Algebraic geometry

for geometric modeling

Ragni Piene

SIAM AG17
Atlanta, Georgia, USA
August 1, 2017
Applied algebraic geometry in the old days:
EU Training networks

I GAIA – Application of approximate algebraic geometry in

industrial computer aided geometry
I GAIA II – Intersection algorithms for geometry based
IT-applications using approximate algebraic methods
I SAGA – ShApes, Geometry, and Algebra

and Centre of Mathematics for Applications (University of Oslo)

Algebraic geometry for geometric modeling?
Ron Goldman: “The main
contribution of algebraic
geometry to geometric
modeling is insight, not
computation.”

His concern is that geometric modeling is interested in metric

properties of curves and surfaces, whereas algebraic geometry is
concerned with affine and projective invariants.
Algebraic geometry can provide constructive tools for
computation. Geometric modeling needs effective numerical
methods.
Classical projective geometry

Classical geometry was real and Euclidean. To get solutions to

polynomial equations, imaginary numbers were introduced, and
projective geometry in order to have stability of intersections.
Complex projective geometry is much easier to work with than
affine real geometry.
For computers: the field Q, for pictures the field R, for proving
theorems the field C.
Projective invariants
Let X ⊂ Pn be a nonsingular projective algebraic variety.
The Chern classes of X are topological invariants.
The Euler–Poincaré characteristic χ(OX ) is a birational
invariant.
Example
If X is a surface, with Chern classes c1 (X) and c2 (X), then
Noether’s formula holds:
1
χ(OX ) = (c1 (X)2 + c2 (X)).
12
If X 0 → X is the blowing up of a point, then
c1 (X 0 )2 = c1 (X)2 − 1 and c2 (X 0 ) = c2 (X) + 1, so
χ(OX 0 ) = χ(OX ).
Polar varieties and Chern classes
Let L ⊂ Pn be a linear subspace of codimension m − k + 2,
where m = dim X.
The kth polar variety of of X ⊂ Pn with respect to L is

Mk (L) := {P ∈ X | dim(TP X ∩ L) ≥ k − 1}.

1 (1)) ∩ [X], hence we get the Todd–Eger
Its class is [Mk ] = ck (PX
relation
k  
X m − i + 1 k−i
[Mk ] = (−1)i h ci (X) ∩ [X], (1)
m−k+1
i=0

where ci (X) are the Chern classes of the tangent bundle of X

and h = c1 (OX (1)) is the class of a hyperplane.
Singular varieties
Conversely, the Chern classes are expressed in terms of the polar
varieties:
k  
i m−i+1
X
ck (X) ∩ [X] = (−1) hk−i ∩ [Mi ].
m−k+1
i=0

This works also for singular varieties: replace Mk by the closure

of Mk |Xns to get the Chern–Mather classes cM k (X).
0
Let π : X → X 0 ⊂ Pn be a (suitably generic) linear projection:
0
cM M
k (X ) = π∗ (ck (X)).
The Chern–Mather classes and the polar classes are projective
invariants (the Chern–Schwartz–MacPherson classes are
topological invariants).
Geometric interpretation: the contour curve
Let X = Z(F ) ⊂ P3 be a surface, P = (p0 : · · · : p3 ) ∈ P3 .
The first polar variety of X wrt P is the intersection of X with
its first polar:
X X
M1 (P ) = X ∩ Z( pi Fi ) = Z(F, pi Fi ),

and M1 (P ) ⊂ X is the contour curve on X under the projection

X → P2 from P .
If X is smooth, of degree d, and P is general, then
deg M1 = d(d − 1), and the projection of M1 has d(d − 1)(d − 2)
cusps.
cf. MS Algebraic Vision yesterday!
Affine space as Euclidean space
as the affine plane minus an arbitrary
plus the line at line `, and then
infinity. A converse regard the
relationship was Euclidean plane as
suggested by the affine plane
Coxeter: “Kepler’s Poncelet (1822) and with a special rule
invention of points von Staudt (1847): for associating pairs
at infinity made it regard the affine of points on ` (in
possible to regard plane as the “perpendicular
the projective plane projective plane directions”).”

Affine space + notion of perpendicularity = “Euclidean space”

Euclidean normal bundle

Let X ⊂ Pn be a variety of dimension m. Define the Euclidean

normal bundle E with respect to a non-degenerate quadric Q in
the hyperplane H∞ .
Use the polarity in H∞ ∼
= Pn−1 induced by Q to define
Euclidean normal spaces at each smooth point P ∈ X \ H∞ :

NP X = hP, (TP X ∩ H∞ )⊥ i

Then E = NX (1)∨ ⊕ OX (1), where NX is the conormal bundle

(or the “Mather–Nash conormal”, if the variety is singular).
Reciprocal polar varieties
Define reciprocal polar varieties
Mk (L)⊥ := {P ∈ X|NP X ∩ L 6= ∅},
for L linear space of codimension n − m + k. Then
k
X
[Mk⊥ ] = sk (E) ∩ [X] = 1
cj (PX (1))c1 (OX (1))k−j ∩ [X].
j=0

In particular,
m
X
⊥
deg[Mm ] = deg[Mj ].
j=0
This is the degree of the end point map and also the Euclidean
distance degree.
MS Euclidean distance degree – Thursday!
Building blocks for modeling
Making pots before the wheel was invented?

Piecewise linear Natural quadrics

Parameterized surfaces

Most algebraic surfaces used in geometric modeling are rational,

meaning that they can be obtained as the image of a map

(s, t) 7→ (p1 (s, t), p2 (s, t), p3 (s, t)) ∈ R3

for (s, t) in some plane domain D ⊂ R2 , where the pi are

polynomials or rational functions with coefficients in Q.
Typically D is either the triangle with vertices (0, 0), (1, 0), (0, 1)
or the square with vertices (0, 0), (1, 0), (0, 1), (1, 1).
Projective “triangular” surfaces

Let Md = {m0 , . . . , mN } be the set of all monomials of degree d

in three variables.
The image of the map P2 → PN given by

(s : t : u) 7→ (m0 (s, t, u) : · · · : mN (s, t, u))

is the Veronese surface of degree d2 .

Any triangular parameterized surface in P3 is a linear projection
of this surface from a center L ⊂ PN . Algebraic geometry gives
insight into the properties of such projections.
The Steiner surface (d = 2)

There are 6 different real Steiner surfaces.

A. x2 y 2 +y 2 z 2 + z 2 x2 − xyz = 0
B. x2 y 2 −y 2 z 2 + z 2 x2 − xyz = 0
C. xyz 2 + xy − x2 − z 4 − 2z 2 − 1 = 0
D. xz 2 − y 2 + z 4 = 0
E. x4 + y 2 + z 2 − 2x2 y − 2x2 z + 2yz − 4yz = 0
F. y 2 + 2yz 2 + z 4 − x = 0
A: Steiner’s Roman surface. Three real double lines meeting in
a triple point. Each line has two real pinchpoints (d = 4,  = 3,
t = 1, ν2 = 6).
B: Three real double lines meeting in a triple point. One line
has two real pinch points (d = 4,  = 3, t = 1, ν2 = 2).
C: One real double line. The line has one real pinch point
(d = 4,  = 1, ν2 = 1).
D: One simple and one double double lines meeting in a triple
point. The simple line has two real pinch points (d = 4,  = 2,
t = 1, ν2 = 2).
E: Somewhat similar to D.
F: One threefold double line containing a triple point.
Projective “tensor” surfaces

Let Ma,b = {m0 , . . . , mN 0 } be the set of all monomials in four

variables, of bidegree (a, b). Then the image of the map
0
P1 × P1 → PN given by

(s : t) × (u : v) 7→ (m0 (s : t; u : v) : · · · : mN (s : t; u : v))

is the Segre surface of degree 2ab.

Any “tensor” surface in P3 is the linear projection of a Segre
surface from some center L.
For computability, a and b should be small!
Example: a = 1, b = 2

A tensor surface of bidegree (1, 2) in 3-space is the projection of

the rational balanced normal scroll X ⊂ P5 of type (2, 2) and
degree 4 from a line L.
A classification is obtained by the position of L with respect to
X, its tangent spaces, its osculating spaces, etc.
A general projection has a twisted cubic as its double curve,
with 4 pinch points and no triple points.
Application: modeling corn leaves (Thi Ha Lê)
Use Bézier surface tensor patches [0, 1]2 → R3 :
P
(i,j) w(i,j) p(i,j) F(i,j) (t1 , t2 )
B(t1 , t2 ) = P ,
(i,j) w(i,j) F(i,j) (t1 , t2 )

where F(i,j) (t1 , t2 ) = 2i ti1 (2 − t1 )2−i 1j tj2 (1 − t2 )1−j .

General and special projections – singularities
Generic projections of the d-uple Veronese surface has
I a double curve of degree  = d (d2 + d − 3)


2
I ν2 = 6(d − 1)2 pinch points
I t = 16 (d6 + 9d3 + 44d2 ) − 2d4 − 12d + 5 triple points

Generic projections of the bidegree (a, b) Segre surface has

I a double curve of degree  = 2ab(ab − 2) + a + b
I ν2 = 4(3ab − 2(a + b) + 1) pinch points
I t = 31 4ab(a2 b2 + 11) − 8a2 b2 + 2ab(a + b) − 8(a + b) + 4
triple points
Insight: these numbers are upper bounds in the real case.
Toric surfaces

Let A = {a0 , . . . , aN } ⊂ Z2 be a lattice point configuration,

K∗ 2 → PNK the corresponding toric embedding, and XA the
closure of the image.
Let A0 be a lattice point configuration obtained from A by
removing N − 3 points.Then the toric embedding XA0 ⊂ P3 is
the (toric) linear projection of XA with center equal to the
linear span of the “removed points”.
Any toric surface in 3-space is a (special) linear projection of a
Veronese or Segre surface.
Example
Take m, n ∈ N, with gcd(m, n) = 1, m ≥ 3, m ≥ n ≥ 2,and
A = {(0, 0), (m, 0), (1, 1), (0, n)}. Then
(u, v) 7→ (1 : um : uv : v n ) gives the surface
XA = Z(xn z m wmn−m−n − y mn ) ⊂ P3 .

The surface has three singular lines: Z(x, y), Z(y, z), Z(y, w).

◦ •

• ◦ ◦ ◦ ◦ •
Real picture for m = 5, n = 2

In the affine space z = 1: In the affine space x = 1:

Toric patches (R. Krasauskas)

A toric surface patch associated with a lattice polygon ∆ ⊂ R2

is a piece of an algebraic surface parameterized by the rational
map B∆ : ∆ → R3 given by
P
b wa pa Fa (t1 , t2 )
B∆ (t1 , t2 ) = Pa∈∆ ,
b wa Fa (t1 , t2 )
a∈∆

where ∆b = ∆ ∩ Z2 , pa ∈ R3 are control points, wa > 0 are

weights, and Fa (t1 , t2 ) = ca h1 (t1 , t2 )h1 (a) · · · hr (t1 , t2 )hr (a) , where
hi are linear forms defining the r edges of ∆.
Monoid surfaces
A monoid surface in P3 of degree d is a surface with a singular
point of multiplicity d − 1.
Let Z(fd−1 ), Z(fd ) ⊂ P2 be plane curves, with no common
components and no common singular points.
The monoid surface Z(fd−1 x0 + fd ) ⊂ P3 has a rational
parameterization P2 99K P3

(a1 : a1 : a3 ) 7→ (fd (a) : fd−1 (a)a1 : fd−1 (a)a2 : fd−1 (a)a3 ).

The point O := (1 : 0 : 0 : 0) has multiplicity d − 1. All other

singular points are contained in lines on the surface through O,
corresponding to the base points Z(fd−1 ) ∩ Z(fd ) of the
parameterization.
Applications

Polynomials defining a monoid of a given degree are “sparse”

compared to all polynomials of that degree. Therefore they have
been used in approximate implicitization problems.
Real quartic monoid surfaces
Classify by the form of the projective tangent cone Z(f3 ) at O.
There are nine cases, giving nine strata in the space of all
quartic monoids, which have been described.

Z(x3 + y 3 + 5xyz − z 3 (x+y)) Z(x3 + y 3 + 5xyz − z 3 (x−y))

Algebraic splines
A draftsman’s spline:
A draftman's spline 11/15/08 4:26 PM

A draftman's spline

 
... and a Boeing draftman using it (view, don't copy! I had to pay
Boeing $100 for the right to display it here)
Algebraic spline rings
Let ∆ ⊂ Rd be a (pure) d-dimensional simplicial complex.
Let C r (∆) denote the set of piecewise polynomial functions
(algebraic splines) on ∆ of smoothness r.
C r (∆) is a ring under the usual pointwise addition and
multiplication.
The (global) polynomial functions R[x1 , . . . , xd ] are r-smooth for
any r, so
C ∞ (∆) := R[x1 , . . . , xd ] ⊂ C r (∆),
for any r.
The vector spaces Ckr (∆)

Let Ckr (∆) ⊂ C r (∆) consist of splines of degree ≤ k.

These subsets are vector spaces over R.
Standard problems, that many people have worked on, are
I to determine (upper and lower bounds for) dimR Ckr (∆),
I to construct a basis for Ckr (∆).

But also: to determine the structure of the rings C r (∆) and

their geometric interpretation.
Classical Stanley–Reisner rings

Let ∆ ⊂ Rd be a simplicial complex, with vertices {v1 , . . . , vn }.

The Stanley–Reisner ring, or face ring, of ∆ is the ring

A∆ := R[Y1 , . . . , Yn ]/I∆ ,

where I∆ is the monomial ideal generated by the products

Yi1 · · · Yij such that {vi1 , . . . , vij } is not a face of ∆.
It is known [Bruns–Gubeladze] that if two simplicial complexes
have isomorphic Stanley–Reisner rings, then they are themselves
isomorphic.
Identify Yi with the Courant function Yi (vj ) = δij extended by
linearity. Then P
i Yi = 1,

and
A∆ /( i Yi − 1) = C 0 (∆)
P

is the ring of (continuous) splines on ∆.

Call this ring the affine Stanley–Reisner ring of ∆. Then

Spec(C 0 (∆)) = Spec(A∆ ) ∩ Z( i Yi − 1) ⊂ AnR = Rn ,

P

and the points that have non-negative coordinates, give a model

of ∆.
Example 1: d = 1
Let ∆ be a one-dimensional simplicial complex with three
vertices v1 , v2 , v3 ∈ R, and assume v1 < v2 < v3 .

v1 v2 v3

We have

C 0 (∆) = A∆ /( 3i=1 Yi −1) = R[Y1 , Y2 , Y3 ]/(Y1 Y3 , Y1 +Y2 +Y3 −1)

P

Spec(C 0 (∆)) = Z(Y1 , Y2 + Y3 − 1) ∪ Z(Y3 , Y1 + Y2 − 1) ⊂ R3

The segments of these two lines contained in the positive octant
mimic the two 1-faces of ∆, and they intersect transversally.
Trivial splines
Let v1 , . . . , vn be the vertices of ∆ ⊂ Rd .
Write vi = (vi1 , . . . , vid ) ∈ Rd .
Set
Hj := v1j Y1 + . . . + vnj Yn .

Then Hj is equal to the jth coordinate function on ∆ ⊂ Rd .

So
R[H1 , . . . , Hd ] ⊂ C r (∆)
is the subring of trivial splines.
Example 1: d = 1
Now
H := v1 Y1 + v2 Y2 + v3 Y3
is the trivial spline: H(x) = x for any point x ∈ |∆|,
and R[H] ⊆ C r (∆), for any r ≥ 0.
Computations show that that H and Y1r+1 generate the subring
C r (∆) of C 0 (∆).
Define
ϕ : R[y1 , y2 , y3 ] → R[Y1 , Y2 , Y3 ]/(Y1 Y3 , Y1 + Y2 + Y3 − 1)
ϕ(y1 ) = (v1 − v2 )r+1 Y1r+1 ,
ϕ(y2 ) = 1 − (H − v2 ) = 1 − (v1 − v2 )Y1 − (v3 − v2 )Y3 ,
ϕ(y3 ) = (v3 − v2 )r+1 Y3r+1 .
Then

ϕ(R[y1 , y2 , y3 ]) = C r (∆)
and
Ker ϕ = (y1 y3 , y1 + y3 − (1 − y2 )r+1 ).

Hence

C r (∆) ∼
= R[y1 , y2 , y3 ]/(y1 y3 , y1 + y3 − (1 − y2 )r+1 ),

and

Spec(C r (∆)) = Z(y1 , y3 − (1 − y2 )r+1 ) ∪ Z(y3 , y1 − (1 − y2 )r+1 ).

For r ≥ 1, both curves have the y2 -axis as tangent at their point
of intersection (the origin), and the tangent intersects each
curve with multiplicity r + 1.

Figure: Spec(C r (∆)) for r = 0, 1, 2 in Example 1.

Example 2: d = 2
Consider a two-dimensional simplicial complex ∆, with four
vertices v1 , . . . , v4 ∈ R2 (no three on a line), vi = (vi1 , vi2 ), and
{v1 , v3 } as the only non-face.

v4
v3
v1
v2
Then
C 0 (∆) = R[Y1 , Y2 , Y3 , Y4 ]/(Y1 Y3 , Y1 + Y2 + Y3 + Y4 − 1),
where the Yi are the Courant functions and
Hj := v1j Y1 + v2j Y2 + v3j Y3 + v4j Y4 , for j = 1, 2
are the trivial splines.
Observe that Y1r+1 , Y3r+1 ∈ C r (∆).
We can deduce the following linear relation
H1 − v21 H2 − v22
− =
v41 − v21 v42 − v22
v11 − v21 v12 − v22
v31 − v21 v32 − v22

− Y1 + − Y3 .
v41 − v21 v42 − v22 v41 − v21 v42 − v22
 C r (∆) = ϕr (R[y1 , y2 , y3 , y4 ]),
where
P
ϕr : R[y1 , y2 , y3 , y4 ] → R[Y1 , Y2 , Y3 , Y4 ]/(Y1 Y3 , i Yi − 1),

v11 −v21 v12 −v22 r+1 r+1

ϕr (y1 ) = v41 −v21 − v42 −v22 Y1
H1 −v21
ϕr (y2 ) = v41 −v21
v31 −v21 v32 −v22 r+1 r+1


ϕr (y3 ) = v41 −v21 − v42 −v22 Y3
H2 −v22
ϕr (y4 ) = v42 −v22
Ker ϕr = (y1 y3 , y1 + y3 − (y2 − y4 )r+1 ), so

C r (∆) ∼
= R[y1 , y2 , y3 , y4 ]/(y1 y3 , y1 + y3 − (y2 − y4 )r+1 ).

Hence

Spec(C r (∆)) =
Z(y1 , y3 − (y2 − y4 )r+1 ) ∪ Z(y3 , y1 − (y2 − y4 )r+1 ).

The intersection of these surfaces is the line Z(y1 , y3 , y2 − y4 ).

The plane Z(y1 , y3 ) is the tangent plane to both surfaces at all
points of their line of intersection. The intersection of this
tangent plane and each surface is the line, with multiplicity
r + 1.
The local spline ring conjecture
Let ∆ be a (general) d-dimensional simplicial complex
consisting of two d-simplices intersecting in a (d − 1)-simplex.
Then we can realize Spec(C r (∆)) ⊂ Rd+2 as the union of two
smooth d-dimensional varieties V1 and V2 intersecting along a
linear (d − 1)-dimensional space L, such that V1 and V2 have the
same d-dimensional linear space T as tangent space at each
point of L and such that Vi and T have order of contact r + 1 at
each point of L.
For a proof, and generalizations:
Nelly Villamizar’s talk tomorrow
MS Multivariate Splines and Algebraic Geometry
References
Thanks to my CMA, GAIA, and SAGA students!

Cathrine Wenche Margrethe Dana Skrbo

Tegnander Oxholm Naalsund


Pål Hermunn Georg
Heidi Mork Johansen Muntingh Nelly Villamizar