Introduction to Projective Geometry Solutions 5.6

Collineations of types II and III

A 17 minute read, posted on 1 Sep 2019
Last modified on 1 Sep 2019

Tags computer vision, projective geometry, problem solution

Please read this introduction first before looking through the solutions. Here’s a quick index to all the problems in this section.

1 2 3 4 5 6 7 8 9 10

1. Show that no collineations of type II can be periodic.

The matrix in standard form for collineations of type II is as follows.

k2a1200k2000k1

Raising this to the nth power we get

⎜ ⎜kn2na12kn1200kn2000kn1⎟ ⎟

As a120 for Ak2I to be of rank 2 and k20 as a collineation is nonsingular by definition, the nth power of a collineation of type II can never be a scalar multiple of I. Hence a collineation of type II can not be periodic.

2. Under what conditions if any, can a collineation of type III be of period 2? Of period 3? Of period 4? Of period n?

The matrix in standard form for collineations of type III is as follows.

k2000k2000k1

Raising this to the nth power we get

⎜ ⎜kn2000kn2000kn1⎟ ⎟

For it to be of period n, it must be a scalar multiple of I. It is no specialization to take the scalar to be 1. So, for the nth power of a collineation of type III to be of period n, k1 and k2 must be distinct pth roots of 1 such that pn and at least one of them is an nth root of 1.

3. Show that the plane perspective transformation described in Chap. 1 is a collineation of type III.

This exercise ties things nicely together.

For a transformation to be a collineation of type III it must be nonsingular and it must have one point-by-point invariant line (corresponding to the repeated characteristic root) accompanied by an invariant point not on the invariant line (corresponding to the simple root).

Both these characteristics are displayed by the transformations described in Chap. 1. In Π2, each point has an image and each image has a pre-image proving that the transformation is nonsingular. The axis of transformation is the point-by-point invariant line and the center of transformation is the other invariant point not on the axis.

In fact if you look at the equations of transformation derived in the solution to exercise 1.2.16, we can express them in matrix form in homogeneous coordinates as follows

ba00c001b

This has characteristic roots (c,b) with multiplicity (1,2) respectively. The invariant point is (a,b+c,1) and the axis is given by the line, x2=0, formed by the characteristic vectors, (1,0,0),(0,0,1), corresponding to the root b.

This matches with the x-axis being the axis of transformation and the invariant point being O after rabbatement.

4. If P is a point on the line F1F2, and if P is its image under a collineation of type II, what is (F1F2,PP).

If the triangle of reference is defined such that F1:(0,0,1) and F2:(1,0,0), any point on the line F1F2 will have 0 as the value of the second coordinate.

Let’s denote the coordinates of a general point P on F1F2 by (p1,0,p3). Its image under a collineation of type II represented by the matrix in #1 will be (k2p1,0,k1p3).

If the line F1F2 is parameterized by the points F1 and F2, the coordinates of P and its image P will be (p1,p3) and (k2p1,k1p3). Then the cross ratio of (F1F2,PP) will be

10p1p301k2p1k1p310k2p1k1p301p1p3

=p3(k2p1)k1p3(p1)

As p10 and p30 (else P would be F2 or F1)

(F1F2,PP)=k2k1

5. Find the fixed points of the transformation X=AX if:

(a) A=311623423 (b) A=433445210

(c) A=9374121249 (d) A=3221276843

This one is pretty straightforward and you can use any software to find the eigenvectors of the given matrices. If the an eigenvalue is repeated then all the points on the line formed by the corresponding eigenvectors will be invariant. For a simple root only the eigenvector corresponding to it is invariant.

6. What is the locus of a point with the property that it and its image under a collineation of type II are collinear with a given point O? What is the locus of the images of such points?

Consider a general point P:(p1,p2,p3) and its image P:(p1,p2,p3).

Using the equations of transformation of a general collineation of type II, we get

p1=k2p1+a12p2,p1=p1k2a12p2k22 p2=k2p2,p2=p2k2 p3=k1p3,p3=p3k1

As per Sec 4.2, Theorem 1 if P and its image are to be collinear with a given point O:(o1,o2,o3), we get

∣ ∣p1p2p3k2p1+a12p2k2p2k1p3o1o2o3∣ ∣=0

Expanding this we get (a12o2k2o1+k1o1)p2p3+(k2k1)o2p1p3a12o3p22=0

Similarly, expressing in terms of image coordinates, we get

∣ ∣ ∣ ∣p1k2a12p2k22p2k2p3k1p1p2p3o1o2o3∣ ∣ ∣ ∣=0

Expanding this we get (a12o2k1+o1k2k1o1k22)p2p3+(k2k1)o2k2p1p3a12o3k1p22=0

In both cases the locus is a conic.

7. What are the equations of a collineation of type II with characteristic values k1, k2, k2 if:
(a) The corresponding fixed points are F1:(0,1,0) and F2:(1,1,0), and the second fixed line is f2:x1x2=0?
(b) The corresponding fixed points are F1:(1,0,0) and F2:(0,1,1), and the second fixed line is f2:x1=0?

Note: The answers for (a) and (b) are flipped in the book.

(a) As the second invariant line goes through F2 it must be the characteristic vector associated with the repeated characteristic root k2.

As F1 and F2 are invariant points, under the transformation, (0,1,0)(0,k1,0) and (1,1,0)(k2,k2,0). Using this information in the equations of the general collineation, we get

a12=a32=0,a22=k1 a11=k2,a21=k2k1,a31=0

Hence the equations of the transformation can be written as x1=k2x1+a13x3 x2=(k2k1)x1+k1x2+a23x3 x3=a33x3

In matrix form this would be A=k20a13k2k1k1a2300a33

As k2 must be a repeated root of AkI, a33 must be k2.

As f2 is the second fixed line, any point of the form (1,1,a) must be transformed to a point of the form (1,1,b). Using this in the equations of transformation we get

k2+a13a=1 k2+a23a=1 k2a=b

Hence, a13=a23. Also, a130 as the rank of Ak2I must be 2. So the equations of transformation will be

x1=k2x1+a13x3 x2=(k2k1)x1+k1x2+a13x3 x3=k2x3

(b) As the second invariant line goes through F2 it must be the characteristic vector associated with the repeated characteristic root k2.

As F1 and F2 are invariant points, under the transformation, (1,0,0)(k1,0,0) and (0,1,1)(0,k2,k2). Using this information in the equations of the general collineation, we get

a21=a31=0,a11=k1 a12=a13 a23=a22k2 a32=a33k2

As f2 is the second invariant line, a point of the form (0,a,b) must have an image of the form (0,c,d). This implies that a12=0.

Putting this all together, we get the matrix of transformation as follows.

A=k1000a22a22k20a33k2a33

The characteristic equation of this matrix will be

(k1k)((a22k)(a33k)(a22k2)(a33k2))=0

As this is a collineation of type II, the two quadratic roots of the second term must be equal to the repeated root of the collineation, k2. This means

a33+a22k2=k2 a33=2k2a22

Hence the equations of transformation will be x1=k1x1 x2=a22x2+(a22k2)x3 x3=(k2a22)x2+(2k2a22)x3 7 As the rank of Ak2I must be 2, a22k2.

8. (a) Show that any collineation of type I can be obtained as the composition of two collineations of type III.
(b) Can every collineation of type III be obtained as the composition of two collineations of type II?

(a) Consider two collineations A and B of type III with characteristic roots a1, a2, a2 and b1, b2, b2, centers FA, FB and axes fa and fb respectively such that FB lies on fa and FA lies on fb. Let the intersection of fa and fb be FAB. So we get

ABFA=A(b2FA)=b2AFA=b2a1FA ABFB=A(b1FB)=b1AFB=b1a2FB ABFAB=A(b2FAB)=b2AFAB=b2a2FAB

Hence the collineation AB is a collineation of type I with characteristic roots a1b2, a2b1 and a2b2 and corresponding fixed points FA, FB and FAB respectively.

Now if we represent the three roots as k1, k2 and k3, we will get

k1=a1b2 k2=a2b1 k3=a2b2

For any given k1, k2 and k3 the values a1=k1k3, a2=1, b1=k2 and b2=k3 satisfy these constraints.

Hence any collineation of type I with characteristic roots k1, k2 and k3 and corresponding fixed points F1, F2 and F3 can be represented as the composition of the following two collineations of type III.

  1. Center F1 and axis F2F3 with characteristic roots k1k3, 1 and 1.

  2. Center F2 and axis F1F3 with characteristic roots k2, k3, and k3.

To verify this, let’s multiply the matrices of the collineations of type III to see if we get back the collineation of type I.

Assuming the coordinate system to be specialized so that F1, F2 and F3 are the vertices of the triangle of reference, the matrices of transformation for the collineations of type III will be

A=⎜ ⎜k1k300010001⎟ ⎟

B=k3000k2000k3

And their composition AB will be

⎜ ⎜k1k300010001⎟ ⎟k3000k2000k3 =k1000k2000k3

which is the collineation of type I that we started out to decompose.

(b) I’ll be honest, this one threw me for a loop. I had to think about it for a couple of days and my brain calling me a dumb-dumb for not being able to solve a teeny-weeny part (b) of an even numbered problem didn’t make things any easier.

It all boils down to whether the collineations of type II can compose to preserve the invariant properties of a collineation of type III while fulfilling their own constraints.

We know that a collineation of type III has a point of fixed lines, the center, and a line of fixed points, the axis. So the composition of the collineations of type II must preserve these two properties.

A collineation of type II alone can’t preserve a point of fixed lines, by definition. So one of the two collineations of type II must be transforming each fixed line through the center to another line, and the other must be reversing this transformation to get back the original line. This means the point of intersection of the original invariant line and its image must be invariant under both the collineations of type II, and this must be true for each fixed line through the center. However, collineations of type II only have 2 fixed points. Thus no collineation of type III can be expressed as a composition of two collineations of type II.

9. If T1 is the harmonic homology whose center is the point C1:(1,1,0) and whose axis is the line l:x1x2=0, and if T2 is the quadratic inversion defined by the point C2:(0,0,1) and the conic Γ:x23x1x2=0, find the equations of T1 and T2 and show that T1 and T2 commute. Show also that T=T1T2 is of period 2.

Let P:(p1,p2,p3) be a general point.

The equation of PC1 is (p1,p2,p3)×(1,1,0)=[p3,p3,p1p2]

The intersection of PC1 with the axis is C1:[p3,p3,p1p2]×[1,1,0]=(p1p2,p1p2,2p3)

If the line PC1 is parameterized by C1 and C1, the parameters of P will be (p1p2,1). By Lemma 2 Section 4.7, the parameters of the harmonic conjugates of P with respect to C1 and C1 will be (p1p2,1). Hence P, the image of P, under this harmonic homology is (p2,p1,p3) and the equations of transformation for T1 will be

p1=p2 p2=p1 p3=p3

As per Sec 3.7, the intersections S and T of PC2 with Γ will be the solution to the quadratic equation

λ2PTAP+2λμPTAC2+μ2CT2AC2=0

where A=⎜ ⎜01201200001⎟ ⎟.

Evaluating, we have PTAP=p23p1p2, PTAC2=p3 and CT2AC2=1. So the equation above becomes λ2(p23p1p2)+2λμp3+μ2=0 The parameters of intersection are thus (1,p3+p1p2), (1,p3p1p2) and the corresponding coordinates are S:(p1,p2,p1p2) and T:(p1,p2,p1p2).

Taking S and T to be the parameters of the line PC2, the parameters of P will be (p3p1p2,p3+p1p2). According to Lemma 2 Section 4.7, the parameters of P the harmonic conjugate of P with respect to S and T should be (p3+p1p2,p3+p1p2). Hence the coordinates of P under this quadratic inversion will be (p1p3,p2p3,p1p2) and the equations of transformation for T2 will be

p1=p1p3 p2=p2p3 p3=p1p2

Clearly, the equations of transformation of T1T2 and T2T1 will be

p1=p2p3 p2=p1p3 p3=p1p2

Hence T1T2=T2T1.

Applying this transformation twice, the coordinates of the image will be (p21p2p3,p22p1p3,p23p1p2) which is (p1,p2,p3). Hence T1T2 has a period of 2.

10. In Exercise 9, what is the image of each side of the triangle of reference under the transformation T=T1T2? What can be said about the curve into which T transforms a conic which is tangent to each side of the triangle of reference? What can be said of the curve into which T transforms a conic which intersects each side of the triangle of reference in two real points? What are the images under T of the vertices of the triangle of reference? What is the image under T of a conic which passes through the vertices of the triangle of reference?

This problem exposed me to the concepts of nonlinear transformations and algebraic curves1. I always wondered what name to give curves that have two parameters x and y, now I know.

To find the images of the sides we first have to find the equations of the inverse of T1T2. These are p21=p2p3p1 p22=p1p3p2 p23=p2p1p3

The equations for the sides of the reference triangle are p1=0 p2=0 p3=0

Squaring these and using the equations of the inverse of T we get p2p3=0 p1p3=0 p2p1=0

So the equations of the sides of the reference triangle become singular conics x2x3=0, x1x3=0, x2x1=0 under the nonlinear transformation T1T2. Basically, a line gets transformed to two lines!

We know from Theorem 13, Section 4.9 that if a line Λ is tangent to a conic with the matrix of quadratic form A then ΛTA1Λ=0.

Applying this to the general matrix A of a conic that is tangent to each side of the triangle of reference and representing its inverse by B, we get b11=b22=b33=0. So B looks like B=0b12b13b120b23b13b230 Inverting this, we get B1=A=⎜ ⎜b223b13b23b12b23b12b23b213b12b13b12b23b12b13b212⎟ ⎟ So the equation of the conic will be $$b^2_{23}p^2_1 + b^2_{13}p^2_2 + b^2_{12}p^2_3 - 2b_{13}b_{23}p_1p_2

  • 2b_{12}b_{23}p_1p_3 - 2b_{12}b_{13}p_2p_3 = 0$$

Using the equations of the inverse transform, we get $$b^2_{23}\frac{p'_2p'_3}{p'_1} + b^2_{13}\frac{p'_1p'_3}{p'_2} + b^2_{12}\frac{p'_2p'_1}{p'_3} - 2b_{13}b_{23}p'_3

  • 2b_{12}b_{23}p'_2 - 2b_{12}b_{13}p'_1 = 0Multiplyingbothsidesby$p1p2p3$,wegetb^2_{23}(p'_2p'_3)^2 + b^2_{13}(p'_1p'_3)^2 + b^2_{12}(p'_2p'_1)^2 \\ - 2b_{13}b_{23}p'_1p'_2p'^2_3 - 2b_{12}b_{23}p'_1p'^2_2p'_3 - 2b_{12}b_{13}p'^2_1p'_2p'_3 = 0$$

This is a quartic curve2. As you can see from the image below, it kind of looks like a cruciform curve3 but it’s not. Maybe someone who’s better at algebraic geometry than me can tell me if this type of curve has a special name.

If a conic intersects each side of the reference triangle in 2 points then each of these quadratic equations must have 2 solutions. a11λ2+2λμa12+a22μ2=0 a22λ2+2λμa23+a33μ2=0 a33λ2+2λμa13+a11μ2=0 So, a212a22a11 a223a22a33 a213a33a11 Well that’s not much to go by so we can just represent such a conic with the general quadratic equation a11p21+$a22p22+a33p23+a12p1p2+a13p1p3+a23p2p3=0 with its image under the given transformation being a11p2p3p1+a22p1p3p2+a33p1p2p3+a12p3+a13p2+a23p1=0 Multiplying by p1p2p3 we get a11(p2p3)2+a22(p1p3)2+a33(p1p2)2+a12p1p2p23+a13p1p22p3+a23p21p2p3=0 which is the equation of a general quartic curve.

The vertices of the triangle of reference do not have images under this transformation.

The matrix of transformation of a conic that is contains the vertices of the triangle of reference will be B as given above and its equation will be b12p1p2+b13p1p3+b23p2p3=0

Using the equations of transformation, the image of this conic will be b12p3+b13p2+b23p1=0 which is the equation of a line. So under this transformation, a conic can get transformed into a line!

References


  1. Wikipedia. Algebraic curves. https://en.wikipedia.org/wiki/Algebraic_curve↩︎

  2. Wikipedia. Quartic plane curve. https://en.wikipedia.org/wiki/Quartic_plane_curve↩︎

  3. Wolfram Mathworld. Cruciform curve. http://mathworld.wolfram.com/Cruciform.html↩︎