ASC Report No.

39/2012

Congruences of Convex Algebras

A. Sokolova, H. Woracek

Institute for Analysis and Scientic Computing Vienna University of Technology TU Wien www.asc.tuwien.ac.at ISBN 978-3-902627-05-6

Most recent ASC Reports

38/2012 37/2012 M. Langer, H. Woracek Stability of the derivative of a canonical product F. Bruckner, C. Vogler, B. Bergmair, T. Huber, M. Fuger, D. Suess, M. Feischl, T. F uhrer, M. Page, D. Praetorius Combining micromagnetism and magnetostatic Maxwell equations for multiscale magnetic simulations M. Bukal, E. Emmrich, and A. J ungel Entropy-stable and entropy-dissipative approximations of a fourth-order quantum diusion equation W. Auzinger and M. Tutz A review of stability and error theory for collocation methods applied to linear boundary value problems F. Bruckner, M. Feischl, T. F uhrer, P. Goldenits, M. Page, D. Praetorius and D. S uss Multiscale Modeling in Micromagnetics: Well-Posedness and Numerical Integration H. Winkler, H. Woracek Symmetry in de Branges almost Pontryagin spaces H. Hofst atter, O. Koch, M. Thalhammer Convergence of split-step generalized-Laguerre-Fourier-Hermite methods for Gross-Pitaevskii equations with rotation term S. Esterhazy and J.M. Melenk An analysis of discretizations of the Helmholtz equation in L2 and in negative norms (extended version) H. Hofst atter, O. Koch An Approximate Eigensolver for Self-Consistent Field Calculations A. Baranov, H. Woracek De Branges theorem on approximation problems of Bernstein type

36/2012

35/2012

34/2012

33/2012 32/2012

31/2012

30/2012 29/2012

Institute for Analysis and Scientic Computing Vienna University of Technology Wiedner Hauptstrae 810 1040 Wien, Austria E-Mail: admin@asc.tuwien.ac.at WWW: http://www.asc.tuwien.ac.at FAX: +43-1-58801-10196 ISBN 978-3-902627-05-6 c Alle Rechte vorbehalten. Nachdruck nur mit Genehmigung des Autors.

TU WIEN

ASC

Congruences of Convex Algebras

Ana Sokolova1 , and Harald Woracek2

Abstract We provide a full description of congruence relations of convex, positive convex, and totally convex algebras. As a consequence of this result we obtain that nitely generated convex (positive convex, totally convex) algebras are nitely presentable. Convex algebras, in particular positive convex algebras, are important in the area of probabilistic systems. They are the Eilenberg-Moore algebras of the subdistribution monad. AMS MSC 2010: Primary 08A30, 52A20, 18C20. Secondary 08A62, 18C05 Keywords: convex algebra, congruence relation, nitely presentable

Introduction

In this paper we present a study of the equational classes CA, PCA, and TCA, of convex, positive convex, and totally convex algebras. We describe all congruence relations of such algebras. Knowing the congruences, we obtain that nitely generated convex (positive convex, totally convex) algebras are nitely presentable. A convex algebra is an algebra (with nonempty carrier set) with an innite set of operations of arbitrary positive arities providing convex combinations of the arguments, which satisfy two axioms (axiom schemes): (1) the projection axiom stating that a convex combination with a single coecient equal to 1 equals the identity map, and (2) the baricenter axiom stating that a convex combination of convex combinations equals the convex combination with suitably multiplied and summed coecients. Positive convex and totally convex algebras are dened in a similar way from larger convex structures (sub-convex combinations for positive convex algebras and linear combinations with coecients whose absolute values are sub-convex for totally convex algebras). The precise denitions and details follow in Section 3. Examples of convex algebras are provided by convex subsets of a vector space over the scalar eld R (a subset of a vector space is convex, if it contains with each two points the whole line segment connecting them): If K is such, then K is a convex algebra with the operations inherited from the vector space. However, these examples do not exhaust the class CA; the major obstacle being possible failure of cancellation laws in general convex algebras. Examples of positive convex algebras are provided by convex subsets of a vector space over R which contain the zero vector. Examples of totally convex algebras are provided by convex subsets of a vector space R which are symmetric around the zero vector. Convex algebras appear in a categorical context. To explain this, e.g. for the PCA-situation, consider the category Vec+ 1 whose objects are regularly ordered
1 University Research

of Salzburg, email: anas@cs.uni-salzburg.at supported by Austrian Science Fund (FWF) grant V00125. 2 Vienna University of Technology, email: harald.woracek@tuwien.ac.at

normed vector spaces over the scalar eld R and morphisms are positive and linear contractions between such spaces. The functor : Vec+ 1 Sets which acts on objects as (V ) := x V | x 1, x 0 , and on morphisms as restriction to (V ), has a left adjoint. It turns out that the algebraic category PCA is the category of Eilenberg-Moore algebras of the monad induced by this adjunction, cf. [Pu84]. Moreover, the monad in question is actually the discrete subprobability distribution monad, hence PCA is the category of Eilenberg-Moore algebras of the subprobability distribution monad [Do06, Do08]. Our aim in this paper is to achieve full understanding of the structure of nitely generated algebras in CA (PCA and TCA). We manage this with Theorem 4.3 and 4.4 below, where we describe the congruences on any polytope in the euclidean space Rn considered as a convex algebra. It is simple to check that, for each n N+ , the free algebra Fn in CA with n generators is given by the standard (n1)-simplex in Rn (a particular polytope). For n = 3, we can picture this algebra as1

(1.1)

Clearly, knowing all congruences of the free algebras Fn , n N+ , is enough to understand all nitely generated algebras in CA. These results on CA can be transferred to PCA and TCA. Therefore, we also achieve full understanding of the structure of nitely generated positive convex or totally convex algebras. Besides its obvious intrinsic interest, our motivation to investigate nitely generated algebras in CA (PCA or TCA) originates in a categorical problem. The probability subdistribution monad arising from the above mentioned adjunction, including the functor , and its Eilenberg-Moore algebras play a crucial role in connection with the axiomatization of trace semantics for probabilistic systems given in [SS11]. There the question arose whether or not each nitely generated algebra in PCA is also nitely presentable. Using the newly established knowledge about congruence relations, we can answer this question armatively, cf. Corollary 5.3. Historically, work on convex algebras can be traced back (at least) to [Se67]. The theory of totally convex algebras (and their analogues allowing innitary operations) started in [PR84], where they were realized to be the Eilenberg-Moore algebras associated with the adjunction induced by the unit ball functor from the category of Banach spaces (with linear contractions) to Sets. A similar treatment of positive convex algebras was given shortly after in [Pu84]. Later on these notions were extensively studied, mainly focussing on the categorical viewpoint
1 Picture

source: http://en.wikipedia.org/wiki/Simplex

and topological questions, see, e.g., [PR85, BK93, Ke98, Pu01a, Pu01b, Pu03] and the references therein. A far reaching generalization, namely the concept of convexity theories, has been developed in a series of papers involving several authors which started with [R o94], and went on (at least) till [R o01]. Previous work which is the closest to our approach is [PR90, Ke99, Ke00], where congruence relations in (innitary or p-) totally convex modules (algebras) are studied. Some parts of our results read similarly and several geometric ideas employed there can also be used in the present setting. In order to prevent confusion concerning terminology, let us note explicitly that in the literature algebras in CA (PCA or TCA) are also called nitely (positively/totally) convex modules. The term nitely thereby refers to the fact that they carry only nitary operations. However, in the present paper we stick to the purely algebraic setting and do not touch upon the possibility of allowing innitary operations. Hence, we omit the prex nitely from the notation. Moreover, we also choose the term algebra over module since it has been used in recent work regarding positive convex algebras [Do06, Do08, SS11]. The structure of the paper is as follows. After the introduction, we recall some notions and facts from convex geometry in Section 2. In Section 3, we present the equational classes of convex, positive convex and totally convex algebras. After collecting some basic facts, we investigate the relationship between CA, PCA, and TCA. Interestingly, it turns out that CA and PCA are in essence just the same, whereas TCA carries a signicantly stronger structure. Section 4 is the core of the paper. There we formulate and prove Theorems 4.3 and 4.4 that describe the congruence relations on a polytope K in euclidean space. It turns out that a congruence on K is fully determined by two ingredients: (1) a family of linear subspaces, describing the congruence classes in the interior of K and in the interior of each of its lower dimensional facets; (2) a graph, describing how the interiors of K and each of its facets are related to the lower dimensional facets forming the respective boundary. Finally, in Section 5, we give the already mentioned application, and show that in each of the algebraic categories corresponding to CA, PCA, and TCA, the notions nitely generated and nitely presentable coincide. Our basic reference concerning terminology and results of universal algebra is the (old but still excellent) book [Gr a68], or the more recent treatment [BS00]. For the (few) notions from category theory which are used in this paper, we refer the reader to [La98]. Our standard reference concerning convex geometry is [Gr u03].

Preliminaries from convex geometry

Basic universal algebra and euclidean topology notions and results will be recalled when they are needed. In this section we explicitly recall some denitions and results from convex geometry. We give even proofs of simple properties, to set the mood for what follows in the paper. We start with the denition of a convex set. Denition 2.1. A subset C of a vector space V over the eld R is convex if for all x, y C and any scalar [0, 1] it holds that x + (1 )y C . Geometrically, this means that C contains, together with each two points, the whole line segment connecting them. 3

Convex sets are mapped by linear functions to convex sets. The following simple property shows that convexity is the same as being closed under arbitrary convex linear combinations. Lemma 2.2. Let C be a subset of a vector space V over R. Then C is convex if and only if
n

pi x i C
i=1

for all n N+ , all xi C , and all pi [0, 1] with i = 1, . . . , n such that n i=1 pi = 1. Proof. It is clear that being closed under convex linear combinations implies convexity. The other direction is proved by induction. Let C be a convex subset n and let xi C , pi [0, 1], for i = 1, . . . , n be given such that i=1 pi = 1. If n = 1 or pi = 1 for some i, the statement is trivial. If n = 2, the statement holds by convexity. Assume that n > 2, p1 = 1, and the statement holds for n 1. Then n n pi pi xi = p1 x1 + (1 p1 ) xi 1 p1 i=1 i=2
pi and i=2 1 p1 = 1. Hence, from the inductive hypothesis n and by convexity also i=1 pi xi C . n n pi i=2 1p1 xi

The following subsets, associated with a nite subset Y of V , play an important role: span Y :=
y Y

y y | y R , y y | y R,
y Y y Y

dir Y := a Y :=
y Y

y = 0 , y = 1 ,
y Y

y y | y R, y y | y [0, 1],
y Y

co Y := co Y :=
y Y

y = 1 ,
y Y

y y | y (0, 1],
y Y

y = 1 .

We refer to co Y as the closed convex hull of Y and co Y as the open convex hull or the interior of co Y . We will see later that this choice of terminology is indeed justied, cf. Lemma 2.5. The linear span span Y is the smallest vector subspace that contains Y . Moreover, we refer to a Y as the ane space generated by Y and dir Y as the directions of a Y . Note that dir Y is a vector subspace. Clearly, for each nite set Y , co Y co Y a Y span Y and dir Y span Y.

If Y contains only one element, then co Y = co Y = a Y = Y and dir Y = {0}. If |Y | 2, then co Y co Y a Y and dir Y = {0}. First, some simple geometric properties of these sets. 4

Lemma 2.3. Let Y be a nite subset of a vector space V over R. Then the following hold: (i) For each z a Y we have a Y = z + dir Y . (ii) For each z a Y , we have dir Y = w z | w a Y . (iii) We have dir Y = w z | z, w a Y . (iv ) And, for every y0 Y , dir Y = span{y y0 | y Y . Proof. Let z a Y be given, and write z = yY y y with Given x a Y , write x = yY y y with yY y = 1. Then xz =
y Y y Y

y = 1.

(y y )y ,

and yY (y y ) = 0. This means that x z dir Y , and we see that the inclusion in item (i) holds. All other inclusions in item (i) - item (iii) follow in the same way. Item (iv ) is straightforward by the denition of dir Y . In the situation that V = Rn some important topological properties hold. These are expressed in the following two lemmas. Lemma 2.4. Let Y be a nite subset of Rn . Then co Y is compact and convex. Moreover, co Y is the closure of co Y. Proof. It is easy to check by the denitions that co Y is convex. Recall that a set is compact if it is closed and bounded. Moreover, continuous functions map a compact set to a compact set. The set = {(y )yY R|Y | | y [0, 1],
y Y

y = 1}

we have y co Y . The elements y tend to y when tends to 0. 5

is compact, the function (y )yY yY y y is continuous, and it maps to co Y . Hence, co Y is compact. Since co Y co Y and co Y is closed, we get Clos(co Y ) co Y . For the opposite inclusion, take an element y co Y . Write Y = {y1 , . . . , yn }, and n n y = i=1 i yi with i [0, 1] and i=1 i = 1. We will construct a sequence of elements in co Y whose limit is y , showing that y Clos(co Y ). Without loss of generality, we may assume that n > 0 and 1 , . . . , k are all coecients that n i yi where (0, n ) and are equal to 0. Then for y = i=1 k i = 1, . . . , k i := i i = k + 1, . . . , n 1 n k i = n

Lemma 2.5. Let Y be a nite subset of Rn with |Y | 2. Then co Y is open considered as a subset of a Y . Proof. Let Y = {y0 , . . . , ym } Rn with m 1. From Lemma 2.3, for each z a Y , a Y = z + span{yk y0 | k = 1, . . . , m}. Let {b1 , . . . , bl } be a maximal linearly independent subset of {yk y0 | k = 1, . . . , m}. Then a Y = z + span{b1 , . . . , bl }. Choose (if necessary) bl+1 , . . . , bn such that {b1 , . . . , bn } is a basis of Rn and consider the norm ||x|| := max |i |, x Rn ,
i=1,...,n

where i are the unique coecients in x = i=1 i bi . For each > 0, the set U = {x Rn | ||x|| < } is open with respect to the euclidean topology in Rn (by equivalence of norms2 ). Thus, for each z a Y , the set (z + U ) a Y is an open subset of a Y . We have
l

(z + U ) a Y = z + {
i=1

i bi | |i | < } ,

where is trivial and follows since {b1 , . . . , bn } is linearly independent. m Assume now that z co Y , and write z = i=0 i yi with i (0, 1). This is always possible since |Y | 2 and hence no i equals 1. Set := 1 min ({i | i = 0, . . . , m} {1 i | i = 0, . . . , m}) . m

Then > 0 and it is not dicult to check that the choice of guarantees that (z + U ) a Y co Y. Hence we have found an open set in a Y containing z and being contained in co Y . Since z co Y was arbitrary, this shows that co Y is open as a subset of a Y . Lemma 2.5 implies an alternative characterization of dir Y . Lemma 2.6. Let Y be a nite subset of Rn . Then we have dir Y = span{y z | y co Y }, and dir Y = span{y2 y1 | y1 , y2 co Y }. Proof. If |Y | = 1, we have dir Y = {0} and co Y = a Y = Y . Hence, the stated equalities hold in this case. Moreover, the second asserted equality will follow immediately once the rst is shown, and the inclusion in the rst one is trivial. Assume that |Y | 2, and let z a Y be given. Since span{y z | y co Y } dir Y and both are vector subspaces, we get that span{y z | y co Y} is a subspace of dir Y . Note that no proper subspace of dir Y contains a nonempty open subset of dir Y (think as an illustration of a line in a plane and
2 n i=1,...,n

z a Y ,

max |i |

i=1

2 i

1 2

n max |i |.
i=1,...,n

a full open disc in this plane). In order to show the needed equality, we will show that span{y z | y co Y } does contain a non-empty open set in dir Y . Consider the translation map Tz : x x z . It is a homeomorphism, i.e., a continuous bijective map whose inverse is also continuous. Homeomorphisms map open subsets onto open subsets. Note that Tz maps a Y onto dir Y . As a consequence, also by Lemma 2.5, Tz (co Y ) is a non-empty open subset of dir Y . Clearly, Tz (co Y ) = {y z | y co Y } span{y z | y co Y }. Let C Rn be convex. A point e C is called an extremal point if e = tx + (1 t)y with x, y C, t (0, 1) x = y = e . Geometrically, this means that e does not lie in the interior of any line segment with endpoints in C . We denote the set of all extremal points of C by ext C . It is an important fact that compact convex sets can be recovered from their extremal points. The Kre n-Milman theorem states in a very general context that each compact convex set is the closed convex hull of its extremal points, see, e.g., [Ru91, 3.23]. The version of this theorem for subsets C of Rn , and this is what we use here, can be found in [Gr u03, 2.4.5]. We mainly deal with a certain kind of geometric objects called polytopes. Denition 2.7. Let K be a subset of the euclidean space Rn . The set K is a polytope if it is of the form K = co Y for some nite set Y Rn . A fundamental example of a polytope is a simplex. Example 2.8 (A d-dimensional simplex). Let a Rn , and let {u1 , . . . , ud } be a linearly independent subset of Rn . Then the polytope K := co {a} {a + ui | i = 1, . . . , d} is called a d-dimensional simplex. For instance, for d = n = 3 and 0 1 0 0 a := 0 , u1 := 0 , u2 := 1 , u3 := 0 , 0 0 0 1

(2.1)

we obtain the pyramid having the triangle with corner points (0, 0, 0), (1, 0, 0), (0, 1, 0) as its base and the point (0, 0, 1) as its apex. Another concrete example of a polytope is an octahedron. Example 2.9 (A d-dimensional octahedron). Let a Rn , and let {u1 , . . . , ud } be a linearly independent subset of Rn . Then the polytope K := co {a + ui | i = 1, . . . , d} {a ui | i = 1, . . . , d}

is called a d-dimensional octahedron.

For instance, if d = n = 3 and a, u1 , u2 , u3 again as in (2.1), we obtain a regular octahedron with center at the origin.3

(2.2)

Polytopes can be dened in several equivalent ways. The denition used in [Gr u03] is presented in the next lemma. The fact that this denition is equivalent to the one above, i.e., the proof of the lemma is, in essence, a consequence of the Kre n-Milman theorem; we skip the details. Lemma 2.10. A subset K Rn is a polytope if and only if K is compact, convex, and the set ext K of its extremal points is nite. Note that if K is a polytope and K = co Y for some nite set Y , then ext K Y and K = co(ext K ). Remark 2.11. The mentioned concrete examples, the simplex (2.1) and the octahedron (2.2), are of particular interest in the present context. They are the free algebras with 3 generators in the equational classes PCA and TCA, respectively (similarly as the standard simplex (1.1) is in CA). This fact (of course for dimension n instead of 3), together with the results of Section 3 below, shows that understanding polytopes as convex algebras suces to understand all nitely generated algebras in CA, PCA, and TCA.

The equational classes CA, PCA, and TCA

In this section we investigate the three convexity theories of convex, positive convex, and totally convex algebras and their induced equational classes. To start with, let us recall the denitions. Denition 3.1. A convex algebra is an algebra (with nonempty carrier set) of type
n n + Tca := (pi )n i=1 R | n N , p1 , . . . , pn 0, i=1

pi = 1 ,

which satises (we denote by f(pi )n , (pi )n i=1 Tca , the operations of the algei=1 bra) (1) The projection axiom: f(ij )n (x1 , . . . , xn ) = xj , i=1 where ij denotes the Kronecker-delta ij :=
3 Picture

n N+ , j = 1, . . . , n,

1, 0,

i=j i=j

source: http://en.wikipedia.org/wiki/Octahedron

(2) The baricenter axiom: f (pi )n f (p1 j )m (x1 , . . . , xm ), . . . , f(pnj )m (x1 , . . . , xm ) = i=1 j =1 j =1 = f(
n i=1

(x1 , . . . , xm ), pi pij )m j =1

m whenever n, m N+ , (pi )n i=1 Tca , and (pij )j =1 Tca , i = 1, . . . , n.

The operation appearing on the right hand side of the baricenter axiom is welldened since
m n n m

pi pij =
j =1 i=1 i=1

pi
j =1

pij ,

(3.1)

and hence (

n i=1

pi pij )m j =1 Tca .

By CA we denote the equational class of convex algebras. Denition 3.2. A positive convex algebra is an algebra (with nonempty carrier set) of type
n n + Tpca := (pi )n i=1 R | n N , p1 , . . . , pn 0, i=1

pi 1 ,

which satises the projection axiom and the baricenter axiom, where in the m latter (pi )n i=1 and (pij )j =1 vary through Tpca . We denote the equational class of all positive convex algebras as PCA. Denition 3.3. A totally convex algebra is an algebra (with nonempty carrier set) of type
n n + Ttca := (pi )n i=1 R | n N , i=1

|pi | 1 ,

which satises the projection axiom and the baricenter axiom, where in the m latter (pi )n i=1 and (pij )j =1 vary through Ttca . We denote the equational class of all positive convex algebras as TCA. Note that, again because of (3.1), the operation appearing on the right hand side of the baricenter axiom is always well-dened, i.e., is of type Tpca or Ttca , respectively (to see this for Ttca , use the triangle inequality). It is obvious that each positive convex algebra can be considered as a convex algebra. To be precise, if P, (f ) Tpca is a positive convex algebra, then P, (f ) Tca is a convex algebra. Similarly: If P, (f ) Ttca is a totally convex algebra, then P, (f ) Tpca is a positive convex algebra, and in turn P, (f ) Tca is a convex algebra. Due to this fact, many results can immediately be transferred from CA to PCA and TCA. Another interesting fact, which we shall explain in the sequel, is that CA and PCA are essentially the same from an algebraic viewpoint, whereas TCA carries more structure than CA (compare Proposition 3.6 with Proposition 3.8 below). When working with algebras of one of the types Tca , Tpca , or Ttca , it is practical (and customary) to write operations as formal sums and/or to use vector notation: x1 n . f (pi )n ( x , . . . , x ) = p x = ( p , . . . , p ) 1 n i i 1 n . . . i=1
i=1

xn

With this notation the projection axiom writes as x1 . (0, . . . , 1 , . . . , 0) . . = xj ,

j -th place

xn

and the baricenter axiom as p11 . . . . (p1 , . . . , pn ) . . pn1 ...

pnm

p1m x1 . . = . . . . xm

In the next lemma we provide some simple but useful identities which follow from the projection and baricenter axioms. For the convenience of the reader, we provide an explicit proof. In the setting of TCA (with innitary operations) these identities were shown in [PR84, Theorem 2.4] using a dierent proof. Lemma 3.4. For items (i)(iii), let P be an algebra in any of the classes CA, PCA, or TCA. For items (iv ) and (v ), assume that P belongs to PCA or TCA. Let c stand for ca, pca, tca, if P is in CA, PCA, TCA, respectively. (i) The above introduced sum notation already suggests that operations are in a sense commutative. They indeed are, we have f (pi )n (x1 , . . . , xn ) = f(p(i) )n (x(1) , . . . , x(n) ), i=1 i=1 whenever n N+ , is a permutation of {1, . . . , n}, and (pi )iI Tc . (ii) The extended projection law: f (pi )n (x1 , . . . , xn ) = f(pik )m (xi1 , . . . , xim ), i=1 k=1 holds whenever (pi )n i=1 Tc and i1 , . . . , im satisfy i 1 < < im , {i1 , . . . , im } {i {1, . . . , n} | pi = 0}.

p11 . = (p1 , . . . , pn ) . .

... ...

pn1

pnm

p1m x1 . . . . . . . xm

(iii) Whenever (pi )n i=1 Tc and x P , we have f (pi )n (x, . . . , x) = f( i=1 (iv ) The elements (x1 , . . . , xn ), f(0)n i=1 n N+ , x1 , . . . , xn P,
n i=1

pi ) (x).

all coincide. We denote this element as 0P .

10

(v ) The element 0P plays the role of a zero element: Let (pi )n i=1 Tc and let i1 , . . . , im satisfy i1 < < im , Then f (pi )n (x1 , . . . , xn ) = i=1 f ( p ik ) m (xi1 , . . . , xim ) , k=1 0P , m1 . m=0 {i1 , . . . , im } {i {1, . . . , n} | xi = 0P }.

Proof. We denote, here and in the sequel, by In the n n identity matrix and by 0n,m the n m zero matrix. (i) Set pij := i,1 (j ) and compute f(p(i) )n x(1) , . . . , x(n) = i=1 =f(p(i) )n f (p1 j )n (x1 , . . . , xn ), . . . , f(pnj )n (x1 , . . . , xn ) = i=1 j =1 j =1 =f(
n i=1

(x1 , . . . , xn ) p(i) pij )n j =1

= f (pj )n (x1 , . . . , xn ). j =1

(ii) By commutativity it is enough to consider the case that ik = k , k = 1, . . . , m. Then we compute, using the projection and baricenter axioms, x 1 . x1 . . . x m Im 0m,nm xm+1 = (p1 , . . . , pm ) . . =(p1 , . . . , pm ) . . xm .
xn

= (p1 , . . . , pm ) Im

0m,nm

(iii) f(pi )n (x, . . . , x) = f(pi )n (f(1) (x), . . . , f(1) (x)) = f( i=1 i=1 (iv ) Apply the extended projection law twice to obtain

x1 . = (p1 , . . . , pm , 0, . . . , 0) . . . xn (p1 ,...,pn )

x1 . . . = xn

n i=1

pi ) (x).

f(0)n (x1 , . . . , xn ) = f(0)n+n (x1 , . . . , xn , x 1 , . . . , xn ) = f(0)n (x1 , . . . , xn ). i=1

i=1 i=1

(v ) Consider rst the case that m = 0. We have, using (iv ), 0P 0P . . . (p1 , . . . , pn ) . = (p1 , . . . , pn ) 0n,n . . = 0P = (p1 , . . . , pn )0n,n 0P 0P

0P 0P . . =0 . . P . = (0, . . . , 0) . . 0P

11

Assume now that m 1. Again, it is enough to consider the case that ik = k , k = 1, . . . , m. Then we can compute x1 x 1
. . Im m (p1 , . . . , pn )x 0P =(p1 , . . . , pn ) 0 . nm,m . . 0P .

0m,nm 0nm,nm

. . xm 0P = . . . 0P

= (p1 , . . . , pn )

Im 0nm,m

0m,nm 0nm,nm

x1
. . . 0P

. x1 . . xm (ii) . =(p1 , . . . , pm , 0, . . . , 0) 0P = (p1 , . . . , pm ) . . . . . xm .

0P

x1

. . x m 0P = .

Note that items (iv ) and (v ) would not at all make sense within CA, since the operations appearing in them do not belong to Tca . Remark 3.5. Let P CA. A consequence of the extended projection law is that the baricenter axiom remains valid when the vectors (pij )m j =1 appearing therein are no more of the same length m and no more bound to the same variables. More precisely, let
n

n, m N+ , (pi )n i=1 Tca ,

Ki N+ , i = 1, . . . , n with
i=1

Ki = {1, . . . , m}, (3.2) (3.3)

(pij )j Ki with pij 0,

j Ki

pij = 1,

i = 1, . . . , n.

i i Set mi := |Ki |, and write Ki = {i k | k = 1, . . . , mi } with k < k+1 . Then n mi m n

pi
i=1 k=1

pii x i = k k
j =1 i=1 j Ki

pi pij xj ,

whenever x1 , . . . , xm P . The same holds when P PCA and the conditions (3.2) and (3.3) are replaced by (pi )n i=1 Tpca , (pij )j Ki with pij 0,
j Ki

pij 1,

i = 1, . . . , n.

Correspondingly, the same holds when P TCA and (pi )n i=1 Ttca , (pij )j Ki with pij R,
j Ki

|pij | 1,

i = 1, . . . , n.

12

This remark also claries the note made in [Pu03, p.110] immediately after the denition of positively convex algebra (in which the above stronger form of the baricenter axiom is required).

3.1

Positive convex algebras vs. convex algebras

In this subsection we make precise what we meant above by claiming that CA and PCA are just the same. Let P, (f ) Tca CA. We call a collection {f | Tpca } of operations on P an operational extension of P, (f ) Tca to PCA, if f = f , Tca , and P, (f ) Tpca PCA . (3.4)

Proposition 3.6. Let P, (f ) Tca CA. Then the set of all operational extensions of P, (f ) Tca to PCA corresponds bijectively to P . In particular, since P = , P can be operationally extended to a positive convex algebra. Proof. In the rst step we show existence of operational extensions. Choose an element z P , and dene for (pi )n i=1 Tpca f (x1 , . . . , xn ) := f(p1 ,...,pn ,p (pi )n ) (x1 , . . . , xn , z ), i=1 where p := 1
i=1 n

pi .

The extended projection law in P, (f ) Tca gives f = f , Tca . Knowing that f = f , T , it is clear that the projection axiom holds ca in P, (f ) ; the operations involved all belong to Tca . To show the PCA Tpca m baricenter axiom, let (pi )n T and ( p ) T , pca ij i=1 pca i = 1, . . . , n, be given. i=1 Denote
n m

p := 1
i=1

pi ,

p i := 1
j =1

pij , i = 1, . . . , n.

Then (ij again denotes the Kronecker-delta) (p1 , . . . , pn , p ), (pi1 , . . . , pim , p i ), i = 1, . . . , n,

+1 (i,m+1 )m i=1

all belong to Tca , and hence we may apply the CA-baricenter axiom. This gives f (x1 , . . . , xm ), . . . , f (x1 , . . . , xm ) = f (pi )n (p1 j )m (pnj )m j =1 j =1 i=1 f (p1 j )m (x1 , . . . , xm ) j =1 . . . = (p1 , . . . , pn , p ) = f (x1 , . . . , xm ) (pnj )m j =1 z p11 . . = (p1 , . . . , pn , p ) . pn1 0 13 ... ... ... p1m . . . pnm 0

x1 p 1 . . . . . . = p n xm z 1

p11 . . = (p1 , . . . , pn , p ) . pn1 0 = = f (

n

p 1 x1 . . . . . . = . . . pnm p n xm ... 0 1 z x1 n n n . . pi pi1 , . . . , pi pim , pi p i + p . = x m i=1 i=1 i=1 z ... p1m . . .

n i=1

(x1 , . . . , xm ). pi pij )m j =1
n

The last equality holds since, due to (3.1),

n

pi p i + p = 1
i=1 i=1

pi pi1 + +
i=1

pi pim .

Moreover, we have for arbitrary x P 0P = f (0) (x) = f(0,1) (x, z ) = z . Second, assume conversely that operations f , Tpca , with (3.4) are given. Then, by Lemma 3.4(v), f (x1 , . . . , xn ) = f (pi )n (p1 ,...,pn ,p ) (x1 , . . . , xn , 0P ) = i=1 = f(p1 ,...,pn ,p (pi )n ) (x1 , . . . , xn , 0P ), i=1 Tpca , n where p := 1 i=1 pi . Hence, the operations f coincide with the operaions constructed in the rst step with the choice z = 0P . Finally, the above computation also shows that an operational extension according to (3.4) is uniquely determined by the corresponding element 0P , showing that P indeed corresponds bijectively to the set of all operational extensions. Next, we show that congruence relations within CA or PCA are just the same. Let P be an algebra in CA or PCA and an equivalence relation on P . Recall that is a congruence if whenever (xi , x i ) , for i {1, . . . , n} then also n n n n ( i=1 pi xi , i=1 pi x ) for ( p ) T i i=1 ca or (pi )i=1 Tpca , respectively. By i ConCA P or ConPCA P we denote the sets of all CA- or PCA-congruence relations on P , respectively. Lemma 3.7. Let P, (f ) Tca CA and let {f | Tpca } be an operational extension of P, (f ) Tca to PCA. Moreover, let be an equivalence relation on P . Then ConCA P, (f ) Tca if and only if ConPCA P, (f ) Tpca . Proof. Clearly, each PCA-congruence on P is also a CA-congruence. Conversely, assume that ConCA P . Let (pi )n i=1 Tpca and (xi , xi ) , i = 1, . . . , n, n be given. Again set p := 1 i=1 pi , so that (p1 , . . . , pn , p ) Tca . Since ConCA P , it follows that
n n

pi x i ,
i=1 i=1

pi x i =

= p1 x 1 + + pn x n + p 0P , p1 x 0P 1 + + pn x n + p

14

3.2

Totally convex algebras vs. convex algebras

In this subsection we characterize those convex algebras which can be operationally extended to totally convex algebras. In view of Proposition 3.6 this problem is equivalent to characterizing those positive convex algebras that can be operationally extended4 to totally convex algebras. It turns out that not every convex algebra has this property, contrasting the situation for passing from CA to PCA. The decisive additional property which allows to operationally extend to TCA is existence of an involutory endomorphism satisfying a particular computation rule. Recall that an endomorphism of an algebra is called involutory if it is inverse to itself, i.e., if 2 = id. Proposition 3.8. Let P, (f ) Tpca PCA. Then there exists an operational extension {f | Ttca } of P to TCA if and only if there exists an involutory PCA-endomorphism of P which has the property that f(p1 ,...,pn ,q1 ,...,qn ) (x1 , . . . , xn , x1 , . . . , xn ) =
(x1 , . . . , xn , x1 , . . . , xn ), = f (p 1 ,...,pn ,q1 ,...,qn )

(3.5)

whenever n N+ ,
(p1 , . . . , pn , q1 , . . . , qn ), (p 1 , . . . , pn , q1 , . . . , qn ) Tpca , pk q k = p k qk ,

k = 1, . . . , n.

Proof. Consider rst the case that |P | = 1. Then P = {0P } and trivially P can be operationally extended to a totally convex algebra. Equally trivially the identity map is an involutory endomorphism and satises (3.5). Hence, in this case, the asserted equivalence holds. For the rest of the proof we may thus assume that P contains more than one element. Necessity. Assume that an operational extension {f | Ttca } is given, and consider the map : P P dened as (x) = f (1) (x). Let (pi )n i=1 Ttca , then f (x1 , . . . , xn ) (1) f(pi )n i=1 = = f (x1 , . . . , xn ) ( pi )n i=1 f f (pi )n (1) (x1 ), . . . , f(1) (xn ) . i=1

This shows that is an endomorphism (even a TCA-endomorphism). Moreover, we have f (1) (f(1) (x)) = f(1) (x) = x, i.e. is involutory. In order to show (3.5), it is certainly sucient to show that f (p1 ,...,pn ,q1 ,...,qn ) (x1 , . . . , xn ,x1 , . . . , xn ) = =f (p q ,...,p q ) (x1 , . . . , xn ),
1 1 n n

(3.6)

4 Operational extensions of an algebra in CA or PCA to an algebra in TCA are dened analogously as operational extensions of an algebra in CA to one in PCA.

15

whenever (p1 , . . . , pn , q1 , . . . , qn ) Ttca . Note here that

n n n n

|pi qi |
i=1 i=1

(|pi | + |qi |) =
i=1

|pi | +
i=1

|qi | 1,

and hence the operation written on the right side is legitimate. To see (3.6), we compute x1 . . x1 . xn In . (p1 , . . . , pn , q1 , . . . , qn ) . = = (p1 , . . . , pn , q1 , . . . , qn ) In . x 1 xn . . . xn x1 x1 In . . . = (p1 , . . . , pn , q1 , . . . , qn ) = (p1 q1 , . . . , pn qn ) . . . In . xn xn

Suciency. Assume that an involutory PCA-endomorphism on P with (3.5) is given. For each (pi )n as i=1 Ttca we dene an operation f(pi )n i=1 (x1 , . . . , xn , x1 , . . . , xn ), f (x1 , . . . , xn ) := f(p+ ,...,p+ (pi )n i=1 n ,p ,...,pn )
1 1

(3.7) (3.8)

where p+ := max{p, 0}, p := min{p, 0}, Note that p = p p . Moreover, we have

n n n +

p R.

p+ i
i=1

+
i=1

p i

=
i=1

|pi |,

and hence the operation in Tpca on the right side of (3.7) is well-dened. If + (pi )n i=1 Tpca , then pi = pi and pi = 0 for all i {1, . . . , n}, and the extended projection law for P, (f ) Tpca gives f (x1 , . . . , xn ) (pi )n i=1 = = f(p1 ,...,pn ,0,...,0) (x1 , . . . , xn , x1 , . . . , xn ) f (pi )n (x1 , . . . , xn ). i=1

It is not dicult to see that the projection axiom holds for P, (f ) Ttca ; we n skip the details. To check the TCA-baricenter axiom, let (pi )i=1 Ttca and (pij )m j =1 Ttca , i = 1, . . . , n, be given. First, compute f (x1 , . . . , xm ) (pij )m j =1 = = = Next, set P + := p+ ij f(p+ ,...,p+
i1 im ,pi1 ,...,pim )

(x1 , . . . , xm , x1 , . . . , xm )

f(p+ ,...,p+
i1 i1

im ,pi1 ,...,pim ) + + im ,pi1 ,...,pim )

(x1 , . . . , xm , x1 , . . . , xm ) (x1 , . . . , xm , x1 , . . . , xm ).
i=1,...,n , j =1,...,m

f(p ,...,p
i=1,...,n , j =1,...,m

P := p ij 16

and use the baricenter axiom for P, (f ) Tpca to compute f f (x1 , . . . , xm ), . . . , f (x1 , . . . , xm ) = (pi )n (p1 j )m (pnj )n i=1 j =1 j =1 x 1
.

+ + P =(p+ , . . . , p , p , . . . , p ) n n P 1 1
+ = (p+ 1 , . . . , pn , p1 , . . . , pn )

P P+

. . x n x1 = . . . xn

P+ P

P P+

x1 . . . xn x1 . . . xn n

n + p+ i pi1 + p i pi1 , . . . , i=1

n p+ i pi1 + i=1

=
i=1

+ p i pi1 , . . . i=1

=
i=1

pi pi1 , . . . ,
i=1

pi pi1 , . . .

where the last equality follows from (3.5) since

n + p+ i pij + i=1 i=1 n n p i pij i=1 n n p+ i pij + n

x1 . . . xn x1 . . . xn

x1 . . . xn x1 . . . xn

+ p i pij = i=1 n n

=
n

pi pij +
i=1 pi 0,pij 0 n

pi pij
+ n

pi (pij )

(pi )pij =

i=1 pi <0,pij <0

i=1 pi 0,pij <0

i=1 pi <0,pij 0

=
i=1

pi pij =
i=1

pi pij

i=1

pi pij

In order to show that TCA is (in the sense of operational extendability) indeed strictly smaller than CA, we need to provide an example of a positive convex algebra which does not admit an involutory endomorphism satisfying (3.5). Such examples can already be found in euclidean space. First, an auxilliary observation. Lemma 3.9. Let P PCA, and let be an involutory PCA-endomorphism on P which satises (3.5). Then has exactly one xed point, namely 0P . Proof. Let x P with x = x. Then, using (3.5),
1 1 (x, x) = f 1 1 (x, x) = f(0,0) (x, x) = 0P . x = f( 2 ,2) (2,2)

To see that 0P indeed is a xed point, compute 0P = f(0) (0P ) = f(0) ( 0P ) = 0P . 17

Proposition 3.10. Let K Rn be a polytope with |K | > 1, 0 K , and let K be endowed with the operations f , Tpca , given as the usual sum of vectors in Rn
n

f (pi )n (x1 , . . . , xn ) := i=1

i=1

pi x i ,

(pi )n i=1 Tpca .

If K can be operationally extended to a totally convex algebra, then K has an even number of extremal points. Proof. Assume that K can be operationally extended to a totally convex algebra and let, by Proposition 3.8, be an involutory PCA-endomorphism of K which satises (3.5). Note that is bijective. The rst step is to show that (ext K ) = ext K . Let x K \ ext K , and choose x1 , x2 K , x1 , x2 = x, and t (0, 1) such that x = tx1 + (1 t)x2 . Then also x = tx1 + (1 t)x2 . Since is injective, we conclude that x cannot be an extremal point of K . Thus we have (K \ ext K ) K \ ext K . Furthermore, K \ ext K = 2 (K \ ext K ) (K \ ext K ), showing that (K \ ext K ) = K \ ext K . Since is bijective, this implies (ext K ) = ext K . Next, consider the restriction |ext K . It is an involutory permutation on ext K . Consider the orbits of |ext K , i.e., the sets {( |ext K )k x | k Z} for x ext K . Each orbit can contain at most two points. Choose x K \ {0}. Then also x = 0. From (3.5), 1 1 1 (x, x) = 0, x + x = f( 1 2,2) 2 2 where the sum on the left denotes the usual vector sum, and we see that 0 ext K . However, we know by Lemma 3.9 that 0 is the only xed point of . Hence, each orbit of |ext K must contain exactly two points. Since ext K is the disjoint union of all orbits, it contains an even number of points. We close the discussion of TCA by making explicit the relationship between TCA-congruences and CA-congruences. Since we already know that CAcongruences are the same as PCA-congruences, we may equally well compare TCA-congruences with PCA-congruences. Denote by ConTCA P the set of all congruences of P in TCA. Lemma 3.11. Let P, (f ) Tpca PCA, let {f | Ttca } be an operational extension to TCA, and set (x) := f ( x ) , x P . Moreover, let be an (1) equivalence relation on P . Then ConTCA P, (f ) Ttca if and only if ConPCA P, (f ) Tpca and ( ) . Proof. Clearly, each TCA-congruence on P is also a PCA-congruence and satises ( ) . We have to show the converse. Let (pi )n i=1 Ttca and

18

(xi , x i ) , i = 1, . . . , n, be given. Using the notation (3.8), we obtain from (3.6) that (x1 , . . . , xn , x1 , . . . , xn ). f (x1 , . . . , xn ) = f(p+ ,...,p+ (pi )n i=1 n ,p ,...,pn )
1 1

The same equation holds for x i instead of xi . Since is a PCA-congruence, xi x i , and xi xi by the assumptions, we see that indeed
f (x1 , . . . , xn ) f (x (pi )n (pi )n 1 , . . . , xn ) . i=1 i=1

Convex equivalences on polytopes

Let K Rn be a polytope. We consider K as a convex algebra endowed with the operations inherited from Rn (as in Proposition 3.10). The following property is a direct consequence of the denitions and Lemma 2.2 but it is an important observation for what follows. Lemma 4.1. An equivalence relation on a polytope K is in ConCA K if and only if it is convex as a subset of K K Rn Rn , with operations dened component-wise. Throughout the paper, we denote VK := P (ext K ) \ {}, where P (M ) denotes the power set of a set M , and consider VK as a joinsubsemilattice of the lattice P (ext K ). Moreover, we denote by Sub Rn the set of all linear subspaces of Rn , and consider Sub Rn as being ordered by inclusion. In the following denition we introduce the crucial concepts for describing ConCA K . Denition 4.2. Let K Rn be a polytope and let ConCA K . We dene a map : VK Sub Rn by (Y ) = span x2 x1 | x1 , x2 co Y, x1 x2 , Y VK .

We dene a graph G with vertices VK and (undirected) edges E given by {Y1 , Y2 } E (co Y1 co Y2 ) = , Y1 , Y2 VK .

We denote by the equivalence relation on VK dened as Y1 Y2 Y1 and Y2 are connected by a path in G . The equivalence classes of are the connected components of G , see, e.g., [Di00]. Note that there is always an edge in G connecting a vertex Y with itself, even if |Y | = 1, due to the denition of co Y. In the following three statements we give a complete description of the congruence lattice ConCA K . Theorem 4.3. Let K be a polytope in Rn and let ConCA K . Then 19

(i) The map is monotone. (ii) Let C be a component of the graph G . Then C contains a largest element with respect to inclusion. Denoting this largest element by Y (C ), we have {Y, Y (C )} E , Y C . (iii) The relation is a congruence of the join-semilattice VK . (iv ) Let C be a component of G and Y (C ) its largest element. Then (Y ) = (Y (C )) dir Y for Y C , for Y C . (4.1) (4.2) (4.3)

co Y + (Y (C )) co Y (C ) = Set Z (C ) :=
Y C

co Y,

then the congruence can be recovered from and G as =

C component of G

(x1 , x2 ) Z (C ) Z (C ) : x2 x1 (Y (C )) .

(4.4)

Let us point out that reconstructing by means of the formula (4.4) only requires knowledge of the classes of and the values of on the respective largest elements of these classes. Theorem 4.4. Let K be a polytope in Rn . Let be a congruence relation of the join-semilattice VK with the property that each congruence class C of contains a largest element, say Y (C ). Moreover, let : {Y (C ) | C class of } Sub Rn be a monotone map such that, for each class C of , (Y (C )) dir Y (C ), co Y + (Y (C )) co Y (C ) = for Y C .

Then there exists a unique congruence ConCA K such that = , (Y (C )) = (Y (C )) for C a class of .

This congruence can be computed from and by means of the formula =

C class of

(x1 , x2 ) Z (C ) Z (C ) : x2 x1 (Y (C )) ,

(4.5)

where again Z (C ) :=

Y C

co Y . Its associated function is given as for Y C , (4.6)

(Y ) = Y (C ) dir Y

and the set of edges E of its associated graph G is given as {Y1 , Y2 } E Y1 Y2 co Y1 + Y ([Y1 ] ) co Y2 = (4.7)

where [Y1 ] denotes the equivalence class of Y1 . 20

Remark 4.5. Note that for any semi-lattice congruence with largest element in every class (as in Theorem 4.4) there is at least one possible choice for the assignment in Theorem 4.4, namely (Y (C )) = dir Y (C ) satises all conditions. We do not have a simple description of all semi-lattices congruences which are admissible in the sense of Theorem 4.4. However, some examples are easily obtained. Let K Rn , |K | > 1, be a polytope, and let Y0 ext K . Dene an equivalence relation on VK by specifying its equivalence classes to be Cy := {y } , y Y0 , C0 := VK \
y Y0

Cy .

Clearly, each of these classes contains a largest element Y (C ) (for C0 it is Y (C0 ) = ext K ), and it is straightforward to verify that is a congruence of the joinsemilattice VK . In order to illustrate these results, let us consider a (toy) example. Example 4.6. Let Y := {0, 1} and K := co Y in R. Then K = [0, 1] and K is in CA. We will show, using Theorem 4.3 and Theorem 4.4, that there are exactly ve CA-congruences on K . These are: 1 2 3 4 5 =

= {(0, 0), (1, 1)} (0, 1) (0, 1) = {(0, 0)} (0, 1] (0, 1] = [0, 1) [0, 1) {(1, 1)}, and = [0, 1] [0, 1].

We have VK = {{0}, {1}, {0, 1}}. To ease the notation we write 0 = {0}, 1 = {1}, 01 = {0, 1}, hence VK = {0, 1, 01}. Next we list all join-semilattice congruences of VK with the property that each class has a largest element. There are four such, given by their partitions: VK / 1 VK / 2 VK / 3 VK / 4 = = = = {{0}, {1}, {01}} {{0}, {1, 01}} {{1}, {0, 01}} {{0, 1, 01}}.

Note that R has only two vector subspaces, the trivial ones, Sub R = {0, R}. Furthermore, dir 0 = 0, dir 1 = 0, and dir 01 = R. For each of the join-semilattice congruences we need to consider all monotone maps mapping the largest elements of each class to Sub R and satisfying the conditions (Y (C )) dir Y (C ) and co Y + (Y (C )) co Y (C ) = for Y C . Consider 1 = . The second condition here is always satised, and the rst implies that (0) = 0 and (1) = 0. Hence, there are two possibilities for dening , namely 1 = (0 0, 1 0, 01 0) and 2 = (0 0, 1 0, 01 R) . 21

From (4.5) we then get two convex congruences on K and these are exactly 1 and 2 . Consider next 2 . Due to the conditions on , there is a unique possibility to dene this map, namely 3 = (0 0, 01 R) and (4.5) then gives 3 . The case 3 is symmetric. Again there is a unique possibility to dene the map , leading to 4 = (1 0, 01 R) and 4 . Finally, consider 4 . There is a unique map 4 = (01 R) which satises the conditions imposed on leading to 5 . By Theorem 4.3, there are no other convex congruences on K . The next theorem shows that also the order relation on the congruence lattice ConCA K can be characterised in terms of and G . Theorem 4.7. Let K be a polytope in Rn . If 1 , 2 ConCA K , then 1 2 E 1 E 2 1 2

where denotes the pointwise order by inclusion. The remainder of this section is entirely devoted to the proof of Theorem 4.3, Theorem 4.4, and Theorem 4.7. Thereby, the hard part is to establish Theorem 4.3. Once this is achieved, Theorem 4.4 and Theorem 4.7 are fairly straightforward. We are going to proceed according to the following schedule: In Subsection 4.1 we provide some simple geometric consequences of convexity which are used throughout. Moreover, we show Theorem 4.3, (i). In Subsection 4.2 we investigate the structure of G , and show Theorem 4.3, (ii). In Subsection 4.3 we investigate the behaviour of a congruence relation in the interior of K (or one of its lower dimensional facets). In Subsection 4.4 we establish the representation (4.4). In Subsection 4.5 we use (4.4) to deduce items (iii) and (iv ) of Theorem 4.3. In Subsection 4.6 we carry out the converse construction and provide evidence to Theorem 4.4. Finally, in Subsection 4.7, we show that the order structure of ConCA K can be characterised as stated in Theorem 4.7. For the rest of this section let a polytope K be xed, and throughout Sections 4.1 4.5 let a congruence ConCA K be xed.

22

4.1

Geometric consequences of convexity

We start with a denition of a useful concept of perspective. Denition 4.8. For each z K we denote by z : [0, 1] Rn Rn the map dened as z (s, x) := sz + (1 s)x, s [0, 1], x Rn . Based on geometric intuition, we refer to z as the perspective with center z . Since K is convex and z K , we have z (s, K ) K for any s [0, 1]. Moreover, z (0, .) is the identity map, and z (1, .) is the constant map with value z . The following observation is simple but important, and we further refer to it as perspective invariance. Intuitively speaking, perspective invariance means that a congruence class cannot split and distribute over several dierent classes when moved with a perspective. Lemma 4.9. Let A K be an equivalence class of . Then, for each z K and s [0, 1], there exists an equivalence class Az,s K of with z (s, A) Az,s . K
A z (s, A) z

Moreover, perspective invariance implies that each equivalence class of is convex. Proof. Let x1 , x2 K with x1 x2 , z K , and s [0, 1] be given. Since z z and (x, y ) sx + (1 s)y is an operation of K CA, we have z (s, x1 ), z (s, x2 ) = (sz + (1 s)x1 , sz + (1 s)x2 ) . To deduce that equivalence classes are convex, let x1 , x2 K with x1 x2 , and s [0, 1] be given. Perspective invariance with z := x1 gives (x1 , sx1 + (1 s)x2 ) = x1 (s, x1 ), x1 (s, x2 ) .

Also the next fact is simple but useful. Lemma 4.10. (i) Let Y1 , Y2 VK be given. Then z co Y1 , x co Y2 : z (s, x) co( Y1 Y2 ), s (0, 1).

23

(ii) Let Y VK be given. Then z co Y, x co Y : z (s, x) co Y, z co Y, x co Y : z (s, x) co Y, Proof. For the proof of (i) write z=
y Y1

s [0, 1). s (0, 1].

1 y y, 2 y y,
y Y2

1 y (0, 1],
y Y1

1 y = 1, 2 y = 1.
y Y2

x= Then z (s, x) = s
y Y1

2 y

(0, 1],

1 y y + (1 s)
y Y2

2 yy =
2 s1 y + (1 s)y y +

=
y Y1 \ Y2

s1 yy +
y Y1 Y2

(1 s)2 y y.
y Y2 \ Y1

All coecients in these sums are positive and they sum up to 1. z (s, x) co( Y1 Y2 ). Item (ii) follows in the same manner.

Hence,

Let us immediately exploit perspective invariance to show that is monotone. Proof (of Theorem 4.3, (i)). Let Y1 , Y2 VK with Y1 Y2 be given. If Y1 = Y2 , there is nothing to prove. Hence, assume that Y1 Y2 . Let w (Y1 ), and write according to the denition of
m

w=
k=1

k k (xk 2 x1 ),

with some k R,
k xk Y1 , 1 , x2 co k xk 1 x2 , k = 1, . . . , m. 1 Set z := |Y2 \ y Y2 \Y1 y , and x s (0, 1). Note that the assumption Y1 Y2 Y1 | ensures that z is well dened. Perspective invariance gives k z (s, xk 1 ) z (s, x2 ),

k = 1, . . . , m.

By Lemma 4.10(i), since z co( Y2 \Y1 ) and Y2 = (Y2 \Y1 )Y1 by the assumption, k we have z (s, xk Y2 , and hence 1 ), z (s, x2 ) co
m

w :=
k=1

k k z (s, xk 2 ) z (s, x1 ) (Y2 ).

However,
k k k z (s, xk 2 ) z (s, x1 ) = (1 s)(x2 x1 ),

and hence w = (1 s)w. Since (Y2 ) is a vector subspace, it follows that also w (Y2 ). The property shown in the next lemma will be used in several instances. 24

Lemma 4.11. Let Y VK , |Y | 2, and let x1 , x2 co Y , x1 = x2 . Consider the map : R a Y given by (t) := tx2 + (1 t)x1 , t R.

Clearly, (R) = a {x1 , x2 } is the line containing x1 and x2 .5 Then (r) (s, (t)) = sr + (1 s)t , r, t R, s [0, 1]. (4.8)

There exist numbers t < 0 and t+ > 1, such that 1 (co Y ) = (t , t+ ), 1 (co Y ) = [t , t+ ]
co Y a Y

(t )

x1

x2

(t+ )

t+

Proof. To show (4.8), we compute (r) (s, (t)) = s(r) + (1 s)(t) = =s rx2 + (1 r)x1 + (1 s) tx2 + (1 t)x1 = = sr + (1 s)t x2 + s(1 r) + (1 s)(1 t) x1 = = sr + (1 s)t x2 + 1 (sr + (1 s)t) x1 = (sr + (1 s)t). Note that R and a Y inherit the euclidean topology from Rn . Moreover they also inherit the euclidean metric from Rn . Also, a line is continuous, i.e., is a continuous function. Consider the set 1 (co Y ). Since is continuous and co Y is an open subset of a Y by Lemma 2.5, this set is an open subset of R. Since is linear and co Y is convex, cf. Section 2, it is convex. We will now show that 1 (co Y ) is bounded. Note that (t) = x1 + t(x2 x1 ). Therefore, for denoting the euclidean norm, using the downward triangle inequality, we get (t) | t(x2 x1 ) x1 | |t| x2 x1 x1 .
+ x1 Hence, for each positive real number R, if |t| > R x2 x1 , then (t) > R. This shows that the inverse image by of a bounded set in a Y is a bounded set in R. Now, since co Y co Y and the latter is bounded by Lemma 2.4, we get
5 Note that actually depends on the points x and x but we prefer a light, overloaded 1 2 notation.

25

that 1 (co Y ) is bounded. Note that in R a set is open, convex, and bounded if and only if it is a nite open interval. Hence 1 (co Y ) = (t , t+ ) with some numbers t , t+ R. Since (0) = x1 co Y and (1) = x2 co Y, we must have t < 0 and t+ > 1. Again by Lemma 2.4, co Y is the closure of co Y and hence continuity of implies that 1 (co Y ) is closed. Since 1 (co Y ) 1 (co Y ) and [t , t+ ] is the closure of 1 (co Y ) = (t , t+ ), we also have [t , t+ ] 1 (co Y ). To show the opposite inclusion, let t R with (t) co Y be given. If t [0, 1], we have (t) co{x1 , x2 } co Y , and hence t (t , t+ ). Next, consider the case that t > 1. Choose t (1, t) and set s := tt . Then s (0, 1) and (t) (s, x1 ) = s(t) + (1 s)x1 = s(tx2 + (1 t)x1 ) + (1 s)x1 = = (st)x2 + (1 st)x1 = t x2 + (1 t )x1 = (t ). By Lemma 4.10, we have (t ) co Y . Thus t (t , t+ ). Since t was arbitrary in (1, t), we have (1, t) (t , t+ ) and hence [1, t] [t , t+ ] showing t [t , t+ ]. The case that t < 0 is analogous.

4.2

The structure of G

In order to understand the structure of G , we need one preparatory result. Lemma 4.12. Let Y1 , Y2 VK , and assume that Y1 Y2 . If {Y1 , Y2 } E , i.e., x1 co Y1 , x2 co Y2 : x1 x2 , then actually
x Y 1 , x Y2 : x 1 co 2 co 1 x2 . Proof. If Y1 = Y2 , we can always choose x Y1 | = 1, 2 := x1 . If |Y1 | = 1, also | co and hence the assertion is trivial. Note that the assertion is not trivial in general if Y1 Y2 as it often happens that then co Y1 co Y2 = , think for example of |Y1 | = 2 and |Y2 | = 3 when co Y2 is the interior of a triangle and co Y1 is the interior of one of its sides. Assume that Y1 Y2 , |Y1 | > 1, and choose x1 co Y1 and x2 co Y2 with x1 x2 . Let x Y1 be given. If x 1 co 1 = x1 , we can take x2 := x2 . Hence, assume that x1 = x1 . Consider the line containing the points x 1 , x1 , i.e.,

(t) := tx1 + (1 t)x 1,

t R.

By Lemma 4.11, we have 1 (co Y1 ) = (t , t+ ) and 1 (co Y1 ) = [t , t+ ] with 1 some t < 0 and t+ > 1. Set s := 1 t , then s (0, 1) and st + (1 s) 1 = 0. By (4.8) since (1) = x1 we have (t ) (s, x1 ) = (st + (1 s) 1) = (0) = x 1. Take x 2 := (t ) (s, x2 ). By perspective invariance and the above equality, we get x Y2 , Lemma 4.10(ii) implies 1 x2 . Since (t ) co Y1 co Y2 and x2 co x2 = (t ) (s, x2 ) co Y2 , which completes the proof. 26

co Y2
x2

(t ) (s, x2 ) x 2

(R)
(t )

x1 x 1 (t ) (s, x1 )

Corollary 4.13. Let Y1 , Y2 , Y3 VK , and assume that {Y1 , Y2 }, {Y2 , Y3 } E . If Y2 Y3 , then also {Y1 , Y3 } E . Proof. Choose x1 co Y1 and x2 co Y2 with x1 x2 . According to Lemma 4.12 there exists x3 co Y3 with x2 x3 . Transitivity gives x1 x3 , and we see that {Y1 , Y3 } E . We can now establish the result on the structure of components of G . Proof (of Theorem 4.3, (ii)). We proceed in three steps. The rst step is to show that {Y1 , Y2 } E {Y1 , Y1 Y2 }, {Y2 , Y1 Y2 } E

To this end, choose x1 co Y1 and x2 co Y2 with x1 x2 . By convexity of equivalence classes, we have 1 (x1 + x2 ) xj , 2 j = 1, 2,

1 1 and since x1 1 Y 1 Y2 ) 2 , x2 = 2 (x1 + x2 ), Lemma 4.10 gives 2 (x1 + x2 ) co( which shows the above property.

Let C be a component of G . The second step is to show that Y 1 , Y2 C Y3 C : Y1 Y2 Y3 , {Y1 , Y3 }, {Y2 , Y3 } E

To established existence of Y3 with the required properties, we use induction on the length of a path in G connecting Y1 and Y2 . If Y1 and Y2 can be connected by a path of length 1, i.e., if {Y1 , Y2 } E , then set Y3 := Y1 Y2 . By the property shown in the rst step, we have {Y1 , Y3 }, {Y2 , Y3 } E . This also implies that Y3 C . For the inductive step, let m N+ be given, and assume that Y1 and Y2 can be connected by a path of length m + 1. This means that there exist
Y0 , . . . , Ym +1 VK , Y0 = Y1 , Ym+1 = Y2 with {Yk , Yk+1 } E for k {0, . . . , m}. Clearly, all Yk belong to C . The inductive hypothesis provides a vertex Y C with Y1 Ym Y , {Y1 , Y }, {Ym , Y } E .

27

 Now, since Ym Y and {Y2 , Ym }, {Ym , Y } E , from Corollary 4.13, we get {Y2 , Y } E . From the property shown in the rst step, we get {Y2 , Y2 Y }, {Y , Y2 Y } E . Another application of Corollary 4.13, this time since {Y1 , Y }, {Y , Y2 Y } E and obviously Y Y2 Y , gives that also {Y1 , Y2 Y } E . Thus the element Y3 := Y2 Y has all required properties. This nishes the proof of the second step.

Let Y C . If there exists Y C with Y Y , the property shown in the second step gives an element Y C with Y Y Y and Y Y (since Y Y Y ). Hence, Y is not maximal in C . We conclude that if an element is maximal in C , then it must also be the largest in C . Since C is nite (the whole graph is nite), there certainly exists a maximal element. Let Y0 be such. Then, since it is the largest, we have Y0
Y C

Y,

and hence Y0 = Y C Y . Let Y C be given. By the property shown in the second step, applied to Y and Y0 , we obtain Y3 C with Y Y0 Y3 , {Y, Y3 }, {Y0 , Y3 } E .

However, since Y0 is the largest element of C , it follows that Y3 = Y0 . Hence, we have shown that {Y, Y0 } E .

4.3

Behaviour of inside of co Y

Let Y VK . Our aim is to describe (co Y co Y ). If Y contains only one element, then co Y = Y and clearly (co Y co Y ) = Y Y . Hence, assume that |Y | 2. The main observation is that an equivalence class of which contains two dierent points of co Y must already stretch all the way to the boundary of co Y . This is the analogue of [PR90, Theorem 1.3], where a similar result was established for congruences of totally convex algebras. However, the proof given there does not immediately carry over to the presently considered situation, since we are bound to operations in Tca , i.e., true convex combinations. Lemma 4.14. Let x1 , x2 co Y , x1 = x2 , with x1 x2 . Then a {x1 , x2 } co Y [x1 ] , where [x1 ] denotes the -equivalence class of x1 . Proof. Consider the line containing the points x1 and x2 , i.e., the map (t) := tx2 + (1 t)x1 , t R,

and let t , t+ be as given by Lemma 4.11. As a rst step we show that, for each t (t , t+ ), the set 1 ([(t)] co Y ) contains an open neighbourhood of t. 1 Consider rst the case that t 2 . Choose s [0, 1) such that st+ +(1s) 1 2 = . Then by t. This is doable since under the assumption for t we have t+ > 1 2

28

(4.8) (t+ ) (s, x1 ) = (t+ ) (s, (0)) = (st+ ), (t+ ) (s, x2 ) = (t+ ) (s, (1)) = (st+ + (1 s)), 1 1 (t+ ) s, (x1 + x2 ) = (t+ ) s, ( )) = (t). 2 2

(t+ )

(t)

x2 x1 (t ) (s, .) +

Projective invariance implies (st+ ) (st+ + (1 s)). Note that co(Y + z ) = co(Y )+ z and for any linear map f it holds that f (co Y ) = co(f (Y )). Since the map t (t) x1 is linear, we have [st+ , st+ + (1 s)] = co (st+ ), (st+ + (1 s)) and we conclude from convexity of equivalence classes that [st+ , st+ + (1 s)] [(st+ )] . Since st+ < t < st+ + (1 s), in particular (t) [(st+ )] and so [(st+ )] = [(t)] . It follows that t (st+ , st+ + (1 s)) 1 [(t)] . We have t < 0 st+ and (since t+ > 1) st+ + (1 s) < st+ + (1 s)t+ = t+ . Therefore, (st+ , st+ + (1 s)) (t , t+ ) co Y. Hence, we have found an open neighbourhood of t which is contained in 1 ([(t)] co Y ). The case that t 1 2 is treated in the same way, using the perspective (t ) . From this it is easy to deduce that actually for each t (t , t+ ) the set 1 ([(t)] co Y ) is open: Let t 1 ([(t)] co Y ). Then (t ) [(t)] co Y and hence, by Lemma 4.11, t (t , t+ ) and (t )(t). Then, by what we showed above, there exists an open neighbourhood U of t with U 1 ([(t )] co Y ). However, [(t )] = [(t)] . Let I (t , t+ ) be such that {(t) | t I } is a set of class representatives of (4.9)

29

the equivalence relation [((t , t+ )) ((t , t+ ))]. Recall that (t , t+ ) = 1 (co Y ) by Lemma 4.11. We have, (t , t+ ) =
tI

1 ([(t)] ) (t , t+ ) ,

where the inclusion is obvious, and the other one is a consequence of 1 . tI ([(t)] ) (t , t+ ) and properties of the inverse image Now, the sets 1 ([(t)] ) (t , t+ ), for t I , are all nonempty (each contains t), disjoint (since the corresponding equivalence classes are disjoint), and open (by the above shown). Since (t , t+ ) is connected, it must be that |I | = 1. This just says that all of ((t , t+ )) is contained in a single class of . On the other hand, (t , t+ ) = 1 (co Y ) = 1 ((R) co Y ) = 1 (a {x1 , x2 } co Y) and hence ((t , t+ )) = (1 (a {x1 , x2 } co Y )) = a {x1 , x2 } co Y where the last equality holds since a {x1 , x2 } co Y is contained in the image of . The statement of the lemma is now a direct consequence of this as x1 a {x1 , x2 } co Y. Next we present a description of inside of co Y. Lemma 4.15. We have (co Y co Y ) = (x1 , x2 ) co Y co Y : x2 x1 (Y ) . Proof. The inclusion is immediate from the denition of , Denition 4.2. We have to show the reverse inclusion. Let x1 , x2 co Y be given, and assume that x2 x1 (Y ). If x1 = x2 , there is nothing to prove. Hence, assume in addition that x1 = x2 . Note that this implies that |Y | > 1. Since x2 x1 (Y ), we can write
m

x2 x1 =
i=1 i xi 1 , x2

i i (xi 2 x1 ),

i with some i R and co Y , xi 1 x2 . Clearly, we may assume that i i always x1 = x2 . Choose > 0, and dene elements zk for k = 0, . . . , m k i i (xi 2 x1 ). i=1

zk := x1 + Note that for k = 1, . . . , m we have

k zk = zk1 + k (xk 2 x1 ).

Then it is easy to see that z0 = x 1 , zk a Y, k = 0, . . . , m. 30

By Lemma 2.5, co Y is an open subset of a Y . Now from x1 co Y , we can make the choice of > 0 such that zk co Y, We will next show that zk 1 zk , k = 1, . . . , m. k = 0, . . . , m.

Let k {1, . . . , m} be given. First consider the case that xk 1 = zk1 . Then we have k k k k zk = x k Y. 1 + k (x2 x1 ) a {x1 , x2 } co By Lemma 4.14, the set on the right side is contained in [xk 1 ] , and thus zk zk1 . Assume now that xk = z . Let be the line containing the points zk1 k 1 1 and xk , that is 1 (t) := txk t R. 1 + (1 t)zk1 , Let t , t+ be as in Lemma 4.11. Choose s (0, 1) such that st + (1 s) = 0 1 (this is always possible, namely s = 1 t is such). Then, using Lemma 4.11, (t ) (s, xk 1 ) = (t ) (s, (1)) = (st + (1 s) 1) = (0) = zk1 . By perspective invariance (t ) (s, xk 2 ) zk 1 . Also
k k k k (t ) (s, xk 2 ) zk1 = (t ) (s, x2 ) (t ) (s, x1 ) = (1 s)(x2 x1 ).

It follows that zk = zk 1 + k (t ) (s, xk 2 ) zk 1 1s a (t ) (s, xk Y [zk1 ] , 2 ), zk1 co

where the last inclusion holds by Lemma 4.14 since (t ) (s, xk Y by 2 ) co k Lemma 4.10 using also Lemma 4.11, (t ) (s, x2 ) zk1 as shown above, and k k (t ) (s, xk 2 ) = zk1 since x2 = x1 as assumed above, and we again have zk zk 1 . By transitivity we have z0 zm , i.e. x1 zm , and another application of Lemma 4.14 gives x2 = x1 + 1 i i (xi Y [x1 ] , 2 x1 ) = x1 + (zm x1 ) a {x1 , zm } co i=1
m

where again x1 = zm by the made assumption x1 = x2 .

31

4.4

Recovering from and G

We are now in position to establish the representation (4.4) of . Lemma 4.16. Let C be a component of G and Y (C ) its largest element. Then (Z (C ) Z (C )) = (x1 , x2 ) Z (C ) Z (C ) : x2 x1 (Y (C )) where Z (C ) is as in (4.3). Proof. As a rst step, we show the following statement: If x Z (C ), x co Y (C ), and xx , then x x (Y (C )). 1 If x = x , this is trivial. Hence, assume that x = x . Let z := 2 (x + x ). Since equivalence classes of are convex, we have z x . Recall that Y (C ) = Y C Y and co is monotone, implying that Z (C ) co Y (C ) co Y (C ). Hence x co Y (C ) and x co Y (C ), which by Lemma 4.10 1 , x ). Lemma 4.15 now implies that z x gives z co Y (C ) as z = x ( 2 (Y (C )). It follows that x x = 2(z x ) (Y (C )) which completes the rst step. We now prove the inclusion in (4.10). Let x1 , x2 Z (C ) with x1 x2 be given. Since, by the already proved Theorem 4.3(ii), for any Y C we have {Y, Y (C )} E , we can choose x Y (C ) with x1 x 1 , x2 co 1 and x2 x2 . Transitivity implies x1 x2 , and Lemma 4.15 thus gives x1 x2 (Y (C )). By the property shown in the rst step, also x1 x 1 , x2 x2 (Y (C )), and together
x2 x1 = (x2 x 2 ) + (x2 x1 ) + (x1 x1 ) (Y (C )).

(4.10)

Finally, we show the inclusion in (4.10). This is done by reversing the argument in the previous paragraph. Let x1 , x2 Z (C ) with x2 x1 (Y (C )) be given. Again choose x Y (C ) with x1 x 1 , x2 co 1 and x2 x2 . Again from the property shown in the rst step, we have x1 x1 , x2 x2 (Y (C )). Hence, also
x 2 x1 = (x2 x2 ) + (x2 x1 ) + (x1 x1 ) (Y (C )), and Lemma 4.15 implies that x 1 x2 . Transitivity gives x1 x2 .

Proof (of Theorem 4.3, relation (4.4)). The main observation to make is that K=
Y VK

co Y =
C Y C

co Y.

Hence, we can write K as the disjoint union K=

C component of G

Z (C )

where disjointness is easy to check: Assume x Z (C ) Z (C ). Then x co Y for some Y C and x co Y for some Y C . From the reexivity of we 32

have (x, x) and from the denition of G this yields {Y, Y } E , which implies that C = C . As a consequence, is the disjoint union =
C ,C components of G

(Z (C ) Z (C )) .

With a similar argument as above, the denition of G ensures that (Z (C ) Z (C )) = , and hence =
C component of G

C = C,

(Z (C ) Z (C )) .

The desired representation (4.4) follows now directly from Lemma 4.16.

4.5

Further properties of and G

The properties stated as (iii) and (iv ) of Theorem 4.3 can now be deduced using the already established relation (4.4). Proof (of Theorem 4.3, (iii)). Let Y1 , Y2 VK , and let C1 and C2 be the components which contain Y1 and Y2 , respectively. Moreover, let C be the component which contains Y (C1 ) Y (C2 ). We show that then Y1 Y2 C . This suces for the proof that is a congruence on the join-semilattice VK since if Y1 Y1 and Y2 Y2 , i.e., Y1 C1 and Y2 C2 , then both Y1 Y2 and Y1 Y2 are in C showing that Y1 Y2 Y1 Y2 . For j = 1, 2 choose xj co Y j , x Y (Cj ) with xj x j co j. Since is convex, we have 1 1 1 1 (x1 + x 1 ), (x2 + x2 ) = (x1 , x1 ) + (x2 , x2 ) . 2 2 2 2 By Lemma 4.10 we have 1 x1 ( , x2 ) 2 1 x ( , x ) 1 2 2 = = 1 (x1 + x2 ) co( Y 1 Y2 ) 2 1 (x + x Y (C1 ) Y (C2 ) 2 ) co 2 1

and hence Y1 Y2 Y (C1 ) Y (C2 ). This just means that Y1 Y2 C . Proof (of Theorem 4.3, (iv )). Let C be a component of G , Y C , and Y (C ) the largest element of C . If |Y | = 1, we have (Y ) = {0} and dir Y = {0}. Hence, in this case, equality (4.1) holds trivially. Assume that |Y | 2. The inclusion in (4.1) follows since is monotone, Y Y (C ), and (Y ) span co Y co Y = dir Y, 33

where the last equality follows by Lemma 2.6. To show the reverse inclusion, let u (Y (C )) dir Y be given. Choose x1 co Y , let > 0, and set x2 := x1 + u. Since u dir Y and x1 a Y , we have x1 + span{u} a Y . Since co Y is an open subset of a Y , we can choose > 0 so small that x2 co Y. We have x2 x1 = u (Y (C )), and hence (4.4) implies that x1 x2 . This 1 is enough to conclude that u = 1 u = (x2 x1 ) (Y ). It remains to prove (4.2). To this end, remember that {Y, Y (C )} E , i.e., there exist x1 co Y , x2 co Y (C ), with x1 x2 . By (4.4), we have x2 x1 (Y (C )), and hence x2 co Y + (Y (C )) co Y (C ).

By now all assertions of Theorem 4.3 are proved.

4.6

The converse construction

In this subsection let a relation and a map as in the statement of Theorem 4.4 be given. For each equivalence class C of , we denote its largest element by Y (C ) and set Z (C ) := co Y.
Y C

Moreover, we dene a relation on K as =

C class of

(x1 , x2 ) Z (C ) Z (C ) | x2 x1 (Y (C )) .

The main step is to show: Lemma 4.17. The relation is a convex congruence on K , i.e., ConCA K , and = , (Y (C )) = (Y (C )) for C class of . Proof. First of all, is an equivalence relation as a direct consequence of (Y (C )) being a linear subspace. Next, notice the following: If x co Y , x co Y , and xx , then x and x must both belong to the same of the sets Z (C ), and hence Y Y . To show that is a convex congruence, let (x1 , x 1 ), (x2 , x2 ) and s (0, 1) be given. Choose Yj , Yj , j = 1, 2, such that xj co Y j , x Yj . Then j co Yj Yj , and hence also Y1 Y2 Y1 Y2 . Let C be the class of which contains Y1 Y2 . Lemma 4.10 gives x := sx1 + (1 s)x2 co( Y1 Y2 ),
x := sx Y1 Y2 ), 1 + (1 s)x2 co(

and hence x, x Z (C ). Let Cj be the class which contains Yj . Since Yj Y (Cj ), it follows that Y1 Y2 Y (C1 ) Y (C2 ), and hence Y (C1 ) Y (C2 ) Y (C ). Since is monotone, we conclude that (Y (Cj )) (Y (C )), j = 1, 2. 34

We compute
x x = s(x 1 x1 ) + (1 s)(x2 x2 ) (Y (C1 )) + (Y (C2 )) (Y (C )),

and this shows that (x, x ) and therefore is a convex congruence. Next, we show that (Y (C )) = (Y (C )) whenever C is a class of . By the denition of , we have x2 x1 | x1 , x2 co Y (C ), x1 x2 (Y (C )), and hence (Y (C )) (Y (C )). Conversely, let u (Y (C )) be given. Choose x1 co Y (C ) a Y (C ), let > 0, and set x2 := x1 + u. Since (Y (C )) dir Y (C ), we have x2 a Y (C ). Since co Y (C ) is an open subset of a Y (C ), we may choose > 0 so small that x2 co Y (C ). Then, by the denition of , we have x1 x2 . It follows that u = 1 (x2 x1 ) (Y (C )). In order to establish the inclusion , it is enough to show that {Y1 , Y2 } E implies Y1 Y2 . This, however, is clear from the note made in the beginning of this proof. We next prove the reverse inclusion. First, let one element Y VK be given, and denote by C the class of which contains Y , i.e., C = [Y ] . By the hypothesis that [co Y + (Y (C ))] co Y (C ) = , we can choose x1 co Y , u (Y (C )), and x2 co Y (C ), such that x2 = x1 + u. By the denition of , we have x1 x2 . This shows that {Y, Y (C )} E . Let now Y1 , Y2 VK with Y1 Y2 , and denote by C the class of which contains Y1 (and hence also Y2 ). By what we just showed, {Y1 , Y (C )}, {Y2 , Y (C )} E . This implies that Y1 Y2 . The fact that can be computed by means of (4.5) is just the above denition of . The fact that is unique, is clear from (4.4). The remaining assertions of Theorem 4.4 now follow easily, as shown below. Proof (of Theorem 4.4, (4.6), (4.7)). Let Y VK , and let C be the class of which contains Y . Using (4.1) and Lemma 4.17, we obtain (Y ) = (Y (C )) dir Y = (Y (C )) dir Y, showing (4.6). For (4.7), it is enough to note that our denition of ensures Y 1 , Y2 V K . {Y1 , Y2 } E Y1 Y2

as already noted in the proof of Lemma 4.17 and Y 1 , Y2 C . {Y1 , Y2 } E co Y1 + (Y (C )) co Y2 =

which is easy to show unfolding the denitions.

4.7

Characterising inclusions

In this subsection we prove Theorem 4.7. Having Theorem 4.3, this is no longer dicult. Proof (of Theorem 4.7). The implication is immediate from the denition of Ej and j . Conversely, assume that E1 E2 and 1 2 . The

35

former implies that each component C2 of G2 is the union of components of G1 . This implies that Z (C2 ) = Z (C1 ).
C1 comp. of G1 C1 C2

Using this fact, the representation (4.4) for 1 and for 2 , monotonicity of 1 and 2 , and the assumption 1 2 , we compute 1 =
C1 comp. of G1

(x1 , x2 ) Z (C1 ) Z (C1 ) : x2 x1 1 (Y (C1 ))

C2 comp. of G2 C1 comp. of G1 C1 C2

(x1 , x2 ) Z (C1 ) Z (C1 ) : x2 x1 2 (Y (C2 ))

C2 comp. of G2

(x1 , x2 ) Z (C2 ) Z (C2 ) : x2 x1 2 (Y (C2 ))

= 2 .

Finitely presentable (positive, totally) convex algebras

We have already mentioned in the introduction that the free algebra Fn (CA) with n generators is the standard (n 1)-simplex in Rn . Moreover, Fn (PCA) is the n-dimensional simplex in Rn generated by the point 0 and the unit vectors e1 , . . . , en , cf. Example 2.8, and Fn (TCA) is the octahedron in Rn centered at 0 and having the 2n corners {ei :, i = 1, . . . , n}, cf. Example 2.9. Hence, all these free algebras are polytopes whose congruences we have fully described. Using this description, we can prove the theorem below. Before that, we recall some more notions: An algebra A is nitely generated if it is a quotient (under a congruence) of a free algebra FX (V ) for X a nite set. A congruence on an algebra A is nitely generated if there exists a nite subset R A A such that is the smallest congruence which contains R. Theorem 5.1. Let V be one of the equational classes CA, PCA, or TCA, and let A V be nitely generated. Then every congruence relation on A is nitely generated. Proof. The proof is carried out in three steps. First, we give a straightforward equivalent formulation of the theorem. Then we verify the assertion of the given reformulation for polytopes. After that we deduce the theorem, using that nitely generated free algebras are polytopes. Reformulation. We show that the assertion of the theorem is equivalent to the following statement: Let n N+ . Let 1 , 2 be congruence relations on the free

36

algebra Fn (V ) in V with n generators, and assume that 1 2 . Then there exists a nite subset R Fn (V ) Fn (V ), such that 2 is generated by 1 R. For the proof of equivalence, assume rst that the assertion of the theorem holds, and let Fn (V ), 1 , 2 be given as in the above statement. Consider the canonical projection : Fn (V ) Fn (V )/1 with ker = 1 . It is a surjective homomorphism, and hence induces an order isomorphism : ConV Fn (V ) | 1 ConV Fn (V )/1 . According to the theorem, as Fn (V )/1 is nitely generated, we can choose a nite subset R of Fn (V )/1 Fn (V )/1 , such that ( )2 is generated by R. Choose a subset R of 2 with |R | = |R| and ( )R = R. If ConV Fn (V ) contains 1 R , then ( ) R, and hence ( ) ( )2 . We are going to show still that then 2 and hence the reformulated assertion holds. Let (x1 , x2 ) 2 , then ( )(x1 , x2 ) ( )2 ( ) and hence there exists a pair (y1 , y2 ) with ( )(x1 , x2 ) = ( )(y1 , y2 ). But this means that (x1 ) = (y1 ) and (x2 ) = (y2 ) which due to ker = 1 implies (x1 , y1 ) 1 and (x2 , y2 ) 1 . But then (x1 , x2 ) 1 1 and since 1 we get (x1 , x2 ) . Assume now that the above reformulated statement holds and let A and ConV A be given according to the theorem. Choose n N+ and 1 ConV Fn (V ), such that A = Fn (V )/1 . Let be the corresponding surjective homomorphism of Fn (V ) onto A, and set 2 := ( )1 . Note that 2 is a convex congruence and 2 ker = 1 . Choose a nite subset R 2 such that 2 is generated by 1 R and set R := ( )R. Note that R is nite and R . If ConV A and R , then ( )1 1 R, and hence ( )1 2 . This implies that . Hence we have shown that is nitely generated, namely it is generated by R . Generators for convex congruences on polytopes. Let K be a polytope in Rn , and consider K as an algebra in CA. Let n N+ , and 1 , 2 ConCA K with 1 2 be given. We will now construct a nite subset R 2 , such that 2 is generated by 1 R. Dene R 2 as the collection of pairs of the following two kinds: (i) For each edge {Y1 , Y2 } E2 \ E1 choose a pair (a, b) 2 (co Y1 co Y2 ). Take this pair into R. (ii) For each Y VK , choose pairs (ai , bi ) 2 (co Y co Y ), i = 1, . . . , dim 2 (Y ) dim 1 (Y ), such that 1 (Y ) + span{bi ai | i = 1, . . . , dim 2 (Y ) dim 1 (Y )} = 2 (Y ). (5.1) Take these pairs into R. Let ConCA K , and assume that 1 R. Then E E1 since 1 , and E E2 \ E1 , since R. Hence E E2 . Similarly, since 1 we have (Y ) 1 (Y ), for all Y VK , and since R contains the pairs with (5.1), also (Y ) 2 (Y ), for all Y VK . By Theorem 4.7, 2 . Hence 2 is indeed generated by 1 R. Proof of the theorem. Let V be one of the classes CA, PCA, or TCA. Then as discussed above Fn (V ) is a polytope in euclidean space (endowed with the 37

usual vector operations). Let 1 , 2 ConV Fn (V ), and assume that 1 2 . We can consider Fn (V ) as an element of CA, and then j ConCA Fn (V ), cf. Lemma 3.7 and Lemma 3.11, respectively. By the second step, there exists a nite subset R 2 , such that 2 is generated by 1 R as an element of ConCA Fn (V ). This immediately implies that 2 is generated by 1 R as an element of ConV Fn (V ). We have thus veried the statement shown to be equivalent to the theorem in the rst step. Let us now recall the (algebraic) denition of a nitely presentable algebra, see, e.g., [AR94, 3.10]. The connection to the categorical concept of presentability is made in [AR94, Theorem 3.12], see also [ARV11, Corollary 11.33]. Denition 5.2. Let V be an equational class, and let A V . Let FX (V ) be the free algebra in V with the set X as free generators. A presentation of A is a pair (X, RX ) where X is a set and RX is a subset of FX (V ) FX (V ) such that A is isomorphic to the quotient FX (V )/ where is the smallest congruence containing RX . An algebra A is nitely presentable if there exists a presentation (X, RX ) with both X and RX nite. Obviously, an algebra is nitely generated if and only if there exists a presentation (X, RX ) with X nite. Hence, trivially, each nitely presented algebra is nitely generated. Having shown the previous theorem, we obtain as an immediate consequence that for CA, PCA, and TCA also the converse holds. Corollary 5.3. Let V be one of the equational classes CA, PCA, or TCA, and let A V . If A is nitely generated, then A is also nitely presentable. Hence a convex, positive convex, or totally convex algebra is nitely presentable if and only if it is nitely generated. Remark 5.4. A sucient condition in order that each nitely generated algebra of an equational class is nitely presentable, appeared recently in a categorical context. This condition is that the subclass of all nitely generated algebras of the equational class is closed under kernel pairs, cf. [BMS11, Lemma 3.21]. Formulated algebraically, it means that each congruence of any nitely generated algebra P is nitely generated as a subalgebra of P P . We see that the condition of being closed under kernel pairs is quite strong. In order to have nitely generated nitely presented, it is certainly sucient that each congruence of a free algebra F with nitely many generators is nitely generated as a congruence of F 6 . Remark 5.5. The question whether or not every nitely generated algebra of an equational class if nitely presented is classical. Some previously known examples where the answer is positive are commutative groups: Due to the classication of nitely generated commutative groups. semimodules over a Noetherian semiring : Due to a kernel pair argument, cf. [BMS11, Proposition 2.3].
6 Being nitely generated as a subalgebra is in general a signicantly stronger property than being nitely generated as a congruence, see the below Proposition 5.6.

38

commutative semigroups: This is essentially a particular case of the previous item (units can be adjoined easily), but has a longer history. It was rst shown by L.Redei, cf. [Re63, Satz 72] or [CP67, 9.3]. A short proof based on Hilberts basis theorem (i.e., a Noetherian argument) is given in [Fr68]. There are many equational classes where the answer is negative. For example, the equational class of all groups: Not every nitely generated group is nitely presented. In fact, there exist only countable many non-isomorphic nitely presentable groups, but already 20 non-isomorphic 2-generator groups. The (probably) simplest example of a nitely generated but not nitely presented group is the standard wreath product Z wr Z, cf. [Ro93, 14.1]. The question whether a specic nitely generated group is nitely presented may already be involved, see for example [RS76], [KC97], or [GS97]. In view of Remark 5.4, it is interesting to observe that for V being any of CA, PCA, TCA, the free algebras Fn (V ) for n N+ (except for F1 (CA) which contains only one element) always contain congruence relations which are not nitely generated as a subalgebra of Fn (V ) Fn (V ), cf. Example 5.8 below. This follows immediately from the next proposition, where we characterise the congruences on a polytope that are nitely generated as subalgebras of K K , showing that a kernel pair argument cannot be applied in these equational classes. Proposition 5.6. Let T be one of Tca , Tpca , or Ttca . Let K be a polytope in Rn . Then K, (f ) T CA ( PCA or TCA, respectively) with operations given by
n

f (pi )n (x1 , . . . , xn ) = i=1

i=1

pi x i ,

(pi )n i=1 T ,

where the linear combination on the right side denotes the usual vector operation on Rn . Let ConCA K ( ConPCA K or ConTCA K , respectively). Then is nitely generated as a CA- ( PCA- or TCA-, respectively) subalgebra of K K if and only if is closed as a subset of Rn Rn . Before we can prove this fact, we describe the closure of in terms of . Lemma 5.7. Let K be a polytope in Rn , and let ConCA K . Then Clos = (x1 , x2 ) K K : x2 x1 (ext K ) . Proof. Denote the set on the right side of the desired equality as 0 . By the representation (4.4) and monotonicity of , we have 0 . It is easy to show that 0 is closed (the limit of any converging sequence of elements of 0 is in 0 ) since K is closed and (ext K ) is closed as a linear subspace. Let (x1 , x2 ) 0 be given. Choose a point z co(ext K ). Then, since x1 , x2 co(ext K ), we obtain from Lemma 4.10 that z (s, x1 ), z (s, x2 ) co(ext K ), s (0, 1), and we have z (s, x2 ) z (s, x1 ) = (1 s)(x2 x1 ) (ext K ). 39

Now using (4.4), since ext K is a vertex of G and hence in some connected component C , and is monotone, we get z (s, x1 ), z (s, x2 ) , s (0, 1).

Letting s tend to 0, z (s, x1 ) tends to x1 and z (s, x2 ) to x2 and it follows that (x1 , x2 ) Clos . Hence 0 Clos and therefore 0 = Clos . This lemma tells us, in particular, that is closed if and only if = 0 . Proof (of Proposition 5.6). Necessity. Assume that R is a nite subset of K K , such that is generated by R as a subalgebra of K K CA (PCA or TCA, respectively). Then we have ( CA situation ) co R = co {0K } R ( PCA situation ) co R (R) ( TCA situation )

where is as in Proposition 3.8. For the PCA or TCA situation, note here that each PCA or TCA linear combination can be written as a convex combination by making use of the zero element 0K and the -images of the involved elements, and that for convex combinations the operations coincide with the usual vector operations. In any case, is a polytope in R2n and thus in particular closed. Suciency. Assume that ConCA K , and that is closed as a subset of Rn Rn . We will now show that is nitely generated as a CA-subalgebra of K K . Set d := n dim (ext K ), and choose a linear and surjective map : Rn Rd whose kernel equals (ext K ). Note that this means (x1 , x2 ) ker (x1 ) = (x2 ) x2 x1 {x | (x) = 0} x2 x1 Ker x2 x1 (ext K ).

We denote by the diagonal in Rn Rn . Since is closed, by Lemma 5.7 and the above, we have = Clos = (K K ) ker = (K K ) ( )1 (). The set K K is a compact (since K is compact) and convex subset of R2n , and one can easily show unfolding the denition of an extremal point that ext(K K ) = ext K ext K, so ext(K K ) is a nite set. Hence, K K is a polytope in R2n . Now we will employ some non-trivial geometric arguments from [Gr u03]. The diagonal is a linear subspace, and hence can be written as a nite intersection of halfspaces, i.e., it is a polyhedral set in the sense of [Gr u03, 2.6]. Since ( ) is linear and surjective, the inverse image of a halfspace is again a halfspace. Hence, ( )1 () is again a polyhedral set.

40

As an intersection of a polytope with a polyhedral set, is a polytope, cf. [Gr u03, 3.1]. Hence, has only nitely many extremal points and = co(ext ). This means that the nite set ext generates as a CA-subalgebra of K K . It remains to consider PCA and TCA congruences. Assume that ConPCA K (or ConTCA K , respectively) and is closed. Then also ConCA K , and hence by what we have proven so far is nitely generated as a CAsubalgebra of K K . In particular, it is therefore also nitely generated as a PCA- (TCA-, respectively) subalgebra of K K . Example 5.8. Let K Rn be a polytope with | ext K | 2, and let Y0 be a nonempty and proper subset of ext K . Consider the join-semilattice congruence on VK dened by specifying its equivalence classes to be Cy := {y } , y Y0 , C0 := VK \
y Y0

Cy ,

and the map : {Y (C )|C class of } Sub Rn dened as Y (Cy ) := {0}, y Y0 , Y (C0 ) := dir(ext K ) , cf. Remark 4.5. Applying Theorem 4.4, we obtain that the relation := (x, x) | x Y0 (x1 , x2 ) Z (C0 ) Z (C0 ) | x2 x1 dir(ext K ) is a congruence of K when K is considered as a convex algebra with the usual vector space operations of Rn . Let y0 Y0 . Since y0 is an extremal point, it cannot be an element of Z (C0 ). However, it can be approximated by elements from Z (C0 ): Choose y (ext K ) \ Y0 , and set x := y + (1 )y0 , (0, 1]. Then one can check from the denitions that (x , y ) and lim0 (x , y ) = (y0 , y ) , and this shows that is not closed. This example applies immediately to the free algebras Fn (CA), n 2, and Fn (PCA), n N+ , and shows that they contain congruences which are not nitely generated as subalgebras. For Fn (TCA), choose Y0 symmetric around the zero vector.

Conclusion

We fully describe the congruence lattice of a polytope in nite-dimensional euclidean space when considered as a convex algebra. This description applies in particular to the free algebras with a nite number of generators in this equational class, and hence claries the structure of nitely generated algebras. In particular, we see that each nitely generated convex algebra is nitely presented. The proofs are algebraic in their nature and use the geometry of euclidean space. We show that the equational classes of positive- or totally convex algebras (and their respective congruence lattices) are closely related with convex algebras. Using this relation, similar structure results for these equational classes follow. The presented results have an interpretation in a categorical context, since the considered equational classes are the categories of Eilenberg-Moore algebras associated with certain monads. 41

Acknowledgements
Our work on convex algebras was initiated by a related problem on which Ana Sokolova worked with Alexandra Silva and Stefan Milius. We thank Alexandra and Stefan for many insightful initial discussions.

References
[AR94] J.Adamek, J.Rosicky: Locally presentable and accessible categories, London Mathematical Society Lecture Note Series 189, Cambridge University Press, Cambridge, 1994.

[ARV11] J.Adamek, J.Rosicky, E.M.Vitale: Algebraic Theories, Cambridge Tracts in Mathematics 184, Cambridge University Press, Cambridge, 2011. [BK93] rger, R.Kemper: Normed totally convex spaces, Comm. Algebra 21 R.Bo (1993), no. 9, 32433258.

[BMS11] M.Bonsangue, S.Milius, A.Silva: Sound and complete axiomatizations of coalgebraic language equivalence, arXiv:1104.2803v1 [cs.LO]. [BS00] S.Burris, H.P.Sankappanavar: A course in Universal Algebra, the millenium edition, http://www.math.uwaterloo.ca/snburris/htdocs/ualg.html. original edition: Graduate Texts in Mathematics 78, Springer, New York, 1981. A.H.Clifford, G.B.Preston: The algebraic theory of semigroups. Vol. II, Mathematical Surveys, No. 7 American Mathematical Society, Providence, R.I., 1967. R.Diestel: Graph Theory, second edition, Graduate Texts in Mathematics 173, Springer, New York, 2000. Ernst-Erich Doberkat: Eilenberg-Moore algebras for stochastic relations, Inf. Comput. 204 (12), p. 1756-1781, 2006. Ernst-Erich Doberkat: Erratum and Addendum: Eilenberg-Moore algebras for stochastic relations, Inf. Comput. 206 (12), p. 1476-1484, 2008. P.Freyd: Redeis niteness theorem for commutative semigroups, Proc. Amer. Math. Soc. 19 (1968), 1003. tzer: Universal algebra, van Nostrand Company, Princeton, 1968. G.Gra nbaum: Convex Polytopes, second edition, Graduate Texts in MathB.Gru ematics 221, Springer, New York, 2003. V.Guba, M.Sapir: Diagram groups, Mem. Amer. Math. Soc. 130 (1997), no. 620. R.Kemper: Positively convex spaces, Appl. Categ. Structures 6 (1998), no. 3, 333344. R.Kemper: p-Banach spaces and p-totally convex spaces, Appl. Categ. Structures 7 (1999), no. 3, 279295. R.Kemper: p-Banach spaces and p-totally convex spaces II, Appl. Categ. Structures 8 (2000), no. 4, 579593. S.Krstic, J.McCool: The non-nite presentability of IA(F3 ) and GL2 (Z[t, t1 ]), Invent. Math. 129 (1997), no. 3, 595606. S.Mac Lane: Categories for the Working Mathematician, second edition, Graduate Texts in Mathematics 5, Springer, New York, 1998.

[CP67]

[Di00] [Do06] [Do08] [Fr68] [Gr a68] [Gr u03] [GS97] [Ke98] [Ke99] [Ke00] [KC97] [La98]

42

[Pu84] [Pu01a]

n: Regularly ordered Banach spaces and positively convex spaces, D.Pumplu Results Math. 7 (1984), no. 1, 85112. n: Absolutely convex modules and Saks spaces, J. Pure Appl. AlD.Pumplu gebra 155 (2001), no. 2-3, 257270.

n: The universal compactication of topological convex sets and [Pu01b] D.Pumplu modules, J. Convex Anal. 8 (2001), no. 1, 255267. [Pu03] [PR84] [PR85] [PR90] [Re63] n: Positively convex modules and ordered normed linear spaces, D.Pumplu J. Convex Anal. 10 (2003), no. 1, 109127. n, H.Ro hrl: Banach spaces and totally convex spaces. I, Comm. D.Pumplu Algebra 12 (1984), no. 7-8, 9531019. n, H.Ro hrl: Banach spaces and totally convex spaces. II, Comm. D.Pumplu Algebra 13 (1985), no. 5, 10471113. n, H.Ro hrl: Congruence relations on totally convex spaces, D.Pumplu Comm. Algebra 18 (1990), no. 5, 14691496. dei: Theorie der endlich erzeugbaren kommutativen Halbgruppen, HamL.Re burger Mathematische Einzelschriften, Heft 41, Physica-Verlag, W urzburg, 1963. English translation: The theory of nitely generated commutative semigroups, edited by N.Reilly Pergamon Press, Oxford-Edinburgh-New York, 1965. U.Rehmann, C.Soule: Finitely presented groups of matrices, Algebraic K theory (Proc. Conf., Northwestern Univ., Evanston, Ill., 1976), pp. 164169. Lecture Notes in Math., Vol. 551, Springer, Berlin, 1976. D.Robinson: A Course in the Theory of Groups, Graduate texts in mathematics, v.80, Springer, New York, 1993. hrl: Convexity theories. 0. Foundations, Appl. Categ. Structures 2 H.Ro (1994), no. 1, 1343. hrl: Convexity theories. VII. Discrete convex modules, Appl. Categ. H.Ro Structures 9 (2001), no. 3, 259283. W.Rudin: Functional Analysis, second edition, McGraw-Hill, New York, 1991. Z.Semadeni: Categorical methods in convexity, 1967 Proc. Colloquium on Convexity (Copenhagen, 1965) pp. 281307, Kobenhavns Univ. Mat. Inst., Copenhagen. A.Silva, A.Sokolova: Sound and complete axiomatization of trace semantics for probabilistic systems, Proceedings of MFPS XXVII, 2011, ENTCS 276 (2011), 291-311.

[RS76]

[Ro93] [R o94] [R o01] [Ru91] [Se67]

[SS11]

43