Afne connections, midpoint formation, and point reection

Anders Kock University of Aarhus

Preface
It is a striking fact that differential calculus exists not only in analysis (based on the real numbers R and limits therein), but also in algebraic geometry, where no limit processes are available. In algebraic geometry, one rather uses the idea of nilpotent elements in the afne line R; they act as innitesimals. (Recall that an element x in a ring R is called nilpotent if x k = 0 for suitable non-negative integer k.) Synthetic differential geometry (SDG) is an axiomatic theory, based on such nilpotent innitesimals. It can be proved, via topos theory, that the axiomatics covers both the differential-geometric notions of algebraic geometry and those of calculus. I shall illustrate this synthetic method, by presenting its application to three particular types of differential-geometric structure, namely that of afne connection, midpoint formation, and point reection (geodesic symmetry). I shall not go much into the foundations of SDG, whose core is the so-called KL1 axiom scheme. This is a very strong kind of axiomatics; in fact, a salient feature of it is: it is inconsistent if you allow yourself the luxury of reasoning with so-called classical logic, i.e. use the law of excluded middle, proof by contradiction, etc. Rather, in SDG, one uses a weaker kind of logic, often called constructive or intuitionist. Note the evident logical fact that there is a trade-off: with a weaker logic, stronger axiom systems become consistent. For the SDG axiomatics, it follows for instance that any function from the number line to itself is innitely often differentiable (smooth); a very useful simplifying feature in differential geometry but incompatible with the law of excluded
Expanded version of [8]. The notion of midpoint formation considered in loc.cit. has been generalized, so that a notion of point reection can be considered, together with their interdependence, cf. Theorem 4.2 and Figure (4.1) below. Also, some proofs have been supplied, using Christoffel symbols for afne connections. 1 for Kock-Lawvere

middle, which allows you to construct the non-smooth function f (x) = 1 if x = 0, . 0 if not

Nilpotents, and neighbours

Nilpotent elements on the number line serve as innitesimals2 , in a sense which is forbidden when the number line is R. Nilpotent innitesimals come in a precise hierachy, since xk = 0 implies x k+1 = 0.

The method of SDG combines the nilpotency ideas from algebraic geometry with category theory, and categorical logic: category theory has provided a sense by which reasoning in (constructive) naive set theory is sound for geometric reasoning. So the following is formulated in such naive set theoretic terms. We plunge directly into the geometry of innitesimals (in the nilpotency sense): let us denote by D R the set of x R with x2 = 0 (the rst order innitesimals), more generally, let Dk R be the set of kth order innitesimals, meaning the set of x R with x k+1 = 0. (So D = D1 .) The basic instance of the KL axiom scheme says that any map Dk R extends uniquely to a polynomial map R R of degree k. Thus, given any map f : R R, the restriction of f to Dk extends uniquely to a polynomial map of degree k, the kth Taylor polynomial of f at 0. For x and y in R, we say that x and y are kth order neighbours if x y Dk , and we write x k y. It is clear that k is a reexive and symmetric relation. It is not transitive. For instance, if x D and y D, then x + y D2 , by binomial expansion of ( x + y)3 ; but we cannot conclude x + y D. So x 1 y and y 1 z imply x 2 z, and similarly for higher k. We now turn to the (rst order) neighbour relations in the coordinate plane R2 . It is, in analogy with the 1-dimensional case, dened in terms of a subset D (2) R2 ; we put
2 2 D (2) = {( x1 , x2 ) R R | x1 = 0, x2 = 0, x1 x2 = 0}.

So D (2) D D. We dene the rst order neighbour relation (or 1 ) by putting x y if x y D (2), where x = ( x1 , x2 ) and y = (y1 , y2 ). Similarly for D (n) Rn , and the resulting rst order neighbour relation on the higher coordinate vector spaces Rn . If x D (n), we have B( x, x ) = 0 for any bilinear B : Rn Rn Rm , and if B is furthermore symmetric, we therefore also have the useful B( x + y, x + y) = 2B( x, y)
2 they

(1.1)

are not to be compared to the innitesimals of non-standard analysis

for x and y in D (n). There is also a kth order neighbour relation k on Rn , dened in a completely analogous manner from the set Dk (n) := {( x1 , . . . , xn ) Rn | the product of any k + 1 of the xi s is 0}, namely x k y if x y Dk (n). A higher dimensional version of the KL axiom scheme says that any map D (n) R extends uniquely to an afne map. More generally, any map Dk (n) R extends uniquely to a poynomial map Rn R of degree k. The codomain R here may be replaced by any other nite dimensional vector space. In particular, if a map D (n) Rm takes 0 to 0, it extends uniquely to a linear map Rn Rm . It is then easy to prove that if a map : D (n) D (n) Rm vanishes on the two axes, i.e. if (d, 0) = 0 = (0, d) for all d D (n), then extends uniquely to a bilinear map Rn Rn Rm ; and this bilinear map is symmetric iff itself is so. The following (cf. [7], Proposition 1.5.1) is another consequence of the KL axiom scheme: Theorem 1.1 Any map f : Rn Rm preserves the kth order neighbour relation, x k y implies f ( x ) k f ( y ).

Proof sketch for n = 2, m = 1, for the rst order neighbour relation 1 . It sufces to see that x 1 0 implies f ( x ) 1 f (0), i.e to prove that x D (2) implies f ( x ) f (0) D. Now from a suitable version of the KL axiom scheme follows that on D (2), f agrees with a unique afne function T1 f : R2 R, so for x D (2), f ( x ) f (0) = a1 x1 + a2 x2 . Squaring the right hand side here yields 0, since not only x1 D and x2 D, but also x1 x2 = 0. So f ( x ) f (0) D. From the Theorem follows that the relation k on Rn is coordinate free, i.e. is a truly geometric notion: any re-coordinatization of Rn (by any map, not just by a linear or afne one) preserves the relation k . For suitable denition of what an open subsets of Rn is, and for a suitable denition of n-dimensional manifold (= something that locally can be coordinatized with open subsets of Rn ), one concludes that on any manifold, there are canonical reexive symmetric relations k : they may be dened in terms of a local coordinatization, but, by the Theorem, are independent of the coordinatization chosen. Any map between manifolds preserves the relations k . We shall mainly be interested in the rst order neighbour relation 1 , (also written just ). In Sections 3 and 4, we study aspects of the second order neighbour relation 2 . 3

So for a manifold M, we have a subset M(1) M M, the rst neighbourhood of the diagonal, consisting of ( x, y) M M with x y. It was in terms of this scheme M(1) that algebraic geometers in the 1950s gave nilpotent innitesimals a rigourous role in geometry. Note that for M = Rn , we have M(1) M D (n), by the map ( x, y) ( x, x y). = Let us consider some notions from innitesimal geometry which can be expressed in terms of the rst order neighbour relation on an arbitrary manifold M. Given three points x, y, z in M. If x y and x z we call the triple ( x, y, z) a 2-whisker at x (sometimes: an innitesimal 2-whisker, for emphasis); since is not transitive, we cannot in general conclude that y z; if y happens to be z, we call the triple ( x, y, z) a 2-simplex (sometimes an innitesimal 2-simplex). Similarly for k-whiskers and k-simplices. A k-simplex is thus a k + 1-tuple of mutual neighbour points. The k-simplices form, as k ranges, a simplicial complex, which in fact contains the information of differential forms, and the de Rham complex of M, see [2], [6], [1], [7]. (When we say that ( x0 , x1 , . . . , xk ) is a k-whisker, we mean to say that it is a k-whisker at x0 , i.e. that x0 xi for all i = 1, . . . , k. On the other hand, in a simplex, none of the points have a special status.) Given a k-whisker ( x0 , . . . , xk ) in M. If U is an open subset of M containing x0 , it will also contain the other xi s, and if U is coordinatized by Rn , we may use coordinates to dene the afne combination

i =0

ti xi ,

(1.2)

(where ti = 1; recall that this is the condition that a linear combination deserves the name of afne combination). The afne combination (1.2) can again be proved to belong to U, and thus it denes a point in M. The point thus obtained has in general not a good geometric signicance, since it will depend on the coordinatization chosen. However (cf. [5], [7] 2.1), it does have geometric signicance, if the whisker is a simplex: Theorem 1.2 Let ( x0 , . . . , xk ) be a k-simplex in M. Then the afne combination (1.2) is independent of the coordinatization used to dene it. All the points that arise in this way are mutual neigbours. And any map to another manifold M preserves such combinations. Proof sketch. This is much in the spirit of the proof of Theorem 1.1: it sufces to see that any map Rn Rm (not just a linear or afne one) preserves afne combinations of mutual neighbour points. This follows by considering a suitable rst Taylor polynomial of f (expand from x0 ), and using the following purely algebraic fact: If x1 , . . . , x k are in D (n), then any linear combination of them will again be in D (n) provided the xi s are mutual neighbours.

Examples. If x y in a manifold (so they form a 1-simplex), we have the afne combinations midpoint of x and y, and reection of y in x,
1 2x 1 + 2y

and

2x y,

respectively. If x, y, z form a 2-simplex, we may form the afne combination u := y x + z; geometrically, it means completing the simplex into a parallelogram by adjoining the point u. Here is the relevant picture: $$q $$$ z q$ q $ y $$$ q x $$ u = yx+z

(1.3) (All four points here are neighbours, not just those that are connected by lines in the gure.) The u thus constructed will be a neighbour of each of the three given points. Remark. If x, y, z and u are as above, and if x, y, and z belong to a subset S M given as the zero set of a function f : M R, then so does u = y x + z; for, f preserves this afne combination.

Afne connections

If x, y, z form a 2-whisker at x (so x y and x z), we cannot canonically form a parallelogram as in (1.3); rather, parallelogram formation is an added structure: Denition 2.1 An afne connection on a manifold M is a law which to any 2whisker x, y, z in M associates a point u = ( x, y, z) M, subject to the conditions ( x, x, z) = z, ( x, y, x ) = y. (2.1)

(Cf. [3].) It can be veried (cf. [7] 2.3) that several other laws follow; in a more abstract combinatorial situation than manifolds, these laws should probably be postulated. Some of the laws are: for any 2-whisker ( x, y, z) ( x, y, z) y and ( x, y, z) z (y, x, ( x, y, z)) = z One will not in general have or require the symmetry condition ( x, y, z) = ( x, z, y); (2.4) (2.2) (2.3)

nor do we in general have, for 2-simplices x, y, z, that ( x, y, z) = y x + z. (2.5)

The laws (2.4) and (2.5) are in fact equivalent, and afne connections satisfying either are called symmetric or torsion free. We return to the torsion of an afne connection below. If x, y, z, u are four points in M such that ( x, y, z) is a 2-whisker at x, the statement u = ( x, y, z) can be rendered by a diagram q u = ( x, y, z) $$$ q $$$$ y $ q x $ $ z q$$ (2.6) The gure3 is meant to indicate that the data of provides a way of closing a whisker ( x, y, z) into a parallellogram (one may say that provides a notion of innitesimal parallelogram); but note that is not required to be symmetric in y and z, which is why we in the gure use different signatures for the line segments connecting x to y and to z, respectively. Here, a line segment (whether single or double) indicates that the points connected by the line segment are neighbours. If x, y, z, u are four points in M that come about in the way described, we say that the 4-tuple form a -parallelogram . The fact that we in the picture did not make the four line segments oriented contains some symmetry assertions, which can be proved by working in a coordinatized situation; namely that the 4-group Z2 Z2 acts on the set of -parallelograms; so for instance (u, z, y, x ) is a -parallelogram, equivalently (( x, y, z), z, y) = x. On the other hand (( x, y, z), y, z) x, (2.7) but it will not in general be equal to x; its discrepancy from being x is an expression of the torsion of . Even when y z (so x, y, z form a simplex), the left hand side of (2.7) need not be x. Rather, we may dene the torsion of to be the law b which to any 2-simplex x, y, z associates (( x, y, z), y, z). Then b( x, y, z) = x for all simplices iff is symmetric, cf. [7] Proposition 2.3.1. There is also a notion of curvature of : Let M be a manifold equipped with an afne connection . Note that ( x, y, ) takes any neighbour v of x into a neighbour of y (-parallel transport of v from x to y). If now x, y, z form an innitesimal 2-simplex, we may successively make three transports,
3 Note the difference between this gure and the gure (1.3), in which y and z are assumed to be neighbours, and where the parallelogram canonically may be formed.

from x to y, then from y to z, and nally from z back to x. Thus the 2-simplex x, y, z gives rise to a permutation of the set of neighbour points v of x, and the connection is at (curvature free) if all permutations arising this way are the identity permutation. More generally, the curvature r of is dened as the law, which to a innitesimal 2-simplex x, y, z provides the permutation of the neighbours of x just described. (In the terminology of [7], r is a group-bundle valued combinatorial 2-form.) We give two examples of afne connections on the unit sphere S. Example 1. The unit sphere S sits inside Euclidean 3-space, S E. Since E is in particular an afne space, we may for any three points x, y, z in it form y x + z E. For x, y, z in S, the point y x + z will in general be outside S; if x, y, z are mutual neighbours, however, y x + z will be in S, cf. Remark at the end of Section 1. What if x, y, z only form an innitesimal 2-whisker? Then we cannot expect y x + z to be in S, but we may project it down to S and dene ( x, y, z) S to be the point, where the half line from the center of S to y x + z meets S. If S is the surface of the earth, this just means that ( x, y, z) which is vertically below y x + z. This afne connection is evidently symmetric in y and z, so is torsion free; it does, however, have curvature, which one can see from the integrated version (holonomy) of the connection, i.e. the parallel transport (according to ) along curves on the sphere: for instance, transporting along a spherical triangle, whose sides each are 90o , will provide a permutation of the neigbour points of any vertex, namely a rotation by 90o (make a picture!). This connection is the Riemann- or Levi-Civita connection on sphere. Example 2. (This example does not work on the whole sphere, only away from the two poles.) Given x, y and z with x z (x y is presently not relevant). Since x and z are quite close, we can uniquely describe z in a rectangular twodimensional coordinate system at x with coordinate axes pointing East and North. Now take ( x, y, z) to be that point near y, which in the East-North coordinate system at y has same coordinates as the ones obtained for z in the coordinate system that we considered at x. The description of this afne connection is asymmetric in y and z, and it is indeed easy to calculate that it has torsion ([7], Section 2.4). It has no curvature. Connections constructed in a similar way also occur in materials science: for a crystalline substance, one may attach a coordinate system at each point, by using the crystalline structure to dene directions (call them East and North and Up, say). The torsion for a connection constructed from such coordinate systems is a measure for the imperfection of the crystal lattice (dislocations), see [10], [4] and the references therein. For calculations, and even for communication, one usually needs coordinates. Coordinate expressions for an afne connection are the Christoffel symbols. Let be an afne connection on an n-dimensional manifold M; assume that we have identied some open subset of M with some open subset of Rn .

If x y in this subset, y = x + d for some d D (n), by denition of the neighbour relation. So a whisker ( x, y, z) at x may be written ( x, x + d1 , x + d2 ) with (d1 , d2 ) D (n) D (n). Dene, for xed x, the function : D (n) D (n) Rn by the rule (d1 , d2 ) = ( x, x + d1 , x + d2 ) x + d1 + d2 . (To record the dependence of on the point x, we may write ( x; d1 , d2 ).) Thus, measures the discrepancy between and the canonical afne connection 0 in the afne space Rn . From the law ( x, x, z) = z follows (0, d2 ) = 0, and similarly ( x, y, x ) = y gives (d1 , 0) = 0. From the discussion prior to Theorem 1.1 follows that extends uniquely to a bilinear map : Rn Rn Rn , which is the Christoffel symbol for at the point x relative to the coordinatization assumed around x. It is clear that is a symmetric (= torsion free) afne connection iff the Christoffel symbols ( x; , ) are symmetric bilinear maps, for all x M. We can rephrase the relation between and the Christoffel symbols (in a coordinatized situation) by ( x, x + d1 , x + d2 ) = x + d1 + d2 + ( x; d1 , d2 ) where d1 and d2 are in D (n). (2.8)

Second order notions; midpoint formation

We describe a geometric notion of midpoint formation structure which can be used to construct torsion free afne connections. Let M(2) M M denote the set of pairs ( x, u) of second order neighbours; M(2) is the second neighbourhood of the diagonal, in analogy with the rst neighbourhood M(1) described in Section 1. We have M(1) M(2) . Recall that for x 1 y in M, we have canonically the afne combination 1 x + 1 y, the midpoint formation for rst order neighbours; it denes a map 2 2 M(1) M. Denition 3.1 A midpoint formation structure on M is a map : M(2) M which extends the midpoint formation M(1) M for pairs of rst order neighbour points.
1 Thus, ( x, u) is dened whenever x 2 u; and ( x, u) = 1 x + 2 u whenever 2 x 1 u. It can be proved that such is automatically symmetric, ( x, u) = (u, x ). (The proof follows the same lines as the symmetry of in the proof of Theorem 4.2 below.) It can also be proved that ( x, u) 2 x and 2 u. Beware that ( x, u) is midpoint of x and u, (where x 2 u) does not imply x 1 ( x, u) 1 u; in fact either of these will hold only if x is already 1 u.

Theorem 3.2 Any midpoint formation structure on M gives canonically rise to a symmetric afne connection on M. 8

Proof. Given , and given an innitesimal 2-whisker ( x, y, z). Since x 1 y, we may form the afne combination 2y x (reection of x in y), and it is still a rst order neigbour of x. Similarly for 2z x. So (2y x ) 2 (2z x ), and so we may form (2y x, 2z x ), and we dene ( x, y, z) := (2y x, 2z x ). The relevant picture is here:

2z x ( x, y, z) $. $$ $ z .$$ $$$ 2y x $$ .$ $$$ y $$$ . x It is symmetric in y and z, by the symmetry of . Also, if y = x, we get
1 1 ( x, x, z) = ( x, 2z x ) = 2 x + 2 (2z x ),

since x 1 2z x and extends the canonical midpoint formation for rst order neighbours. But this equals z, by evident equations for afne combinations. This proves the rst equation in (2.1), and the second one then follows by symmetry. Theorem 3.3 Any symmetric afne connection on M gives canonically rise to a midpoint formation . This is less evident. First it requires an interpolation axiom, consistent with the KL axiomatics, namely that any D2 (n) may be written d1 + d2 for suitable d1 and d2 in D (n). This implies that for any manifold M, and any x 2 u in M, we may interpolate a y, in the sense that x 1 y 1 u. If M is equipped with an afne connection , we may therefore also for x 2 u nd a -parallelogram ( x, y, z, u) (take z = (y, x, u)). Now given a -parallelogram ( x, y, z, u), and given a scalar t R. Since x y, we may form the afne combination xt := (1 t) x + ty, and similarly, given s R, we may form zs := (1 s) x + sz; both these points are x, and so we may form the point ( x, (1 t) x + ty, (1 s) x + sz). See picture: u $$q $$$ z q$ $ $ q $$ zs $$q y q $$$ t y q x $ (3.1) 9

The picture suggests that for s = t, we could dene (1 t) x + tu as ( x, yt , zt ) (this will certainly be correct in an afne space). Thus we have described a candidate for this afne combination of x and u, even though x and u are in general only second order neighbours. In particular, taking t = 1 , we would 2 get a candidate for the midpoint ( x, u). The problem with this denition is of course that it depends not only on x and u (and on t, of course), but also (seemingly) on the interpolating y and z. By working in coordinates, using the Christoffel symbol = ( x; , ) for at x, we shall prove that this dependence is only apparent for symmetric . We may write y = x = d1 and u = y + d2 = x + d1 + d2 . It is not in general true that z = x + d2 , but rather z = x + d2 ( d1 , d2 ). (3.2)

For, since ( x, y, ) maps the the set of neighbours of x bijectively to the set of neighbours of y, it sufces to see that ( x, y, x + d2 (d1 , d2 )) = x + d1 + d2 . Calculating the left hand side with yields (since y x = d1 ) x + d1 + d2 (d1 , d2 ) + (d1 , d2 (d1 , d2 ))

= x + d1 + d2 (d1 , d2 ) + (d1 , d2 ) (d1 , (d1 , d2 )),

using bilinearity of , so we end up with x + d1 + d2 (d1 , (d1 , d2 )). Now the last term vanishes: d1 here occurs in a quadratic fashion, and d1 D (n); more precisely, since (, (, r )) is bilinear, it vanishes if a vector from D (n) is put in both the empty slots. Thus we nally end up with x + d1 + d2 , which is u. Substituting the expression (3.2) for z gives ( x, (1 t) x + ty, (1 t) x tz) = ( x, x + td1 , x + t(d2 (d1 , d2 )); let us calculate this using ; we get that it equals x + td1 + td2 t(d1 , d2 ) + (td1 , td2 t(d1 , d2 )). As before, the nested appearance of vanishes since d1 D (n), and we are left with x + td1 + td2 t(d1 , d2 ) + (td1 , td2 ). (3.3) Now we use that was assumed symmetric, so that is a symmetric bilinear form. Then by (1.1), (d1 , d2 ) = 1 (d1 + d2 ), and similarly for (td1 , td2 ). Thus 2 the expression (3.3) only depends on x and d1 + d2 , that is, on x and u only. This proves the desired independence of d1 and d2 individually. Let us show that one gets the symmetric afne connection back from the midpoint formation to which it gives rise. Let be the afne connection constructed from , so for a whisker x, y, z at x, use x as interpolating point between 2y x and 2z x; so
1 ( x, y, z) = (2y x, 2z x ) = ( x, 1 x + 2 (2y x ), 1 x + 1 (2z x )), 2 2 2

10

but 1 x + 1 (2y x ) = y and 1 x + 1 (2z x ) = z, by purely afne calculations; 2 2 2 2 so we get ( x, y, z) back. In [5], it is shown how a Riemannian metric geometrically gives rise to a midpoint formation (out of which, in turn, the Levi-Civita afne connection may be constructed, by the process given by the Theorem).

Point reection (geodesic symmetry)

For a pair of rst order neighbours, x 1 y, on a manifold M, one has canonically the point reection of y in x, namely the afne combination 2x y; it thus denes a map M(1) M. Denition 4.1 A point reection on a manifold M is a map : M(2) M, which extends the point reection M(1) M for pairs of rst order neighbour points. We write x y for the values of this map, x y is the reection of y in x. It should be thought of as an innitesimal geodesic symmetry. Problem: One would like to investigate the conditions when a point reection structure satises the equation,

( x y) ( x z) = x (y z)
for symmetric spaces in the sense of Loos (cf. [9]), whenever x, y, z are mutual 2-neighbours. The equation does not immediately make sense in our context, since we cannot assert that x 2 (y z). However, it can be be proved that x 3 (y z), and one can prove that : M(2) M extends uniquely to a map : M(3) M satisfying x ( x u) = u for all u 3 x, and with this extended , the expression x (y z) makes sense. Theorem 4.2 Any point reection structure on M gives canonically rise to a symmetric afne connection on M. Proof. Given , and given an innitesimal 2-whisker ( x, y, z). Since x 1 z, we may form the afne combination z := 2x z (reection of z in x), and it is still 1 a rst order neigbour of x. Also, we may form m := 2 x + 1 y, likewise a rst 2 order neighbour of x, so m 2 z , and therefore we may form m z . We dene ( x, y, z) to be this point. The relevant picture is here: ( x, y, z) $ $ q 7 $$$ 7 z q$ 7 7 $q y q $$$ 7 q x $$ m 7 7 7 7 q z

(4.1)

11

It is easy to see that ( x, x, z) = z and ( x, y, x ) = y, so is indeed an afne connection. The fact that is symmetric is not immediate from the denition; we prove it by considering an coordinatized situation, so identify an open neighbourhood of x, y, z with an open subset of Rn , so y = x + d1 , z = x + d2 for d1 and d2 in D (n). Now, since extends the canonical point reection for rst order neighbours, it is by KL of the form x u = 2x u + Q x (u x ) where Q x : Rn Rn is a quadratic map, i.e. Q x (v) = ( x; v, v) for some (unique) symmetric bilinear map ( x; , ) : Rn Rn Rn . The symmetry of the constructed will now follow by proving that the Christoffel symbols for are the s. Here is the calculation. Note that m = x + 1 d1 , and that z = x d2 - Also, 2 1 z m = ( x d2 ) m = 2 d1 d2 , so ( x, y, z) = m (2x z) = 2m ( x d2 ) + ( x + 1 d1 ; 1 d1 d2 , 1 d1 d2 ). 2 2 2 Now the terms before the term yield by simple additive calculation x + d1 + d2 = y x + z. The term may be rewritten by (1.1) using symmetry and 1 bilinearity of ( x + 1 d1 ; , ); it gives ( x + 2 d1 ; d1 , d2 ). Here, d1 appears as an 2 argument in a linear position (after the semicolon), and then a Taylor expansion argument gives that the x + 1 d1 in front of the semicolon may be replaced 2 by x. (This is because d1 is in D (n); for a precise formulation of this Taylor principle, see [7] (1.4.2).) Putting things together, we thus have ( x, y, z) = y x + z + ( x; d1 , d2 ), proving that ( x; , ) is indeed the Christoffel symbol at x for . This proves the symmetry. Theorem 4.3 Any symmetric afne connection on M gives canonically rise to a point reection structure (geodesic symmetry) . Just as Theorem 4.3, the construction depends on interpolating a -parallelogram x, y, z, u between x and u for x 2 u (see gure (3.1): one then puts x u := ( x, 2x y, 2x z), and the proof that this is independent of the choice of the interpolation is as for Theorem 4.3. Again, the constructions are inverse of each other. We may summarize the results of the last two sections in Theorem 4.4 On any manifold M, there are canonical bijective correspondences between the following three kinds of geometric structure: symmetric afne connections midpoint formations : M(2) M point reection structures : M(2) M.

12

Any map f : M M between manifolds preserves the neighbour relations 1 and 2 (Theorem 1.1); therefore if M and M are equipped with afne connections and , respectively, it makes sense to ask whether f is connection preserving f ( ( x, y, z)) = ( f ( x ), f (y), f (z)), for any x 1 y, x 1 z in M . If M and M are equipped with midpoint formation structures and , respectively, it makes sense to ask whether f preserves midpoint formation, f ( ( x, u)) = ( f ( x ), f (u)) for pairs of second order neighbours x 2 u in M . Similarly if M and M are equipped with point reection structures. Symmetric afne connections, midpoint formation structures, and point reection structures correspond, by Theorem 4.4; the correspondences are constructed using afne combinations of rst order neighbour points, preserved by any f by Theorem 1.2. Therefore it follows that the assertions f is connection preserving , f is midpoint preserving, and f is point-reection preserving are equivalent (for symmetric afne connections, and the corresponding and ). Such maps f : M M deserve the name geodesic maps. In particular, the number line R (being an afne space) carries canonical structures , , (which correspond to each other), namely ( x, y, z) = y x + z,
1 ( x, u) = 2 x + 1 u, 2

x u = 2x u,

(in fact without any restrictions like x 2 u). A map f : R M into a manifold M equipped with a symmetric afne connection deserves the name (parametrized) geodesic curve if f is geodesic in the general sense just dened. This is equivalent to f preserving midpoint formation or point reection. In particular, f is geodesic if f ( x + 2d) = f ( x + d) f ( x ) for d D2 . A subset C M deserves the name unparametrized curve if there is an embedding f : R M mapping bijectively R onto C. In this case, C deserves the name geodesic if x 2 u in C implies x u C, in other words, if C is stable under point reection. Acknowledgement. Most of the content of Sections 1-3 of the present article was presented as an invited lecture at the 15th IAPR International Conference, DCGI 2009 (Montr al September/October 2009), cf. [8]. e

References
[1] L. Breen and W. Messing, Combinatorial differential forms, Advances in Math. 164 (2001), 203-282. [2] A. Kock, Synthetic Differential Geometry, London Math. Soc. Lecture Notes Series 51, Cambridge University Press 1981 (Second edition: London Math. Soc. Lecture Notes Series 333, Cambridge University Press 2006). 13

[3] A. Kock, A combinatorial theory of connections, in Mathematical Applications of Category Theory, Proceedings 1983 (ed. J. Gray), AMS Contemporary Math. 30 (1984), 132-144. [4] A. Kock, Introduction to Synthetic Differential Geometry, and a Synthetic Theory of Dislocations, in Categories in Continuum Physics, Proceedings Buffalo 1982 (ed. F.W. Lawvere and S. Schanuel), Springer Lecture Notes Vol. 1174 (1986). [5] A. Kock, Geometric construction of the Levi-Civita parallelism, Theory and Applications of Categories 4 (1998), 195-207. [6] A. Kock, Differential forms as innitesimal cochains, Journ. Pure Appl. Alg. 154 (2000), 257-264. [7] A. Kock, Synthetic Geometry of Manifolds, Cambridge Tracts in Mathematics Vol. 180 (2010). [8] A. Kock, Afne connections, and midpoint formation, in Discrete Geometry for Computer Imagery, Proceedings Montr al 2009, Springer LNCS 5810, 13e 21. [9] O. Loos, Symmetric Spaces, 1-2, Benjamin (1969). [10] W. Noll, Materially uniform simple bodies with inhomogeneities, Arch. Rat. Mech. Anal. 27 (1967), 1-32.

14