You are on page 1of 9

Partial Ordering

Notes by U. J. R uetschi, 20060412


Partial orders are special relations are special sets.
are structurally the same as directed acyclic graphs.
Synonyms: partial order, partially ordered set, poset, or simply order.
Partial orders can be used to formally capture many informal concepts such as parthood (the
hall is part of the house), precedence (rst peel the orange, then eat it), inferiority (3 is less
than 5), dependence (before assembling the car, build the engine), etc.
Properties of relations
A relation A A B from A to B is the mathematical way to express connections among the
elements of two sets A and B. We only consider the case A = B; then R A A is a relation
on set A. We write xRy as an abbreviation for (x, y) R.
Some typical properties of binary relations are (for all x, y, z in A):
reexive xRx
irreexive xRx
symmetric xRy y Rx
asymmetric xRy y Rx
identitive xRy yRx x = y aka antisymmetric
transitive xRy yRz xRz
connected xRy yRx aka total; all elems comparable
These properties are not independent of each other: an irreexive and transitive relation is
always asymmetric and every asymmetric relation is always irreexive. Note that irreexive is
the negation of reexive, but neither asymmetric nor antisymmetric is a negation of symmetric!
Types of posets
A poset (X, P) is a relation P on a set X having some of the above properties:
reexive poset, type identitive, transitive, and reexive
strict poset, type < identitive, transitive, and irreexive (implies asymmetric)
total = linear poset a reexive or strict poset that is connected
If the term poset is used without qualication, we have to guess its type from context. Note,
however, that reexive and strict orders can be constructed from each other.
In total orders, each element is comparable to all others.
This allows for a linear representation of the elements:

We write x y if x and y are comparable, i.e., either xP y or yP x,


and x y if x and y are non-comparable, i.e., neither xP y nor y P x.
Whenever xP y we say x is a part of / is contained in / is less than (or equal) y.
Frequently, the relation P is clear without explicit mention. We then simply refer to the
poset X instead of the poset (X, P).
Typical examples of posets: the well-known total (reexive or strict <) ordering on
numbers; the subset ordering a is a subset of b; the divisibility ordering n divides m.
Visualisation: Hasse diagrams
The elements of a poset (P, ) are represented by dots and if a b
a
b
c
d
e
f
then the dot for a is lower on the sheet than the dot for b. Dots a
and b that are in an immediate relation (i.e., there is no x such that
a x and x b; this is also known as the cover relation: a is a
lower cover of b and b an upper cover of a) are connected by a line.
The other pairs in correspond to paths and circles in the diagram.
Technically, they can be reclaimed by closure operations (explained
below). The reason posets are visualised bottom-up is that terms
like maximal element or greatest lower bound (both dened below)
have an intuitive graphical rendering. Hasse diagrams are also known
as poset diagrams.
Some terminology
The Hasse diagram is essentially a directed acyclic graph (or dag for short): just turn each edge
into an arrow pointing downwards. Now note that dags are a generalisation of trees: branches
may grow together. However, the usual terminology for trees (children, parents, siblings, depth,
etc.) is misleading when applied to posets.
We use the following terms instead: a is a (transitive) predecessor of b, c, d, e, f but an immediate
predecessor of only b and c. Similarly, c is an immediate successor of b and a (transitive) succes-
sor of b and a. The terms ancestor/descendent are also used in place of predecessor/successor.
What does the relation mean?
Anything consistent with the properties of an order relation may be considered a partial order.
Examples include:
geometrical containment (e.g., in the work by W. Kainz)
membership among building parts (my work about scene space)
being earlier in a sequence (e.g., in Kuipers 1979)
The rst two are specic instances of a general part-of relation; the third example is not related
to part-of: it represents partial knowledge of a sequence and illustrates the importance of partial
orders in knowledge representation. Partonomies are so important that they are the subject of
a whole branch of philosophy: mereology. Another important class of partial orders is that of
taxonomies, or type-of hierarchies:
Taxonomy
kind-of
inference of properties, inference of
class inclusion: e.g., robin is-a bird,
is-an animal; hence if animals breathe,
then so do robins
Partonomy
part-of
property inferences not allowed, even the
transitivity of part-of is controversial; they
do promote (but not promise) inference from
appearance to function
The nature of the ordering relation determines if a poset is reexive or strict, but it may also
questions transitivity. For example, is a cars doors handle a functional part of the car? But this
is a problem for philosophers and mereology.
Special elements
An element s X of a poset P = (X, ) is. . .
least x X : s x the least of all, at most one
greatest x s the greatest of all, at most one
minimal x X : x s the least of those comparable
maximal s x the greatest of those comparable
There can be any number of minimal/maximal elements: none, one, many.
If a least/greatest element exists, then it is also the only minimal/maximal element.
Note that the denition of minimal elements is equivalent to x X : s x s x.
Note that if both x, y X are minimal in P, then x y.
The set of all minimal elements of P is referred to as min(P) or min(X, )
and the set of all maximal elements is max(P) or max(X, ).
Special elements s X with respect to a subset T X:
lower bound of T s T smaller than any elem in T: x T: s x
upper bound of T T s greater than any elem in T: x T: x s
inmum

T or x y if T = x, y the greatest lower bound, g.l.b.
supremum

T or x y if T = x, y the least upper bound, l.u.b.
Suprema and inma are unique, if they exist at all.
For two subsets A and B, if

A

B exists, then it is the supremum of A B.


Subsets of posets
Given a poset (X, P), any subset Y X is again a poset under the relation P[
Y
, which is the
restriction of P to Y (remove all pairs in P containing an element not in Y ).
If such a subset is linearly ordered, then its called a chain. A chain M is maximal if there is no
other chain C with M C (i.e., M is contained in no other chain). A subset of X consisting of
elements that are mutually non-comparable is called an anti-chain.
Let p X, then the lower closure of p is the set of all elements that are contained in or are
less than p. This notion can be extended for subsets S X. The dual is called upper closure.
lower closure: p := x [ x p S := x [ x p, p S =

pS
p
upper closure: p := x [ p x S := x [ p x, p S =

pS
p
Alternative notation:
p := x [ x < p and p := p p
p := x [ p < x and p := p p
Lower closure is sometimes also called ideal or down-set.
Upper closure is sometimes also called lter or up-set.
If a lower closure of a set S equals S, then its called a lower set. Similarly,
if an upper closure of a set S equals S, then its called an upper set.
An interval [p, q] := x [ p x q consists of all elements between p and q.
Note that [p, q] = p q.
Measures in and of posets
The number of elements in a poset is called its order. A poset of order N can have at most
1
2
N (N 1) pairs in its order relation. This limit is achieved for linearly ordered sets. True
partial orders have smaller order relations.
Other than for trees, there is no natural concept of levels in posets; nevertheless, we may dene
a posets height to be the size of its longest chain. Similarly, the width is the size of its largest
anti-chain.
The rank of the elements p in a poset P is obtained by assigning rank 0 to all minimal elements
in P, then removing these elements, assigning rank 1 to the minimal elements in that smaller
poset, and so on [Schroder 2003]:
rank(p) :=

0 if p is minimal in P
n if p is minimal in P ` q [ rank(q) < n
The rank of an element is the length of the longest chain between that element and a minimal
element. It is possible for adjacent elements to have ranks that dier by more than one. In the
example poset below we have rank(e) = rank(f) = 0, rank(d) = rank(c) = 1, rank(b) = 2, and
rank(a) = 3.
If we substitute maximal for minimal in the denition of rank above, then we get a rank
measured from the top of the poset instead of from the bottom. Therefore, each element p can
be characterised by two ranks, rank
min
(p) and rank
max
(p) [Joslyn 2004].
The dimension of a poset is the minimal number of linear extensions such that their intersection
is the poset (see below).
Dilworths theorem. A poset (X, P) can be decomposed into w chains such that X = C
1
C
2

C
w
where w is the posets width; similarly, it can be decomposed into h antichains such that
X = A
1
A
2
A
h
where h is the poests height.
Closure and the skeleton
The Hasse diagram only represents the skeleton of the poset, that is, a
a
b
c
d
e
f
minimal subset of the order relation P from which the whole of P can
be recovered by applying closure operations:
transitive closure t(P) :=

(n=1,2,3...)
t
n
(P)
where t
n
(P) := P t
n1
(P) and t
0
(P) := P
If the graph is reexive (instead of strict), also apply
reexive closure r(P) := P (x, x) [ x X
Reexive closure is adding circles to all nodes in the dag and transitive closure is adding all
paths in the dag as edges (see gure). The original graph may also be called the transitive
reduction graph because it includes only immediate relationships and drops all those with
intermediate elements. Hasse diagrams thus are special dags. There is no closure operation
corresponding to antisymmetry (or asymmetry); rather, it is the property that forbids cycles.
Another closure operation on posets is called st-closure and adds a greatest element (top)
and a least element (bottom), if they do not yet exist.
Linear extension
The elements of any poset P can be arranged in a sequence such that whenever x y in P
then x precedes y in the sequence. For example, a set of tasks (where some tasks depend on
the completion of some other tasks) have to be solved by just one workforce, they have to be
arranged in such a sequence. This process is called topological sorting or linear extension.
x < y = x
L
y = x < y x y
A linear extension is an embedding of a poset into a total order over the same elements, such
that the left implication holds (
L
means precedes in the linear extension L). But note that
x
L
y only tells us that x > y is not in the original poset, and x > y means x < y or x y.
Topological sort algorithm. For topologically sorting a poset P, proceed as follows:
1. choose a minimal element m from P, output m, remove m from P
2. continue with step 1 until P is empty
Knuth presents a very fast algorithm and implementation for the computation of a topological
sort of a poset in Fundamental Algorithms, volume 1 of The Art of Computer Programming.
Counting linear extensions. Every poset that is not a chain has more than one linear extension.
Let L(P) be the number of linear extensions for a given poset P. Then we have [Schroder 2004]
L(P) =

m minimal
L(P ` m)
If P is our example poset and L
xy...
is a shorthand notation for L(P ` x, y, . . .), then
L(P) = L
e
+L
f
= (L
ec
+L
ef
) +L
fe
=

(L
ecb
+L
ecf
) + (L
efc
+L
efd
)

+ (L
fec
+L
fed
)
=

L
ecbf
+ (L
ecfb
+L
ecfd
)

(L
efcb
+L
efcd
) +L
efdc

(L
fecb
+L
fecd
) +L
fedc

= L
ecbfd
+L
ecfbd
+L
ecfdb
+L
efcbd
+L
efcdb
+L
efdcb
+L
fecbd
+L
fecdb
+L
fedcb
= L
ecbfda
+L
ecfbda
+L
ecfdba
+L
efcbda
+L
efcdba
+L
efdcba
+L
fecbda
+L
fecdba
+L
fedcba
= 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 = 9
Posets with no elements are vacuously linearly ordered and thus have exactly one linear exten-
sion, the empty set. Note that the superscripts in the second-to-last line in the example above
are precisely the linear extensions: e < c < b < f < d < a, e < c < f < b < d < a, etc.
A poset that is a chain has just one linear extension (namely itself); a poset that is an antichain
of n elements has n! linear extensions. All other posets with n are between these bounds. (Why?
Because every pair (x, y) in a poset means that there cannot be a linear extension in which y
precedes x, thus reducing the number of possible linear extensions.)
The restriction of a linear extension to a subposet is again a linear extension of the subposet; and
any linear extension of a subposet can be lifted to a linear extension of the parent poset.
Dimension of a poset
The intersection of all linear extensions of a given poset results in the original poset. Frequently,
the intersection of just a few linear extensions is enough to reconstruct the original poset. Such a
set of linear extensions is called a realiser of the poset. The minimal number of linear extensions
whose intersection is the poset they extend is called the posets dimension. (Remember that an
ordering is a set of pairs; the intersection of two orderings consists of those pairs that are in both
orderings.)
Continuing the previous example, the three linear extensions
a
b
c
d
e
f
L
1
: e < c < b < f < d < a
L
2
: e < f < d < c < b < a
L
3
: f < e < c < b < d < a
form a realiser of the poset, but
L
1
: e < c < b < f < d < a
L
2
: f < e < d < c < b < a
is also a realiser and it is the one with the fewest linear extensions.
Consequently, the posets dimension is dim(P) = 2.
If L
1
, L
2
is a realiser of some poset P, then x < y in P i x
L
1
y and x
L
2
y. This is
making realisers interesting for representing posets on computers: testing the order relation
x y can be done in O(dimP).
Unfortunately, posets exist with few elements but high dimension.
b
1
b
2
b
3
b
4
a
1
a
2
a
3
a
4
The upper bound for a poset (X, P) with [X[ = n elements is given
by dim(X, P)
n
2
|. The so called standard example S
n
is a worst
case: it has 2n elements and dimS
n
= n; the gure shows S
4
.
If (X, P) is a poset with dim(X, P) = n and Y X, then dim(Y, P[
Y
) n,
i.e., a posets dimension cannot be increased by removing elements.
An element is least (greatest) i it is least (greatest) in all linear extensions of a realiser.
Proof: by the denition of least and greatest element.
An element is minimal (maximal) if it is least (greatest) in at least one linear extensions of a
realiser of the poset. The converse does not hold (think of antichains for a counterexample).
Proof: Let m least in L, then m L, implying m p m p for all p P, which states
precisely that m is minimal in P (P a poset, L a linear extension thereof).
If R = L
1
, L
2
, . . . L
n
is a realiser of a poset (X, P) and Y X,
then S = L
1
[
Y
, L
2
[
Y
, . . . L
n
[
Y
is a realiser of the sub-poset (Y, P[
Y
).
Tree-shaped posets are 2-dimensional.
Proof: If T is a tree with subtrees T
1
, T
2
, . . . T
n
, then (T, T
1
, T
2
, . . . T
n
) and (T, T
n
, T
n1
, . . . T
1
)
are both linear extensions of the poset and their intersection contains the pairs (T, T
1
),
(T, T
2
),. . . (T, T
n
), that is, the original tree-shaped poset.
Realisers are nice, but how can they be compted?
Lattices
A non-empty poset in which any two elements x and y have a least upper bound (x y) and a
greatest lower bound (x y) is called a lattice.
In lattices, least upper bound is a binary operation on the elements of the lattice L and usually
called join and similarly for greatest lower bound, which is usually called meet. Both operations
are idempotent (x x = x), commutative, and associative.
: L
2
L : (x, y) sup(x, y) = join(x, y) = x y
: inf meet x y
Finite lattices always have a greatest and a least element; frequent synonyms are:
greatest: top unit 1
least: bottom zero 0
A prime example of a lattice are the subsets of a given set S under set inclusion: (power S, ).
Here, 0 = and 1 = S, join corresponds to union and meet to intersection .
An upper (lower) semilattice is a lattice with only joins (meets).
A truncated lattice is a lattice without greatest and least element.
Lattices can also be dened algebraically in terms of the join and meet operations.
Then the (reexive) ordering relation can be recovered by these equivalences:
x y x = x y and x y x = x y
An interesting application of lattices is the concept lattice (dt. Begrisverband), developed by
Wille [1982]. Lattices also occur in the analysis of the semantics of programming languages.
Quasi orders
A binary relation that is reexive and transitive (but not necessarily antisymmetric) is called
a quasi order. The semantic relation of hyponymy is an example: synonymy can be dened as
symmetrical hyponymy (A is synonymous to B i A is a hyponym of B and B is a hyponym of
A). But then, antisymmetry would imply that A and B are equivalent, thus we do not require
the relation to be antisymmetric [Heller 1994].
Any quasi order can be turned into a partial order by taking the equivalence classes induced by
the synonymy relation: [a] = x P [ a synonymous b , then dene [a] [b] in the poset i
a b in the quasi order.
Supplements
If a relation is irreexive and transitive, then it is asymmetric.
Proof: Let x, y P, xRy and y Rx (that is, not asymmetric). Then xRx by transitivity, which
contradicts irreexivity, hence it must be asymmetric.
If a relation is asymmetric, then it is irreexive.
Proof: Let x P, xRx (that is, not irreexive). Then xRx implies xRx by asymmetry, which is
a contradiction. Hence it R must be irreexive.
A note on proofs by contradiction.
If some statement W holds, then this means that true W. But then the contrapositive
holds, that is, W false. Here false is just another word for contradiction. Therefore, to
prove statement W, we can assume that W is wrong and show that this leads necessarily to a
contradiction. If the statement to prove is of the form A B, then we have to show that B
and A results in a contradiction. Reason: (A B) false A B false.
Antisymmetry of strict posets is vacuous.
This is because xRy and y Rx implies by transitivity xRx, which contradicts irreexivity, hence
the precondition of antisymmetry never holds (and ex falso quodlibet, so we dont care).
Mappings between orders. . .
have to preserve the order relation in the sense that whenever xRy in the domain then x

in
the codomain. Such a function is said to be monotone.
Hasse diagrams are named after the German mathematician Helmut Hasse (18981979).
More non-mathematical examples of partial orders: (1) Eric Levenezs Unix and programming
languages timelines are nice examples of partial orders represented graphically as Hasse diagrams
(but from left to right), http://www.levenez.com/. (2) Dependence among chapters in text-
books. (3) American architect Christopher Alexander claims in his paper A city is not a tree
(Architectural Forum, 1965) that the fundamental units of a historically grown city (like a side-
walk, a neighbourhood, a drugstore, a newsrack, a trac light, a district, or a square) are not
strictly hierarchical, but rather overlap, which results in an upper semilattice structure with the
city as its greatest element. In contrast, articially designed cities exhibit a strict hierarchy, and
thus are missing the essential ingredient that Alexander thinks makes natural cities so much
more enjoyable.
Orders can be constructed from orders: the dual order is the order we get by the converse
relation, the union of two orders is again an order (but the union of two lattices is no longer a
lattice), even a product of orders can be dened based on the Cartesian product of the carrier
sets. Normal completion is an operation that transforms a poset into a lattice by adding some
strategic elements. And nally, strict orders can be derived from reexive orders and vice versa.
Category Theory
Posets can be considered categories and there is a category of posets and monotone mappings.
The former is particularly helpful in understanding category theory because universal mapping
properties are so obvious. Products are joins, coproducts are meets, the terminal object is the
least element, if it exists, and dually, the initial objects is the greatest element, if it exists.
References and Bibliography
Alexander, Christopher: A city is not a tree.
Architectural Forum, 1965.
Birkho, Garrett: Lattice Theory, 3rd edition.
American Mathematical Society, Providence, Rhode Island, 1967.
Hein, James L.: Discrete Mathematics, 2nd edition.
Jones and Bartlett Publishers, Sudbury, Massachusetts, 2003.
Joslyn, Cli: Poset Ontologies and Concept Lattices as Semantic Hierarchies.
Conceptual Structures at Work, LNAI 3127, Springer-Verlag, 2004.
Kainz, Wolfgang: Application of lattice theory to geography.
Proc. 3rd Int. Symposium on Spatial Data Handling, Sydney, 1988.
Kuipers, Benjamin J.: On representing commonsense knowledge.
Associative Networks: The Representation and Use of Knowledge by Computers,
Academic Press, New York, 1979.
Schroder, Bernd S. W.: Ordered Sets.
Birkhauser, Boston, 2003.
Trotter, William T.: Combinatorics and partially ordered sets: dimension theory.
Johns Hopkins University Press, Baltimore.
Varzi, Achille: Mereology.
Stanford Encyclopedia of Philosophy, http://plato.stanford.edu/
Wikipedia, entries on orders and lattices.
http://www.wikipedia.org/
Wille, Rudolf: Restructuring lattice theory: An approach based on hierarchies of concepts.
Pages 445470 in Ordered sets, edited by I. Rival, Dordrecht: Reidel, 1982.

You might also like