Differential Geometry, John W.barrett

DIFFERENTIAL GEOMETRY
John W. Barrett
I.
II.
III. IV.
0. Preface Special geometries 1. Vector space geometry 2. A ne geometry 3. Euclidean geometry 4. The projective line 5. Projective geometry Smoothness 6. Smooth functions 7. The derivative 8. Open sets 9. Inverse function theorem 10. Catastrophe theory Manifolds 11. Manifolds 12. Discrete quotients 13. The tangent space Groups 14. Lie groups 15. Flows 16. One-parameter subgroups 17. Subgroups of GL(n) 18. The commutator 19. Computer vision
Copyright c John W. Barrett 1996 1998 1999. This book has been downloaded from http://www.maths.nott.ac.uk/personal/jwb/. It may be copied under licence under the conditions currently displayed in a copyright notice at that website. If you Typeset by AMS-TEX
Preface
This book is an introduction to aspects of di erential geometry in an elementary and novel manner. The main idea is to introduce the concept of a manifold and apply it to the study of Lie groups. The book was developed as a set of lecture notes for a third-year undergraduate lecture course at the University of Nottingham. Riemannian geometry is not treated.1 The emphasis is on de nitions and examples, and the abstract formalism has been reduced to a minimum. Most textbooks take the line that a di erentiable manifold is a topological manifold with a di erential structure. To eliminate bizarre examples, the topological space is required to be Hausdor and have a countable base of open subsets. All this is extremely technical, and a theorem of Whitney asserts that any manifold of this type is isomorphic to a submanifold of a Euclidean space, R n 1]. Moreover, most examples of manifolds occur naturally in this way. The de nition of a manifold used here is a locally Euclidean subset of R n . This clear and concise de nition was given by John Milnor in a book which was the inspiration for the present work 2]. The sections on the special geometries, and the style generally was in uenced by Elmer Rees' book on geometry 3]. Aside from the main development of the book, there are brief excursions into singularity and catastrophe theory, and into the subject of computer vision, based on a specialist book on the subject 4]. There are many exercises distributed in the text. The simpler examples are designed to be done immediately in the lecture room. Students nd this gives an opportunity to review and digest what has just been said. Also it gives the lecturer a valuable opportunity for feedback. The longer exercises are there to change the learning from passive to active mode.
References 1. G. de Rham, Di erentiable Manifolds, Grundlehren der mathematischen Wissenschaften 266, Springer, 1984. 2. J.W. Milnor, Topology from the di erentiable viewpoint, University Press of Virginia, 1965. 3. E.G. Rees, Notes on Geometry, Universitext, Springer, 1983.
do not nd the copyright notice or are unwilling or unable to abide the conditions, then a licence to copy this work is not granted. 1 It overlaps too much with general relativity.
4. J. Mundy and A. Zisserman, eds, Geometric invariance in computer vision, MIT Press, 1992.
tion of vectors, and multiplication by scalars. In this book the scalars are always taken to be the real numbers, R . The rst example of interest is the Euclidean space V = R n . This has a standard basis set of vectors (1 0 0 . . . 0) (0 1 0 . . . 0) . . . (0 0 0 . . . 1): A second set of examples of vector spaces arises by taking V to be a linear subspace of R k . In this example, V does not have a standard, or uniquely speci ed, basis. Examples of this type turn out to be of major importance in di erential geometry. Therefore it is important to understand the precise nature of the di erence between the vector space R n , and these more general examples. In general, a vector space possesses many di erent basis sets of vectors. Every basis contains the same number of elements, called the dimension of V . Linear mappings of vector spaces : V ! W are those that preserve the structure, i.e., the addition of vectors and multiplication by scalars. An isomorphism is a linear mapping which has an inverse. Suppose V is a vector space of dimension n. Choosing an ordered basis set of vectors e1 e2 . . . en for V is the same thing as specifying a linear isomorphism R n ! V . Given the basis, the linear isomorphism is de ned to be the linear map speci ed by mapping (1 0 0 . . . 0) 7! e1 , etc., . . . , (0 0 0 . . . 1) 7! en . This isomorphism R n ! V can also be written n X (a1 a2 . . . an ) 7! ai ei : This is a group, called the general linear group GL(n). Let us check that GL(n) is a group. (1) The composition of two isomophisms 1 and 2 is the map v 7! 2 ( 1 (v)). This linear map is called 2 1 and is also an isomorphism. (2) The composition of maps is associative (3) The identity mapping e: R n ! R n is the identity element of the group. (4) The inverse of an element 2 GL(n) is the inverse mapping, i.e., ;1 = ;1 = e:
Vector spaces. A vector space is a set V , with two operations: addi-
1. Vector space geometry
The general linear group. Consider the set of all isomorphisms

Rn
i=1
! Rn.
A linear map R n ! R n has a matrix, and conversely, an n n matrix determines a linear map. The linear map determined by the matrix of numbers 0 1 11 12 . . . @ 21 22 . . . A ... takes the point x = (x1 x2 . . . xn ) 2 R n to (x) = (y1 y2 . . . yn ), where
yi =
n X j =1
ij xj
1 i n:
So using this identi cation of linear maps on R n with matrices, GL(n) can be thought of as the group of invertible n n matrices. Now suppose V is a general nite-dimensional vector space. Then there is an isomorphism V ! R n , but this is not unique. Suppose 1 , 2 are any two isomorphisms. If v 2 V , how are x = 1 (v ) and ;1 (x), i.e. they are related by y = 2 (v) related? Obviously y = 2 1 a mapping in GL(n). Any element of GL(n) could arise in this way. Thus elements of the group GL(n) relate the coordinate representations x = 1 (v) and y = 2 (v) of V . Notation. The general convention will be used that if a 2 R n , then the letters a1 . . . an will be used for the coordinates of a, i.e., a = (a1 a2 . . . an ): Vectors (elements of R n ) are generally written as horizontal row vectors. Sometimes they are written vertically, as column vectors. There is no di erence in meaning intended it is conventional to do this when multiplying a matrix with a vector. Exercise 1. De ne the map y = (x) by b x1 (y1 y2) = a ad ; bc 6= 0: c d x2 Which equation in y1 and y2 de nes the image of the points satisfying 2 x1 + x2 2 = 1 under the map ? Geometry. As a general principle: \The objects of a geometry are those which retain their form under the transformations being considered." Some examples: (1) Circles are not objects of vector space geometry. As we saw in exercise 1, the image of a circle need not be a circle under a linear isomorphism.
(2) A linear subspace S R n has the de ning property s1 s2 2 S ) s1 + s2 2 S This property is preserved under linear mappings , so the set (S ) = f (x) j x 2 S g is a linear subspace if S is. Therefore linear subspaces are objects of vector space geometry. Solution to exercise 1. The values of x1 and x2 for a given point (y1 y2) can be calculated using the inverse matrix: 1 d ;b y1 : (x1 x2 ) = ad ; y2 bc ;c a 2 Substituting these values into the equation x2 1 + x2 = 1 gives ;(dy ; by )2 + (;cy + ay )2 = 1: 1 1 2 1 2 (ad ; bc)2 This is a more complicated equation than that for (x1 x2 ). In general it will not be a circle but will give an ellipse.
2. Affine Geometry
The a ne geometry is the second of the `special geometries' which we are going to study. As in vector space geometry, there is a group of transformations which characterises a ne geometry. A ne geometry is not too di erent to vector space geometry. In vector space geometry, the origin plays a distinguished role. However in many applications there are objects in Euclidean space R n (e.g. ordinary `space' of physics) but it does not really matter where the origin is. An a ne subset A R n is a subset of R n with the property that if a b 2 A, then a + b 2 A for all 2 R such that + = 1. In other words, the straight line through any two points of A is also in A. Some examples: (1) A linear subspace L 2 R n . ( a + b 2 L without any condition on ). (2) The set x + L = fx + l j l 2 Lg for a linear subspace L. This follows from the calculation (x + l1 ) + (x + l2) = x + l1 + l2 2 x + L:
A ne subsets of R n .
The second example is more general than the rst because x + L need not contain the origin. It is in fact the most general type of example. All a ne subsets are of the form of example 2. To show that this is true, x an element a 2 A and de ne TAa = A ; a = fx ; a j x 2 Ag: This is called the tangent space to A at a. What we have to show is that this tangent space is a linear subspace of R n . Proof. Suppose 2 R and x ; a 2 TAa . Then the scalar multiple (x ; a) 2 TAa because (x ; a) + a = x + (1 ; )a 2 A: Now suppose x ; a, y ; a 2 TAa. Then the sum (x ; a) + (y ; a) is in TAa because y ; a 2 A: x;a+y;a+a=x+y;a=2 x+ 2 This shows that a ne subsets are just `linear subspaces with the origin shifted'. This readily suggests that if a b 2 A, then TAa = TAb. This is true because b ; a 2 TAa, and so if x ; a 2 TAa, so is (x ; a) ; (b ; a) = x ; b: Exercise 1. Show that for each k 1,
k X i=1 i ai
2A
if ai 2 A and
k X
1
i = 1:
A ne maps. An a ne map R n ! R m is de ned to be a map satisf ( a + b) = f (a) + f (b) + = 1: The a ne maps preserve a ne subsets, and in particular they map straight lines to straight lines. If an a ne map is invertible, it is called an a ne isomorphism. The set of all a ne isomophisms R n to itself is a group, called the a ne group A(n). This group characterises a ne geometry. It is possible to de ne a notion of an à ne space' by giving axioms, in a similar way to the de nition of a vector space, such that the a ne subsets of R n are examples. We do not need to go into this. There are two important examples of a ne isomorphisms of R n (1) translations x 7! x + c, for a constant c 2 R n (2) linear isomorphisms in GL(n)
fying
More examples can be made by combining these two. In fact, every a ne map R n ! R m is a linear map R n ! R m composed with a translation of R m . The proof is an exercise: Exercise 2. Prove that if f is an a ne map, then the map df de ned by x 7! f (x) ; f (0) is linear. Example. The most general a ne map f : R ! R is given by the formula f (x) = ax + b. Then df (x) = ax. The coe cient a in this f linear map is d dx , for any x.
c a b d
2 x2 1 + x2 = constant. The line intersects the region between the circles in two segments it is easy to see that the lengths of the two segments are equal. This is because the diagram is symmetrical about an axis, vertical in the diagram. Newton wanted to know if the same property is true for the concentric ellipses shown on the right. These are given by (x1 =e)2 + x2 2 = constant for some e 2 R . He proved that they are by applying a linear transformation L (x1 x2 ) 7! (x1 =e x2) to the right hand diagram. Since this transformation is a ne, the line is mapped to a line. From the left-hand diagram
Example. In the gure on the left, there are two concentric circles,
Since L is linear, a ; b = c ; d, and so the lengths of the two segments on the right are equal. Newton used this to show that there is no gravitational force inside an ellipsoidal shell of matter.
L(a) ; L(b) = L(c) ; L(d):
Solution to exercise 1. There are two strategies for this

(1) Induction. Assume it is true for k ; 1. Then
k X i=1 i ai = (1
k)
k ;1 X i=1
i ai + a 1 k k k
2 A:
The special case k = 1 can be treated by picking another term to leave out of the sum, since not all i can equal 1. The induction starts with k = 2 being the de nition. (2) Tangent space.
i ai ; b =
i (ai
; b) 2 TAb
Solution to exercise 2.
since TAb is a linear subspace. Hence
P a 2 A. i i
df ( x + y) = f ( x + y + (1 ; ; )0) ; f (0) = f (x) + f (y) + (1 ; ; )f (0) ; f (0) = df (x) + df (y)
Exercise 3. Let a b c 2 R 2 be three points which are not collinear

(do not lie on a line). Let A B C be any three points in R n . Work out how to de ne an a ne map R 2 ! R n such that
a 7! A
b 7! B
c 7! C:
What can you say about the image of your map? Exercise 4. Colours can be obtained by mixing red, green and blue in any desired proportions, which are given by numbers adding up to 1. Draw a diagram to represent the possible colours, and plot a point to represent 1=2 blue, 1=4 red, 1=4 green. Explain how the point is plotted.
10
Euclidean geometry is the study of gures in R n using the concept 2 2 1=2 , of distance. The length of a vector is jxj = (x2 1 + x2 + . . . xn ) and the distance between two points x and y is de ned to be d(x y) = jx ; yj. The latter concept does not require the origin as a distinguished point. To be more precise, the concept of distance is invariant under translations, d(x + a y + a) = d(x y). The space R n with the distance function d is called n-dimensional Euclidean space. The 3-dimensional Euclidean space is very familiar as it is a mathematical model of space for the positions of objects in physics. Euclidean space is a metric space. Mappings f : R n ! R m which preserve the distance function,
3. Euclidean Geometry
d(f (x) f (y)) = d(x y)

are called isometries. The motions of rigid bodies in physics are examples of isometries. An isometry is injective, since f (x) = f (y) implies
d(x y) = d(f (x) f (y)) = 0

which implies that x = y. Moreover, Theorem. An isometry R n ! R m is an a ne map Proof. Let p be an a ne linear combination of x y 2 R n , i.e., p = x + y, with + = 1. Then p is the unique point such that
d(x p) = j jd(x y) and d(y p) = j jd(x y):
If f is an isometry, then it follows that
d(f (x) f (p)) = j jd(f (x) f (y)) and d(f (y) f (p)) = j jd(f (x) f (y))
which shows that f (p) = f (x) + f (y), i.e., f is a ne. From this theorem, one can easily deduce that any isometry f : R n ! R n must be invertible. Recall that an a ne map is a composition of the linear map df with a translation. Since translations are always invertible isometries, it hinges on whether df is invertible. However df is linear and injective, so it must be invertible.
11
Euclidean space (to itself) is called the Euclidean group, E(n). It is a subgroup of A(n). There are a ne isomorphisms which are not isometries. For example, in 1 dimension, f (x) = ax + b is not an isometry unless a = 1. The linear transformation L in Newton's example above, or the transformation in exercise 1 of section 1 are examples in more than one dimension. Clearly, a Euclidean transformation is a composition of a linear isometry followed by a translation. Hence it is important to characterise the linear isometries. Linear isometries preserve the square of the distance of a point from the origin. This quadratic form can be written in various ways:
The Euclidean group. The set of all isometries of n-dimensional
d(x 0) = jxj =
2 2
n X i=1
x2 i =x x
Orthogonal group. An orthogonal transformation of R n is a linear
jL(x)j = jxj for all x: The set of all linear maps L: R n ! R n which preserve the distance to
map L: R n ! R n which preserves distances to the origin, i.e.
the origin is a group called the orthogonal group, O(n). For linear maps, preserving the distances to the origin actually implies that all distances are preserved, since jL(x) ; L(y)j = jL(x ; y)j = jx ; yj. Thus orthogonal transformations are isometries (and hence invertible). Polarization. Orthogonal transformations actually preserve all dot products. This follows from the `polarization identity' (x ; y) (x ; y) = x2 + y2 ; 2x y
This shows that L(x) L(y) = x y. Summary. We have seen in three cases, vector space geometry, a ne geometry and Euclidean geometry, that the geometry is characterised by a group of transformations which preserves the structure of interest (such as linearity or distances). \The properties of objects in a geometry are those which do not change under the transformations of the geometry."
12
2
Exercise 1. Let x 6= y
both equations if + = 1. Be sure to include the cases where or is negative. Exercise 2. Let L: R n ! R n be an orthogonal map. Write the relation L(x) L(y) = x y in terms of the vector components xi and yi and the matrix Lij Show that this orthogonal matrix satis es
n X i=1
2 R . Draw a diagram to show the set of points p satisfying d(x p) = j jd(x y) and the set of points p satisfying d(y p) = j jd(x y). Explain why there is only one point p satisfying
Lij Lik = jk
where ik is the identity matrix. Conversely, check that any such matrix gives an orthogonal transformation. Take the determinant of each side of this equation to nd the possible values for det L, and give an example of a 3 3 orthogonal matrix exhibiting each of these values. Exercise 3. Show that all elements of O(2) are either cos ; sin sin cos
for some value of or have another form, which you should nd. Give a geometrical interpretation of both forms. Find an element of GL(2) with determinant 1 which is not in O(2).
4. The projective line
The last of the four special geometries studied here is projective geometry. The space for this projective geometry is projective space, which is distinct from Euclidean space, and there is one such space in each dimension. The one-dimensional space is called the projective line. Motivation. The projective line can be introduced by considering projections, such as arise in perspective drawing. Consider two (straight) lines m and l in R 2 , which meet, and a point p which lies on neither line. The projection from line l to line m through the point p takes a 7! b, as shown in the diagram.
m a p b c l
13
This does not quite de ne a map l ! m because the point c, lying on a parallel to m through p, does not project onto m. However consider the map l f1g ! m f1g which takes a 7! b as shown, for a 6= c 1 c 7! 1 1 7! d: This map is a bijection, the inverse being the projection from m to l. In nity is to be regarded as an extra abstract point, which is added to each line in order to make the projection well-de ned. Example. A concrete example is given by considering p to be the ; origin, l to be the line 7! x0 (1 ; )y0 , with x0 , y0 not both 0, and m to be the line 0 7! ( 0 1), for 0 2 R . The projection from l to m is given by ; x (1 ; )y 7! x0 1 0 0 (1 ; ) y0 with also (x0 0) 7! 1 and 1 7! (;x0 =y0 1). The parameter for each line gives an identi cation of each line with R . The projection takes parameter on line l to parameter x0 0= (1 ; ) y0 on line m.
14
equivalence classes of the plane with the origin removed, R 2 n f0g, under the relation v v, for 6= 0 2 R . If a point p 2 P 1 is the equivalence class of (x y), this is written p = x : y], and is called the ratio of x and y. The pair of numbers (x y) are also called homogeneous coordinates for the points p 2 P 1. It is worth noting that these are not coordinates in the strict sense to be used later on, because (x y) is not uniquely determined by a given point p 2 P 1. However the term `homogeneous coordinates' is standard in this situation, and so we shall use it. Each point of the projective line can be identi ed with a unique line through the origin in R 2 , and this is a useful way of thinking of the projective line. The projective line P 1 can be viewed as R f1g in many di erent ways. Take any two linearly independent vectors u and v. Then the line l: R ! R 2 7! u + (1 ; )v is not through the origin. This determines a map R taking equivalence classes.
l
The de nition. The projective line P 1 is de ned to be the set of
!R !P
2
by
There is one point of P 1 not in the image of this map, namely the equivalence class of vectors parallel to l. This point is identi ed with 1. Projective transformations. Elements of GL(2) preserve the equivalence relation v v in the de nition of P 1, i.e. if A 2 GL(2), then Av Av is true whenever v v. Therefore A determines a trans-
15
b , then the transformation is formation of P 1. If A = a c d
Inhomogeneous coordinates. As noted above, the homogeneous co-
x : y] 7! ax + by : cx + dy]
ordinates (x y) of a point x : y] in P 1 are not uniquely determined. However the inhomogenous coordinate z = x=y is uniquely determined by x : y]. This is a real number if y 6= 0. If y = 0, it can be taken to be 1. Using the inhomogeneous coordinate to identify P 1 with R f1g is the same as considering the identi cation determined by the line m given by 7! ( 1) in the example above. This is because
x : y] = x=y : 1]:
Using the inhomogeneous coordinate, the formula for a projective transformation becomes (ax + by) = (a + b) : =x ! 7 y (cx + dy) (c + d) Taking the special case a = x0 =y0, b = 0, c = ;1, d = 1 gives the formula given in the example. Question. Why should the formula in the example be a projective transformation? This point has not been made clear so far, but is worth thinking about. Exercise 1. Which matrices A 2 GL(2) give projective transformations for which 1 : 0] 7! 1 : 0]? Give the formula for the transformation using the inhomogeneous coordinate. What type of transformation of R is this? Determine all the projective transformations for which both 0 : 1] 7! 0 : 1] and 1 : 0] 7! 1 : 0].
5. Projective Geometry
The considerations for the projective line extend to the case of more than one dimension in a straightforward way. The n-dimensional projective space P n is de ned to be the set of equivalence classes in R n+1 n f0g under the relation v v, for 6= 0 2 R . The general linear transformations A 2 GL(n + 1) give n +1 mappings of R which respect the equivalence relation, and so determine bijections of P n . These are called projective transformations, and
16
the set of all projective transformations is called the projective group, PGL(n + 1). More than one element of the group GL(n +1) gives rise to the same projective transformation. This means that PGL(n + 1) is not equal to GL(n + 1), but is a quotient group. Matrices that give the same projective transformation as A 2 GL(n + 1) are the scalar multiples, A, for 6= 0. A point x1 : x2 : . . . : xn+1 ] 2 P n has homogeneous coordinates (x1 x2 . . . xn+1 ), and inhomogeneous coordinates (x1 =xn+1 x2 =xn+1 . . . xn =xn+1 ) which are valid if xn+1 = 6 0. The points x1 : x2 : . . . : xn : 0] are called `points at in nity', and there are, for n > 1, more than one of them. Example. In the projective plane P 2, the points 1 : 0 : 0] and 1 : 1 : 0] are both points at in nity, but (1 0 0) 6= (1 1 0) for any , so 1 : 0 : 0] 6= 1 : 1 : 0]. Historically, one of the reasons for the development of projective geometry was its use in simplifying equations. Consider for example the inhomogeneous equation
Equations in projective geometry.
au2 + buv + cv2 + du + ev + f = 0

whose solutions are points (u v) 2 R 2 . This equation can be rewritten in terms of 3 variables by the substitution u = x=z, v = y=z as
Q(x y z) = ax2 + bxy + cy2 + dxz + eyz + fz2 = 0:

This equation is homogeneous, meaning that all the terms are of the same degree. This implies that if (x y z) is a solution, then so is ( x y z). Therefore (x y z) can be regarded as the homogeneous coordinates in projective space P 2, and the solutions de ne a subset of P 2. The importance of writing the equation in this form is that projective transformations can be applied to homogenous equations, which allows the coe cients (a b c d e f in this example) to be simpli ed.
R3
17
The function Q in the homogeneous equation is a quadratic form on . A projective transformation of P 2 is given by a linear transformation L of R 3 . If the image of (x y z) is (x0 y0 z0) = L(x y z) the equation for x y z can be reexpressed as an equation amongst the new variables x0 y0 z0. This new equation is also homogeneous of degree two, and is given by a new quadratic form Q0 on R 3 obtained from Q by a similarity transformation: Q0(x0 y0 z0) = Q(L;1 (x0 y0 z0)) = 0. Quadratic forms are similar to a nite number of canonical forms. Example. If the quadratic form Q is positive de nite, then it is similar to x2 + y2 + z2 = 0 and the original equation in inhomogeous coordinates becomes u2 + v 2 + 1 = 0 after the projective transformation. In this case there are no solutions for x : y : z] 2 P 2. Exercise 1. The points of P n = f x1 : x2 : . . . : xn+1 ]g can be split into two disjoint subsets, as xn+1 6= 0 or xn+1 = 0. Give a bijection of the rst subset with R n . The second subset is called the `points at in nity'. Find a bijection of the points of in nity with P n;1. Give a decomposition P n = R n R n;1 R n;2 . . . R 1 R 0 :
Exercise 2. Explain how the equation for a line in R 2

ax1 + bx2 + c = 0 for constants a b c 2 R can be written as an equation in P 2. How many points in P 2 solve the corresponding equation which do not correspond to solutions in R 2 ? Give the equations of two parallel lines in R 2 . Where do these lines meet in P 2? Exercise 3. Show that the equation of the line joining two distinct points a1 : b1 : c1], a2 : b2 : c2 ] in P 2 is 0x y z1 det @ a1 b1 c1 A = 0: a2 b2 c2
18
6. Smooth functions
The functions we are mainly concerned with are those that can be di erentiated. One could work with the set of all di erentiable functions, but this is inconvenient, because the derivative of a di erentiable function need not be itself di erentiable. Therefore we shall work with functions which can be di erentiated any number of times. These are called smooth functions. exists for every x 2 R and is a continuous function of x. Often you can show a function is smooth by giving the formula for all the derivatives. Examples. Most familiar functions from calculus are smooth: n f (1) Polynomials are smooth, as d dxn = 0 for n greater than the order of the polynomial. (2) The standard functions sin(x), cos(x), exp(x) etc., are smooth. (3) More generally, an analytic function de ned on C de nes two smooth functions, its real and imaginary parts, when its domain of de nition is restricted to the real axis. Exercise 1. Are these functions smooth? p (1) jxj x 2 R (2) cos;1 x ;1< x<1 (3) ;1=x2 x > 0 e x 7! 0 x 0
dxn
Functions of one variable. De nition. A function f : R ! R is smooth if f and each derivative dn f
The second example illustrates a use of the inverse function theorem for one variable dy = 1= dx dx dy this will subsequently be generalised to R n . It is also worth noting that it makes sense to ask about the smoothness of a function de ned only on a part of R , namely the open interval ;1 < x < 1. The properties of di erentiable functions carry over in a straightforward way to the case of smooth functions. Given smooth functions f , g, then the following are smooth: (1) Linear combinations x 7! f (x) + g(x)
19
(2) Product fg: x 7! f (x)g(x) (3) Composite f g: x 7! f (g(x)) The second of these is proved by using the Leibnitz rule, which gives a formula for the derivative: d(fg) = f dg + g df : dx dx dx This can be used to give a proof that fg is smooth if f and g are. Iterating Leibnitz rule gives
n dn (fg) = X n dk f dn;k g : k dxk dxn;k dxn k=0
The right-hand side is the sum of a product of continuous functions and so is continuous. The third of these is similarly related to the chain rule, which gives the derivative of a function of a function. Let f be given as a function of y, g be given as a function of x. d(f g) = df (g(x)) dg dx dy dx f The notation is potentially confusing. The formula d dy (g (x)) means write f as a function of y, di erentiate it with respect to y, then substitute g(x) everywhere for y. For example, if f (y) = y3 + y and g(x) = cos(x), then df (g(x)) = g(x)2 + 1 = cos2(x) + 1: df = y 2 + 1 dy dy
fi: R n ! R , i = 1 . . . m, given by f (x) = (f1(x) f2 (x) . . . fm(x)) : De nition. A function f : R n ! R m is smooth if the components (f1 f2 . . . fm ) of f and each of their partial derivatives 2f @ @f i = 1 ...m i i ... fi @x @x @x j k = 1 ...n j j k exist and are continuous functions of x.
Functions of several variables.

If f : R n
! R m is a function, then it has m component functions
20
Exercise 2. Work out the partial derivatives for the components of a

linear map f : R n ! R m given by
fi : (x1 . . . xn ) 7!
n X j =1
Lij xj :
Show that a linear map is smooth. Smooth functions can be combined in a similar way to the onevariable case, by linear combinations f + g of two functions f , g from R n to R m , by the product fg of two functions R n ! R , and by composition of f : R n ! R m with g: R m ! R k to give a smooth function g f: Rn ! Rk. Exercise 3. Which of the following functions R 2 ! R are smooth? p (1) px2 + y2 (2) 1 + x2 + y2 (3) ey sin x Let c: R ! R m and h: R m ! R be smooth functions. Then the composite is ; h c: t 7! h c1 (t) c2 (t) . . . cm (t) : The chain rule for di erentiating this is
m @h ; d(h c) = X dci : c ( t ) dt dt i=1 @xi
The chain rule for several variables. First, consider a special case.
This formula is a sum of a product of continuous functions and is therefore a continuous function. By repeated application of this formula, and the Leibnitz rule, one can arrive at a formula for the n-th derivative, thus showing that h c is smooth. Exercise 4. Put c(t) = (t2 t3), and suppose h is a smooth function R 2 ! R such that h(c(t)) = t. (1) Di erentiate both sides of h(c(t)) = t, using the chain rule. (2) Set t = 0. What does this tell you about h? How does this relate to the image of c?
21
The general case. If g: R n ! R m is a smooth function of x and f : R m ! R k a smooth function of y, then

m @f ; @ (f g ) = @ ( f g ) = X i g(x) @gl i i @xj @xj @xj l=1 @yl
This follows from setting c(t) = g(x1 . . . xj + t . . . . . . xn ), h = fi , and computing d ;h c dt at t = 0. The formula can be interpreted as the multiplication of a k m matrix with an m n matrix. 0 @f1 . . . @f1 1 0 @g1 . . . @g1 1 @ym @x1 @xn 1 @ @y ... ... A@ ... ... A @fk . . . @fk @gm . . . @gm @y1 @ym @x1 @xn
Exercise 5. Consider the functions
g :U ! R 2 U = f(r )jr > 0g (r ) 7! (r cos r sin ) f :R 2 ! R 2 (x y) 7! (x2 ; y2 xy)
(1) Give reasons to show that f and g are smooth. (2) Calculate
@g1 @r @g 2 @r @g1 @ @g 2 @ @f1 @x @f2 @x @f1 ! @y @f2 @y
(3) Work out a formula for f g, and its matrix of partial derivatives, without using the chain rule. Show the chain rule is satis ed by multiplying matrices. A smooth functions f with a smooth inverse f ;1 is called a di eomorphism. This satis es
Di eomorphisms.
f f ;1 = identity
and
f ;1 f = identity:
22
It can be regarded as de ning new coordinates. The gure shows the (x1 x2 )-coordinate axes of R 2 and the new axes for (u1 u2 ) = f (x1 x2 ), for some di eomorphism f . Geometric objects de ned in terms of the old coordinates can be de ned in terms of the new coordinates, and vice versa, using (x1 x2 ) = f ;1(u1 u2 ).
x2 u2
x1 u1
Exercise 6. Suppose f : R n ! R n is a di eomorphism, with inverse g. Solution to exercise 1.

(1) The function is not di erentiable at x = 0:
Apply the chain rule to f g. What can you say about the matrix of partial derivatives of f ?
f (x) ; f (0) = lim jxj = lim x!0 x x
are zero.
x 2 (2) If x = cos y, then d The indy dy ;1 verse function theorem tells us that d x = p1;x2 . This formula can be repeatedly di erentiated to give formulae for the n-th derivative, since x2 < 1. (3) All the derivatives of the function exp(;1=x2 ) converge to 0 as x ! 0, so the given function is smooth. @fi @xj = Lij , a constant. All further derivatives
1: p = ; sin y = ; 1 ; x 6= 0.
23
(1) (2) Setting t = 0 gives 0 = 1. Hence h cannot be smooth. (However 3 x .) a continuous h can be found. An example is h(x1 x2 ) = p 2 The image of c has a cusp at t = 0, a point where the curve is not smooth. The curve is called the semicubical parabola.
1
@h + 3t2 @h = 1: 2t @x @x
1 2
0.5
y 0
-0.5
-1 -1
-0.5
0 x
0.5
7. The derivative A function f : R k R l has a derivative dfx at a given point x R k if there is a linear map dfx: R k R l and a continuous function : R k R l such that
f (x + h) ; f (x) = dfx (h) + jhj (x + h)
and
(x) = 0: The formula can be solved for as long as h 6= 0, and this formula for is continuous where h 6= 0 (assuming f is). So the crucial point of the de nition is that the error converges to 0 as h ! 0. Lemma. The derivative is unique. Proof. Suppose another linear map L also satis es the de nition, with error 0. Then
; dfx (h) ; L(h) = jhj (x + h) ; 0(x + h) :
24
t dfx(y) ; L(y) = tjyj (x + ty) ; 0(x + ty) : For t 6= 0, the number t cancels on both sides, giving ; dfx(y) ; L(y) = jyj (x + ty) ; 0(x + ty) : But as and 0 are continuous, this equation is also true at t = 0. At t = 0, the right-hand side is zero. It follows that dfx = L. For a xed point p, the map R k ! R l x 7! f (p) + dfp (x ; p) is an a ne map which approximates f at p. Clearly, if f is itself a ne, then the approximation is exact, and the derivative is the linear map df de ned earlier for a ne maps, and does not depend on p. However if f is not a ne, then the linear map dfp does vary with p.
y y=f(x)
Consider h = ty for a xed vector y and t ! 0 2 R . Then this becomes
The a ne approximation to f at p Exercise 1. Let f : R n ! R m be a linear map. Use the de nition to show that dfx = f , for any x. Let h: R n ! R m be an a ne map. Show that dhx = dhy , for any x y 2 R n , and that this linear map coincides with the map named dh in the section on a ne maps. Lemma. If f : R k ! R l is a smooth function, then the derivative dfx exists for every x, and has as its matrix the partial derivatives 0 @f1 . . . @f1 1 @xk 1 @ @x ... ... A: @fl . . . @fl @x1 @xk
25
Proof. Given h 2 R k , de ne a sequence of vectors
h0 = 0 h1 = (h1 0 0 . . . 0) h2 = (h1 h2 0 . . . 0) .. . hk = (h1 h2 h3 . . . hk ) = h The superscript is used to distinguish di erent vectors hn , as distinct from the subscript, which denotes the components hn of h as usual. The sequence of vectors is constructed so that it interpolates between 0 and h, with each successive pair hn , hn;1 di ering only in one coordinate, hn . Then, using the mean value theorem for this coordinate, @fi (x + cn ) fi(x + hn ) ; fi (x + hn;1) = hn @x n where the vector cn lies on the line between hn and hn;1 .
n-th axis hn cn 0 h n-1 other axes
Summing this equation over n gives
k X @fi (x + cn ): fi (x + h) ; fi(x) = hn @x n n=1
This expression gives the i-th component of a vector obtained by a matrix of partial derivatives acting on the vector h. The partial derivatives are evaluated at points x + cn . For jhj 6= 0, de ne = ( 1 2 . . . k ) by the equation ()
k X @fi (x) + jhj : fi(x + h) ; fi(x) = hn @x i n n=1
26
Then
i=
which converges to zero as h ! 0, because the di erence of partial derivatives converges to zero, and ;1 hn=jhj 1. This shows that the vector is continuous if (x) is de ned to be zero. Equation ( ) is just the i-th component of the equation de ning the derivative, with the linear map dfx given by the matrix of partial derivatives at x. Chain rule for derivatives. In the light of this lemma, results about the partial derivatives can be transcribed into the new notation. The chain rule becomes d(f g)x = dfg(x) dgx :
f Exercise 2. Let f : R ! R . What is df1(2) in the usual d dx notation? Exercise 3. Let f (x y) = (x2 + y2 x2 ; y2 ). Calculate df(x y) using
X hn @fi n ) ; @fi (x) ( x + c jhj @xn @xn

n
partial derivatives. Work out a formula for in the de nition of the derivative of f , and verify that ! 0 as h ! 0. Exercise 4. Give a proof of the chain rule by applying the de nition of the derivative to d(f g)x , dfg(x) and dgx , and manipulating the three error terms.
8. Open sets
Quite often we need to de ne smooth mappings not on the whole of 7! cos x is a di eomorphism of the open interval 0 < x < to the open interval ;1 < y < 1. Another requirement is to discuss the behaviour of a function is a `small region around a point'. Balls. In Euclidean space R n , the distance d(x y) = jx ; yj can be used to say when points are close. If r is a number greater than zero, then the ball of radius r at x 2 R n is de ned to be Br (x) = fy 2 R n j d(x y) < rg The ball is the subset of points closer to x than the radius r. However, as we have already seen, there are many maps of interest, a ne maps for example, which do not preserve distances. Therefore it is necessary to have adopt a sophisticated approach to the idea of closeness of points.
R n , but on certain subsets. For example x
27
Open subsets of Examples.
subset which is a union of balls in R n , Br (x), for points x 2 R n , and radii r > 0. Note that the union can be over any set of balls, not restricted to nite or countable, and of course x and r vary.
Rn
. An open subset P of R n is de ned to be a
(1) A ball (2) The intersection of two balls (3) The empty set (4) The whole of R n The second example is an open subset for the following reason. If p 2 P = Br (x) \ Br (x0 ), then p is the centre of a ball of radius min(r ; d(p x) r0 ; d(p x0 )), which is contained in P .
0
So P is the union of all the balls constructed in this way for each p 2 P. A point y 2 Y is called an interior point of Y if there is an r such that Br (y) Y . The set Y is an open set if and only if every point of Y is interior. Exercise 1. Which of the following subsets of the plane R 2 are open sets? Points (x y) such that: (1) x 0 (2) x > 0 (3) x + 2y = 0 (4) x2 > y3 (5) xn + yn > 1 for every positive even integer n. If U R n is an open subset, then a function de ned on U , rather than the whole of R n , can be di erentiated in just the same way for any point x 2 U . For example, to de ne a partial derivative of f : U ! R , @f = lim f (x1 . . . xj + h . . . xn ) ; f (x) @xj h!0 h the function f needs to be evaluated at points close to x along a line parallel to the xj axis. Such points are in the set U , because it contains the whole of a ball centred on x.
28
This means that all of the previous de nitions for di erentiation also make sense for functions de ned on an open subset of R n . Example. The function f : R 2 ! R 2 de ned by (x y) 7! (x2 y) is a smooth map but is not a di eomorphism. However if U is the open subset of R 2 de ned by x > 0, then the same formula gives a di eop morphism U ! U . The inverse function is (x y) 7! ( x y). Open subsets of X R n . Now let X be a subset of R n , not necessarily an open subset. For example, X could be an a ne subset of a lower dimension, or a manifold (see below). An open subset of X is simply any set of the form X \ U , where U is an open subset of R n . Now consider X R n , Y R k and f : X ! Y a continuous mapping. If V Y , its inverse image, f ;1(V ) is the set of all points x 2 X which map into V . Continuous mappings respect open sets in the following way: Proposition. For f : X ! Y , the inverse image of an open subset V Y is an open subset of X . The situation is simpler if f has a continuous inverse mapping f ;1. Proposition. Let X R n and Y R k . If the continuous function f : X ! Y has a continuous inverse f ;1, then U X is an open subset of X if and only if f (U ) Y is an open subset of Y . is an invertible linear map for each x. Is the converse true? Suppose f is a smooth map and dfx is invertible for all x. Is it a di eomorphism? If n = 1, this is true. The inverse function theorem for one variable is f Theorem. If f : R ! R has non-zero derivative d dx for every x, then the inverse function exists and is di erentiable. It is not too hard to extend this to proving that the inverse is smooth. If n > 1, it is not true. Consider the mapping
9. Inverse function theorem Suppose f : R n R n is a di eomorphism. Then the derivative dfx
g(x y) = (ex cos y ex sin y):

This has derivative
x cos y ;ex sin y e dg(x y) = ex sin y ex cos y
29
which is invertible. However, g(x y) = g(x y + 2 ). This shows the inverse can not exist on the whole of R 2 . However, the idea of the de nition of the derivative is that a smooth function f is approximated near a point x by an a ne map. If x is a regular point, this a ne map is invertible, so we expect the original function f to be invertible when its domain and range are restricted to some su ciently small regions around x and f (x). Inverse Function Theorem. If U R n and V R n are open subsets, f : U ! V is smooth, x 2 U and dfx is invertible, then there are open subsets U 0 U , V 0 V such that x 2 U 0 and f restricted to U 0 is a di eomorphism to V 0 . Exercise 1. Are the following functions R 2 ! R 2 di eomorphisms? If not, is there an open set containing the origin on which the function is a di eomophism to its image? Give the reasons for your answers. (1) (x y) 7! (x + y3 y) (2) (x y) 7! (x + x3 x) (3) (x y) 7! (x + x3 y) (4) (x y) 7! (x2 + yx y) (5) (x y) 7! (x2 + yx + x y) Exercise 2. Consider g(x y) = (ex cos y ex sin y). Verify that dg(x y) is invertible for all (x y) 2 R 2 . State how the inverse function theorem applies to the behaviour of g near to the point (x y). Now verify that the inverse function theorem is true in this case by giving explicit formulae and a domain for the inverse function.
10. Catastrophe theory Let f be a smooth function from R n to R n . The points x R n for which the linear map dfx is not invertible are called singular points of
f . The corresponding y = f (x) are called singular values of f . Points which are not singular are called regular points, and values of y which are not singular values are called regular values of f . We shall study two basic examples in R 2 . (1) The standard fold (y1 y2) = f (x1 x2 ) = (x2 1 x2 ): The matrix of partial derivatives is 2x1 0 0 1
30
This is non-invertible (singular) for x1 = 0, giving the singular points. The singular values are on the line y1 = 0. The behaviour of f can be plotted in the y1-y2 plane. For each (y1 y2), the number of points (x1 x2 ) such that (y1 y2) = f (x1 x2 ) are plotted. It is noteworthy that the singular values lie where the number changes. The resulting diagram is called a bifurcation diagram. (2) The standard cusp (y1 y2) = g(x1 x2 ) = (x3 1 + x2 x1 x2 ): tives of the standard cusp g to nd an equation in (x1 x2 ) for the singular points of g, and an equation in (y1 y2 ) for the singular values of g. Verify that the number of points mapping to (y1 y2) is the number of roots of the cubic t3 + y2 t = y1: Compute the stationary points of h(t) = t3 + y2t, and draw its graph for y2 < 0, y2 = 0 and y2 > 0. For each of the three cases, work out the ranges of y1 for which h(t) = y1 has one, two or three solutions. Draw the bifurcation diagram for g. Exercise 2. Give an example to show that a function can map a regular point to a singular value. If g: R 2 ! R 2 is the standard cusp, with the coordinates named as y = g(x), then using new coordinates x0 = h;1 (x) results in a more general form of cusp, y = g(h(x0 )), with the same bifurcation diagram but di erent regular points. This follows from the chain rule: the matrix @ (g h)i @x0j is singular at a point x0 if and only if the matrix @gi @xj is at the corresponding value x = h(x0 ). Similarly, one can change the y coordinates, by y0 = l(y) which results in the cusp y0 = l(g(x)) with di erent singular values, but corresponding smoothly with the singular values for the standard cusp g.
Exercise 1. Calculate the determinant of the matrix of partial deriva-
31
Whitney's theorem. Whitney showed that a generic smooth func-
tion F : R 2 ! R 2 has singular points which are all either cusps or folds, expressed in some coordinates x0 , y0, which are related to the standard x, y coordinates of the standard cusp or fold by a di eomorphism. The adjective generic refers to a `typical' function F . This means that if a function f does not obey Whitney's theorem, then there is a function with arbitrarily small values, such that F = f + does. This result is fairly di cult to state precisely, and also hard to prove. We shall just explore its content with examples and applications of the idea. Surface projections. The standard fold and cusp can be viewed as projections of smooth surfaces in three-dimensional space. The fold and cusp are the composite mappings R 2 ! R 3 ! R 2
2 (x1 x2 ) ! (x2 1 x2 x1 ) ! (x1 x2 ) 3 (x1 x2 ) ! (x3 1 + x2 x1 x2 x1 ) ! (x1 + x2 x1 x2 )
The rst map in each line parameterises a smooth surface in threedimensional space, and the second map projects it onto the plane R 2 by ignoring the third coordinate value, as one would see by viewing the surface from a direction along the third coordinate axis. Here are two views of the cusp surface in R 3 . The second view is the projection given above.
0.5
-1 -0.5z 0 00 0.5 x y
-0.5 1 1.5 2
-1
-1
-2
32
-1
-0.5
0.5 y
1.5
-1
-2
According to Whitney's theorem, the projection of any surface to the plane will generically have singularities of these types. This can be seen by examples, or by experimenting with real objects. The gure shows the projection of a glass torus, in which four cusp points can be seen, with the lines corresponding to folds.
Exercise 3. List a number of qualitatively di erent ways that the
pattern of folds and cusps can change when you rotate an object. Hint: start by considering walking past a camel with two humps.
Other patterns of change can be obtained by viewing the camel humps in di erent ways.
33
Exercise 4. Consider the stationary points of the function

for constants a b 2 R . How does the number of these vary as the parameters a b vary? Draw the graph of V for each qualitatively di erent set of parameters (a b). Find the equation satis ed by the stationary points of V . How do these values (a b x) relate to the standard cusp? The minimum points can be thought of as places where a ball rolling in the potential V (x) comes to rest. Suppose the ball sits at a minimum point, and the parameters (a b) are varied slowly and smoothly. How do the minima of V behave as (a b) vary? Answer this question by drawing various possible trajectories for a curve in the a-b plane (the bifurcation diagram). Assume that if the minimum point at which the ball sits disappears, then the ball jumps to a new minimum point by rolling downhill. Show how the following phenomena occur (1) (Catastrophes) Discontinuous jumps in the position of the ball (as just described). (2) (Memory) The position of the ball depends on its past history as well as the values of (a b). (3) (Hysteresis) Reversing the path of the parameters (a b) does not reverse the path of the ball. (4) (Divergence) The nal position of the ball depends not only on the initial position and the nal parameters, but also on the path taken by the parameters in the a-b plane Exercise 5. Use the library to nd practical applications of the cusp catastrophe described in the previous question. Which of the described phenonema occur in your examples? 1 x4 + 1 ax2 + bx V (x) = 4 2
34
Examples of surfaces. Some examples of surfaces have already been
11. Manifolds
used, for example in the discussion of the cusp singularity. If f : R 2 ! R is a smooth function, then the graph of f , namely the set of points (x y z) 2 R 3 such that z = f (x y), is a surface. Let us name this set M. The set M has parameters and coordinates, de ned as follows. The ; 2 function : R ! M de ned by (x y) 7! x y f (x y) is called a parameterisation of M . The function : M ! R 2 given by
;x y f (x y) 7! (x y)
is called a coordinate function (or just coordinates) for M . Each component of this function is called a coordinate, i.e., the functions giving the value of x or y. The parameterisation and coordinate functions just de ned are inverses of each other,
R2
; !M ; !R
2
is the identity map. There are more general surfaces than graphs of functions, however. Just one example will su ce for now: the sphere S 2 R 3 de ned by
f(x y z) j x
+ y2 + z2 = 1g:
Clearly, the sphere is not the graph of any function. Also, the sphere does not have a parameterisation in the same way there is no smooth bijection R 2 ! S 2 . The property the sphere does have is that there are locally parameterisations and coordinates. For each point p 2 S 2 , there is an open subset V of the sphere containing p, and an open subset U R 2 , and a parameterisation : U ! V: Example. The spherical polar parameterisation is the map U ! V S 2 de ned by ( ) 7! (sin sin sin cos cos ) and is de ned on the set U R 2 given by 0 < < , 0 < < 2 . The image V is the open subset of the sphere given by excluding the points (x y z) such that x = 0 and y > 0.
35
Spherical polar coordinates are the inverse of this mapping, V ! U . The de nition of a manifold will generalise this idea. A manifold is a subset of R n which has local parameterisations and coordinates from open subsets of R k . We say that a manifold is a `locally Euclidean' subset of R n . Smooth functions have been de ned already when the domain is an open subset of R n . Now follows the de nition for an arbitrary subset. De nition. Let X be any subset of R n . A function f : X ! Y R m is called smooth if for every point x 2 X there is an open subset U R n so that U contains x, and a smooth function F : U ! R m which agrees with f on the open subset V = U \ X of X .
U x X
Functions de ned on subsets of R n .
As before, if f is smooth and f ;1 is also smooth, then f is called a di eomorphism. Exercise 1. Let L R n be a linear subspace, and f : L ! R m a linear map. Is f smooth? Exercisep2. De ne X = f(x y)jxy = 1g R 2 , and f : X ! R by (x y) 7! x2 + y2. Is f smooth? Exercise 3. De ne X = f(t2 t3) j t 2 R g, and the function f : X ! R by (t2 t3 ) 7! t. Is f smooth? Exercise 4. A function is de ned on the line L R 2 which passes through two distinct points a b 2 R 2 by
f ( a + (1 ; )b) =
What is the de nition of a smooth function de ned on L? Construct a function F on the whole plane which agrees with f on L.
36
Rk
De nition of a manifold. A set X
is called a manifold of dimension n if for every x 2 X there is an open subset U X containing x, and an open subset V R n , such that there is a di eomorphism from U to V . This di eomorphism is called a coordinate function, and its inverse a parameterisation of X .
X in the de nition of a manifold, together with the di eomorphism to V R n , are often called coordinate charts, or just charts. A collection of charts is said to cover X if every point x 2 X is in at least one of them. Such a collection is called an atlas for X . Coordinate charts are also called coordinate patches.
Terminology. The open subsets U
Part of an atlas
Examples of manifolds.
Open subsets. The most elementary examples of manifolds are given by taking X R n to be an open subset. Then U = V = X and the di eomorphism in the de nition is the identity map. These examples include R n itself. Graphs. A large class of examples of manifolds are given by the graph of a smooth function de ned on an open subset W R n , f : W ! ; R m . This is the set X = f w f (w) jw 2 W g R n+m . The diagram shows a one-dimensional example. To show that this set is a manifold, take U = X and V = W in the ; de nition ; of a manifold. The parameterisation is x 7! x f (x) and its inverse is x f (x) 7! x.
y y=f(x)
37
The graph of a function
Example. The upper hemisphere U n

n +1 X i=1
R n+1
is the subset
x2 i =1
xn+1 > 0:
This is the graph of the function B1 (0) ! R de ned by
v u n X u (x1 . . . xn ) 7! t1 ; x2 i:
i=1 n +1 X i=1
Spheres. The n-dimensional sphere S n of all points satisfying
R n+1
is de ned as the set
x2 i = 1:
S n is a manifold. This is because any point x 2 S n is contained in a hemisphere for one of the axes. For example, if xn+1 > 0 then x 2 U n . If xn+1 < 0, then x is contained in the lower hemisphere Ln = ;U n . If xn+1 = 0, there is another coordinate, xk , which is not zero, and then x is contained in a hemisphere de ned by taking xk to be the independent variable, instead of xn+1 . Therefore S n is the union of a number of open subsets U (the hemispheres) which are di eomorphic to an open subset of R n , as required in the de nition of a manifold. Torus. Manifolds of dimension 2 are called surfaces. These arise most naturally as subsets of R 3 , according to our everyday experience. An example we have already met is S 2 R 3 . Another example is the torus. This can be drawn by rst drawing a circle of radius 2 in the
38
x ; y plane, then taking each of these points as the centre of another circle of radius 1, lying in a plane through the z-axis. Following this idea through, leads to a de nition of T R 3 as the subset of points satisfying
x +y
2
;2
+ z2 = 1
It is possible to show that T is a manifold directly, by nding a set of coordinate charts which cover T , as was done for the sphere. It also follows from an exercise below.
z y
Groups. The groups which arise in the special geometries have sets of elements which are labelled by continuous parameters. In this situation, it is natural to ask whether these sets are manifolds. The group GL(n) is the set of all invertible matrices thus the matrix entries can be taken 2 2 n as n coordinates. This is a map GL(n) ! R given by 0 a11 . . . a1n 1 @ ... A 7! ;a11 a12 . . . ann an1 . . . ann
It is often convenient to regard GL(n) as a subset of R n2 , as the rearrangement of the matrix entries as a vector is of little consequence. The condition on the matrix A for it to be invertible is det A 6= 0, which is a polynomial equation in the matrix entries. For example, for n = 2, this is a11a22 ; a12a21 6= 0:
39
This de2 nes an open subset of R n2 . In general, this is because the map det: R n ! R is continuous, and GL(n) is the inverse image of R nf0g, an open set. Thus GL(n) is an n2 -dimensional manifold. Various subgroups of GL(n) are also manifolds. For example, the group of n n orthogonal matrices, O(n), is a manifold of dimension 1 2 n(n ; 1). Alternative de nition. Many books use an alternative de nition of manifold, as a set X which is a topological space, and an atlas of coordinate charts for X . Thus X is not regarded as a subset of R n . There are conditions on the atlas for this de nition to make sense, and the main disadvantage of this method is that these conditions require much more technical e ort to explain. There is no more generality in this as the set of manifolds obtained by the alternative de nition is equivalent to the set of manifolds de ned here. Exercise 5. Let U R 3 be the upper hemisphere
f(x y z)jx
+ y2 + z2 = 1 z > 0g
The map f : U ! R 2 is de ned by (x y z) 7! (x=z y=z). Give a formula for the inverse function f ;1. Is f ;1 smooth? Exercise 6. The spiral S R 2 is the set of points of the form (e cos e sin ) for 2 R . Sketch S . Now consider the function f : S ! R de ned by (e cos e sin ) 7! . Consider a point x 2 S , and suppose the ball Br (x) does not contain the origin (r jxj). Show that there is a map F : Br (x R 2 ) ! R which agrees with f by giving an explicit formula. What goes wrong if r > jxj? Is the original function f smooth? Is S a manifold? Exercise 7. Give a set of parameters for the elements of the a ne group A(n). Hence describe A(n) as a subset of R k for some k. Exercise 8. In complex analysis, it is often convenient to add a single extra point àt in nity' to the complex plane, so that any set of points of increasing radius from the origin converges to the extra point 1. What manifold do think C f1g should be?
40
If M R m and N R n are manifolds, then we have essentially already de ned a smooth map M ! N : it is a smooth map M ! R n , as de ned previously, such that the image lies in N . If f : M ! N is a smooth map with a smooth inverse, then it is called a di eomorphism, and M and N are said to be di eomorphic. For example, S 2 is di eomorphic to an ellipsiod f(x y z) j ax2 + by2 + cz2 = 1g determined by constants a b c > 0. The di eomorphism is the linear map x p y p z : (x y z) 7! p a b c The torus and the sphere are an example of two manifolds which are not di eomorphic. This is intuitively obvious as the torus has a hole through the middle but the sphere does not, but proving it requires some thought. Here is an argument which can be made into a proof. Suppose f : T ! S 2 is a di eomorphism. The torus T has two circles on it which cross at only one point (e.g., given by z = 0 and y = 0). The images of these circles would be two circles on the sphere which cross at only one point. Draw one circle on a sphere. It is òbvious' that one cannot draw a second circle to cross it at only one point. Product of manifolds. If M R k and N R l are manifolds, then so is M N R k+l . As an example, S 1 S 1 is the subset of R 4 given by x2 + y2 = 1 z2 + t2 = 1: Exercise 9. How can S 1 S 1 R 4 be parameterised? Find a di eomophism S 1 S 1 ! T . Exercise 10. Give a di eomorphism between the manifold D R 5 given by w2 + x2 + y2 + z2 ; t2 = 1 and S 3 R . Solution to exercise 1. Let L0 be a complementary subspace, so that R n = L + L0 . Decompose a vector 2 R n as = l + l0 , l 2 L, l0 2 L0 , and set F ( ) = f (l). This map is de ned on the whole of R n (certainly an open subset), agrees with f , and is smooth. Note that since there are many choices of complementary subspaces L0 , there are many linear extensions F . There are also choices of F which are not linear.
Maps of manifolds.
41
Let M ! we are interested in are maps which are isomorphisms on a su ciently small scale. De nition. A local di eomorphism f : M ! N is a map such that for each x 2 M there is an open subset U M containing x and an open subset V N , so that f restricts to a di eomorphism U ! V . If the local di eomorphism f is onto, we say that N is a discrete quotient of M . In the rest of this section, this is abbreviated to quotient, though the term quotient can apply to more general onto maps. If y 2 N , we say that f identi es the points f ;1(y) M in the quotient. Example. The simplest example is the map R ! S 1 given by
12. Discrete Quotients Rm, N R n and f : M N a smooth map. The maps
sin ): The points + 2 n, n 2 Z, are identi ed to a single point in S 1 . There is a purely topological notion of a quotient space. Namely, given a topological space X and an onto map f to a set Y , then there is the quotient topology of Y de ned by making V Y open whenever f ;1(V ) is. This could give a possibly con icting notion of quotient, so it is important to show that these coincide. Theorem. Let f be a local di eomorphism of M onto N . Then N has the quotient topology. Proof. This is proved by showing that f maps open sets to open sets. This is called an open mapping. Then the result follows easily from this. The details are in the following two lemmas Lemma. A local di eomorphism is an open mapping. Proof. For each x 2 M , let Ux be the open set containing x on which f is a di eomorphism. Then if U M is any open set,
7! (cos
U = x2U Ux \ U
and
f (U ) = x2U f (Ux \ U ): Since f is a homeomorphism Ux ! f (Ux ), f (Ux \ U ) is an open subset of f (Ux ) and hence an open subset of N . Hence f (U ) is an open subset of N .
42
the quotient topology. Proof. Let V N be a subset such that U = f ;1 (V ) is open. Then f (U ) is open since f is an open mapping, and V = f (U ) since f is onto. Hence a subset of N which is open in the quotient topology is open in the given topology of N . The converse follows from the fact that f is continuous. However the condition of local di eomorphism is much stronger than being a purely topological statement. For example x 7! x3 is a homeomorphism but not a local di eomorphism. But the topological notion of a quotient space is a useful half-way house to give a description of N. De nition. A fundamental domain for a quotient f : M ! N is a closed subset D M such that f (D) = N but f only identi es points on the boundary @D of D. For example, the interval 0 2 ] is a fundamental domain for the map R ! S 1 above. The boundary is the endpoints of the interval, f0 2 g and these two points are identi ed to one point in the quotient. This example can be generalised to products of the circle, for example R 2 ! S 1 S 1 . A fundamental domain is a square with opposite edges identi ed, and all four corners identi ed to one point. Projective space. So far, P n has been described simply as a set. Now it can be described as a manifold, obtained as a quotient of the sphere. Recall that P n is de ned as the set of lines through the origin in R n+1 . Each line intersects the sphere S n in exactly two points, x.
Lemma. If f : M ! N is an open mapping and is onto, then N has
So to construct projective space as a manifold, it is su cient to nd a quotient of the sphere, f : S n ! Qn R k . This map to R k has
43
the property that f (x) = f (;x). So to construct such a map, a good starting point is to nd a set of functions S n ! R with this property. Let x = (x1 . . . ) denote points on S n . Then the functions xj xk for any choice 1 j 1k n + 1 give a function with same value on x. De ne f : S n ! R 2 (n+1)(n+2) to be the map 2 (x1 x2 . . . xn+1 ) 7! (x2 1 x1 x2 x1 x3 . . . xn+1 ) and de ne Qn to be the image of f . Now a number of questions need to be settled to show this construction works. (1) Is f a local di eomorphism? (2) Is Qn a manifold? (3) Are any other points identi ed, beside x 2 S n ? The answer to (1) is yes due to the following construction. The map f can be locally inverted by explicit formulae, which will be given here for n = 2. The function f : S 2 ! Q2 is the map 2 2 (x1 x2 x3 ) 7! (x2 1 x1 x2 x1 x3 x2 x2 x3 x3 ) = (y1 y2 y3 y4 y5 y6 ): Assume that x3 > 0. Then x 2 S 2 can be calculated from q2 x x 1 x3 2 x3 x1 = p 2 x2 = p 2 x3 = x3: x3 x3 Therefore the inverse of f is locally y3 y5 py : (y1 y2 y3 y4 y5 y6) 7! p y6 py6 6 This formula clearly extends to points in an open subset of R 6 and so indeed provides the smooth inverse to f for the hemisphere x3 > 0. The answer to the question (2) is yes by the following general result. Lemma. Let M be a manifold, and N a subset of R k . If there is a local di eomorphism f : M ! N , then N is manifold. Proof. Suppose f restricts to a di eomorphism U ! V for open subsets U and V . Let x 2 U . Then there is a coordinate chart de ned on an open subset U 0 M , with x 2 U 0. Of course, U and U 0 need not coincide, but one can restrict both maps to the intersection, U 00 = U \ U 0. Then U 00 is di eomorphic to an open subset of R n , and to an open subset f (U 00 ) of N . This provides a coordinate chart around the point f (x) 2 N . Finally, it is an algebraic exercise to settle question (3).
44
are x. What is a fundamental domain for this quotient? How are the boundary points in your fundamental domain identi ed? Exercise 2. Show that the inhomogeneous coordinates for P n give coordinates for the manifold Qn . These calculations show that Qn really ìs' P n, viewed as a manifold. Calculations can be done using the inhomogeous coordinates, as previously, but now with the understanding that these are coordinate charts on a manifold. Exercise 3. Draw the subset Q1 R 3 . Exercise 4. Find a map S 1 ! S 1 which is a local di eomorphism, but is not a bijection. Is this possible for a map R ! R ? Exercise 5. De ne M P 4 by the equation Find an onto map S 2 S 1 ! M , and describe the quotient of S 2 S 1 this de nes. is an open interval, say fa < t < bg. Its tangent vector at a given parameter t is dc = dc1 . . . dck : dt dt dt Now suppose f : R k ! R l is any smooth function. It carries the curve c into a curve c0 in R l , namely c0(t) = f (c(t)). The tangent vector for curve c0 = f c can be calculated by the chain rule by
R
Exercise 1. Show that the only points identi ed in the map S n ! Qn
x2 + y2 + z2 ; t2 ; w2 = 0:
Tangent vectors.
13. The tangent space A curve is a map c: I R k , where I
dc0t = dfc(t) dct so that dc : dc0 = dc0 (1) = df (dc (1)) = df t c (t) c (t) t dt dt Therefore the derivative of f , dfx, can be interpreted as a linear mapping of tangent vectors, for curves passing through the point x 2 Rk.
45
be a manifold, and x 2 M . Then a vector v 2 M at x if there is a curve c c: I ! M such that c(t) = x for some t 2 I , and d dt = v . De nition. The tangent space at x 2 M , denoted TMx, is the subset of R k consisting of all vectors which are tangent to M at x. Exercise 1. Show that 0 2 TMx . Lemma. TMx R k is a linear subspace. Proof. Let p: U ! M be a parameterisation, for an open subset U R m , such that u 2 U is mapped to x. It will be shown that TMx = Image (dpu ) which is a linear subspace. Firstly, if c is a curve in M with c(0) = x, then for values of t su ciently close to 0, c(t) = p p;1 c(t), so at t = 0 dc = dp d(p;1 c) 2 Image(dp ) : u u dt dt ; Conversely, if v 2 Image dpu , then v = dpu ( ) for some 2 R m . The curve c(t) = p(u + t ) has tangent vector dc = dp ( ) = v u dt at the parameter value t = 0. Exercise 2. Use the chain rule to show that the dimension of TXx is the dimension of X . Use the fact that a parameterisation has an inverse. Examples of tangent spaces. The tangent space to R n at any point is R n itself, as the identity map is a parameterisation, and the derivative of this is also the identity map. The tangent space to an a ne subset A R k at any point x 2 A is the subspace TA = fa ; xja 2 Ag introduced earlier. It is independent of x. This follows because an a ne subset can be parameterised by an a ne map : R n 7! A R k . Then d x is just the map d : y 7! (y) ; (x) introduced earlier. The image of this map is the set of points f (y) ; (x)g for a xed x and all y 2 R n , which is the de nition of TA. Manifolds M R k can occur as the solutions to an equation F = 0, where F : R k ! R l . Then dFx (v) = 0 for any vector v 2 TMx. If c dF c v=d dt , then F c = 0, so dt = dFx (v ) = 0. This gives a method of determining linear equations for TMx .
Tangent space. Let M
R k is said to be tangent to
Rk
46
n +1 X i=1
Example. The sphere is given by F : R n+1 ! R

F= x2 i
; 1 = 0:
Then dFx (v) = 2( xi vi ) = 2x v = 0. This is one linear equation in R n+1 and hence determines an n-dimensional linear subspace. Since n also has dimension n, it follows that these two spaces are equal. TSx n is the set of all vectors satisfying x v = 0. Hence TSx Exercise 3. Show that the maps S 3 ! R 4 given by
v: (x1 x2 x3 x4 ) 7! (x3 ;x4 ;x1 x2 ) w: (x1 x2 x3 x4 ) 7! (;x2 x1 ;x4 x3 ) determine three vector elds on S 3 . Do u(x), v(x), w(x) form a basis 3? for TSx
the derivative of f at x 2 M is the linear map dfx : TMx which satis es c = d(f c) : dfx d dt dt
u: (x1 x2 x3 x4 ) 7! (;x4 ;x3 x2 x1 )
Di erentiation on manifolds. De nition. Let M R k , N R l be manifolds, and f : M ! N . Then
! TNf x
( )
where c is any curve through x. In this de nition, it is necessary to check that a linear map with these properties exists. Suppose F is de ned on an open subset of R k and agrees locally with f . Then we can use the chain rule for R k on F c) = dF dc = dF (v): dfx (v) = d(fdt c) = d(F x c(t) dt dt
This shows that the derivative in this more general situation (manifolds) is just the restiction of dFx to TMx. This proves that a linear map with these properties exists. One might worry that this depends on the choice of F however the de nition of dfx determines its values on all vectors tangent to M uniquely without reference to F . Exercise 4. Show that the chain rule holds for maps of manifolds.
47
Vector elds. A vector eld on a manifold is a choice of a tangent
vector at each point of M . More precisely, a vector eld on a manifold M R k is a smooth map v: M ! R k such that v(x) 2 TMx for each x 2 M. The tangent bundle. The tangent spaces for the di erent points x 2 M are generally di erent subspaces of R k . In general, the tangent spaces can be `glued together' to form the tangent bundle of a manifold, TM R 2k = R k R k . This is de ned to be the set of all points (x v), for x 2 M and v 2 TMx . A vector eld can be described as a map M ! TM , as
x 7! (x v(x)):
explicitly by giving two equations for the subset. Give a vector eld on S 1 which is nowhere zero, i.e., the tangent vector at every point is not 0. If V is a one-dimensional vector space, explain how a choice of vector in V determines a linear isomorphism R ! V . Use your vector eld to give a di eomorphism S 1 R ! TS 1. Exercise 6. Let f : R n ! R m be a smooth function. Show that the graph G R n+m of f is a manifold. For a point g 2 G, which function is TGg the graph of? Explain how you would show that the subset of R 3 given by
R4
Exercise 5. Describe the tangent bundle TS 1
x4 + y 4 + z 4 = 1
is a manifold. Give an explicit description of the tangent space at the point x = 1 y= p 4 2 , z = 0, with numerical coe cients in the equation that you use. Exercise 7. A robot arm in the plane has its elbow at x 2 R 2 and hand at y 2 R 2 . These are constrained by jxj = 1 and jy ; xj = 1. Let X = f(x y)g R 4 be the set of con gurations for the arm. Which standard manifold is X di eomorphic to? The arm is controlled by a motor which sets the angle of the upper arm x relative to a xed axis, and a second motor which sets the angle of the lower arm (y ; x) relative to the upper arm. Explain how the hand y can be moved in a given direction in the plane given by a tangent vector v by giving a direction in the space of angles controlled by the
48
motors. At what points y does this control mechanism fail to work for some tangent v? How would you draw a circle of very small radius at the point y = (1 0)? Explain why this does not work at the origin y = 0. Exercise 8. Two solid bodies touch at a single point p. Assume the boundaries of the solid bodies can be modelled as smooth surfaces. What can you say about the relation between the tangent spaces of the two boundary surfaces at p? Give at least one concrete example.
49
A group is a set G with maps : G G ! G and : G ! G giving the multiplication and inverse of group elements, and an element e 2 G, the identity, all satisfying the usual axioms. The usual notation is (a b) = ab and (a) = a;1 . De nition. A Lie group is a group in which G is a manifold and and are smooth maps. In the same way, one can de ne `Lie' versions of all the elementary de nitions in group theory. For example, a subgroup H G which is also a Lie group is called a Lie subgroup of G. A homomorphism of Lie groups F ! G is a group homomorphism which is also a smooth map.
14. Lie Groups
Examples of Lie Groups.
The group GL(n) is a Lie group. The coordinates for GL(n) are the matrix entries. The map is smooth because a matrix entry for the product (a b) is a polynomial in the matrix entries for a and b. Also, is smooth because the matrix entries for a;1 are polynomials divided by det a, which is never zero for elements of GL(n). The vector space R n is a Lie group, with (a b) = a + b, (a) = ;a. This group is called the translation group, T(n). Exercise 1. Show A(n) is a Lie group. Further examples of Lie groups arise as subgroups of GL(n). These will be discussed later. The Euclidean groups E(n) and the projective groups PGL(n) are also Lie groups.
Actions of Lie groups.
Let M be a manifold, and G a Lie group. An action of G on M is a map : G M ! M satisfying (1) ( (g h) x) = (g (h x)) (2) (e x) = x Each element g 2 G provides a smooth map
g: x
7!
(g x)
called the action of g. The conditions (1) and (2) can be written
(gh)
= g h e = identity :
50
;
;
Since e = g 1g = g 1 g , the action of g;1 is the inverse of the action of g. Therefore, the action of g 2 G is a di eomorphism. The conditions satis ed by an action can be stated alternatively as saying that there is a homomorphism from G to the group of all di eomorphisms of M . Classi cation of actions. There are three properties an action may have (1) E ective. If g 2 G is such that g = identity, then g = e. (2) Free. If (g x) = x for some x 2 M , then g = e. (3) Transitive. For all x y 2 M , there exists g 2 G such that (g x) = y. The rst property, e ective, is that every element of G except the identity does `something somewhere'. This property is automatic for any action de ned as the set of all transformations of a manifold of a particular kind. The action of GL(n) on P n;1 is not e ective because all multiples of the identity in GL(n) act as the identity in P n;1. A xed point for an element g 2 G is a point x 2 M such that (g x) = x. For example, the rotations of a sphere about the z-axis have the north and south poles as xed points. An action is free if every element except e has no xed points. For example, the action of T(n) on R n by translations is free. By contrast, the action of O(3) on S 2 is not, because of the xed points for rotations just mentioned. The orbit of a point x 2 M is the set of all points f (g x)jg 2 Gg. An action is transitive if there is only one orbit. For example, the action of T(n) on R n by translations is transitive, as is the action of O(3) on S 2. However if you take the subgroup of rotations about the z-axis, this is not transitive.
fixed point for every g
orbits
Exercise 2. Are the following actions e ective, free or transitive?

(1) GL(n) acting on R n
51
(2) The group of rotations about the z axis in R 3 acting on the sphere, S 2. (3) A Lie group G acting on G by multiplication in the group: (g h) = gh. (4) A Lie group G acting on G by conjugation: (g h) = ghg;1 . A ow is an action of the group T(1) on a manifold. Let : R M ! M be the action. Then each point x 2 M gives a curve x: R ! M t 7! (t x): Since x(0) = (0 x) = x, the curve gives all points on the orbit of x under the ow. The curve has tangent vector x d v(x) = dt (0) 2 TMx: A vector eld, called the velocity vector eld of the ow is de ned by the function x 7! v(x). Example. A river ows smoothly (of course) along a waterway M . The function (t x) gives the position at time t of the molecule of water which is at the point x at time 0. The vector v(x) gives the velocity of the water passing the point x at any value of the time parameter. This follows from the property of a group action that x (t) = y (T + t) for x = (T y) i.e., the curve through x is the same as the curve through y with the parameter shifted by t 7! T + t. Then d y (T ) = d x (0) = v(x): dt dt
15. Flows
and a point x 2 M , the problem is to nd a curve c: I ! M such that c(0) = x, I R is an open interval containing 0, and the tangent vector at any parameter t agrees with the vector eld, i.e. dc (t) = v;c(t) : dt Such a curve is called an integral curve of the ordinary di erential equation. If it exists, the integral curve is unique. (The proof of this, not given here, involves some analysis). If the vector eld v is the velocity of a ow, then the solution is given by c(t) = x (t). Therefore, the velocity vector eld of a ow speci es the ow uniquely.
Ordinary di erential equation. Given a vector eld v on a manifold
52
Example. The di erential equation on R

with a 2 R a constant, has solution c(t) = xeat , which is determined by the ow (t x) = eatx. Exercise 1. Solve the equation dc = ac2 dt and show that the solutions do not determine a ow. Example. In mechanics, Newton's equations for the position of n particles x 2 R 3n can be written as the ordinary di erential equation dx = p dp = f (x): dt dt The function f gives the forces on the particles as a function of the positions, and is determined by the particular problem. This is an ordinary di erential equation in a subset of R 6n . In Newton's theory of gravity, there is a formula for f determined by the inverse square law. The integral curves of this equation account for the orbits of the planets, amongst other phenomena. For n 4, the integral curves cannot always be de ned for all t 2 R . There is, suprisingly, a con guration of 4 bodies for which the orbits become progressively more violent, and one of the bodies `reaches in nity' in a nite interval of time. Vector eld as an operator. Suppose M R k is a manifold, and v is a vector eld on M . If : M ! R is a function, then the vector eld can be regarded as a di erential operator which acts on to give a new function, called Dv .
k @ X ; Dv (x) = d x v(x) = @x vi (x): i i=1
dc = ac dt
This operation can be thought of as di erentiating in the direction given by v.
53
This can be substantiated as follows. Let : R M ! M be a map, which could be a ow. Then it de nes a vector eld by
x d v(x) = dt (0)
as before. Now di erentiating along each curve x , one nds d( dt

x x ) (0) = d x d dt = D v
Exercise 2. Take M = S 2 , and the rotation : R S 2 ! S 2

(t x1 x2 x3 ) 7! (x1 cos t + x2 sin t ;x1 sin t + x2 cos t x3 ) Calculate the vector eld v and an expression for Dv . Suppose : M ! R is a function which is constant along a ow , ; x i.e., (t) = (x) for each x, t. This gives the rst order partial di erential equation k @ X vi @x = 0: i i=1 This equation can be used to solve the ordinary di erential equation given by v. The integral curves of the ordinary di erential equation must lie in the subset (x) = constant, for each solution of the partial di erential equation. Example. In R 2 , the equation is
Partial di erential equation.
@ + v (x x ) @ = 0: v1 (x1 x2 ) @x 2 1 2 @x2 1
In most cases, a solution to this equation will give a 1-manifold as the set of points (x1 x2 ) = constant, which can be taken as the image of a curve, given by this implicit equation. The di erence between a solution to this equation and a solution to the ordinary di erential equation is that no parameterisation of the curve is speci ed.
54
Symmetries of a di erential equation. Lie groups of symmetries

for a di erential equation can often be used to reduce the number of equations or independent variables. In the simplest cases, the equations will reduce to an equation in one variable which can be solved by integration. A transformation : M ! M is a symmetry of an equation if it transforms solutions to solutions. For an ordinary di erential equation, this means that if the curve c: R ! M is a solution, then so is c. Suppose that these solutions arise from a ow, so that c = x . Then c must be the curve determined by the ow through the point (x), i.e., (x) . This condition can be written ( (t x)) = t ( (x))
: Now suppose is itself any one of the transformations of a second ow 0. Then 0 t = t 0: s s for all (s t) 2 R 2 . Two ows which satisfy this condition are said to commute. The map on either side of this equality can be taken to de ne an action of the group T(2) = (R 2 +) on M . Exercise 3. Show that (s t) (u v) = (s+u t+v) , a condition for this to be an action. Exercise 4. Write down two ows on R 2 which are distinct and (1) commute (2) do not commute Solution to exercise 1. The equation has the solution c(t) = x=(1 ; atx), which is de ned only as long as at < 1=x. Clearly c(t) is in nitely large as this limit is reached, and the solution does not exist for all t. Hence the solutions are not determined by a ow.
t= t
for all x, t, or
A one-parameter subgroup of a Lie group is a homomorphism h: R ! G i.e., ancurve where h(0) = e and h(s)h(t) = h(s + t). Since h(t) = ;h,(t=n ) , the one-parameter subgroup is determined by its values for t arbitrarily close to 0. This section will show that it is in fact determined by its tangent vector at h(0) = e.
16. One-parameter subgroups
55
One-parameter subgroups play a central role in the theory of Lie groups. For example, if G acts on a manifold M by : G M ! M , then a one-parameter subgroup determines a ow by For a xed x 2 M , de ne x : G ! M by;g 7! (x g). Then the integral curve of the ow through x is c(t) = x h(t) . The velocity vector is dc (0) = d x dh (0) e dt dt (t x) 7! h(t) x :
2 TMx:
This vector eld on M is determined completely by the tangent vector dh (0) 2 TG : e dt Lemma. A one-parameter subgroup h is determined uniquely by its tangent vector dh=dt at the origin. Proof. Let M = G and the action (g x) = gx be the group multiplication. Then the ow is (t x) 7! h(t)x and the integral curve through e is just h itself. However, the ow is determined uniquely by its velocity vector eld, which by the preceding argument is determined by dh=dt(0). Tangents to GL(n). Recall that GL(n) R n2 , the latter regarded as the set of all n n matrices. Since it is an open subset, T GL(n)x = R n2 . A tangent vector can likewise be regarded as an n n matrix. For example, if c: R ! GL(n) is a curve, with
0 c11 c12 . . . 1 c(t) = @ c21 . . . . . . A

.. . .. .
11 12
then
.. .. For GL(n), it is easy to show the converse of the preceding lemma
0 dc dc . . . 1 t c t .d dc = B dd .. ... C dt @ A 2 T GL(n)c(t) : dt . .
21
56
Lemma. Any vector A 2 T GL(n)e is tangent to a one-parameter subgroup. Proof. The one-parameter subgroup is given by
h(t) = exp(tA) =
1 tn An X
In this formula, An refers to the matrix product. The sum is easily seen to converge, and
1 ntn;1 1 tn;1 X dh (t) = X n A = A (n ; 1)! An;1 = Ah(t) dt n ! n=1 n=1 so that dh=dt = A when t = 0. Exercise 1. Multiply the exponential series to show that h(s)h(t) = h(s + t). This shows that for GL(n), the one-parameter subgroups are in 1 ; 1 correspondence with elements of the tangent space at e 2 GL(n).
17 Subgroups of GL(n)
n=0
n! :
The following subgroups are all Lie groups. Special linear group. The special linear group SL(n) GL(n) is the subgroup of matrices with determinant equal to one. Since det (exp(A)) = exp (trace(A)) for any matrix A, then exp(A) 2 SL(n) if and only if trace(A) = 0. Hence T SL(n)e T GL(n)e is the linear subspace given by the linear condition trace(A) = 0: Orthogonal group. The orthogonal group O(n) is the group of n n matrices which satisfy the condition M T M = e: Suppose c: R ! O(n) is a curve with c(0) = e and tangent equal to A at e. Then at t = 0,
T ; d c d c = AT e + eA = AT + A T 0 = dt c(t) c(t) = dt c(0) + cT (0) d dt
57
so that A is an antisymmetric matrix. Conversely, if A is an antisymmetric matrix, then (exp A)T = exp AT = exp(;A) = (exp A);1 so exp A is orthogonal. This shows that T O(n)e T GL(n)e is the linear subspace given by the condition AT + A = 0. The special orthogonal group, SO(n), is de ned to be the intersection O(n) \ SL(n), the orthogonal matrices of determinant 1. Since det is continuous, and the only values it takes in O(n) are 1, any curve c which passes through the point e must have det(c(t)) = 1 for all t. Therefore, T SO(n)e = T O(n)e . The group SO(3) is called the rotation group. To justify this name, we prove the following Theorem. Every rotation has an axis. Proof. Let M 2 SO(3). The characteristic polynomial of M has at least one real root, so that M has an eigenvector v. Since M is an isometry, the eigenvalue is 1. If the eigenvalue is 1, then v is the axis. The matrix M acts in the plane orthogonal to v. In an orthonormal basis which includes v, M is
01 0 01 @0 a b A:
0 c d
b , Since the determinant of M is 1, so is the determinant of a c d which gives an element of SO(2), for which the formulae give explicitly a rotation in this plane. If the eigenvalue is ;1 then M acts in the orthogonal plane again, but with determinant ;1. This gives a re ection in this plane, and the explicit formulae show that there is a re ection axis, Mv = v. This returns to the previous case with v the axis.
By taking real and imaginary parts of the components of a complex vector, C n can be regarded as R 2n . Thus an invertible matrix with complex entries determines an element of GL(2n). Example. An invertible 1 1 matrix is just a non-zero complex number c = c1 + ic2 . This acts on x = x1 + ix2 2 C by multiplication of
Unitary group.
58
1 (x1 x2 ) 7! c c2
complex numbers. This amounts to the formula
;c
c
2 1
x1 x2
which de nes an element of GL(2). The unitary group, U(n), is de ned to be the group of unitary n n complex matrices. These satisfy U ;1 = U T . The special unitary group SU(n) U(n) is the subgroup of unitary matrices with determinant one. Exercise 1. What complex numbers correspond to elements of U(1)? Exercise 2. Show that elements of SU(2) can be written in the form a b , where a and b are complex numbers. Give a di eomorphism ;b a 3 SU(2) ! S .
18 The commutator
The condition that two ows commute can be written entirely in terms of their velocity vector elds. Let v and v0 be the velocity vector elds of ows , 0 on a manifold M . De nition. Let M R k , and v v0 : M ! R k be vector elds on M . The commutator of v and v0 is de ned to be the function v v0 ]: M ! R k given by Dv v0 ; Dv v:
0
The coordinate expression is that the j -th component of v v0 ] is
X @vj0 0 @vj vi @x ; vi @x : i i i
Theorem. The two ows and
mutator v v0 ] of their velocity vector elds is zero. Proof. Pick x 2 M , and let y(s t) = 0 (s (t x)). The behaviour of y for small values of s and t near 0 is determined by the second derivative @ 2 y at (0 0). This is calculated in the following way. @s@t
0 commute if and only if the com-
@y (0 t) = v0 ( (t x)) @s
59
@ 2 y (0 0) = dv0 ; d = dv0 ;v(x) x dt x @t@s Likewise if z(s t) = (t 0 (s x)), then @ 2 z = dv ;v0 (x) : x @s@t If the ows commute, then y(s t) = z(s t). Since these mixed second order partial derivatives are equal, the result that v v0 ] = 0 follows. Conversely, assume that v v0 ] = 0. According to a previous argument, it is su cient to show that each transformation 0s is a symmetry of the ordinary di erential equation determined by the vector eld v, d ( 0 x ) = v 0 x for all s. s dt s Note that at s = 0, this equation reduces to the de ning equation for v, namely d x =v x dt and so certainly holds. Using the chain rule, the condition is equivalent to d ( 0s )x (v(x)) = v ( 0s (x)) for all s 2 R x 2 M . Consider (s) = d ( 0s )x (v(x)) ; v ( 0s (x)) Some di erentiation, and using the hypothesis v v0 ] = 0 shows that d = dv 0 (s x) ( (s)) : ds Since (0) = 0, this ordinary di erential equation has the unique solution (s) = 0 for all s 2 R . Example. For the ow given by rotations of S 2 about the x3 -axis, the vector eld is v(x1 x2 x3 ) = (x2 ;x1 0). Consider a second ow given by rotations about the x1 -axis. This vector eld is v0 (x1 x2 x3 ) = (0 x3 ;x2 ). The commutator is given by
0
and so
! 0 0 0 X X X @v1 @v @v @v @v @v vi @x2 ; vi0 @x2 vi @x3 ; vi0 @x3 = vi @x ; vi0 @x1 i i i i i i i i i
(;x3 0 x1 ):
60
This is not zero, so the ows do not commute. The preceding theorem gives an interpretation of the vanishing of the commutator of two vector elds, in the case when they generate ows. The commutator of two vector elds is also important when it does not vanish. The main fact is Theorem. The commutator of two vector elds on M is also a vector eld on M . This fact will be proved below. Note that in the example, calculating Dv v0 (x) gives (0 0 x1 ) and Dv v(x) = (x3 0 0), neither of which are tangent to S 2 . Only the di erence of these gives a vector eld on S 2 . The commutator has an interpretation in terms of vector elds acting on functions. If v and v0 are thought of as operators acting on functions, then taking an arbitrary function : M ! R , we calculate the di erence of v0 acting followed by v acting and v acting followed by v0 acting. This is also called a commutator, namely the commutator of the rst order di erential operators given by v and v0 . X @ 0@ @ v @ Dv (Dv ) ; Dv (Dv ) = vi @x vj @x ; vi0 @x j @x i j i j ij 0 @ @vj j @ =D = vi @x @x ; vi0 @v vv] @xi @xj i j as the terms involving second derivatives of cancel. So the commutator of the rst order di erential operators acting on is just v v0 ] acting on . This gives an argument about why the commutator is a vector eld. If the manifold is de ned by an equation = 0, then certainly Dv = 0 and Dv (Dv ) = 0, and so D v v ] = 0. Hence v v0 ] is tangent to M . For S 2, = x x ; 1, which explains the example. More generally, we could take M R k to be given by an equation = 0 for some : R k ! R l . This would give a proof if we knew that all manifolds can be de ned in this way (which they can, at least locally). A slight modi cation of this idea gives the proof. Proof of theorem. For each point x 2 M R k , there is locally a coordinate function f from M to R m , and its inverse, a parameterisation , from R m to M . (`Locally' means that these are de ned on open subsets around x or its image.) Let : R k ! M be F , where F is any local extension of f to R k . Then a vector v is a tangent vector to M if and only if d x (v) = v
0 0 0 0 0 0 0
61
Let v be a vector eld on M , and f : M ! N be a mapping, and w a vector eld on N . ; Then w is said to be f -related to v if w(f (x)) = dfx v(x) . For example, if f is a di eomorphism, then given v, there is a unique f related vector eld on N called the induced vector eld ; w(y) = dfx v(x) where x = f ;1(y): Exercise. Show that if g: N ! P is another di eomorphism and z is the vector eld induced on P from w, then z is equal to the vector eld induced by g f from v. The mapping formula for the commutator is: Lemma (Mapping formula). If f : M ! N is a mapping, v v0 are vector elds on M , and w on N is f -related to v, w0 f -related to v0 , then w w0 ] is f -related to v v0 ]. Proof. dfx( v v0 ](x)) = D v v ] f evaluated at the point x. But this function is D v v ] f = Dv (Dv f ) ; Dv (Dv f ) = Dv (w0 f ) ; Dv (w f ) At the point x, the right-hand side is d(w0 f )x (v(x)) ; d(w f )x (v0 (x)) ;dw df (v0(x)) 0 d f ( v ( x )) ; = dwf x x f (x) (x) = w w0](f (x)):
0 0 0 0 0
The mapping formula.
This is because is the identity map on M , and TMx is the image of d x . This equation is equivalent to writing Dv = v for a vector eld de ned locally (i.e., on the open subset of M where is de ned). Then D v v ] = Dv (Dv ) ; Dv (Dv ) = Dv v0 ; Dv v = v v0 ] so that v v0 ](x) 2 TMx.
0 0 0 0
62
19. Computer vision
An optical image is the projection of a three-dimensional scene under a smooth map : R 3 ! R 2 . A variety of possible 's can occur, depending on the camera. It is assumed that various features (points, lines, curves, surfaces, corners, smooth singularities, etc.) can be recognised in the image. The `3-d recovery problem' is to say where these features are in R 3 which give rise to the optical image. An assumption is usually made about the nature of the scene, as a hypothesis which can be then given a mathematical formulation. For example, it could be that the scene is a face, a microscope slide containing cells, an aerial photograph, or a stack of books to be counted. From this assumption, you have a hypothesis about the scene containing a number of continuous parameters. For example, if the scene is a cell which is a assumed to be spherical, then the parameters might be the radius of the cell and the position of its centre. If the cell is not assumed spherical, then additional parameters would be needed for its shape and its angular orientation in space. The 3-d recovery problem can be thought about in two ways (1) 3-d Euclidean geometry. Find the set of all objects in R 3 which could give rise to the image. For example, if the scene consists of rigid bodies (bodies for which the distance between the constituent parts does not change) then the Euclidean group E (3) acts on the set of all possible positions for each body, and the recovery problem would reduce to nding the possible Euclidean transformations which take each object from a standard position to its actual position. (2) 2-d Non-Euclidean geometry. For each possible three-dimensional object, nd the range of possible images. See which of these ts the given image. For example, if the camera gives a projection along straight lines, then the image can be regarded as a parameterisation of part of the projective plane, P 2. Projective transformations can be applied to the images. More information can be gained if the image varies with time. One of the ways of converting this information into an easily usable form is to look at the velocity vector for each point in the scene, or image. The velocity vector of the motion of the points in the image is called the optical ow. The geometry of the scene and the camera projection places constraints on the set of possible optical ows. The 3-d recovery problem in this instance is the determination of the
63
scene at time t given the image at time t and the optical ow at time t. This problem is often called `shape from motion'. More information can be gained about the scene if one knows the optical ow at a point in time as well as just the image. In practice, the calculations would be done with a computer. In all but the simplest situations, the equations are exceedingly complicated and do not have a simple solution which can be obtained on paper. A computer program might also take into account other attributes for the image, such as colour, texture, shading, shadows, or statistical data.
Shape from motion.
The motion of a rigid body in in R 3 is given by transformations of R 3 in the Euclidean group E (3). This is
where M 2 O(3) and a 2 R 3 . Suppose that these vary with time t 2 R , such that M = identity, and a = 0 at t = 0. Clearly, a curve in the Euclidean group E (3) is equivalent to a curve M (t) in O(3) and a curve a(t) in R 3 . The curve through point x is c(t) = M (t)x + a(t), which has tangent dc (0) = dM (0) x + da (0): dt dt dt De ne the matrix and the vector
x 7! Mx + a
M (0) = dd t a (0): =d dt
As is a tangent to O(3) at e it is an antisymmetric matrix. As there is a curve in O(3) with any antisymmetric matrix as tangent, can be any antisymmetric matrix. Likewise, is tangent to R 3 and can be any vector. Thus the vector elds which can arise as velocity vector elds for rigid body motions are v: R 3 ! R 3
Exercise 1.
v: x 7! x + :
0 0 ! ;! 1 3 2 @ (1) Write = ;!3 0 !1 A. Write out x in components

!2 ;!1 0 and show it is the vector cross product of ! and x.
64
The scene is a surface in R 3 , the plane (an a ne subset) given by
Z = pX + qY + r and the camera is the projection : R 3 ! R 2 given by (X Y Z ) 7! (X Y ):

(2) The points in the plane move by a rigid body motion depending on a parameter t, time. Explain why the plane remains a plane for all times t. It is assumed that the plane can always be described by Z = pX + qY + r. (3) Does this assumption place any restriction on the rigid body motions? The points in the plane can be parameterised by the corresponding points in the optical image, R 2 , by a map : R 2 ! R 3 , so that ( (X Y )) = (X Y ): (4) Write an explicit formula for . As the points in the plane move, so do the points in the optical image. This is given by (X Y ) ! 7 ;M (X Y ) + a : (5) Explain why this is the correct formula. Di erentiate this expression with respect to t, at t = 0, assuming as above the M (0) = identity, and a(0) = 0, and obtain a vector eld w on R 2 , the optical ow. As a special case, you should get for r = 0 0 X 1 !3 ;!2 @ Y A : (X Y ) 7! ( 1 2 ) + ;0 !3 0 !1 pX + qY (6) Explain why it should be impossible to determine r from the optical image or the optical ow. Why does r appear in your formula?
65
Now make the simplifying assumption that r = 0 for all time. From measuring the optical ow you can determine the parameters A, B , C , D, E , F in an optical ow
A B (X Y ) 7! (E F ) + C D
X : Y
(7) Look at the four diagrams of an optical ow. Give parameters which give formulae for these optical ows. (8) Express A B C D E F in terms of p q and !. Show that these equations are solved by 1 arg(S ) ; 1 arg(2! ; R ; iT ) !1 + i!2 = k exp i 4 + 2 3 2 1 S exp i ; 1 arg(S ) + 1 arg(2! ; R ; iT ) p + iq = k 3 4 2 2 where
!3 = 1 2 R
jSj ; T
2
T =A+D R=C;B S = (A ; D) + i(B + C ) k is indeterminate, and there are two solutions, , for each choice of k. (9) Interpret R, S , T in the four diagrams of optical ow. (10) Suppose the mapping of the surface to the optical image had not been non-singular. Would you expect the optical ow to be a smooth vector eld?

Differential Geometry, John W.barrett

Uploaded by

Copyright:

Available Formats

Differential Geometry, John W.barrett

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Differential Geometry, John W.barrett

Uploaded by

Copyright:

Available Formats

DIFFERENTIAL GEOMETRY

Vector spaces. A vector space is a set V , with two operations: addi-

1. Vector space geometry

The general linear group. Consider the set of all isomorphisms

L(a) ; L(b) = L(c) ; L(d):

Solution to exercise 1. There are two strategies for this

since TAb is a linear subspace. Hence

df ( x + y) = f ( x + y + (1 ; ; )0) ; f (0) = f (x) + f (y) + (1 ; ; )f (0) ; f (0) = df (x) + df (y)

Exercise 3. Let a b c 2 R 2 be three points which are not collinear

d(f (x) f (y)) = d(x y)

d(x y) = d(f (x) f (y)) = 0

d(x p) = j jd(x y) and d(y p) = j jd(x y):

If f is an isometry, then it follows that

The Euclidean group. The set of all isometries of n-dimensional

Orthogonal group. An orthogonal transformation of R n is a linear

map L: R n ! R n which preserves distances to the origin, i.e.

The de nition. The projective line P 1 is de ned to be the set of

b , then the transformation is formation of P 1. If A = a c d

Inhomogeneous coordinates. As noted above, the homogeneous co-

Equations in projective geometry.

au2 + buv + cv2 + du + ev + f = 0

Q(x y z) = ax2 + bxy + cy2 + dxz + eyz + fz2 = 0:

Exercise 2. Explain how the equation for a line in R 2

Functions of one variable. De nition. A function f : R ! R is smooth if f and each derivative dn f

Functions of several variables.

! R m is a function, then it has m component functions

Exercise 2. Work out the partial derivatives for the components of a

The general case. If g: R n ! R m is a smooth function of x and f : R m ! R k a smooth function of y, then

Exercise 5. Consider the functions

g :U ! R 2 U = f(r )jr > 0g (r ) 7! (r cos r sin ) f :R 2 ! R 2 (x y) 7! (x2 ; y2 xy)

Exercise 6. Suppose f : R n ! R n is a di eomorphism, with inverse g. Solution to exercise 1.

f (x) ; f (0) = lim jxj = lim x!0 x x

f (x + h) ; f (x) = dfx (h) + jhj (x + h)

; dfx (h) ; L(h) = jhj (x + h) ; 0(x + h) :

Consider h = ty for a xed vector y and t ! 0 2 R . Then this becomes

Proof. Given h 2 R k , de ne a sequence of vectors

Summing this equation over n gives

k X @fi (x + cn ): fi (x + h) ; fi(x) = hn @x n n=1

X hn @fi n ) ; @fi (x) ( x + c jhj @xn @xn

Open subsets of Examples.

. An open subset P of R n is de ned to be a

g(x y) = (ex cos y ex sin y):

Exercise 1. Calculate the determinant of the matrix of partial deriva-

Whitney's theorem. Whitney showed that a generic smooth func-

Exercise 3. List a number of qualitatively di erent ways that the

Exercise 4. Consider the stationary points of the function

Examples of surfaces. Some examples of surfaces have already been

Functions de ned on subsets of R n .

De nition of a manifold. A set X

Terminology. The open subsets U

The graph of a function

Example. The upper hemisphere U n

This is the graph of the function B1 (0) ! R de ned by

Spheres. The n-dimensional sphere S n of all points satisfying

is de ned as the set

12. Discrete Quotients Rm, N R n and f : M N a smooth map. The maps

Lemma. If f : M ! N is an open mapping and is onto, then N has

Exercise 1. Show that the only points identi ed in the map S n ! Qn

13. The tangent space A curve is a map c: I R k , where I