spaces. We pick a regular point B of M, a regular 
point C of N , and g, h ∈ O(n)

such that g(B) = A and h(C) = A. Then we define σ̄(M, N ) := σ̄ DB g∗ (NB M), DC h∗ (NC N ) .
It is important to note that σ̄(M, N ) does not depend on the choice of the regular points
B, C and the maps g, h and hence is well defined. This is a consequence of Lemma 3.12.
Remark 3.14. Suppose that the normal spaces NB M of M at regular points B of M are tensor
products NB M = VB ⊗ WB , with subspaces VB ⊆ B and WB ⊆ B ⊥ of a fixed dimension µ1
and µ2 respectively. Similarly, we assume that at regular points C of N we have NC N = VC′ ⊗WC′
with subspaces VC′ ⊆ C and WC′ ⊆ C ⊥ of fixed dimensions ν1 and ν2 , respectively. Then
σ̄(M, N ) only depends on the dimensions µ1 , µ2 , ν1 , ν2 and k, n; see Remark 3.11. For instance,
this assumption is satisfied if we are considering semialgebraic subsets of the projective space
RPn−1 = G(1, n).
The following result follows by combining Howard [26, Theorem 3.8] with Lemma 3.7.
Theorem 3.15. Let M and N be semialgebraic subsets of G(k, n) of codimension µ and ν
respectively, with transitive action on their tangent spaces. Then we have:
1
E g∈O(n) |M ∩ gN | = σ(M, N ) · · |M| · |N |.
|G(k, n)|
Here, |M ∩ gN | denotes the volume in the expected dimension k(n − k) − µ − ν, |M| and |N |
denote the volumes in the dimensions of M and N , respectively.
 PROBABILISTIC SCHUBERT CALCULUS 17

Remark 3.16. Howard [26, Theorem 3.8] states his result under the assumption of two smooth
compact submanifolds of G(k, n) (possibly with boundary). However, the compactness assump-
tion is not needed so that we can apply his result to the regular loci Reg(M) and Reg(N ) of
M and N , respectively. For this, note that the singular locus Sing(M) has dimension strictly
smaller than M (see §2.2). Moreover, E g∈G |M∩gN | = E g∈G |Reg(M)∩gReg(N )| since, almost
surely, the dimension of the difference (M ∩ gN ) \ (Reg(M) ∩ gReg(N )) is strictly smaller than
the typical dimension of M ∩ gN .
Example 3.17. Here is a typical application of Theorem 3.15 in the special case G(1, n) = RPn−1 .
It is clear that any semialgebraic subset M of G(1, n) has transitive action on its tangent spaces.
Let m := dim M. Applying Theorem 3.15 to M and N := RPn−m−1 , we obtain
1
E g∈O(n) #(M ∩ gRPn−m−1 ) = σ · · |M| · |RPn−m−1 |
|G(1, n)|
where the average scaling factor σ := σ(M, N ) only depends on m and n; see Remark 3.14. In
particular, this gives for M = RPm ,
1
E g∈O(n) #(RPm ∩ gRPn−m−1 ) = σ · · |RPm | · |RPn−m−1 |,
|G(1, n)|
the left-hand side of which clearly equals one. Solving this for σ and plugging it in in the previous
equation gives the well known formula (compare [9, A.55])
|M|
(3.4) E g∈O(n) #(M ∩ gRPn−m−1 ) = .
|RPm |
3.4.2. Intersecting many random coisotropic hypersurfaces. The previous discussion extends to
the situation of intersecting many semialgebraic sets; we focus here on the special case of inter-
secting coisotropic hypersurfaces in the Grassmannian, which is particularly simple. The general
case is discussed in the Appendix.
Let E be a Euclidean vector space of dimension Ps n. Let V1 , . . . , Vs be linear subspaces of the
dimensions m1 , . . . , ms , respectively, such that j=1 mj ≤ n. We define the quantity

(3.5) σ(V1 , . . . , Vs ) := kv11 ∧ . . . ∧ v1m1 ∧ . . . ∧ vs1 ∧ . . . ∧ vsms k,

where vj1 , . . . , vjmj are orthonormal bases of Vj . (This is clearly independent of the choice of
the orthonormal bases.) Note that σ(V1 , . . . , Vs ) ≤ 1. In the special case where mj = 1 for all j,
we can interpret σ(V1 , . . . , Vs ) as the volume of the parallelepiped spanned by the unit vectors
v11 , . . . , vs1 . In order to deal with the coisotropic case, we introduce the following definition.
Definition 3.18. For k, m ≥ 1 we define the (real) average scaling factor α(k, m) as
α(k, m) := E k(u1 ⊗ v1 ) ∧ . . . ∧ (ukm ⊗ vkm )k,
where uj ∈ S(Rk ) and vj ∈ S(Rm ), for 1 ≤ j ≤ km, are independently and uniformly chosen at
random.
Note that α(k, m) is nothing but the generalization to the case of many coisotropic submani-
folds of the quantity σ(M, N ) introduced in Definition 3.13. It has the special property that in
the coisotropic case, it does not depend on the choice of the hypersurfaces; compare Remark 3.14.
We will need the following result in the spirit of Theorem 3.15 about intersecting codimension
many coisotropic hypersurfaces in G(k, n).
18 PETER BÜRGISSER AND ANTONIO LERARIO

Theorem 3.19. Let H1 , . . . , HN be coisotropic hypersurfaces of G(k, n), where N := k(n − k).
Then we have:
N
Y |Hi |
E (g1 ,...,gN )∈O(n)N # (g1 H1 ∩ · · · ∩ gN HN ) dg1 · · · dgN = α(k, n − k) · |G(k, n)| · .
i=1
|G(k, n)|

This theorem is a consequence of a generalized Poincaré formula in homogeneous spaces,

stated as Theorem 7.2 in the Appendix 7.5. Its statement and proof is very similar to Howard [26,
Theorem 3.8]. Since we have been unable to locate this result in the literature, we provide a
proof in the Appendix.

4. Probabilistic enumerative geometry

4.1. Special Schubert varieties. By the special real Schubert variety associated with a linear
subspace B ∈ G(n − k, n) we understand the subvariety

Ω(B) := A ∈ G(k, n) | A ∩ B 6= 0 .
If the dependence on B is not relevant, which is mostly the case, we write Σ(k, n) := Ω(B). The
Schubert cell1 associated with B is the open subset

e(B) := A ∈ G(k, n) | dim(A ∩ B) = 1
of Ω(B). It is a well known fact that e(B) is the regular locus of Ω(B), e.g., see [36, 57].
Moreover, e(B) is a hypersurface in G(k, n). Its normal spaces can be described by the following
result, which is a special case of [45, §2.7].
Lemma 4.1. Let A ∈ e(B) and a1 , a2 , . . . , ak be an orthonormal basis of A such that A ∩ B =
Ra1 . Moreover, let (A + B)⊥ = Rf with kf k = 1. Then α ∈ Hom(A, A⊥ ) lies in the normal
space NA Ω(B) iff α(a1 ) ∈ Rf and α(ai ) = 0 for i > 1. In other words, f ∧ a2 ∧ · · · ∧ ak spans
the normal space NA Ω(B), when interpreted as a subspace of Λk Rn as in §2.1.
4.2. Volume of special Schubert varieties. We can determine now the volume of the spe-
cial Schubert varieties Σ(k, n). It is remarkable that the result is completely analogous to the
corresponding result (2.7) over C.
Theorem 4.2. The volume of the special Schubert variety Σ(k, n) satisfies:
Γ k+1 Γ n−k+1



|Σ(k, n)| 2
2
|RPk−1 | |RPn−k−1 |
= k
· n−k
=π· · .
|G(k, n)| Γ 2 Γ 2 |RPk | |RPn−k |

Proof. The isometry G(k, n) → G(n − k, n), A 7→ A⊥ maps Ω(B) to Ω(B ⊥ ). Therefore, we may
assume that k ≤ n/2 without loss of generality.
The proof relies on a general principle. Let M be a compact Riemannian manifold, H be a
smooth hypersurface of M , and K be a nonempty compact subset of H. We define the ǫ-tube
T ⊥ (K, ǫ) of K in M by
T ⊥ (K, ǫ) := expx (ν) | x ∈ K, ν ∈ Nx H, kνk ≤ ε ,


where expx : Tx M → M denotes the exponential map of M at x and Nx H is the (one dimen-
sional) orthogonal complement of the tangent space Tx H in Tx M ; cf. [9, §21.2] and [49]. Let

1We observe, as pointed out by an anonymous referee, that in general e(B) is not a topological cell, but instead
a vector bundle over a product of Grassmannians, see [57, Theorem 2.1].
 PROBABILISTIC SCHUBERT CALCULUS 19

m := dim H. The m-dimensional volume |K| of K can be computed from the m + 1-dimensional
volumes of the tubes T ⊥ (K, ǫ) as follows (see [23]):
1
(4.1) |K| = lim |T ⊥ (K, ǫ)|.
ǫ→0 2ǫ

We shall apply this formula to a Schubert cell H = e(B) for a fixed B ∈ G(n − k, n) in the
Grassmann manifold M = G(k, n) and K belonging to an increasing family of compact sets
exhausting e(B) (see below).
Let us denote by ϑ1 , ϑ2 : G(k, n) → R the functions giving, respectively, the smallest and the
second smallest principal angle between A ∈ G(k, n) and the fixed B. We note that e(B) =
{A ∈ G(k, n) | ϑ1 (A) = 0}. For 0 < δ < π/2, we consider the following compact subset of e(B):

e(B)δ := A ∈ G(k, n) | ϑ1 (A) = 0, ϑ2 (A) ≥ δ .
S
Note that e(B)δ2 ⊆ e(B)δ1 if δ1 < δ2 and e(B) = δ>0 e(B)δ . Therefore, we have |Ω(B)| =
|e(B)| = limδ→0 |e(B)δ |. Combining this with (4.1), we can write the volume of the Schubert
variety as the double limit
1
(4.2) |Ω(B)| = lim |e(B)δ | = lim lim |T ⊥ (e(B)δ , ǫ)|.
δ→0 δ→0 ǫ→0 2ǫ
We need the following technical result, whose proof will be provided in Appendix 7.2.
Lemma 4.3. In the above situation, we have for 0 < ε ≤ δ < π/2 that
T ⊥ (e(B)δ , ǫ) = {A ∈ G(k, n) | ϑ1 (A) ≤ ǫ, ϑ2 (A) ≥ δ .
Let now p(θ1 , . . . , θk ) denote the joint density of the (ordered) principal angles θ1 ≤ · · · ≤ θk
between A ∈ G(k, n) and B, as in Theorem 3.2. Due to Lemma 4.3 we can write for ǫ ≤ δ < π/2,
|T ⊥ (e(B)δ , ǫ))|
Z

= Prob A ∈ G(k, n) | ϑ1 (A) ≤ ǫ, ϑ2 (A) ≥ δ = p(θ) dθ,
|G(k, n)| R(ǫ,δ)

where Prob refers to the uniform distribution on G(k, n) and

R(ǫ, δ) := θ ∈ Rk | 0 ≤ θ1 ≤ . . . ≤ θk ≤ π/2, 0 ≤ θ1 ≤ ǫ, δ ≤ θ2 .


Thus, from (4.2) we see that

|Ω(B)| 1 1
Z
= lim lim p(θ) dθ,
|G(k, n)| 2 δ→0 ǫ→0 ǫ R(ǫ,δ)

and we have reduced the problem to the evaluation of the last limit. We have
!
1 1 ǫ
Z Z Z
lim lim p(θ) dθ = lim lim p(θ) dθ2 · · · dθk dθ1
δ→0 ǫ→0 ǫ R(ǫ,δ) δ→0 ǫ→0 ǫ 0 θ1 ≤δ≤θ2 ≤...≤θk ≤π/2
Z
= lim p(0, θ2 , . . . , θk ) dθ2 · · · dθk
δ→0 0≤δ≤θ ≤...≤θ ≤π/2
2 k
Z
= p(0, θ2 , . . . , θk ) dθ2 · · · dθk .
0≤θ2 ≤...≤θk ≤π/2

Theorem 3.2 on the joint density p gives

k
Y k
Y Y
(cos θj )n−2k 1 − (cos θj )2 (cos θi )2 − (cos θj )2



p(0, θ2 , . . . , θk ) = ck,n−k,n ·
j=2 j=2 2≤i<j≤k
ck,n−k,n
= · p̃(θ2 , . . . , θk ),
ck−1,n−k+1,n
20 PETER BÜRGISSER AND ANTONIO LERARIO

where p̃(θ2 , . . . , θk ) is the joint density of the principal angles between a random (k − 1)-plane
and a fixed (n − k − 1)-plane in Rn . We conclude that
Γ k+1 Γ n−k+1



|Ω(B)| 1 ck,n−k,n 2
2

= = · ,
|G(k, n)| 2 ck−1,n−k−1,n Γ k2 Γ n−k
2
where the last equality follows by a tedious calculation after plugging in the formulas for the
constants ck,n−k,n and ck−1,n−k−1,n from Theorem 3.2. 
4.3. Chow hypersurfaces. We generalize now the definition of special Schubert varieties of
codimension one. Let X ⊆ RPn−1 be a semialgebraic subset of dimension n − k − 1. We
associate with X the set of (k − 1)-dimensional projective subspaces of RPn−1 that intersect X.
We interpret this as a subset of G(k, n) and define
Z(X) := {A ∈ G(k, n) | P(A) ∩ X 6= ∅} .
In analogy with the situation over C, we call Z(X) the Chow hypersurface associated with X.
(The fact that Z(X) is a hypersurface, i.e., a semialgebraic subset of codimension one, will be
proved in Lemma 4.5 below.) We note that Z(X) = Ω(B) if X is projective linear and B the
corresponding linear space.
Example 4.4. Let X ⊆ RP3 be the real twisted cubic, i.e., the image of the map RP1 →
RP3 , (s : t) 7→ (s3 : s2 t : st2 : t3 ). We claim that Z(X) is not an algebraic subset of G(2, 4). For
seeing this, we intersect the affine part Xaff := {(t, t2 , t3 ) | t ∈ R} of X with the open subset
U := {Lu1 u2 v1 v2 | (u1 , u2 , v1 , v2 ) ∈ R4 } ⊆ G(2, 4) of lines
Lu1 u2 v1 v2 := {(0, u1 , u2 ) + s(1, v1 , v2 ) | s ∈ R} .
We have
Xaff ∩ U 6= ∅ ⇐⇒ ∃t ∈ R t2 − v1 t − u1 = 0, t3 − v2 t − u2 = 0.
The right-hand condition is equivalent to v22 + 4u2 ≥ 0, res = 0, where
res := u1 v12 v2 − u2 v13 − u31 + 2u21 v2 − 3u1 u2 v1 − u1 v22 + u2 v1 v2 + u22 = 0,
is the resultant of the above quadratic and cubic equation. Intersecting with the 2-plane u2 =
v2 = 0 gives {(u2 , v2 ) ∈ R2 | v22 + 4u2 ≥ 0}, which is not algebraic.
It is essential that the tangent spaces of Chow hypersurfaces coincide with the tangent spaces
of closely related special Schubert varieties. We postpone the somewhat technical proof to
Appendix 7.3.
Lemma 4.5. The Chow hypersurface Z(X) ⊆ G(k, n) is semialgebraic of codimension one; if
X is compact, then Z(X) is compact and if X is connected, then Z(X) is connected. Moreover,
for generic points A ∈ Z(X) the following is true: the intersection P(A) ∩ X consists of one
point only; let us denote this point by p; then p is a regular point of X and, denoting by B ⊆ Rn
the linear space of dimension n − k corresponding to the tangent space Tp X, we have:
(4.3) TA Z(X) = TA Ω(B).
The following result makes sure that we can apply the methods from integral geometry to
Chow hypersurfaces.
Proposition 4.6. Chow hypersurfaces are coisotropic and hence have transitive action on their
tangent spaces. In particular, this applies to special Schubert varieties of codimension one.
Proof. For codimension one special Schubert varieties, the assertion follows from Lemma 4.1.
The assertion on Chow hypersurfaces is a consequence of Lemma 4.5. 
 PROBABILISTIC SCHUBERT CALCULUS 21


Remark 4.7. Special Schubert varieties of higher codimension are defined by Ω(B, m) := A ∈
G(k, n) | dim(A ∩ B) ≥ m , where B is an ℓ-dimensional subspace such that k + ℓ − n ≤ m ≤ ℓ.
The codimension of Ω(B, m) equals m(n+ m− k − ℓ). For incorporating touching conditions, one
can extend the definition of the Chow hypersurface and study higher associated semialgebraic
sets,
Zm (X) := {A ∈ G(k, n) | ∃p ∈ Reg(X) dim Tp X ∩ Tp (P(A)) ≥ m}.
This is done in [22, §3.2, §4.3], [51], and [29] for complex projective varieties X in the case
dim X = n − k − 1 + m, where Zm (X) typically is a hypersurface. Note that Z0 (X) = Z(X).
The proof of Proposition 4.6 generalizes in a straightforward way to show that Ω(B, m) and
Zm (X) have transitive action on their tangent spaces.
As an application, we derive the following result that relates the volume of a semialgebraic
subset X ⊆ RPn−1 of dimension n − k − 1 with the volume of the associated Chow hypersurface
Z(X) ⊆ G(k, n). (This is analogous to equation (2.8) holding over C.)
Proposition 4.8. Let X be a semialgebraic subset of RPn−1 of dimension n − k − 1. Then we
have:
|Z(X)| |X|
= .
|Σ(k, n)| |RPdim X |
Proof. By (3.4) we can express the volume ratio |X|/|RPdim X | as the expectation E g∈O(n) #(X ∩
gRPk ). Consider now the subvariety Ω := {A ∈ G(k, n) | Rk−1 ⊆ A ⊆ Rk+1 } from Example 3.5,
which has transitive action on its tangent spaces. For almost all g, we have a finite intersection
X ∩ gRPk = {p1 , . . . , pd }. On the other hand, A ∈ gΩ implies A ⊆ gRk+1 , hence P(A) ∩ X =
{p1 , . . . , pd }. Therefore, we have for almost all g,
#(X ∩ gRPk ) = #(Z(X) ∩ gΩ),
and it remains to determine the expectation of the right-hand side.
The Chow hypersurface Z(X) has transitive action on its tangent spaces by Proposition 4.6.
Applying Theorem 3.15 to Z(X) and Ω, we obtain:
1
E g∈O(n) #(Z(X) ∩ gΩ) = σ(Z(X), Ω) · · |Z(X)| · |Ω|.
|G(k, n)|
Applying this formula to a linear space X yields:
1
(4.4) E g∈O(n) #(Σ(k, n) ∩ gΩ) = σ(Σ(k, n), Ω) · · |Σ(k, n)| · |Ω|.
|G(k, n)|
However, it is clear that the left-hand side equals one. According to Definition 3.13, the average
scaling factor σ(Z(X), Ω) is defined in terms of tangent (or normal) spaces of σ(Z(X)) and Ω.
Taking into account Lemma 4.5, we see that
σ(Z(X), Ω) = σ(Σ(k, n), Ω),
hence we can solve for this quantity and plugging this into (4.4) yields the assertion. 
4.4. Random incidence geometry. As already indicated in the introduction, we study the
following problem. We fix semialgebraic subsets X1 , . . . , XN ⊆ RPn−1 of the same dimension
dim Xi = n − k − 1, where 0 < k < n and N := k(n − k) and we ask how many (k − 1)-
dimensional projective linear subspaces intersect all random translations of the X1 , . . . , Xk(n−k) .
More specifically, the task is to determine the expectation E # (g1 Z(X1 ) ∩ · · · ∩ gN Z(XN )) with
respect to independent uniformly distributed gi ∈ O(n). In the case where all Xi are linear
subspaces, we call the answer the expected degree of G(k, n).
22 PETER BÜRGISSER AND ANTONIO LERARIO

Definition 4.9. The expected degree of G(k, n) is defined as the average number of intersection
points of N = k(n − k) many random copies of Σ(k, n), i.e.,
edeg G(k, n) := E #(g1 Σ(k, n) ∩ . . . ∩ gN Σ(k, n))
with respect to independent uniformly distributed gi ∈ O(n).
We note that edeg G(1, n) = 1. In Theorem 6.5 we will provide an asymptotically sharp upper
bound on edeg G(k, n).
Theorem 3.19 combined with the fact that Σ(k, n) is coisotropic implies (recall that α(k, n−k)
was defined in Definition 3.18)
 k(n−k)
|Σ(k, n)|
(4.5) edeg (G(k, n)) = α(k, n − k) · |G(k, n)| ·
|G(k, n)|

!k(n−k)
Γ k+1 n−k+1


2
Γ 2

= α(k, n − k) · |G(k, n)| · ,
Γ k2 Γ n−k2

where the last line follows from Theorem 4.2. Combining Theorem 3.19 with Proposition 4.6
and Proposition 4.8, we obtain a real version of Theorem 2.3.
Theorem 4.10. Let X1 , . . . , XN be semialgebraic subsets of RPn−1 of dimension n − k − 1,
where N := k(n − k). Then we have
N
Y |Xi |
E # (g1 Z(X1 ) ∩ · · · ∩ gN Z(XN )) = edeg G(k, n) · n−k−1 |
,
i=1
|RP

with respect to independent uniformly distributed g1 , . . . , gN ∈ O(n).

This result allows to decouple a random incidence geometry problem into a volume compu-
tation in RPn−1 and the determination of the expected degree of the Grassmann manifold (the
“linearized” problem).

4.5. Classical Schubert calculus revisited. In Definition 3.18 we defined the real average
scaling factor α(k, m). Following the reasonings in §3.3.2, it is natural to define its complex
variant as follows.
Definition 4.11. For k, m ≥ 1 we define the complex average scaling factor αC (k, m) as
αC (k, m) := E k(u1 ⊗ v1 ) ∧ . . . ∧ (ukm ⊗ vkm )k2
where ui ∈ S(Ck ) and vi ∈ S(Cm ) are independently and uniformly chosen at random, for
i = 1, . . . , km.
Unlike for α(k, m), we can give a closed formula for its complex variant αC (k, m).
Proposition 4.12. We have αC (k, m) = N !/N N , where N = km.
The proof follows immediately from the following lemma.
Lemma 4.13. Let Z = [zij ] ∈ CN ×N be a random matrix such that for all i, j, k
1
E zij |2 = and E zij z̄ik = 0 if j 6= k.
N
Then E | det Z|2 = N !/N N .
 PROBABILISTIC SCHUBERT CALCULUS 23

Proof. We have
X N
Y
| det Z|2 = sgn(π) sgn(σ) ziπ(i) z̄iσ(i) .
π,σ∈SN i=1

Taking expectations yields

N
X Y X N!
E | det Z|2 = sgn(π) sgn(σ) E ziπ(i) z̄iσ(i) = E |ziπ(i) |2 = ,
i=1
NN
π,σ∈SN π∈SN

which completes the proof. 

The following theorem is analogous to Theorem 3.19 and can be proved similarly. However, we
can easily derive this result from Theorem 2.3, which even shows that the coisotropy assumption
is not needed here.
Theorem 4.14. Let H1 , . . . , HN be irreducible algebraic hypersurfaces of GC (k, n), where N :=
k(n − k). Then we have:
N
1 |Hi |
Z Y
# (g H
1 1 ∩ · · · ∩ g H
N N ) dg 1 · · · dg N = αC (k, n − k) · |GC (k, n)| · .
|U (n)|N U(n)N i=1
|GC (k, n)|

Proof. By Theorem 2.3, the left-hand side in the assertion of Theorem 4.14 equals deg GC (k, n) ·
Q
i rdegHi . We analyze now the right-hand side. We note that |Hi | = rdegHi · |ΣC (k, n)| since
the proof of Equation (2.8) works for any irreducible hypersurface in GC (k, n). Using (2.7) we
get
|CPN −1 |
|Hi | = rdegHi · |GC (k, n)| · .
|CPN |
Moreover, |GC (k, n)| = deg GC (k, n) · |CPN | by (2.4). Therefore, it suffices to verify that
1  |CPN | N
αC (k, n − k) = · .
|CPN | |CPN −1 |
This follows from Proposition 4.12 using |CPN | = π N /N !. 

This argument showed that complex algebraic geometry implies the integral geometry formula
of Theorem 4.14. One can also argue in the reverse direction. For instance, we can give a new,
integral geometric derivation of the formula for deg GC (k, n). (This formula was first obtained
by Schubert [42].)
0!1!···(k−1)!
Corollary 4.15. We have deg GC (k, n) = (n−k)!(n−k+1)!···(n−1)! · (k(n − k))!.

Proof. We apply Theorem 4.14 in the case Zi := ΣC (k, n) and argue similarly as for Theo-
rem 2.3. Let ι denote the Plücker embedding. Then ι(ΣC (k, n)) is obtained by intersecting
ι(GC (k, n)) with a hyperplane Hi . The translates g1 H1 , . . . , gN HN meet transversally for al-
most all g1 , . . . , gn ∈ U (n). (This follows from Sard’s Theorem as for [9, Prop. A.18].) By
Bézout’s theorem, the left-hand side of Theorem 4.14 equals deg GC (k, n).
From (2.7) we know that ΣC (k, n)| = |GC (k, n)| · π1 · N (and we already mentioned that this
formula was obtained in the first version of [8] without algebraic geometry). Plugging in the
right-hand side of Theorem 4.14 the formula (2.14) for the volume GC (k, n) and the formula for
αC (k, n − k) from Proposition 4.12, we obtain after some calculations the claimed expression for
deg GC (k, n). 
24 PETER BÜRGISSER AND ANTONIO LERARIO

5. Random convex bodies

In this section we will analyze the average scaling factors α(k, m) using properties of the
singular value decomposition and concepts from the theory of random convex bodies.

5.1. The Segre zonoid. We consider the embedding Rk × Rm → Rk×m , (u, v) 7→ uv T and
assume k ≤ m. Recall that in §2.7, we associated a zonoid CY with a random vector Y .
Definition 5.1. The Segre zonoid C(k, m) is the zonoid associated with the random variable
xy T ∈ Rk×m , where x ∈ Rk and y ∈ Rm are independent and standard Gaussian vectors.
If we sample instead x ∈ S k−1 and y ∈ S m−1 in the spheres independently and uniformly,
then the corresponding zonoid equals (ρk ρm )−1 C(k, m), compare Example 2.6.
Corollary 5.2. We have
(k(n − k))!
edeg G(k, n) = |G(k, n)| · · |C(k, n − k)|.
2k(n−k)
Proof. Recall Equation (4.5), which states:
 k(n−k)
|Σ(k, n)|
edeg G(k, n) = α(k, n − k) · |G(k, n)| · .
|G(k, n)|
|Σ(k,n)|
Moreover, Theorem 4.2 states that |G(k,n)| = 12 ρk · ρn−k (recall the definition of ρk in (2.15)).
By the definition of α(k, m) (Definition 3.18), Theorem 2.7 (Vitale), the definition of C(k, m)
(Definition 5.1), and the comment following it (exchanging Gaussian by uniform distributions
on spheres), we obtain

(km)!
α(k, m) = · |C(k, m)|.
(ρk ρm )km
Combining all this, the assertion follows. 

The group O(k) × O(m) acts on the space Rk×m of matrices X via (g, h)X := gXhT . It is
well known that the orbits under this action are determined by the singular values of X. More
specifically, we denote by sv(X) the list of singular values of X (in any order). Let the matrix
diagk,m (σ) be obtained from the diagonal matrix diag(σ) by appending a zero matrix of format
k × (m − k). Then X is in the same orbit as diagk,m (σ) iff σ equals sv(X) up to a permutation
and sign changes.
The distribution underlying the definition of the Segre zonoid C(k, m) is O(k) × O(m)-
invariant, which implies that C(k, m) is invariant under this action. Therefore, C(k, m) is
determined by the following convex set, which we call the singular value zonoid:

D(k) := σ ∈ Rk | diagk,m (σ) ∈ C(k, m) .



We will see in Proposition 5.8 below that D(k) does not depend on m, which justifies the
notation. Moreover, D(k) can be realized as a projection of C(k, m), hence it is indeed a zonoid
(see Remark 5.9). The hyperoctahedral group Hk is the subgroup of O(k) generated by the
permutation of the coordinates and by the sign changes xi 7→ εi xi , where εi ∈ {−1, 1}. Clearly,
the convex set Dk is invariant under the action of Hk .
We note that C(k, m) is the union of the O(k) × O(m)-orbits of the diagonal matrices
diagk,m (σ), where σ ∈ D(k).
 PROBABILISTIC SCHUBERT CALCULUS 25

We next determine the support function of the convex body C(k, m). By the invariance, this
is an O(k) × O(m)-invariant function and hence it can only depend on sv(X). For describing
this in detail, we introduce the function gk : Rk → R by

21
gk (σ1 , . . . , σk ) := E σ12 z12 + . . . + σk2 zk2 ,
where z1 , . . . , zk are i.i.d. standard normal.
Lemma 5.3. The function gk √ is a norm on Rk √ . It is invariant under permutation of its argu-
ments. If kσk = 1, we have ρ1 / k ≤ gk (σ) ≤ 1/ k and the maximum of gk on S k−1 is attained
exactly on the Hk -orbit of the point σu := √1k (1, . . . , 1).
1 1
Proof. The Cauchy-Schwarz inequality gives i σi τi zi2 ≤ ( i σi2 zi2 ) 2 · ( i τi2 zi2 ) 2 . This easily
P P P
implies gk (σ + τ ) ≤ gk (σ) + gk (τ ). It follows that gk is a norm on Rk . The Sk -invariance is clear.
1 1 1
Suppose now kσk = 1. Then there exists i with |σi | ≥ k − 2 and hence gk (σ) ≥ k − 2 E |zi | = k − 2 ρ1 .
1
We consider now the auxiliary function g̃k (σ) := E (σ12 u21 + . . . + σk2 u2k ) 2 , where u ∈ S k−1
is chosen uniformly at √ random. Then we have gk (σ) = ρk g̃k (σ). Thus we need to show that
maxkσk=1 g̃k (σ) = 1/ k, and that the maximum is attained exactly on the Hk -orbit of of σu .
For this, note first that g̃(σu ) = √1k kσu k = √1k . Moreover, by the Cauchy-Schwarz inequality,
k  21 k  21
kσk
 X X
2 2 2 2
g̃k (σ) ≤ E σi ui = σi E ui = √ ,
i=1 i=1
k
since E u21 = . . . = E u2k = 1/k. Equality holds iff σ12 u21 + . . . + σk2 u2k is constant for u ∈ S k−1 .
This is the case iff σ is in the Hk -orbit of of σu . 
Remark 5.4. The function gk (σ) can be given the following interpretation in terms of the ellipsoid
E(σ) := {x ∈ Rk | σ12 x21 + · · · + σk2 x2k ≤ 1}:
ρk |∂E(σ)|
gk (σ) = · .
k |E(σ)|
|∂E(σ)|
This follows from [39, eq. (3)]. The isoperimetric ratio |E(σ)| can be expressed in terms of
certain hypergeometric functions, see [39, eq. (13)].
Lemma 5.5. The support function of the convex body C(k, m) is given by:
1
hC(k,m) (X) = √ gk (sv(X)) for X ∈ Rk×m .
2π
Proof. Let h denote the support function of C(k, m) and define η(σ) := h(diagk,m (σ)). We have
h(X) = η(sv(X)) since X is in the same orbit as diagk,m (sv(X)) and h is O(k)× O(m)-invariant.
From (2.10) we obtain
k
1 1 X
η(σ) = E hdiagk,m (σ), xy T i = E σi xi yi ,
2 2 i=1

where x ∈ Rk and y ∈ Rm are independent standard normal. For fixed x, the random variable
Pk Pk 2 2
i=1 σi xi yi is normal distributed with mean zero and variance
Pk i=1 σi xi . Therefore, the
1 p
expectation of its absolute value equals ( i=1 σi2 x2i ) 2 ρ1 , where we recall that ρ1 = E |z| = 2/π,
z ∈ R denoting a standard Gaussian. Summarizing, we obtain
k  12
1 X 1
η(σ) = ρ1 E σi2 x2i = √ gk (σ),
2 i=1
2π
26 PETER BÜRGISSER AND ANTONIO LERARIO

which finishes the proof. 

Corollary 5.6. The zero matrix is an interior point of C(k, m).

Proof. By Lemma 5.5 and Lemma 5.3 we have

1 1 ρ1 1 1
min hC(k,m) (X) = √ min gk (σ) ≥ √ √ = √ .
kXk=1 2π kσk=1 2π k π k
1
Hence C(k, m) contains the ball of radius π −1 k − 2 around the origin. 

We need a convexity property of singular values in the spirit of the Schur-Horn theorem for
eigenvalues of symmetric matrices.

Theorem 5.7. Let A = (aij ) ∈ Rk×m , k ≤ m, with singular values σ1 ≥ . . . ≥ σk ≥ 0.

Then the diagonal (a11 , . . . , akk ) lies in the convex hull of the Hk -orbit of (σ1 , . . . , σk ) of the
hyperoctahedral group Hk .

Proof. Let B ∈ Rk×k be obtained from A ∈ Rk×m by selecting the first k columns. Let τ1 ≥
. . . ≥ τk ≥ 0 denote the singular values of B. The Courant-Fischer min-max characterization
of singular values (e.g., see [52, Exercise 1.3.21]) implies that τi ≤ σi for all i. Hence the
convex hull of the Hk -orbit of τ := (τ1 , . . . , τk ) is contained in the convex hull of the Hk -orbit of
σ := (σ1 , . . . , σk ).
A result by Thompson [53] implies that the diagonal (a11 , . . . , akk ) of B lies in the convex
hull of the Hk -orbit of τ . Hence it lies in the convex hull of the Hk -orbit of σ. 

We can now prove that the support functions and radial functions of C(k, m) and D(k) are
linked in a simple way.

Proposition 5.8. The following properties hold:

(1) The support function of D(k) equals √12π gk . In particular, D(k) does not depend on m.
(2) The radial function of C(k, m) is given by Rk×m → R, X 7→ rk (sv(X)), where rk denotes
the radial function of D(k).

Proof. (1) We have σ ∈ D(k) iff diagk,m (σ) ∈ C(k, m), which, by Lemma 5.5, is equivalent to
k
X 1
(5.1) ∀Y ∈ Rk×m σi Yii ≤ √ gk (sv(Y )).
i=1
2π

We need to prove that this is equivalent to

k
k
X 1
(5.2) ∀τ ∈ R σi τi ≤ √ gk (τ ).
i=1
2π

One direction being trivial, assume that σ satisfies (5.2) and let Y ∈ Rk×m . Then we have for
all π ∈ Hk ,
1 1
hσ, πsv(Y )i ≤ √ gk (πsv(Y )) = √ gk (sv(Y )).
2π 2π
Theorem 5.7 states that (Y11 , . . . , Ykk ) lies in the convex hull of the Hk -orbit of sv(Y ). Therefore,
(5.1) holds, which proves the first assertion.
 PROBABILISTIC SCHUBERT CALCULUS 27

(2) For the second assertion, let R denote the radial function of C(k, m). By invariance, it is
sufficient to prove that R(diagk,m (σ)) = rk (σ). We have, using the above equivalences,

R(diagk,m (σ)) = max t | t diagk,m (σ) ∈ C(k, m)
k
(5.1)  X 1
= max t | ∀Y tσi Yii ≤ √ gk (sv(Y ))
i=1
2π
(5.2)  1
= max t | ∀τ htσ, τ i ≤ √ gk (τ ) = rk (σ)
2π
and the proof is complete. 
Remark 5.9. Proposition 5.8 easily implies that D(k) is the image of C(k, m) under the projection
taking the diagonal of a rectangular matrix:
Rk×m → Rk , X = (xij ) 7→ (x11 , . . . , xkk ).
In particular, D(k) is a zonoid in Rk .
Lemma 5.10. The maximum of the radial function rk of D(k) on S k−1 equals
1 ρk
Rk := max rk (σ) = √ √
kσk=1 2π k
and the maximum is attained exactly on the Hk -orbit of the point σu := √1k (1, . . . , 1). In partic-
ular, C(k, m) is contained in the ball B(k, m) of radius Rk around in the origin, and this is the
ball of smallest radius with this property.
Proof. According to Proposition 5.8, √12π gk is the support function of D(k). According to
Lemma 5.3, the maximum of gk on S k−1 is attained exactly on the Hk -orbit of σu . We conclude
now with Lemma 2.4. 
Remark 5.11. Corollary 5.6 implies that 0 is an interior point of D(k). This can also be seen
in a different way as follows. A standard transversality argument shows that g1 Σ(k, n) ∩ . . . ∩
gN Σ(k, n) 6= ∅ for all (g1 , . . . , gN ) in a nonempty open subset of O(n)N ; cf. [9, Prop. A.18]. This
implies that edeg G(k, n) is positive. Via Theorem 5.13 below we conclude that D(k) has full
dimension. Since D(k) is symmetric with respect to the origin, it again follows that 0 is an
interior point of D(k).
5.2. Volume of the Segre zonoid. We keep assuming k ≤ m. We first state how to integrate
a O(k) × O(m)-invariant function of matrices in Rk×m in terms of their singular values. This is
certainly known, but we include the proof for lack of a suitable reference.
Let us denote by S km−1 ⊆ Rk×m the unit sphere with respect to the Frobenius norm. More-
over, it will be convenient to abbreviate
k−1
S+ := {σ ∈ S k−1 | σ1 ≥ · · · ≥ σk ≥ 0}.
Also, recall the Stiefel manifold S(k, m) := {Q ∈ Rm×k | QT Q = 1}. We postpone the proof of
the following technical result to Appendix 7.4.
Proposition 5.12. Let f : S km−1 → R be a continuous O(k) × O(m)-invariant function and
put g(σ) := f (diagk,m (σ)). Then we have
k
|O(k)||S(k, m)|
Z Z Y Y
f (X) dS km−1 = k
g(σ) σim−k (σi2 − σj2 ) dS k−1 .
S km−1 2 k−1
S+ i=1 1≤i<j≤k
28 PETER BÜRGISSER AND ANTONIO LERARIO

We can now derive a formula for the volume of C(k, m). We introduce the Hk -invariant
functions pk : Rk → R and qk : Rk → R by
k
Y Y
(5.3) pk (σ1 , . . . , σk ) := |σi |, qk (σ1 , . . . , σk ) := pk (σ1 , . . . , σk )−k · |σi2 − σj2 |.
i=1 i<j

Theorem 5.13. The volume of C(k, m) equals

|O(k)| |S(k, m)|
Z

m
|C(k, m)| = pk rkk qk dS k−1 ,
km 2k k−1
S+

where rk denotes the radial function of the singular value zonoid D(k).
Proof. According to Lemma 5.10, the radius function of C(k, m) is given by X 7→ rk (sv(X)).
The volume of C(k, m) satisfies (see [41, Equation (1.53)])
1
Z
|C(k, m)| = (rk ◦ sv)km dS km−1 .
km S km−1
Since the radial function is continuous and O(k)×O(m)-invariant, we can apply Proposition 5.12
to obtain
1 |O(k)||S(k, m)|
Z Y
|C(k, m)| = k
rk (σ)km pk (σ)m−k (σi2 − σj2 ) dS k−1
km 2 k−1
S+ 1≤i<j≤k
|O(k)||S(k, m)|
Z
m
rk (σ)k pk (σ) qk (σ) dS k−1


=
km2k k−1
S+

as claimed. 
Remark 5.14. The proof of Theorem 5.13 works for any O(k) × O(m)-invariant convex body.
In particular, for the ball B(k, m) := {X ∈ Rk×m | kXkF ≤ Rk } containing C(k, m) (see
Lemma 5.10), we obtain:
km
π 2 |O(k)| |S(k, m)|
Z

m
|B(k, m)| = Rkkm km

= k
pk Rkk qk dS k−1 .
Γ 1+ 2 km 2 k−1
S+

6. Asymptotics
6.1. Laplace method. In this section we recall a useful result for the asymptotic evaluation
of integrals depending on a large parameter. This technique will be crucial in the proofs of
Theorem 6.3 and Theorem 6.8 and goes under the name of Laplace’s method. The following
statement is classical (see for example [56, Section II, Theorem 1]).
Theorem 6.1. Consider the integral
Z t2
I(λ) = e−λa(t) b(t)dt,
t1
where the functions a, b : [t1 , t2 ] → R satisfy the following conditions:
(1) The function a is smooth on a neighborhood of t1 and there exist µ > 0 and a0 6= 0 such
that as t → t1
a(t) = a(t1 ) + a0 (t − t1 )µ + O(|t − t1 |µ+1 ).
(2) The function b is smooth on a neighborhood of t1 and there exist ν ≥ 1 and b0 6= 0 such
that as t → t1
b(t) = b0 (t − t1 )ν−1 + O(|t − t1 |ν ) with b0 6= 0.
 PROBABILISTIC SCHUBERT CALCULUS 29

(3) We have a(t) > a(t1 ) for all t ∈ (t1 , t2 ) and for every δ > 0 the infimum of a(t) − a(t1 )
in [t1 + δ, t2 ) is positive.
(4) The integral I(λ) converges absolutely for all sufficiently large λ.
Then we have the asymptotic
 
ν
1 b 0 Γ µ
 
−λa(t1 )
(6.1) I(λ) = e · ν/µ · ν/µ · 1 + O(λ−(1+ν)/µ ) as λ → ∞.
λ a0 µ
Remark 6.2. (1) The hypotheses of [56, Section II, Theorem 1] are weaker than those of
Theorem 6.1: it is only required that a and b are continuous in a neighborhood of t1
(except possibly at t1 ) and that they have asymptotic series at t1 (which is certainly
true if they are both smooth). Moreover the conclusion given there is also stronger than
what we stated here: it is proved that I(λ) also has an asymptotic series in descending
powers of λ, while we are just writing the leading order term and the order of the error
of this series.
(2) Note that the exponent of the exponential growth in (6.1) is determined by the minimum
a(t1 ), the growth of the pre-exponential factor is determined by the orders µ, ν of the
expansions of the functions a and b, and the leading constant in (6.1) involves the
constants a0 and b0 .
6.2. An asymptotically sharp upper bound for the expected degree. From Lemma 5.10,
we see that the Segre zonoid C(k, m) is contained in the ball
n 1 ρk o
B(k, m) := X ∈ Rk×m | kXkF ≤ √ √ .
2π k
In particular, |C(k, m)| ≤ |B(k, m)|, which implies an upper bound for α(k, m). We next show
that this inequality, for fixed k, is asymptotically sharp up to a subexponential factor.
Theorem 6.3. For fixed k we have log |C(k, m)| = log |B(k, m)| − O(log m) for m → ∞.
Proof. We introduce first some useful
√ notation that we will use in the proof. Recall from
1 k−1
Lemma 5.10 that Rk = (2π)− 2 ρk / k is the maximum of the radial function rk on S+ , and
that the maximum is attained exactly at σu := √1k (1, . . . , 1). It is convenient to define the
normalized radius function r̃k := rk /Rk , which has the maximum value 1.
Theorem 5.13 and Remark 5.14 imply now that we can rewrite the two volumes |C(k, m)| and
|B(k, m)| as
|O(k)| |S(k, m)| km |O(k)| |S(k, m)| km
(6.2) |C(k, m)| = Rk ·IC (k, m), |B(k, m)| = Rk ·IB (k, m),
km 2k km 2k
where
Z Z

m
(6.3) IC (k, m) := r̃kk pk qk dS k−1
and IB (k, m) := pm
k qk dS
k−1
.
k−1 k−1
S+ S+

The idea of the proof is now to compute the asymptotics of two integrals IB (k, m) and
IC (k, m), the first one explicitly and the second one using Laplace’s method, and then to compare
them. Because of |C(k, m)| ≤ |B(k, m)|, and since the factors in front of the two integrals in (6.2)
are the same, we just have to verify that
(6.4) log IC (k, m) ≥ log IB (k, m) − O(log m) for m → ∞.
We first look at the asymptotic growth of IB (k, m). Note that by rearranging (6.2), we get
km2k
IB (k, m) = · |BRkm (0, 1)|,
|O(k)| |S(k, m)|
30 PETER BÜRGISSER AND ANTONIO LERARIO

since Rk is the radius of the ball B(k, m). Using the explicit formulas
π km/2 2k π km/2
|BRkm (0, 1)| =
, |S(k, m)| =
Γ 1 + km Γk m


2 2
and the asymptotics log Γ(z) = z log z − z + O(log z) and log Γk (z) = k(z log z − z) + O(log z)
for z → ∞, we easily deduce that
 
1
(6.5) log IB (k, m) = km log √ + O(log m) as m → ∞.
k
For the asymptotic growth of IC (k, m) we will need the following result, whose proof we postpone.
Lemma 6.4. There exist κ0 , κ1 , δ > 0, depending on k, such that κ1 δ ≤ 1 and for all m,
Z δ m k2 +k−4
k
IC (k, m) ≥ κ0 k − 2 (1 − κ1 ρ) ρ 2 dρ.
0
Rδ
We rewrite the right-hand side integral in Lemma 6.4 as 0 0 e−ma(ρ) b(ρ)dρ with the smooth
functions
  k k2 +k−4
a(ρ) := − log k −k/2 (1 − κ1 ρ) = log k + κ1 ρ + O(ρ2 ), b(ρ) := ρ 2 ,
2
where the expansion is for ρ → 0.
Note now that both functions a and b are smooth (b is a polynomial since for k ≥ 2 the
2
exponent k +k−4
2 is a positive natural number). We can then apply Theorem 6.1 with the
2
choices a0 = κ1 , µ = 1, b0 = 1, ν = k +k−4
2 + 1 and obtain for m → ∞
Z δ0   km
k
 m k2 +k−4 1 1 Γ(ν)  
k − 2 (1 − κ1 ρ) ρ 2 dρ = √ 1 + O(m −(ν+1)
) .
0 k mν κν1
In particular, combining this asymptotic with Lemma 6.4, we deduce that for m → ∞
Z δ0  !
k
m k2 +k−4
log IC (k, m) ≥ log κ0 k − 2 (1 − κ1 ρ) ρ 2 dρ
0


1
≥ km log √ − O(log m).
k
Together with (6.5), this shows (6.4) and proves Theorem 6.3. 
Proof of Lemma 6.4. It will be convenient to replace the singular value zonoid D(k) by a convex
body Y ⊆ D(k) that is simpler to analyze. By Corollary 5.6 (see also Remark 5.11), the zero
vector is an interior point of D(k), hence there exists ǫ > 0 such that BRk (0, ǫ) ⊆ D(k). We
define Y as the convex hull of BRk (0, ǫ) and y := Rk σu ∈ D(k). Thus Y is the cone with
base BRk (0, ǫ) and apex y. Note that Y is rotation symmetric with respect to the line passing
through 0 and σu .
In the definition (6.3) of IC (k, m), we can can integrate over the full sphere S k−1 ,
1
Z Z
k

m k−1

m
IC (k, m) = r̃k pk qk dS = k
r̃kk pk qk dS k−1 ,
S+k−1 k!2 S k−1

by the Hk -invariance of the integrand. Recall that r̃k = Rk−1 rk is the normalized radial function
of D(k). Denoting by rY the radial function of Y and setting r̃Y := Rk−1 rY , we have r̃k ≥ r̃Y ,
since Y ⊆ D(k). Replacing r̃k by the smaller r̃Y , we obtain
1
Z

m
IC (k, m) ≥ r̃k pk qk dS k−1 .
k!2k S k−1 Y
 PROBABILISTIC SCHUBERT CALCULUS 31

For the integration of the right-hand side, we use a coordinate system adapted to the situation.
The exponential map
x ∈ Tσu S k−1 | kxk < π → S k−1 \ {−σu }, x 7→ σ(x)


is a diffeomorphism that maps 0 to σu . We describe tangent vectors in Tσu S k−1 with polar
coordinates (ρ, θ) ∈ [0, ∞) × S k−2 , where S k−2 stands here for the unit sphere of Tσu S k−1 . This
way, we obtain a parametrization σ(ρ, θ) of the points in S k−1 \ {−σu }. It is easy to see that
its Jacobian is of the form ρk−2 w(ρ) with a smooth function w satisfying w(0) > 0. By the
transformation theorem we have
Z Z πZ

m
m
r̃Yk pk qk dS k−1 = r̃Y (σ(ρ, θ))k pk (σ(ρ, θ) qk (σ(ρ, θ)) ρk−2 w(ρ) dS k−2 dρ.
S k−1 0 S k−2

Claim. There are c1 , c2 , c3 , c4 > 0, π > δ1 > 0 and a nonzero continuous function η : S k−2 → R
such that for all ρ ≤ δ1 we have w(ρ) ≥ c1 and
k(k−1)
r̃Y (σ(ρ, θ)) ≥ 1 − c2ρ, pk (σ(ρ, θ)) ≥ k −k/2 1 − c3ρ2 , qk (σ(ρ, θ)) ≥ ρ 2



|η(θ)| − c4 ρ .
Proof. 1. The lower bound on w(ρ) follows from the continuity of w and w(0) > 0.
2. By the axial symmetry of Y , the function r̃Y (σ(ρ, θ)) only depends on ρ. The lower bound
on r̃Y (ρ) easily follows from the construction of Y .
3. The function pk (σ(x)) is smooth in a neighborhood of the zero vector and has the expansion
1
pk (σ(x)) = pk (σ(0)) + xHxT + O(kxk3 )
2
k−1
for x ∈ Tσu S , x → 0. The matrix H is negative semidefinite since x = 0 is a local maximum
of pk (σ(x)). Note that pk (σ(0)) = pk (σu ) = k −k/2 . In particular, there exists a constant c′ > 0
such that for small enough ρ = kxk we have pk (σ(x)) ≥ pk (0) − c′ kxk2 , which shows the stated
lower bound on pk .
Recall from (5.3) that the function qk (σ) is defined as the modulus of the function fk (σ) :=
Qk −k Q
i=1 σ
i · i<j (σi2 − σj2 ). The function fk is smooth at σu and it vanishes at σu to order
k
ℓ := 2 . We consider the Taylor expansion fk (σ(x)) = h(x) + O kxkℓ+1 , where h is a nonzero

homogeneous polynomial of degree ℓ. In particular, we get in polar coordinates

fk (σ(ρ, θ)) = ρℓ η(θ) + O ρℓ+1

where the function η : S k−2 → R is nonzero. The stated lower bound on qk follows and the
claim is shown. 
k−2
R
n we put κo:= S k−2 |h|dS
To complete the proof of Lemma 6.4, , which is positive. Using
κ
the claim, we obtain for 0 < δ ≤ min δ1 , 2c4 |S k−2 | that
Z

m
r̃Yk pk qk dS k−1
S k−1
Z δ  m k(k−1) Z  
≥ (1 − c2 ρ)k k −k/2 (1 − c3 ρ2 ) ρ 2 +k−2 c1 |h| − c4 ρ dS k−2 dρ
0 S k−2
δ m k2 +k−4
c1 κ
Z 
≥ k −k/2 (1 − c2 ρ)k (1 − c3 ρ2 ) ρ 2 dρ
2 0
δ m k2 +k−4
c1 κ
Z 
≥ k −k/2 (1 − 2kc2 ρ) ρ 2 dρ,
2 0
32 PETER BÜRGISSER AND ANTONIO LERARIO

where the last inequality holds if ρ is sufficiently small. The assertion follows with κ0 := 12 c1 κ
and κ1 := 2kc2 . 
Finally, we arrive at the main result of this section, an asymptotically sharp upper bound
for the expected degree of real Grassmannians. The√result should be compared with the corre-
sponding statement (2.4). Recall the quantity ρk ≤ k from (2.15) and (2.16).
Theorem 6.5. Let N := k(n − k) = dim G(k, n). The following statements are true:
(1) For all n ≥ k > 0 we have
r N
|G(k, n)| π ρk
edeg G(k, n) ≤ E(k, n) := · √ .
|RPN | 2 k
(2) For fixed k and n → ∞, the above inequality is asymptotically sharp in the sense that
log edeg G(k, n) = log E(k, n) − O(log n).
(3) For fixed k and n → ∞, we have
√ k+1

!
πΓ
log E(k, n) = kn log k

2 − O(log n).
Γ 2

Proof. Recall the formula for edeg G(k, n) given in Corollary 5.2. In this formula, we replace
|C(k, n − k)| by the larger volume of the ball B(k, n − k), obtaining the quantity
N!
(6.6) E(k, n) := |G(k, n)| · N · |B(k, n − k)|,
2
which satisfies edeg G(k, n) ≤ E(k, n). Theorem 6.3 implies that log E(k, n) − log edeg G(k, n) =
log |B(k, n − k)| − log |C(k, n − k)| = O(log n) for fixed k and n → ∞.
We verify now that the E(k, n) defined in (6.6) indeed has the form √ stated in assertion (1).
Recall that B(k, n − k) is the ball in RN with the radius Rk = ρk / 2πk; cf. Lemma 5.10. We
have |B(k, n − k)| = RkN |BRN (0, 1)| and
N N
1 N −1 2 π2 π2
|BRN (0, 1)| = |S |= = .
N N Γ(N/2) Γ( N2+2 )
N
2
Plugging this into (6.6), and simplifying with the help of the duplication formula N ! = √ π
·
N +1 N +2



Γ 2 · Γ 2 , yields

 r N N
Γ N2+1
r
π 1 π ρk
E(k, n) = |G(k, n)| · N +1 · ρk = |G(k, n)| · · √ ,
π 2 2k |RPN | 2 k
as claimed in assertion (1).
It remains to prove the asymptotic of E(k, n) stated in assertion (3). By the first assertion
and (2.15)
√

!N 
k+1
|G(k, n)| 1 π Γ
log E(k, n) = log  · ·
2 
|RPN | k N/2 Γ k2
√
!
π Γ k+1
 
2 |G(k, n)| 1
= kn log + log · + O(log n).
Γ k2 |RPN | k N/2

Lemma 2.8 states that for fixed k and n → ∞

1 1
(6.7) log |G(k, n)| = − kn log n + kn log(2eπ) + O(log n).
2 2
 PROBABILISTIC SCHUBERT CALCULUS 33

Moreover, for |RPN | = 12 |S N | we obtain that, using log Γ(x) = x log x − x + O(log x) for x → ∞,
 
N 1 1 2eπ
(6.8) log |RP | = − kn log n + kn log + O(log n)
2 2 k
 
Combining (6.7) and (6.8), it follows that log |G(k,n)| 1
|RPN | · kN/2 = O(log n), which completes the
proof. 
We compare now the expected degree of the real Grassmannian with the degree of the corre-
sponding complex Grassmannian.
Corollary 6.6. For k ≥ 2 we have
πρ2k
 
log edeg G(k, n)
εk := lim p = logk .
n→∞ log deg GC (k, n) 2
Moreover, the sequence εk is monotonically decreasing and limk→∞ εk = 1.
Proof. The formula for εk follows from Theorem 6.5 and the asymptotic edeg GC (k, n) = kn log k+
O(log n) from Lemma 2.8. We have limk→∞ εk = 1 due to (2.16). The verification of the mono-
tonicity of εk is left to the reader. 
Corollary 6.6 implies that for fixed k:
1
edeg G(k, n) = deg GC (k, n) 2 εk +o(1) for n → ∞.
This means that for large n, the expected degree of the real Grassmannian exceeds the square
root of the degree of the corresponding complex Grassmannian and in the exponent, the deviation
from the square root is measured by 21 (εk −1). For example, ε2 = 2 log2 (π/2) ≈ 1.30. For large k,
the exponent εk goes to 1 so that we get the square root law
1
edeg G(k, n) = deg GC (k, n) 2 +o(1) for k, n → ∞.
6.3. The Grassmannian of lines.
6.3.1. The expected degree of G(2, 4). We can express this as an integral of the modulus of the
function

a(t1 , t2 , t3 , s1 , s2 , s3 ) = cos s2 sin s1 sin s3 sin t2 sin(t1 − t3 )

− (sin s2 ) (cos s1 sin s3 sin t1 sin(t2 − t3 ) + cos s3 sin s1 sin t3 sin(t1 − t2 )) ,
which can be written in the following symmetric way (this obervation is due to Chris Peterson)
 
sin t1 sin s1 sin t2 sin s2 sin t3 sin s3
|a(t, s)| = det  cos t1 sin s1 cos t2 sin s2 cos t3 sin s3  .
sin t1 cos s1 sin t2 cos s2 sin t3 cos s3
Proposition 6.7. We have edeg G(2, 4) = 2113 [0,2π]6 |a(t, s)| dtds = 1.72.....
R

Proof. Formula (4.5) implies that

 4
|Σ(2, 4)|
edeg G(2, 4) = |G(2, 4)| · · α(2, 2).
|G(2, 4)|
|Σ(2,4)|
By (2.14) we have |G(2, 4)| = 2π 2 and from Theorem 4.2 we deduce |G(2,4)| = π4 . By defini-
tion, we have α(2, 2) = E k(u1 ⊗ v1 ) ∧ . . . ∧ (u4 ⊗ v4 )k, where the u1 , u2 , u3 , v1 , v2 , v3 ∈ S 1 are
34 PETER BÜRGISSER AND ANTONIO LERARIO

independently and uniformly chosen at random, and we can assume u4 = v4 = (1, 0) by the
invariance of the problem. We can thus write ui = (cos ti , sin ti ) and vi = (cos si , sin si ) where
(ti , si ) ∈ [0, 2π]2 are random with the uniform density. Setting wi := ui ⊗ vi , we have by (2.1)
that k(u1 ⊗ v1 ) ∧ . . . ∧ (u4 ⊗ v4 )k2 = det (hwi , wj i). Moreover,
hwi , wj i = hui , uj ihvi , vj i = (cos ti cos tj + sin ti sin tj )(cos si cos sj + sin si sin sj ).
Expanding the determinant, we see after a few simplifications that det (hwi , wj i) = a2 with the
function a defined above. Consequently:
 π 4 1
Z
edeg G(2, 4) = 2π 2 · · |a(t, s)| dtds,
4 (2π)6 [0,2π]6
which proves the claim. 
6.3.2. The general case. In the case of G(2, n + 1), Theorem 6.5 provides the asymptotic
π
log edeg G(2, n + 1) = 2n log + O(log n) as n → ∞.
2
We sharpen this by proving the following result.
Theorem 6.8. The expected degree of the Grassmannian of lines satisfies
 2 n
8 1 π
1 + O(n−1 ) .


edeg G(2, n + 1) = 5/2 √
3π n 4
Proof. By Corollary 5.2 we have
(2n − 2)!
(6.9) edeg G(2, n + 1) = |G(2, n + 1)| ·· |C(2, n − 1)|.
22n−2
The volume of the Grassmannian can be computed using (2.14):
1
|O(n + 1)| |S n | · |S n−1 | π n− 2 (2π)n−1
(6.10) |G(2, n + 1)| = = = n

n+1

= ,
|O(2)| · |O(n − 1)| |O(2)| Γ 2 Γ 2
(n − 1)!
where we used the duplication formula for the last inequality. Moreover,
(2.11) 4π n−1 2n−1 π n−2
|S(2, n − 1)| = 1 = .
π Γ( n−1
2
n−2
2 )Γ( 2 )
(n − 3)!
Theorem 5.13 describes the volume of C(2, n − 1):
Z π
|O(2)| · |S(2, n − 1)| 4
n−1 (cos θ)2 − (sin θ)2
|C(2, n − 1)| = r2 (θ)2 cos θ sin θ dθ
(2n − 2)2 2
0 (cos θ sin θ)2
Z π4
2n−2 π n−1
n−1 (cos θ)2 − (sin θ)2
(6.11) = r2 (θ)2 cos θ sin θ 2 dθ,
(n − 1)Γ(n − 2) 0 (cos θ sin θ)
where here and below, abusing notation, we denote by r2 (θ) the function r2 (cos θ, sin θ).
Plugging (6.10) and (6.11) into (6.9), and simplifying, we obtain:
Z π4
π 2n−2 Γ(2n − 1)
n−1 (cos θ)2 − (sin θ)2
(6.12) edeg G(2, n+1) = r2 (θ)2 cos θ sin θ 2 dθ.
(2n − 2)Γ(n − 2)Γ(n) 0 (cos θ sin θ)
We will now use Laplace’s method for the evaluation of the integral in the above equation; first
we will write it as:
Z π4
I(λ) = e−λa(θ) b(θ)dθ,
0
 PROBABILISTIC SCHUBERT CALCULUS 35

where λ = n − 1 and where we have set

(cos θ)2 − (sin θ)2
a(θ) := − log(r2 (θ)2 cos θ sin θ), b(θ) := 2 .
(cos θ sin θ)
In order to apply Laplace’s method, we verify now that the hypotheses (1)–(4) of Theorem 6.1
are satisfied. (Note that the minimum of a is attained at the right extremum of the interval of
integration, which leads to minor modifications.)
(1) By Proposition 6.9 below, the function r2 (θ)2 is smooth in a neighborhood of π4 and has
the following asymptotic expansion for θ → π4
1 1 π  π 
r2 (θ)2 = − (θ − )2 + O (θ − )3 .
8 8 4 4
2
Consequently the function a(θ) = log(r2 (θ) cos θ sin θ) is smooth in a neighborhood of
π π
4 and, as θ → 4 , we can easily deduce
 π 2  π 
a(θ) = 4 log 2 + 3 θ − + O (θ − )3 .
4 4
(2) The function b is smooth in a neighborhood of π4 , and, as θ → π4 ,
 π  π 
b(θ) = −8 θ − + O (θ − )2 .
4 4
(3) The function cos θ sin θ is monotonically increasing on [0, π/4], hence it has a unique
maximum at π4 ; moreover, by Lemma 5.10, the function r2 (θ) also has a unique maximum
on [0, π/4] at π4 . Therefore, a has a unique minimum at θ = π4 on the interval [0, π/4].
(4) The integrand in I(λ) is nonnegative and I(λ) < ∞ for every λ, since edeg G(2, n + 1)
is finite.
We can now apply Theorem 6.1 with µ = ν = 2, a0 = 3 and b0 = 8 (we have to change
the sign of b0 because the minimum is now attained at the right extremum of the interval of
integration). This implies for λ → ∞:
1 8  
I(λ) = e−λ4 log 2 · · 1 + O(λ−3/2 )
λ 3·2
−4(n−1) 1 4  
(6.13) =2 · · 1 + O(n−3/2 ) .
n−1 3
We plug in now (6.13) into (6.12), obtaining
π 2n−2 Γ(2n − 1) 1 4  
(6.14) edeg G(2, n + 1) = · 2−4(n−1) · · 1 + O(n−3/2 ) .
(2n − 2)Γ(n − 2)Γ(n) n−1 3
We use the duplication formula for the Gamma function and write
Γ(n)Γ n + 21 22n−1


Γ(2n)
(6.15) Γ(2n − 1) = = √ .
(2n − 1) π(2n − 1)
Using (6.15) in(6.14), we get
8 1 Γ n + 12
 π 2n

1 + O(n−1 )


edeg G(2, n + 1) = 5/2 3
2 3π n Γ(n − 2)
 π 2n 8 1
√ 1 + O(n−1 ) ,


=
2 3π 5/2 n
Γ(n+ 12 )
where in the last step we have used the asymptotic Γ(n−2) = n5/2 (1 + O(n−1 )). This concludes
the proof. 
36 PETER BÜRGISSER AND ANTONIO LERARIO

Figure 1. The convex body D(2) on the left and its polar D(2)◦ on the right (the
bodies are centered at the origin, and the two pictures have the same scale).

It remains to show the proposition used in the proof of Theorem 6.8.

Proposition 6.9. The radial function r2 (θ) of the singular value zonoid D(2), parameterized
by the angle θ, is smooth in a neighborhood of π4 . Moreover, we have the following expansion
1 1 π
2  π
4  π
r2 (θ)2 = − θ − +O θ− as θ → .
8 8 4 4 4
Proof. Let us first outline the idea. The main difficulty is that r2 (θ) is only implicitly defined.
However, by Proposition 5.8, we explicitly know the support function h of D(2). (See Figure 1
for the shape of D(2) and its polar body).
Below we will show that h is smooth on R2 \{0}, hence on this open set its gradient ∇h
exists and it is a smooth function. Proposition 2.5 tells us that im(γ) ⊆ ∂D(2), where the
function γ : (0, π2 ) → R2 is defined by

γ(t) := (∇h)(cos t, sin t).

We use γ to obtain local parametrization of ∂D(2) around π/4. Note that γ(π/4) = 41 (1, 1)
by Lemma 5.10. Let θ(t) denote the angle that γ(t) makes with the x-axis, that is, θ(t) :=
arctan(γ2 (t)/γ1 (t)). Then we have
(6.16) r2 (θ(t)) = kγ(t)k
and γ(π/4) = π/4. We will show that the derivative of θ(t) does not vanish at π/4. Then
the implicit function theorem implies that we can locally invert the function θ(t) around π/4
to obtain a smooth function t(θ), and it follows from (6.16) that r2 (θ) is indeed smooth in a
neighborhood of π/4. The Taylor expansion of r2 (θ) around π/4 will be derived by calculating
the expansions of the gradient of h and the resulting expansion of θ(t) around π/4.
We proceed now to prove the smoothness of h. By Proposition 5.8, h can be written as
1 1
Z
2 2

1
2 2 2 − 12 (u21 +u22 )
h(σ1 , σ2 ) = √ g2 (σ1 , σ2 ) = σ 1 u 1 + σ2 u 2 e du1 du2 .
2π (2π)3/2 R2
The above integral can be expressed in terms of the complete elliptic integral of the second kind
Z π/2 p
E(s) := 1 − s(sin t)2 dt, 0 ≤ s ≤ 1,
0
which is a smooth function of s. We obtain after a short calculation that

σ22
 
1
h(σ1 , σ2 ) = |σ1 | E 1 − 2 .
π σ1
 PROBABILISTIC SCHUBERT CALCULUS 37

This formula, combined with h(σ1 , σ2 ) = h(σ2 , σ1 ), implies that h is a smooth function on
R2 \{0}.
The elliptic integral E(s) has the following expansion for s → 0
π π 3π 2 5π 3
E(s) = − s− s − s + O(s4 ).
2 8 128 512
√
We can derive from this the third order Taylor expansion of h at (µ, µ), where µ := 2/2.
Denoting xi := σi − µ and x := (x1 , x2 ), we obtain after some calculation the third order Taylor
expansion of h around x = (0, 0):
µ 1 1 µ µ µ
h(x1 + µ, x2 + µ) = + x1 + x2 + x21 − x1 x2 + x22
2 4 4 16 8 16
1 1 1 1
− x31 + x21 x2 + x1 x22 − x32 + O(kxk4 ).
32 32 32 32
From this, we obtain the Taylor expansion for the gradient ∇h around x = (0, 0):
1 µ µ 3 1 1
∂x1 h(x1 + µ, x2 + µ) = + x1 − x2 − x21 + x1 x2 + x22 + O(kxk3 )
4 8 8 32 16 32
1 µ µ 1 1 3
∂x2 h(x1 + µ, x2 + µ) = − x1 + x2 + x21 + x1 x2 − x22 + O(kxk3 ).
4 8 8 32 16 32
Taking into account the definition γi (t) = ∂xi h(cos t, sin t) and substituting x1 = cos t − µ,
x2 = sin t − µ, we obtain from the above expansion of ∇h by a straightforward calculation (best
done with a computer) the following expansions for τ → 0,
1 1 1 1 1 1
(6.17) γ1 (τ + π/4) = − τ − τ 2 + O(τ 3 ), γ2 (τ + π/4) = + τ − τ 2 + O(τ 3 ).
4 8 16 4 8 16
This implies γγ12 (τ + π/4) = 1 + τ + 21 τ 2 + O(τ 3 ) for τ → 0, and hence
 
γ2 (τ + π/4) π 1
(6.18) θ(τ + π/4) = arctan = + τ + O(τ 3 ).
γ1 (τ + π/4) 4 2
In particular, we see that θ′ (π/4) = 1/2 6= 0, hence the function θ(t) has a smooth inverse
function t(θ) around π/4. Equation (6.17) implies with r2 (θ(t))2 = γ1 (t)2 + γ2 (t)2 (see (6.16))
that for τ → 0
1 1
(6.19) r2 (θ(τ + π/4))2 = − τ 2 + O(τ 4 ).
8 32
From the expansion (6.18) we obtain for θ → π/4
π  π  π 
τ (θ) := t(θ) − = 2 θ − + O (θ − )3 .
4 4 4
Plugging this into (6.19), we finally arrive at
1 1 π 2  π  π
r2 (θ)2 = − θ− + O (θ − )4 as θ → ,
8 8 4 4 4
which concludes the proof. 

7. Appendix
7.1. Proof of Theorem 3.2. Since the density of A is invariant under the orthogonal group
O(n), we can assume that B is spanned by the first l standard vectors, i.e.,
   
1 A1
B= and A = ,
0 A2
38 PETER BÜRGISSER AND ANTONIO LERARIO

where 1 is the ℓ × ℓ identity matrix, A1 is an ℓ × k matrix and A2 an (n − ℓ) × k matrix. (Again,

abusing notation, we denote by A and B also matrices whose columns span the corresponding
spaces.) If we sample A with i.i.d. normal gaussians, the corresponding probability distribution
for the span of its columns is O(n) invariant, and consequently it coincides with the uniform
distribution. In order to compute the principal angles between A and B using (3.1), we need to
orthonormalize the columns of A. Defining
 
A1
 := (AT1 A1 + AT2 A2 )−1/2
A2

we see that the span of the columns of A and  is the same, and the columns of  are orthonormal.
The cosines of the principal angles between A and B are the singular values 1 ≥ σ1 ≥ · · · ≥ σk ≥ 0
of the matrix:
ÂT B = (AT1 A1 + AT2 A2 )−1/2 AT1 .
The σ1 , . . . , σk coincide with the square roots of the eigenvalues 1 ≥ u1 ≥ · · · ≥ uk ≥ 0 of the
positive semidefinite matrix
(AT1 A1 + AT2 A2 )−1/2 AT1 A1 (AT1 A1 + AT2 A2 )−1/2 ,
which are the same as the eigenvalues of N = (AT1 A1 + AT2 A2 )−1 AT1 A1 (eigenvalues are invariant
under cyclic permutations). Consider the Cholesky decomposition M T M = AT1 A1 + AT2 A2 .
Then the eigenvalues of N equal the eigenvalues of
U = (M T )−1 AT1 A1 M −1 .
We use now some facts about the multivariate Beta distribution (see [37, §3.3]). By its definition,
the matrix U has a Betak ( 12 l, 12 (n − l)) distribution. [37, Theorem 3.3.4] states that the joint
density of the eigenvalues of U is given by
k2 k
π 2 Γk n2


(l−k−1)/2
Y Y
(n−l−k−1)/2
(7.1) u j (1 − u j ) (ui − uj ).
Γk k2 Γk 2l Γk n−l




2 j=1 i<j

Recall now that ui = (cos θi )2 for i = 1, . . . , k, which implies the change of variable dui =
−2 cos θi sin θi dθi , and thus (7.1) becomes the stated density in (3.2).

7.2. Proof of Lemma 4.3. We begin with a general reasoning. Assume that A ∈ e(B). The
unit normal vectors of e(B) at A, up to a sign, are uniquely determined by A. Lemma 4.1
provides an explit description for them as follows. Let a1 , a2 , . . . , ak and a1 , b2 , . . . , bn−k be the
orthonormal bases given by Lemma 3.1 for A and B, respectively; note that dim(A ∩ B) = 1
since A ∈ e(B). In particular, hai , bi i = cos θi for i ≤ k, where θ1 ≤ . . . ≤ θk are the principal
angles between A and B. Let Rf = (A + B)⊥ with kf k = 1. According to Lemma 4.1, the unit
vector ν := f ∧a2 ∧· · ·∧ak spans the normal space of e(B) at A. (We use here the representation
of elements of G(k, n) and its tangent spaces by vectors in Λk Rn ; cf. §2.1.)
Fix now B ∈ G(n − k, n) and recall that the functions ϑ1 , ϑ2 : G(k, n) → R give the smallest
and second smallest principal angle, respectively, between A ∈ G(k, n) and B. Let 0 < ǫ ≤ δ <
π/2 and put T := {A ∈ G(k, n) | ϑ1 (A) ≤ ǫ, ϑ2 (A) ≥ δ}. We first prove that T ⊥ (e(B)δ , ǫ) ⊆ T .
For A ∈ e(B) we consider the curve NA (t) := (a1 cos t + f sin t) ∧ a2 ∧ · · · ∧ ak for t ∈ R. We
note that NA (t) arises from A by a rotation with the angle t in the (oriented) plane spanned
by a1 and f , fixing the vectors in the orthogonal complement spanned by a2 , . . . , ak . We have
NA (0) = A and ṄA (0) = ν. It is well known (see, e.g., [13]) that NA (t) is the geodesic through
 PROBABILISTIC SCHUBERT CALCULUS 39

A with speed vector ν, that is, NA (t) = expA (tν) and NA (0) = A. From the definition of the
ǫ-tube, we therefore have
(7.2) T ⊥ (e(B)δ , ǫ) = {NA (t) | A ∈ e(B)δ , |t| ≤ ǫ}.
Since both a1 and f are orthogonal to b2 , . . . , bn−k , the principal angles between NA (t) and B are
|t|, θ2 , . . . , θk , where haj , bj i = cos θj . By our assumption A ∈ e(B)δ , we have δ ≤ θ2 ≤ . . . ≤ θk .
Hence we see that |t| is the smallest principle angle between NA (t) and B if |t| ≤ δ. We have
thus verified that T ⊥ (e(B)δ , ǫ) ⊆ T .
For the other inclusion, let C ∈ T ; assume C ∈ / e(B)δ , otherwise clearly C ∈ T ⊥ (e(B)δ , ǫ).
Let (c1 , . . . , ck ) and (b1 , . . . , bn−k ) be orthonormal bases of C and B, respectively, as provided
by Lemma 3.1. So we have hbj , cj i = cos ϑj (C) for all j, where ϑ1 (C) ≤ ǫ and δ ≤ ϑ2 (C) ≤
. . . ≤ ϑk (C) by the assumption C ∈ T . In particular, b1 is orthogonal to c2 , . . . , ck and b1 , c1
are linearly independent. We define A ∈ G(k, n) as the space spanned by b1 , c2 , . . . , ck . By
construction, ϑ1 (A) = 0 and ϑj (A) = ϑj (C) for j ≥ 2. Hence, A ∈ e(B)δ . Let NA (t) ∈ G(k, n)
be defined as above as the space resulting from A by a rotation with the angle t in the oriented
plane spanned by b1 and c1 . By construction, we have C = NA (τ ) where τ = ϑ1 (C) ≤ ǫ.
Therefore, we indeed have C ∈ T ⊥ (e(B)δ , ǫ) by (7.2). 
7.3. Proof of Lemma 4.5. More generally, we consider a semialgebraic set Y ⊆ RPn−1 of
dimension d ≤ n − k − 1. Consider the semialgebraic set
C(Y ) := {(A, y) ∈ G(k, n) × Y | y ∈ P(A)},
together with the projections on the two factors: π1 : C(Y ) → G(k, n) and π2 : B(Y ) → Y . Since
Z(Y ) = π1 (C(Y )), the set Z(Y ) is semialgebraic. Note that C(Y ) is compact if Y compact,
and C(Y ) is connected if Y is connected (since π1 is continuous). In order to determine the
dimension of C(Y ), we note that the fiber π2−1 (y) over y ∈ Y is isomorphic to G(k − 1, n − 1).
As a consequence, we get
(7.3) dim C(Y ) = dim Y + (k − 1)(n − k).
The fibers π1−1 (A) over A ∈ Z(Y ) consist of exactly one point, except for the A lying in the
exceptional set
Z (2) (Y ) := {A ∈ Z(Y ) | P(A) ∩ Y consists of at least two points}.
Note that P(A) ∩ Y consists of one point only, for all A ∈ Z(Y ) \ Z (2) (Y ).
In order to show that dim Z (2) (Y ) < dim Z(Y ), we consider the semialgebraic set
C (2) (Y ) := {(A, y1 , y2 ) ∈ G(k, n) × Y × Y | y1 , y2 ∈ P(A), y1 6= y2 }
with the corresponding projections π3 : C (2) (Y ) → G(k, n) and π4 : C (2) (Y ) → Y × Y . Note
that Z (2) (Y ) = π3 (C (2) (Y )). For (y1 , y2 ) ∈ Y × Y such that y1 6= y2 , we have dim π4−1 (y1 , y2 ) =
k(n − k) − 2(n − k), since the fibers of π4 are isomorphic to G(k − 2, n − 2). Therefore,
dim C (2) (Y ) = 2 dim Y + k(n − k) − 2(n − k) ≤ dim Y + n − k − 1 + k(n − k) − 2(n − k)
(7.3)
(7.4) = dim Y + (k − 1)(n − k) − 1 = dim C(Y ) − 1,
and we see that dim(C(Y ) \ C (2) (Y )) = dim C(Y ). For the projection f : C (2) (Y ) → C(Y )
defined by f (A, y1 , y2 ) := (A, y1 ), we have π1−1 (Z (2) (Y )) = f (C (2) (Y )), hence
dim π1−1 (Z (2) (Y )) ≤ dim f (C (2) (Y )) ≤ dim C (2) (Y ) < dim C(Y ).
Moreover, the projection C(Y ) \ C (2) (Y ) → Z(Y ) \ Z (2) (Y ), (A, y) 7→ y is bijective, hence
(7.4)
dim(Z(Y ) \ Z (2) (Y )) = dim(C(Y ) \ C (2) (Y )) = dim C(Y ).
40 PETER BÜRGISSER AND ANTONIO LERARIO

Using dim Z(Y ) ≤ dim C(Y ), we see that

(7.5) dim Z(Y ) = dim C(Y ) = dim Y + (k − 1)(n − k),

(7.6) dim Z (2) (Y ) ≤ dim C (2) (Y ) < dim C(Y ) = dim Z(Y ).
In the special case where dim X = n − k − 1, we conclude that Z(X) is a hypersurface.
We consider now the following set of “bad” A ∈ Z(X):

S(X) := Z (2) (X) ∪ Z(Sing(X)) ∪ Sing(Z(X)) ∪ π1 (Sing(C(X))).

We claim that this semialgebraic set has dimension strictly less than Z(X). This follows for
Z (2) (X) from (7.6), for Z(SingX) from (7.5) applied to Y = Sing(X), for π1 (Sing(C(X)) from
Proposition 2.2 and (7.5), and finally for Sing(Z(X)) by Proposition 2.2.
Thus generic points A ∈ Z(X) are not in S(X) and hence satisfy the following:
(1) the intersection P(A) ∩ X consists of one point only (let us denote this point by p);
(2) the point p is a smooth point of X;
(3) the point A is a regular point of Z(X),
(4) the point (A, p) is a regular point of C(X).
It remains to prove that for every A ∈ Z(X)\S(X) we have (4.3). Let us take (A, p) ∈ C(X)
with A 6∈ S(X). We work in local coordinates (w, y) ∈ RN × Rn−1 on a neighborhood U of
(A, p) in G(k, n) × RPn−1 , where N = k(n − k). For simplicity we center the coordinates on the
origin, so that (0, 0) are the coordinates of (A, p). In this coordinates, the set C(X) ∩ U can be
described as:
(7.7) C(X) ∩ U = {(w, x) ∈ RN × Rn−1 | F (w, x) = 0, G(x) = 0},
where F (w, x) = 0 represents the reduced local equations describing the condition x ∈ P(W )
and G(x) = 0 the reduced local equations giving the condition x ∈ X. Since (A, p) is a regular
point of C(X), the tangent space of C(X) at (A, p) is described by:
T(A,p) C(X) = {(ẇ, ẋ) | (D(0,0) F )(ẇ, ẋ) = 0, (D0 G)ẋ = 0}.

Note that p is a smooth point of X, hence (D0 G)ẋ = 0 is the equation for the tangent space
to X at p. On the other hand, since Z(X) = π1 (C(X)), we have:


(7.8) TA Z(X) = D(0,0) π1 T(A,p) C(X) .
Let now B ⊆ Rn be the linear space corresponding to Tp X and let us write the equations for
C(B) in the same coordinates as above (by construction we have (A, p) ∈ C(B)):
C(B) ∩ U = {(w, x) | F (w, x) = 0, (D0 G)x = 0}.
Note that the same equation F (w, x) = 0 as in (7.7) appears here (recall that this is the equation
describing x ∈ P(W )), but now G = 0 is replaced with its linearization at zero. In particular:
T(A,p) C(B) = {(ẇ, ẋ) | (D(0,0) F )(ẇ, ẋ) = 0, (D0 G)ẋ = 0},
which coincides with (7.8). Since Ω(B) = π1 (C(B)), this finally implies:



TA Ω(B) = D(0,0) π1 T(A,p) C(B) = D(0,0) π1 T(A,p) C(X) = TA Z(X),
which finishes the proof. 
 PROBABILISTIC SCHUBERT CALCULUS 41

7.4. Proof of Proposition 5.12. We extend the function f to Rk×m \ {0} by setting f (X) :=
f (X/kXk), denoting it by the same symbol. Similarly, we extend g by setting g(σ) := f (diagk,m (σ)).
We assume now that X ∈ Rk×m has i.i.d. standard Gaussian entries. Then we can write
1
Z Z
km−1
f (X) dS = E f (X) = g(σ) pSVD (σ1 , . . . , σk ) dσ1 · · · dσk ,
|S km−1 | S km−1 σ1 >...>sk

where pSVD denotes the joint density of the ordered singular values of X. This density can be
derived as follows. The joint density of the ordered eigenvalues λ1 > · · · > λk > 0 of the Wishart
distributed matrix XX T is known to be [37, Corollary 3.2.19]
Pk k
1
Y m−k−1 Y
p(λ1 , . . . , λk ) = ck,m · e− 2 i=1 λi
λi 2
(λi − λj ),
i=1 i<j

where
k2
2k π 2
ck,m := km
.
2 2 Γk k2 Γk m


2

From this, using λi (XX T ) = σi (X)2 and the change of variable dλi = 2σi dσi , we obtain
Pk k
− 12 σi2
Y Y
pSVD (σ1 , . . . , σk ) = ck,m · e i=1 σim−k (σi2 − σj2 ).
i=1 i<j

As a consequence, we obtain
Z Pk k
− 21 σi2
Y Y
E f (X) = ck,m g(σ) e i=1 σim−k (σi2 − σj2 ) dσ1 · · · dσk
σ1 >···>σk >0 i=1 i<j
Z ∞ Z k
1 2 Y Y
= ck,m rkm−1 e− 2 r dr g(θ) θim−k (θi2 − θj2 ) dS k−1 (θ),
k−1
0 θ∈S+ i=1 i<j

where in the second line we have switched to polar coordinates σ = rθ with θ ∈ S k−1 and r ≥ 0.
Note that the power of the r-variable arises as
 
k
k(m − k) + 2 + k − 1 = km − 1.
2
R∞ 1 2 km
Using 0 rkm−1 e− 2 r dr = Γ( km2 )2
2 −1 , we obtain

  k
km
Z
km
Y Y
E f (X) = ck,m Γ 2 2 −1 g(θ) θim−k (θi2 − θj2 ) dS k−1 .
2 k−1
S+ i=1 1≤i<j≤k

It is immediate to verify that:

 
km km |O(k)||S(k, m)|
|S km−1 | · ck,m Γ 2 2 −1 = ,
2 2k
which completes the proof. 

7.5. Generalized Poincaré’s formula in homogeneous spaces. The purpose of this section
is to prove the kinematic formula in homogeneous spaces for multiple intersections and to derive
Theorem 3.19. The proofs are similar to [26], to which we refer the reader for more details.
42 PETER BÜRGISSER AND ANTONIO LERARIO

7.5.1. Definitions and statement of the theorem. In the following, G denotes a compact Lie group
with a left and right invariant Riemannian metric.2 See [6] for background on Lie groups. We
denote by e ∈ G the identity element and by Lg : G → G, x 7→ gx the left translation by g ∈ G.
The derivatives of Lg will be denoted by g∗ : Te G → Tg G. By assumption, this map is isometric.
In (3.5) we defined a quantity for capturing the relative position of linear subspaces of a
Euclidean vector space. We can extend this notion to P linear subspaces Vi ⊆ Tgi G in tangent
spaces of G at any points g1 , . . . , gm ∈ G, assuming i dim Vi ≤ dim G. This is done by
left-translating the gi to the identity e: so we define
σ(V1 , . . . , Vm ) := σ (g1 )−1 −1


∗ V1 , . . . , (gm )∗ Vm .

Let now K ⊆ G be a closed Lie subgroup and denote by p : G → G/K the quotient map.
We endow K with the Riemannian structure induced by its inclusion in G, and G/K with the
Riemannian structure defined by declaring p to be a Riemannian submersion. For example,
when G = O(n) with the invariant metric defined in Section 2.8 and K = O(k) × O(n − k), then
G/K with the quotient metric is isometric to the Grassmannian G(k, n) with the metric defined
in Section 2.1.
Note that G acts naturally by isometries on G/K; if g ∈ G and y ∈ G/K, we denote by gy
the result of the action. Further, we denote by y0 = p(e) the projection of the identity element.
The multiplication with an element k ∈ K fixes the point y0 ; as a consequence, the differential
of k induces a map denoted k∗ : Ty0 G/K → Ty0 G/K, so that we have an induced action of K
on Ty0 G/K.
Given a submanifold X of a Riemannian manifold M , we denote by N X its normal bundle
in M (i.e., for all x ∈ X the vector space Nx X is the orthogonal complement to Tx X in Tx M ).
Also, the restriction of the Riemannian metric of M to X allows to define a volume density
on RX; if f : X → R is an integrable function, we denote its integral with respect to this density
by X f (x) dx.
Definition 7.1. For given submanifolds Y1 , . . . , Ym ⊆ G/K, we define the function
σK : Y1 × · · · × Ym → R
as follows. For (y1 , . . . , ym ) ∈ Y1 × · · · × Ym let ξi ∈ G be such that ξi yi = y0 for all i. We define
σK (y1 , . . . , ym ) := E (k1 ,...,km )∈K m σ(k1∗ ξ1 ∗ Ny1 Y1 , . . . , km∗ ξm ∗ Nym Ym ),
where the expectation is taken over a uniform (k1 , . . . , km ) ∈ K × . . . × K.
The reader should compare this definition with [26, Def. 3.3], which is just a special case.
The main result of this section is the following generalization of Poincaré’s kinematic formula
for homogeneous spaces, as stated in [26, Thm. 3.8] for the intersection of two manifolds. We
provide a proof, since the more general result is crucial for our work and we were unable to find
it in the literature.
Pm
Theorem 7.2. Let Y1 , . . . , Ym be submanifolds of G/K such that i=1 codimG/K Yi ≤ dim G/K.
Then, for almost all (g1 , . . . , gm ) ∈ Gm , the manifolds g1 Y1 , . . . , gm Ym intersect transversely, and
1
Z
E (g1 ,...,gm )∈Gm |g1 Y1 ∩ · · · ∩ gm Ym | = σK (y1 , . . . , ym ) dy1 · · · dym ,
|G/K|m−1 Y1 ×···×Ym
where the expectation is taken over a uniform (g1 , . . . , gm ) ∈ G × · · · × G.

2The compactness assumption is not essential, but simplifies the statements, for example the modular function
of G is constant and does not have to be taken into account, see [26, §2.3].
 PROBABILISTIC SCHUBERT CALCULUS 43

We note that when G acts transitively on the tangent spaces to each Yi (formally defined
as in Definition 3.4), then the function σK : Y1 × · · · × Ym → R introduced in Definition 7.1 is
constant. As a consequence we obtain:
Corollary 7.3. Under the assumptions of Theorem 7.2, if moreover G acts transitively on the
tangent spaces to Yi for i = 1, . . . , m, then we have
n
Y |Yi |
E (g1 ,...,gm )∈G |g1 Y1 ∩ · · · ∩ gm Ym | = σK (y1 , . . . , ym ) · |G/K| ·
m ,
i=1
|G/K|
where (y1 , . . . , ym ) is any point of Y1 × · · · × Ym .
Let us look now at the special case G = O(n) and K = O(k) × O(n − k). If Y1 , . . . , Ym
are coisotropic hypersurfaces of G(k, n), then G acts transitively on their tangent spaces by
Proposition 4.6. Moreover, when m = k(n − k), it is easy to check that the constant value of σK
equals the real average scaling factor α(k, n − k) defined in Definition 3.18. Hence in this case,
the statement of Corollary 7.3 coincides with the statement of Theorem 3.19.
7.5.2. The kinematic formula in G. As before, G denotes a compact Lie group. We derive first
Theorem 7.2 in the special case K = {e}, which is the following result (we can w.l.o.g. assume
gm = e).
Pm
Lemma 7.4. Let X1 , . . . , Xm be submanifolds of G such that i=1 codimG Xi ≤ dim G. Then
Z Z
|g1 X1 ∩· · ·∩gm−1 Xm−1 ∩Xm | dg1 · · · dgm−1 = σ(Nx1 X1 , . . . , Nxm Xm ) dx1 · · · dxm ,
Gm−1 X1 ×···×Xm

In the special case of intersecting two submanifolds, this is an immediate consequence of the
following “basic integral formula” from [26, §2.7] (take h = 1).
Proposition 7.5. Let M1 , M2 be submanifolds of G such that codimG M1 +codimG M2 ≤ dim G.
For almost all g ∈ G, the manifolds M1 and gM2 intersect transversely, and if h is an integrable
function on M1 × M2 , then
Z Z Z
h(ϕg (y)) dy dg = h(x1 , x2 ) σ(Nx1 M1 , Nx2 M2 ) dx1 dx2 ,
G gM1 ∩M2 M1 ×M2

where ϕg : gM1 ∩ M2 → M1 × M2 is the function given by ϕg (y) := (g −1 y, y).

In order to reduce the general case to that of intersecting two submanifolds, we first establish
a linear algebra identity.
Lemma 7.6. For subspaces V1 , . . . Vℓ , W, Z of a Euclidean vector space, we have
σ(V1 , . . . , Vℓ , W, Z) = σ(V1 , . . . , Vℓ , W + Z) · σ(W, Z).
Proof. We may assume that W ∩ Z = 0, since otherwise both sides of the identity are zero. Let
us denote by (vi,1 , . . . , vi,di ), (w1 , . . . , wa ), and (z1 , . . . , zb ) orthonormal bases of Vi , W , and Z
respectively. Moreover, we denote by (w1 , . . . , wa , z̃1 , . . . , z̃b ) an orthonormal basis for Z + W
obtained by completing (w1 , . . . , wa ). By definition we have:
^

(7.9) σ(V1 , . . . , Vℓ , W + Z) = vi,j ∧ w1 ∧ · · · ∧ wa ∧ z̃1 ∧ · · · ∧ z̃b .
i,j

On the other hand, since W + Z = span{w1 , . . . , wa , z1 , . . . , zb }, we have

w1 ∧ · · · ∧ wa ∧ z1 ∧ · · · ∧ zb w1 ∧ · · · ∧ wa ∧ z1 ∧ · · · ∧ zb
w1 ∧· · ·∧wa ∧ z̃1 ∧· · ·∧ z̃b = ± =± .
kw1 ∧ · · · ∧ wa ∧ z1 ∧ · · · ∧ zb k σ(W, Z)
44 PETER BÜRGISSER AND ANTONIO LERARIO

Substituting the last line into (7.9) and recalling the definition of σ(V1 , . . . , Vℓ , W, Z) from (3.5),
the assertion follows. 

Corollary 7.7. Let M1 , M2 be submanifolds of G and V1 , . . . , Vℓ be linear subspaces

P of tangent
spaces of G (possibly at different points), such that codimG M1 +codimG M2 + i dim Vi ≤ dim G.
Then we have
Z Z Z
σ(V1 , . . . , Vℓ , Ny (gM1 ∩ M2 )) dy dg = σ(V1 , . . . , Vℓ , Nx1 M1 , Nx2 M2 ) dx1 dx2 .
G gM1 ∩M2 M1 ×M2

Proof. We apply Proposition 7.5 with the function h : M1 × M2 → R defined by

h(x1 , x2 ) := σ V1 , . . . , Vℓ , (x1 )−1 −1


∗ Nx1 M1 + (x2 )∗ Nx2 M2 .

When gM1 and M2 intersect transversely, we have Ny (gM1 ∩ M2 ) = Ny (gM1 ) + Ny (M2 ), which
implies h(ϕg (y)) = σ(V1 , . . . , Vℓ , Ny (gM1 ∩ gM2 )) for y ∈ gM1 ∩ M2 . Hence we obtain with
Proposition 7.5,

Z Z Z Z 
σ(V1 , . . . , Vℓ , Ny (gM1 ∩ M2 )) dy dg = h(ϕg (y)) dy dg
G gM1 ∩M2 G gM1 ∩M2
Z
= h(x1 , x2 )σ(Nx1 M1 , Nx2 M2 ) dx1 dx2
M1 ×M2
Z
= σ(V1 , . . . , Vℓ , Nx1 M1 , Nx2 M2 ) dx1 dx2 ,
M1 ×M2

where the last equality is due to Lemma 7.6. 

Proof of Lemma 7.4. Recall that we already established Lemma 7.4 in the case m = 2 as a
consequence of Proposition 7.5. Let m ≥ 3 and abbreviate Y := g2 X2 ∩ Z, where Z := g3 X3 ∩
. . . ∩ gm−1 Xm−1 ∩ Xm . Then we have
Z Z 


g1 X1 ∩ g2 X2 ∩ . . . gm−1 Xm−1 ∩ Xm dg1 dg2 . . . dgm−1
Gm−2 g1 ∈G
Z Z Z 
= σ(Nx1 X1 , Ny Y ) dx1 dy dg2 . . . dgm−1
Gm−2 X1 Y
Z Z Z Z 
= σ(Nx1 X1 , Ny (g2 X2 ∩ Z) dg2 dy dg3 . . . dgm−1 dx1
X1 Gm−3 g ∈G Y
Z Z Z 2 Z 
= σ(Nx1 X1 , Nx2 X2 , Nz Z) dg2 dg3 . . . dgm−1 ,
X1 Gm−3 x2 ∈X2 z∈Z

where we first applied Lemma 7.4 in the case m = 2, then interchanged the order of integration,
and after that used Corollary 7.7. Proceeding analogously, we see that the above integral indeed
equals
Z
σ(Nx1 X1 , . . . , Nxm Xm ) dx1 · · · dxm ,
X1 ×···×Xm
which completes the proof. 

Proof of Theorem 7.2. We consider Xi := p−1 (Yi ), which is a submanifold, since the projec-
tion p : G → G/K is a submersion, cf. [9, Thm. A.15]. Moreover, g1 X1 , . . . , gm Xm intersect
 PROBABILISTIC SCHUBERT CALCULUS 45

transversally if Y1 , . . . , Ym do so. We can rewrite the integral in the statement as

E := E (g1 ,...,gm )∈Gm |g1 Y1 ∩ · · · ∩ gm Ym |
= E (g1 ,...,gm−1 )∈Gm−1 |g1 Y1 ∩ · · · ∩ gm−1 Ym−1 ∩ Ym |
1 1
Z
= |g1 X1 ∩ · · · ∩ gm−1 Xm−1 ∩ Xm | dg1 · · · dgm−1 ,
|K| |G|m−1 Gm−1
For justifying the last equality, note that the coarea formula [9, Thm. 17.8] yields |p−1 (Y )| =
|K||Y | for any submanifold Y of G/K, since p is a Riemannian submersion. Applying Lemma 7.4
to the integral in the last line, we obtain
1 1
Z
E= σ(Nx1 X1 , . . . , Nxm Xm ) dx1 · · · dxm .
|K| |G|m−1 X1 ×···×Xm
The projection P : X1 × · · · × Xm → Y1 × · · · × Ym defined by p(x1 , . . . , xm ) := (p(x1 ), . . . , p(xm ))
is a Riemannian submersion with fibers isometric to K m . Using the coarea formula
Z
σ(Nx1 X1 , . . . , Nxm Xm ) dx1 · · · dxm
X1 ×···×Xm
Z Z
= σ(Nx1 X1 , . . . , Nxm Xm ) dx dy
y∈Y1 ×···Ym x∈P −1 (y)
Z
= |K|m σK (y1 , . . . , ym ) dy,
y∈Y1 ×···Ym

where the last equality is due to Definition 7.1. (Note that here is where we use the right
invariance of the metric under the action of the elements in K, as one can verify by a careful
inspection of the last steps. The reader can see the proof of [26, Thm. 3.8], which is almost
identical and where all the details of the calculation are shown). This completes the proof. 

References
[1] Pierre-Antoine Absil, Alan Edelman, and Plamen Koev. On the largest principal angle between random
subspaces. Linear Algebra Appl., 414(1):288–294, 2006.
[2] Dennis Amelunxen. Geometric analysis of the condition of the convex feasibility problem. PhD thesis, Insti-
tute of Mathematics, University of Paderborn, 2011.
[3] Saugata Basu, Antonio Lerario, Erik Lundberg, and Chris Peterson. Random fields and the enumerative
geometry of lines on real and complex hypersurfaces. arXiv:1610.01205, 2016.
[4] Jacek Bochnak, Michel Coste, and Marie-Françoise Roy. Real algebraic geometry, volume 36 of Ergebnisse
der Mathematik und ihrer Grenzgebiete (3). Springer-Verlag, Berlin, 1998.
[5] Paul Breiding. The expected number of Z-eigenvalues of a real gaussian tensor. SIAM Journal on Applied
Algebra and Geometry, 1(1):254–271, 2017.
[6] Theodor Bröcker and Tammo tom Dieck. Representations of compact Lie groups, volume 98 of Graduate
Texts in Mathematics. Springer-Verlag, New York, 1995. Translated from the German manuscript, Corrected
reprint of the 1985 translation.
[7] Peter Bürgisser. Average Euler characteristic of random real algebraic varieties. C. R. Math. Acad. Sci.
Paris, 345(9):507–512, 2007.
[8] Peter Bürgisser. Condition of intersecting a projective variety with a varying linear subspace. SIAM Journal
on Applied Algebra and Geometry, 1(1):111–125, 2017.
[9] Peter Bürgisser and Felipe Cucker. Condition: The geometry of numerical algorithms, volume 349 of
Grundlehren der Mathematischen Wissenschaften. Springer, Heidelberg, 2013.
[10] Alexandru Dimca. Topics on real and complex singularities. Advanced Lectures in Mathematics. Friedr.
Vieweg & Sohn, Braunschweig, 1987. An introduction.
[11] Jan Draisma and Emil Horobeţ. The average number of critical rank-one approximations to a tensor. Linear
Multilinear Algebra, 64(12):2498–2518, 2016.
[12] Jan Draisma, Emil Horobeţ, Giorgio Ottaviani, Bernd Sturmfels, and Rekha R. Thomas. The Euclidean
distance degree of an algebraic variety. Found. Comput. Math., 16(1):99–149, 2016.
46 PETER BÜRGISSER AND ANTONIO LERARIO

[13] Alan Edelman, Tomás A. Arias, and Steven T. Smith. The geometry of algorithms with orthogonality
constraints. SIAM J. Matrix Anal. Appl., 20(2):303–353, 1999.
[14] Davis Eisenbud and Joe Harris. 3264 and all that: a second course in algebraic geometry. Cambridge Uni-
versity Press, 2016.
[15] S. Finashin and V. Kharlamov. Abundance of 3-planes on real projective hypersurfaces. Arnold Math. J.,
1(2):171–199, 2015.
[16] Watson Fulks and J. O. Sather. Asymptotics. II. Laplace’s method for multiple integrals. Pacific J. Math.,
11:185–192, 1961.
[17] William Fulton. Intersection theory, volume 2 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3.
Folge. A Series of Modern Surveys in Mathematics. Springer-Verlag, Berlin, second edition, 1998.
[18] Yan V. Fyodorov, Antonio Lerario, and Erik Lundberg. On the number of connected components of random
algebraic hypersurfaces. J. Geom. Phys., 95:1–20, 2015.
[19] Damien Gayet and Jean-Yves Welschinger. Lower estimates for the expected Betti numbers of random real
hypersurfaces. J. Lond. Math. Soc. (2), 90(1):105–120, 2014.
[20] Damien Gayet and Jean-Yves Welschinger. Expected topology of random real algebraic submanifolds. J.
Inst. Math. Jussieu, 14(4):673–702, 2015.
[21] Damien Gayet and Jean-Yves Welschinger. Betti numbers of random real hypersurfaces and determinants
of random symmetric matrices. J. Eur. Math. Soc. (JEMS), 18(4):733–772, 2016.
[22] Israel M. Gelfand, Mikhail M. Kapranov, and Andrei V. Zelevinsky. Discriminants, resultants and multidi-
mensional determinants. Modern Birkhäuser Classics. Birkhäuser Boston, Inc., Boston, MA, 2008. Reprint
of the 1994 edition.
[23] A. Gray and L. Vanhecke. The volumes of tubes in a Riemannian manifold. Rend. Sem. Mat. Univ. Politec.
Torino, 39(3):1–50 (1983), 1981.
[24] Phillip Griffiths and Joseph Harris. Principles of algebraic geometry. Wiley Classics Library. John Wiley &
Sons, Inc., New York, 1994. Reprint of the 1978 original.
[25] Joe Harris. Algebraic geometry, volume 133 of Graduate Texts in Mathematics. Springer-Verlag, New York,
1995. A first course, Corrected reprint of the 1992 original.
[26] Ralph Howard. The kinematic formula in Riemannian homogeneous spaces. Mem. Amer. Math. Soc.,
106(509):vi+69, 1993.
[27] Steven L. Kleiman. The transversality of a general translate. Compositio Math., 28:287–297, 1974.
[28] Steven L. Kleiman and Dan Laksov. Schubert calculus. Amer. Math. Monthly, 79:1061–1082, 1972.
[29] Kathlén Kohn. Coisotropic hypersurfaces in the Grassmannian. arXiv:1607.05932, 2016.
[30] Eric Kostlan. On the distribution of roots of random polynomials. In From Topology to Computation: Pro-
ceedings of the Smalefest (Berkeley, CA, 1990), pages 419–431. Springer, New York, 1993.
[31] S. E. Kozlov. Geometry of real Grassmannian manifolds. I, II, III. Zap. Nauchn. Sem. S.-Peterburg. Otdel.
Mat. Inst. Steklov. (POMI), 246(Geom. i Topol. 2):84–107, 108–129, 197–198, 1997.
[32] Antonio Lerario. Random matrices and the average topology of the intersection of two quadrics. Proc. Amer.
Math. Soc., 143(8):3239–3251, 2015.
[33] Antonio Lerario and Erik Lundberg. Gap probabilities and Betti numbers of a random intersection of
quadrics. Discrete Comput. Geom., 55(2):462–496, 2016.
[34] Stanislaw Lojasiewicz. Ensembles semi-analytiques. IHES Lecture Notes, 1965.
[35] Laurent Manivel. Symmetric functions, Schubert polynomials and degeneracy loci, volume 6 of SMF/AMS
Texts and Monographs. American Mathematical Society, Providence, RI; Société Mathématique de France,
Paris, 2001.
[36] John W. Milnor and James D. Stasheff. Characteristic classes. Princeton University Press, Princeton, N. J.;
University of Tokyo Press, Tokyo, 1974. Annals of Mathematics Studies, No. 76.
[37] Robb J. Muirhead. Aspects of multivariate statistical theory. John Wiley & Sons, Inc., New York, 1982.
Wiley Series in Probability and Mathematical Statistics.
[38] David Mumford. Algebraic geometry. I. Springer-Verlag, Berlin-New York, 1976. Complex projective vari-
eties, Grundlehren der Mathematischen Wissenschaften, No. 221.
[39] Igor Rivin. Surface area and other measures of ellipsoids. Adv. in Appl. Math., 39(4):409–427, 2007.
[40] Luis A. Santaló. Integral geometry and geometric probability. Addison-Wesley Publishing Co., Reading,
Mass.-London-Amsterdam, 1976. With a foreword by Mark Kac, Encyclopedia of Mathematics and its
Applications, Vol. 1.
[41] Rolf Schneider. Convex bodies: the Brunn-Minkowski theory, volume 151 of Encyclopedia of Mathematics
and its Applications. Cambridge University Press, Cambridge, expanded edition, 2014.
[42] H. Schubert. Anzahl-Bestimmungen für Lineare Räume beliebiger Dimension. Acta Math., 8(1):97–118, 1886.
 PROBABILISTIC SCHUBERT CALCULUS 47

[43] Igor R. Shafarevich. Basic algebraic geometry. 2. Springer, Heidelberg, third edition, 2013. Schemes and
complex manifolds, Translated from the 2007 third Russian edition by Miles Reid.
[44] Michael Shub and Stephen Smale. Complexity of Bézout’s theorem II: volumes and probabilities. In F. Eys-
sette and A. Galligo, editors, Computational Algebraic Geometry, volume 109 of Progress in Mathematics,
pages 267–285. Birkhäuser, 1993.
[45] Frank Sottile. Pieri’s formula via explicit rational equivalence. Canad. J. Math., 49(6):1281–1298, 1997.
[46] Frank Sottile. The special Schubert calculus is real. Electron. Res. Announc. Amer. Math. Soc., 5:35–39,
1999.
[47] Frank Sottile. Enumerative real algebraic geometry. In Algorithmic and quantitative real algebraic geometry
(Piscataway, NJ, 2001), volume 60 of DIMACS Ser. Discrete Math. Theoret. Comput. Sci., pages 139–179.
Amer. Math. Soc., Providence, RI, 2003.
[48] Frank Sottile. Real solutions to equations from geometry, volume 57 of University Lecture Series. American
Mathematical Society, Providence, RI, 2011.
[49] Michael Spivak. A comprehensive introduction to differential geometry. Vol. I. Publish or Perish, Inc.,
Wilmington, Del., second edition, 1979.
[50] Jacek Stasica. Smooth points of a semialgebraic set. Ann. Polon. Math., 82(2):149–153, 2003.
[51] Bernd Sturmfels. The Hurwitz form of a projective variety. J. Symbolic Comput., 79(part 1):186–196, 2017.
[52] Terence Tao. Topics in random matrix theory, volume 132 of Graduate Studies in Mathematics. American
Mathematical Society, Providence, RI, 2012.
[53] R. C. Thompson. Singular values, diagonal elements, and convexity. SIAM J. Appl. Math., 32(1):39–63,
1977.
[54] Ravi Vakil. Schubert induction. Ann. of Math. (2), 164(2):489–512, 2006.
[55] Richard A. Vitale. Expected absolute random determinants and zonoids. Ann. Appl. Probab., 1(2):293–300,
1991.
[56] Roderick Wong. Asymptotic approximations of integrals, volume 34 of Classics in Applied Mathematics.
Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2001. Corrected reprint of the
1989 original.
[57] Yung-chow Wong. Conjugate loci in Grassmann manifolds. Bull. Amer. Math. Soc., 74:240–245, 1968.

Institute of Mathematics, Technische Universität Berlin

E-mail address: [email protected]

SISSA (Trieste)
E-mail address: [email protected]