Quick viewing(Text Mode)

On the Fractional Dimension of Partially Ordered Sets

On the Fractional Dimension of Partially Ordered Sets

On the Fractional Dimension of

Partially Ordered Sets

STEFAN FELSNER

Bel l Communications Research South Street Morristown NJ USA and Technische Uni

versitat Berlin Fachbereich Mathematik Strasse des Juni Berlin Germany

Email address felsnermathtub erlinde

WILLIAM T TROTTER

Bel l Communications Research South Street L Morristown NJ USA and Depart

ment of Arizona State University Tempe AZ USA

Email address wttb ellcorecom

Date Septemb er

Abstract We use a variety of combinatorial techniques to prove several theorems concerning fractional

dimension of partially ordered sets In particular we settle a conjecture of Brightwell and Scheinerman

by showing that the fractional dimension of a p oset is never more than the maximum degree plus one

Furthermore when the maximum degree k is at least two we show that holds if and only if one

of the comp onents of the p oset is isomorphic to S the standard example of a k dimensional

k +1

p oset When w  the fractional dimension of a p oset P of width w is less than w unless P contains

S If P is a p oset containing an antichain A and at most n other p oints where n  we show that the

w

fractional dimension of P is less than n unless P contains S If P contains an antichain A such that all

n

antichains disjoint from A have size at most w  then the fractional dimension of P is at most w and

this b ound is b est p ossible

Mathematics Sub ject Classication A C

Key words Partially p oset dimension fractional dimension degree

Intro duction

In this pap er we consider a partial ly ordered P as a pair X P and refer to X as the

ground set and P as the partial order We nd it convenient to use the short form poset

for a and we use the term subposet to refer to a p oset induced by a

of the ground set

Let P X P b e a p oset and let F fM M g b e a multiset of linear extensions

1 t

of P Brightwell and Scheinerman call F a k fold realizer of P if for each incomparable

pair x y there are at least k linear extensions in F which reverse the pair x y ie

jfi i t x y in M gj k The fractional dimension of P denoted by fdimP is

i

then dened in as the least real numb er q for which there exists a k fold realizer

F fM M g of P so that k t q it is easily veried that the least upp er b ound

1 t

of such real numb ers q is indeed attained Using this terminology the dimension of P

denoted by dimP is just the least t for which there exists a fold realizer of P It

follows immediately that fdimP dimP for every p oset P

In this pap er we will nd it convenient to use the following probabilistic interpretation

of this concept Let P X P b e a p oset and let F fM M g b e a multiset of

1 t

linear extensions of P We consider the linear extensions of F as outcomes in a uniform

sample space For an incomparable pair x y the probability that x is over y in F is

FRACTIONAL DIMENSION

given by

fi i t x y in M g Prob x y

i F

t

The fractional dimension of P is then the least rational numb er q so that there exists

a multiset F fM M g of linear extensions of P with Prob x y q for

1 t F

every incomparable pair x y

For each n the height two p oset containing n minimal elements fx x x g

1 2 n

and n maximal elements fy y y g with x y if and only if i j for all i j

1 2 n i j

n will b e denoted by S The p oset S is known as the standard example of a

n n

ndimensional p oset As noted in fdimS dimS n

n n

This pap er is organized as follows In Section we intro duce some terminology nec

essary for our theorems and review facts ab out lexicographic sums In Section we

formulate the results relating the fractional dimension of a p oset to its degree In Sec

tion we present a key technical lemma involving the algorithmic transformation of linear

orders into linear extensions and present the pro of for the main theorem of Section In

Section we present the forbidden subp oset characterization of the degree inequality In

Sections and we derive two other forbidden subp oset characterizations of inequalities

for fractional dimension In b oth cases equality is only p ossible when P contains a full

dimensional standard example as a subp oset

In Section we present an inequality relating the fractional dimension of a p oset

P X P to the width of the subp oset induced by X A where A is an antichain

The b ounds obtained in this case are slightly stronger than the corresp onding b ounds for

dimension We also prove that our b ounds are b est p ossible In Section we characterize

Hiraguchis inequality for fractional dimension Finally in Section we prop ose some

new problems for fractional dimension

Notation and Terminology

Let P X P b e a p oset We denote the set of incomparable pairs of P by incP Now

let F fM M M g b e a nonempty multiset of linear orders on X We call F a

1 2 t

multirealizer of P if P F ie each M is a linear extension of P and prob x y

i

F

for every x y incP

When F is a multirealizer of P we dene the value of F denoted value F as the

P

least rational numb er q so that prob x y q for every x y incP The

F

fractional dimension of P is just the least rational q so that P has a multirealizer F

of value q

Recall that an incomparable pair x y in a p oset P X P is critical if any p oint

less than x is less than y and any p oint greater than y is greater than x We denote the set

of all critical pairs of P by critP The following elementary prop osition rst noted by

Rabinovitch and Rival is stated formally to emphasize the fundamentally imp ortant

role played by critical pairs in dimension theory

PROPOSITION Let P X P be a poset and let F be a multiset of linear exten

sions of P Then F is a multirealizer of P if and only if prob x y for every

F

x y critP 

We say that a linear order L reverses the critical pair x y if x y in L Similarly we

say L reverses the set S crit P if L reverses each pair in S

The following remark is noted in

S FELSNER AND W T TROTTER

PROPOSITION Let P X P be poset and let q be a rational number Then

fdimP q if and only if there exists a multiset F of linear extensions of P so that

Prob x y q whenever x y critP 

F

Let P X P b e a p oset and let D fY Y Q x X g b e a family of p osets

x x x

indexed by the p oint set X of P The lexicographic sum of D over P is the p oset S whose

p oint set consists of all pairs of the form x y where x X and y Y The partial

x x

in S if and only if x y order on S is dened by setting x y

2 x 1 x

2 1

x x in P or

1 2

in Q y x x and y

x x 1 2 x

1 2 1

Both dimension and fractional dimension are well b ehaved with resp ect to lexicographic

sums The statement for dimension is due to Hiraguchi and the statement for frac

tional dimension is given in

PROPOSITION Let S be the lexicographic sum of D fY x X g over P

x

X P Then

dimS maxfdimP max dimY g

xX x

fdimS maxffdimP max fdimY g 

xX x

A p oset S is decomposable with resp ect to lexicographic sums if it is isomorphic to a

lexicographic sum of a family D fY x X g over a p oset P X P where jX j

x

and jY j for at least one x X Otherwise we say S is indecomposable

x

Given a p oint x in a p oset P X P let D x denote the set of all p oints which are

P

less than x in P and U x the set of p oints which are greater than x in P Whenever

P

the meaning is clear from the context we will drop the subscripts and just write D x

and U x Two p oints x and y are said to b e interchangeable in a p oset P X P if

D x D y and U x U y Thus x and y are interchangeable if and only if b oth

x y and y x are critical pairs Note that a p oset with an interchangeable pair is

decomp osable with resp ect to lexicographic sums unless it is a element antichain

Fractional Dimension and Degree

For a p oint x in a p oset P X P dene the degree of x denoted deg x as the numb er

of p oints comparable but not equal to x in P Then let P denote the maximum degree

of P ie P maxfdeg x x X g

The following theorem is proved by Brightwell and Scheinerman in

THEOREM If P X P is a poset then fdimP P 

In the next section we prove the following improved b ound which was conjectured in

THEOREM If P X P is a poset and is not an antichain then fdimP

P 

As noted in fdimS n S for all n so Theorem is b est p ossible

n n

In Section we show that when the the maximum degree of a connected p oset P is

at least the inequality in Theorem is strict unless P is the standard example of

dimension P This result has much the same avor as Bro oks theorem which

asserts that the chromatic numb er of a connected graph is at most one more than the

maximum degree with equality holding if and only if the graph is a complete graph or an

o dd cycle It also provides another instance in which the role played by standard examples

FRACTIONAL DIMENSION

in dimension theory is analogous to the role played by complete graphs in the theory of

graph coloring This parallel is explored in greater detail in the monograph

We will in fact prove a result that is slightly stronger than Theorem Given a p oint

x in a xed p oset P X P let D x D x fxg Then let deg x jD xj and

D

dene P maxfdeg x x X g The set U x and the quantities deg x P

D U

D U

are dened dually

With this notation we can now state the result we actually prove

THEOREM If P X P is a poset then fdimP P

D

Note that Theorem has Theorem as an immediate corollary Since fractional

dimension is dual ie a p oset and its dual have the same fractional dimension the

following result will also b e an immediate corollary to Theorem

COROLLARY If P X P is a poset then fdimP minf P P g

D U

Linear Orders and Linear Extensions

In Furedi and Kahn showed that dimension theory problems could b e formulated in

terms of linear orders on the ground set instead of linear extensions Let P X P b e a

p oset and let L b e any linear order on the ground set X When S X and x X S we

write x S in L when x s in L for every s S We let C L fx y x y crit P

and x D y in Lg Furedi and Kahn noted that the following statement holds

PROPOSITION Let P X P be a poset and let L be any linear order on X Then

there exists a linear extension M of P with C L C M 

Given a p oset P X P and a linear order L on X we let C L C L fx y x y

is a critical pair x D y in L and jD y j Pg

D

LEMMA Let P X P be a poset with no interchangeable pairs and let L be any

linear order on X Then there exists a linear extension M of P with C L C M

Proof In order to simplify arguments to follow we present an algorithmic pro of We

describ e a deterministic algorithm LINEX which takes an arbitrary linear order L on X

as input and outputs a linear extension M LINEX L satisfying the conclusion of the

lemma

Set L L for j we obtain L from L by moving an element of X to an

0 j +1 j

alternate p osition according to rules which we describ e b elow The pro cedure halts when

we have a linear extension M L satisfying the conclusion of the lemma We divide the

j

pro cess into two phases The rst phase just transforms L into a linear extension N with

C L C N and C L C N so in this phase we will already b e proving a mo dest

extension of Prop osition

Phase Obtaining a Linear Extension This phase ends when L is a linear exten

j

sion of P If it is not among all pairs u v with u v in P and v u in L cho ose one

j

for which the numb er of p oints b etween them in L is minimum If there is more than one

j

such pair we cho ose one for which v is as low as p ossible in L

j

Claim If w is b etween u and v in L then w is incomparable with b oth u and v

j

Proof Supp ose to the contrary that w is b etween u and v in L but that w is compa

j

rable to u Now w u in L Also there are fewer p oints b etween w and u than b etween

j

u and v If u w in P then we contradict our choice of the pair u v as we would

S FELSNER AND W T TROTTER

prefer the pair u w On the other hand if w u in P then w v in P and we would

prefer v w to u v The argument when a p oint b etween u and v is comparable to v is

dual

Now form L from L by moving v to a p oint immediately over u From Claim

j +1 j

we note that L preserves exactly one more comparable pair of P than do es L This

j +1 j

insures that Phase will terminate

Claim C L C L

j j +1

Proof Supp ose to the contrary that there exists a critical pair x y C L C L

j j +1

and supp ose v was placed over u in the construction of L Then v D y v x in L

j +1 j

but x v in L However u v in P now implies that u y in P Now x y C L

j +1 j

requires x u in L This already gives x v in L a contradiction

j j +1

The argument for Claim also establishes the next claim which we state for emphasis

Claim C L C L

j j +1

We may now assume that Phase terminates with a linear extension N of P

Phase Doing More Set N N This phase pro duces a sequence fN j g of

0 j

linear extensions of P and terminates when M LINEXL N satises C M C M

j

At each intermediate stage we will preserve the key prop erties that C N C N and

j j +1

C N C N

j j +1

Supp ose that there is some pair x y C N C N Cho ose one such pair and

j j

note that x y in N but z x in N for each z X with z y in P

j j

Claim No p oint b etween x and y in N is comparable to y

j

Proof Supp ose to the contrary that u is b etween x and y in N but that u is comparable

j

to y Since N is a linear extension of P we know that u y in P However this

j

contradicts the assumption that x y C L

Form the linear order N from N by moving y immediately b elow x In view of

j +1 j

Claim we know that N is a linear extension of P

j +1

Claim C N C N

j j +1

Proof Supp ose to the contrary that u v C N C N Then it follows that

j j +1

u y Since jD y j is maximum y is maximal element in P However since y v is a

critical pair we know D y D v and thus D y D v So v is also a maximal p oint

and U y U v We conclude that y and v are interchangeable

Exactly the same argument establishes the following claim

Claim C N C N

j j +1

Phase terminates b ecause C N is a prop er subset of C N as the second set

j j +1

contains x y but the rst do es not This completes the pro of of the lemma 

We now present the pro of of Theorem

Proof Clearly we may assume that P is indecomp osable with resp ect to lexicographic

sums in particular P has no interchangeable pair

Let jX j n and set t n Then consider the set F fL L L g of all linear

1 2 t

orders on the ground set X Note that for a p oint x and a set S with x S the fraction of

linear orders in which x is higher than all p oints in S is exactly jS j It follows that

FRACTIONAL DIMENSION

a critical pair x y with jD y j k is in C L for njD y j dierent p ermutations

L Also a critical pair x y with jD y j k is in C L for njD y j dierent

p ermutations L

Apply the algorithm LINEX dened in the previous section and let S fM M

1 2

M g where M LINEX L for each i t Lemma implies Prob x y

t i i S

jD y j when jD y j k and Prob x y jD y j when jD y j k

S

Thus Prob x y k for every critical pair x y in P 

S

We remark that lo cal exchange arguments of the typ e used in this argument have also

b een applied in

Characterizing the Case of Equality

As noted by Brightwell and Scheinerman the standard example of an ndimensional p oset

satises fdimP P so Theorem is b est p ossible In this section we will

show that the standard examples are the only connected p osets for which the inequality

in Theorem is tight Before presenting the details of the argument we need a little

more notation

Let P X P b e a p oset and let F b e a multirealizer of P with valueF r We

dene ess F fx y critP prob x y r g The pairs in ess F are called

F

the essential pairs of F

We call a multirealizer F of a p oset P X P an optimal family for P if valueF

fdimP Let F and F b e multirealizers of P We denote by F F the union of the

1 2 1 2

two multisets dened so that if a linear order app ears n times in F and m times in F it

1 2

app ears n m times in F F The following elementary prop osition is stated formally

1 2

for emphasis

PROPOSITION If F and F are optimal multirealizers of a poset P then F F

1 2 1 2

is also an optimal multirealizer Furthermore

ess F F ess F ess F 

1 2 1 2

We may then say that a critical pair x y critP is an essential pair of P if x y

ess F for every optimal multirealizer F We let ess P denote the set of all essential

pairs of P Note that ess P is always nonempty when P is not a chain When L is a

linear order on X we let ess L C L ess P and ess L C L ess P The next

statement is an immediate consequence of the preceding denitions

PROPOSITION Let P X P be a poset and let F be an optimal multirealizer

for P If M M F and ess M ess M then ess M ess M 

1 2 1 2 1 2

We are now ready to present our characterization of Theorem

THEOREM If P X P is a poset with P then fdimP P

D D

unless P contains a standard example of a poset of dimension P

D

Proof Let P X P b e a p oset k P fdimP k and jX j n Without

D

loss of generality we may assume that P is indecomp osable with resp ect to lexicographic

sums in particular we may assume P has no interchangeable pair Again let S denote

the family of linear extensions of P constructed in the pro of of Theorem given in the

preceding section Then S is optimal We now derive some elementary statements ab out

the essential pairs reversed by a linear extension in S

S FELSNER AND W T TROTTER

Claim Let L b e any linear order and let M LINEX L Then ess L ess M

Proof Let x y crit P and let L b e a random linear order on X The probability

that x y b elongs to ess L is at least k Should this inequality not b e tight

then it follows that the pair x y do es not b elong to ess P If there is any linear order

L on X for which there exists a pair x y ess LINEX L ess L then we would

conclude that Prob x y k and thus x y ess P

S

Claim If x y ess P then jD y j P

D

Proof Supp ose to the contrary that r jD y j P k If r k then the

D

probability that x y C L for a random linear order on X is r which is greater

than k Hence x y ess P a contradiction

Now supp ose r k and let D y fy y y g Also let L b e any linear

1 2 k 1

order on X so that the lowest k elements of L are fxg D y with y y

1 2

y y x in L Let L b e obtained from L by switching the p ositions of x and y Then

k 1

x y ess L We conclude from Claim and Prop osition that there is a pair a b

in ess L ess L It is easy to see that this requires a b y z for some z which is

higher than x in L This is in turn requires deg z k and hence D z fxg D y

D

Let L b e obtained from L by moving y immediately ab ove x Now ess L $ ess L

r

which is a contradiction

Now x an essential pair y z of P and let D z fy y y g Then

k +1 k +1 k +1 1 2 k

let R b e any linear order on X in which the k lowest elements are fy y y g

1 2 k +1

with y y y in R Note that y z ess R

1 2 k +1 k +1 k +1

For each i k let R b e the linear order on X obtained by moving y im

i i

mediately over y in R Now cho ose an integer i f k g Then y z

k +1 k +1 k +1

ess R It follows that there is a pair in ess R ess R Claim implies that this pair

i i

has the form y z with jD z j k and hence D z fy y y y y g

i i i i 1 2 i1 i+1 k +1

We conclude that the subp oset of P induced by fy y y g fz z z g is

1 2 k +1 1 2 k +1

a standard example of dimension k 

We note that Theorem in is a sp ecial case of the preceding theorem We also

note that the characterization for Theorem follows immediately

COROLLARY If P X P is a poset with fdimP P then one

D

of the connected components of P is isomorphic to the standard example of dimension

P 

D

Fractional Dimension and Width

The following inequality is due to Dilworth

THEOREM Let P X P be a poset Then dimP widthP 

The forbidden subp oset characterization problem for the inequality in Theorem remains

an unsolved and we susp ect very challenging problem although in some innite

families of examples are constructed which must b e present in the solution Also it is not

known whether it is NPcomplete to determine whether the dimension of a p oset is less

than its width Using the techniques develop ed in the preceding sections we can answer

these questions completely for fractional dimension

Let P X P b e a p oset and let S and T b e disjoint of X When M is a

linear extension of P we say S is over T in M and write ST in M when x y in M

for every x y incP with x S and y T The following elementary result is due

to Hiraguchi

FRACTIONAL DIMENSION

PROPOSITION Let P X P be a poset and let C X be a chain Then there

exist linear extensions M M of P with C X C in M and X C C in M 

1 2 1 2

Recall that Dilworths theorem asserts that a p oset P of width w can b e partitioned

into w chains As a consequence the inequality in Theorem is an immediate corollary

to Prop osition

In what follows we will concentrate on chain decompositions also called chain covers

ie we will write the ground set as a union of not necessarily disjoint subsets each of

which is a chain

THEOREM Let P be a poset Then fdimP widthP Furthermore fdimP

widthP unless

widthP or

widthP w for some w and P contains the standard example of a t

dimensional poset as a subposet

Proof The rst statement is trivial Supp ose the second is false and cho ose a coun

terexample P X P with X as small as p ossible Let w widthP fdimP

The minimality of jX j requires that fdimQ w for any prop er subp oset Q of P

Now supp ose that P has fewer than w maximal elements Then there exist a maximal

element y and a chain decomp osition X C C C of P so that y b elongs to two

1 2 w

or more of the chains in the decomp osition Let Q Y Q b e the subp oset obtained by

removing y from P and let fdimQ r w Let G b e an optimal realizer of Q and

supp ose that G consists of t linear extensions For each M G form a linear extension M

of X by adding y at the top of M Let G denote the resulting family of linear extensions

of P

Then for each i w let M b e a linear extension of P with X C C in

i i i

M Let F fM M M g and recall that valueF w Then for each j

i 1 2 w

let j F denote the multirealizer of P consisting of j copies of each linear extension from

F

Let H G j F and observe that for j suciently large valueH w which is

j j

a contradiction

The same argument with an appropriate F shows that for each maximal element y of

P there is some x X so that x y ess P

Let fy y y g denote the set of maximal elements of P and cho ose an integer

1 2 w

i from f w g Then cho ose an element x X for which x y ess P Since

i i i

x y incP we know that x C Cho ose an integer j j so that x C Then

i i i i i i j

x y We now show that x y in P for every k w with k i

i j i k

Supp ose to the contrary that there is some integer k f w g with k i for

which x  y It follows easily that there is a linear extension M of P with X C C

i k k k

k

and x y in M However this implies that x y ess P The contradiction

i i i i

k

completes the pro of of our assertion that x y in P for every k w with

i k

k i

Now it follows that the p oints fx x x g form an antichain and the order induced

1 2 w

on fx x x g fy y y g is isomorphic to S 

1 2 w 1 2 w w

The Complements of Antichains I

The following theorem was proved indep endently by Kimble and Trotter

S FELSNER AND W T TROTTER

THEOREM Let P X P be a poset let A X be an antichain and let n

jX Aj Then dimP maxf ng 

In a forbidden subp oset characterization of this inequality is obtained The case

where jX Aj is trivial and the case jX Aj includes some pathology asso ciated

with the family of irreducible p osets However for jX Aj n there are exactly

n forbidden subp osets in the list

For fractional dimension the result is even more elegant

THEOREM Let P X P be a poset let A X be an antichain and let n

n2

g unless jX Aj Then fdimP maxf ng Furthermore fdimP maxf n

n1

n and P contains a element antichain or

n and P contains the standard example of an ndimensional poset as a sub

poset

Proof Again the rst statement is trivial We provide a sketch of the pro of of the

second omitting details for some routine parts of the argument When n it is proved

in that the dimension of P is at most n unless the n p oints of X A constitute

an antichain on one side of A that is completely ab ove or completely b elow A When

n we leave it to the reader to show that fdimP unless the p oints in X A

are an antichain and all three are on the same side of A Note that this statement is not

true for ordinary dimension

To complete the pro of it it sucient to show that the following p oset has fractional

dimension at most n n The antichain A is the set of minimal elements

of P The maximal elements of P are the n p oints in Y X A fy y y g

1 2 n

Furthermore for each nonempty subset S f ng with S f n g there

is an element a A with a y in P if and only if i S Note that by requiring

S S i

S f n g we are excluding the app earance of S as a subp oset of P

n

Here is the basic idea We describ e two families of linear extensions M fM M

1 2

M g and R fR R R g such that F M n R is a n fold realizer

n 1 2 n1

of P

To dene these extensions we rst describ e how they order the p oints of Y Once this

is done we have the gaps b etween these p oints into which the p oints of the antichain A

will b e inserted These insertions will b e done according to a sp ecied set of rules Finally

we will provide a rule for ordering elements of A which are inserted into the same gap

Let a a note that a y in P if and only if i n Here are the rules for

n n i

fng

constructing M fM M M g

1 2 n

For each i n the p oint y is b elow all other elements of Y in M

i i

The largest element of Y in M is y

1 n

For each i n we insert a A fa g into the highest gap in M

n i

consistent with the partial order P

For each i n if a a A fa g b oth b elong to the same gap in M

n i

and jU aj jU a j then a a in M

i

The p oint a is the second largest element in M and the rst element of M for

n 1 i

each i n

Here are the rules for constructing R fR R R g

1 2 n1

For each i n the p oint y is b elow all other elements of Y in R

i i

For each i n the p oint y is the second lowest element of Y in R

n i

FRACTIONAL DIMENSION

For each i n and each a A we insert a into the highest gap in M

i

consistent with the partial order P

For each i n if a a A b oth b elong to the same gap in R and

i

jU aj jU a j then a a in R

i

We leave it to the reader to verify that a critical pair of the form a a with n S

S n

is reverted in M for i n while every other critical pair x y is reversed in a

i

linear extension M M and in a linear extension R R This shows that F is a

2

n fold realizer Since F consists of n linear extensions we conclude that

fdimP n n as claimed 

It is interesting to note that we have b een able to provide a discrete jump in the

fractional dimension in the characterization of equality in Theorem However there is

no such jump in the characterizations of equality for Theorem

REMARK For al l k and every there exists a poset P with P k

D

and k fdimP k

Let P b e the incidence p oset of a k regular k partite hyp ergraph where the sides if the

k partition all are of size n for some large n n 

The Complements of Antichains II

There is another inequality b ounding the dimension of the complement of an antichain

This is another result of Trotter Let A X b e an antichain in P with X A

and let Q b e the subp oset of P induced by X A

THEOREM Let P X P be a poset with A and Q as above Then dimP

widthQ 

In it is shown that this inequality is b est p ossible The pro of makes use of the Product

Ramsey Theorem see Chapter of We now show that in the case of fractional

dimension the situation changes slightly For a p oset P X P and a subset Y X

we let P Y denote the restriction of P to Y so that Q Y P Y is the subp oset of P

induced by Y

THEOREM Let P X P be a poset let A be an antichain with X A nonempty

and let Q Y P Y Then let w widthQ

If w or w then fdimP w

If w then fdimP w

If w then fdimP w

Furthermore for w and w the inequalities are asymptotical ly best possible

+

Proof Let C C C b e a chain partition of Q With C we denote the sub chain

1 2 w

i

+

of C ab ove A and with C the sub chain b elow A We may assume that C C C

i i

i i i

+

for all i Also let Y X A and dene Y Y U A and Y Y D A

To prove the theorem we will construct a multirealizer F of P This multirealizer is

the union of three blo cks of linear extensions ie F B B B

1 2 3

Blo ck For each i w let L and M b e linear extensions with C X C and

i i i i

X C C Let L fL L g and M fM M g Dene B as pL pM

i i 1 w 1 w 1

the value of p pw to b e sp ecied later

S FELSNER AND W T TROTTER

Blo ck In B we have for each i w a subblo ck C of linear extensions ie

2 i

+ +

+

B C C Each linear extension in C has Y C C and C Y C

2 1 w i

i i i i

Moreover for every subset B of A with jB j jAj we may assume jAj to b e even there

is a linear extension L in C in which elements of B are as low as p ossible while elements

B i



jAj

denote the numb er of linear extensions of A B are as high as p ossible Let q

jAj2

in C

i

Blo ck In B we assemble r copies of a linear extension which simultaneously reverses

3

all critical pairs x y with x y A Assuming that there are no interchangeable pairs

this can easily b e done

To count how often a critical pair x y of P is reversed in F we have to distinguish

three typ es

Typ e x C and y C The pair is reversed whenever C X C and when

i j i i

X C C Since i j this gives p reversals in B

j j 1

Typ e x C and y A The pair is reversed whenever C X C giving p reversals

i i i

in B However x C and the pair is also reversed in a linear extension L C when

1 B i

i

y B Together this gives p q reversals By the dual argument the same numb er

counts the reversals when x A and y C

i

Typ e x A and y A The pair is reversed r times in B Also each C gives more

3 i

than q reversals together this makes more than r w q

Note that jF j p q w r We now give the prop ortions of p and r relative to q

w p r q

w p q r q

w p q r q

w p q r

A simple calculation shows that the multirealizers built with these prop ortions have

the right values Note that in the rst three cases the values p and r where chosen as to

balance the prop ortion of reversals of Typ e with the prop ortion of reversals of Typ es

and In the calculation we only counted w q reversals of Typ e in B As already

3

noted the actual numb er is slightly larger This proves that the given b ounds can not b e

attained

We now sketch the argument showing that for every there is a p oset P containing

an antichain A such that widthP A w and the fractional dimension of P is at least

w

For given w and n let P X P b e the p oset dened by

A is an antichain in P

C C is a chain partition of X A

1 w

+

jC j jC j n for all i

i i

+ + +

C is incomparable to C C is incomparable to C and C C for all i j

i j i j i j

For every downset S D A and every upset T U A there is an element

a A with D a S and U a T

ST ST ST

Now let F b e an optimal multirealizer of P in what follows we consider F as a probability

space We cho ose a and round all probabilities to multiples of Consider an

+

antichain B consisting of one p oint from C for each i For each in the symmetric

i

group on w ob jects let p b e the rounded probability that in a linear extension in F the

FRACTIONAL DIMENSION

1

w !

elements of B app ear in the order sp ecied by There are fewer than p ossible

functions p and we color the antichains by these functions Provided n is large the

Pro duct Ramsey Theorem guarantees the existence of a twoelement sub chain E in each

i

+

C such that all antichains consisting of one p oint from each E receive the same color

i

i

Note that this implies that in all but a negligible p ortion of linear extensions in F the

sub chains E b ehave as p oints relative to one another ie if x E y E and x y

i i j

then E E

i j

The same argument allows us to cho ose two element sub chains F in C for each i

i

i

such that the F b ehave as p oints relative to each other almost all linear extensions in F

i

Now let a A b e the element with a max E ajj min E ajj max F and a min F

i i i i

for all i w Observe that when the E and F b ehave as p oints in a linear

i i

extension L then only one of the w events a min E a max F for i w

i i

can hold in L Hence the probability of the least probable of these events is at most

w where the probability that the E or the F do not b ehave as p oints

i i

only dep ends on the accuracy of the rounding ie on Now can b e chosen so that

w w This completes the pro of 

Note that the examples given in the pro of of Theorem have already b een used in

to show that the inequality of Theorem is tight

Hiraguchis Inequality and Removal Theorems

We now give two further characterizations for inequalities involving fractional dimension

Both results will b e easy corollaries of previous results and b oth show that some problems

are much easier when stated in terms of fractional dimension The rst example is known

as Hiraguchis inequality In this case when jX j the characterizing p osets are

the same for b oth ordinary dimension and fractional dimension For ordinary dimension

the pro of is only mo derately complicated when jX j is even see However for ordinary

dimension there is no transparent pro of when jX j is o dd The most frequently cited source

for the pro of is Kimble but to the b est of our knowledge the arguments therein are

incomplete

By way of contrast the pro of of the characterization for fractional dimension is very

easy in b oth cases for an analogous situation see the characterization given for

dimension in

THEOREM If P X P and jX j then fdimP bjX jc Furthermore if

jX j then fdimP bjX jc only if jX j is even and P is a standard example or jX j

is odd and P contains a standard example on jX j points

Proof If P contains an antichain A with jAj jX j then Theorem implies

fdimP jX Aj jX j Now assume that there is no antichain of size more then

jX j then widthP bjX jc In this case Theorem implies fdimP bjX jc

The characterization is an immediate consequence of the characterization of Theorem



THEOREM If P X P is a poset and x X then fdimX x P X x

fdimP Furthermore fdimX x P X x fdimP for each x X only if

P is a standard example

Note that the condition dimX x P X x dimP for each x X gives a

denition for irreducible p osets A characterization of irreducible p osets has only b een

obtained for dimensions two and three An indication for the abundancy of irreducible

S FELSNER AND W T TROTTER

p osets is the following estimate For each t if n t then the numb er of t

irreducible p osets on n p oints is at least as large as the numb er of p osets on n p oints

having dimension at most t

Proof of Theorem In it is shown that a p oset P X P on n elements contains

n2

fdimP Hence when fdimP an element x such that fdimX x P X x

n

jX j there is an element x X such that fdimX x P X x fdimP The

result now follows from Theorem 

Op en Problems for Fractional Dimension

In most instances we have b een able to show that our inequalities for fractional dimension

are b est p ossible In two cases there is still some uncertainty

PROBLEM Are the fol lowing two inequalities of Theorem best possible

Let P X P be a poset let A be an antichain with X A nonempty and let

Q Y P Y Then let w widthQ

If w then fdimP w

If w then fdimP w 

For the inequality ofTheorem see the remarks at the end of Section we make

the following conjecture

CONJECTURE

For each there is an integer w so that for every w w there exists a poset

P so that w fdimP w widthP

For every positive integer w there is an so that fdimP w for

w w

every poset P with fdimP w widthP 

Our results and techniques suggest that many dimension theoretic problems have analo

gous versions for fractional dimension which may b e somewhat more tractable Here are

two such problems The rst problem is just a restatement of a problem rst p osed by

Peter Fishburn see

PROBLEM For each t let f t be the minimum number of incomparable pairs

2

in a poset P with fdimP t Is it true that f t t 

We note that the chevron has dimension and has only incomparable pairs

However an easy exercise shows that a p oset with fractional dimension at least must

have at least incomparable pairs The case t was resolved by Qin He showed

that a p oset must have at least incomparable pairs in order to have dimension

PROBLEM Given rational numbers p and q what is the minimum value of the

fractional dimension of P Q where fdimP p and fdimQ q 

We were unable to make any progress on determining whether a p oset on three or more

p oints always contains a pair whose removal decreases the fractional dimension by at

most As is the situation with ordinary dimension this app ears to b e a very dicult

problem Perhaps the following restricted version is accessible

PROBLEM Does there exist an absolute constant so that any poset with or

more points always contains a pair whose removal decreases the fractional dimension by

at most 

FRACTIONAL DIMENSION

Acknowledgement

The authors would like to express their appreciation to Peter Winkler for numerous stim

ulating conversations on the topics discussed in this pap er He also gave a careful reading

to a preliminary version and made many helpful suggestions for improving the exp osition

References

G Brightwell and E Scheinerman Fractional Dimension of Partial Orders Order

K P Bogart and W T Trotter Maximal dimensional partially ordered sets I I Discrete Math

K P Bogart and W T Trotter Maximal dimensional partially ordered sets I I I A characterization

of Hiraguchis inequality for interval dimension Discrete Math

R L Bro oks On colouring the no des of a network Proc Cambridge Philos Soc

R A Brualdi H C Jung and W T Trotter On the p oset of all p osets on n elements Discrete

Applied Math to app ear

R P Dilworth A decomp osition theorem for partially ordered sets Ann Math

P C Fishburn and W T Trotter Posets with large dimension and relatively few critical pairs

submitted

Z Furedi and J Kahn On the dimension of p osets of b ounded degree Order

R L Graham J Sp encer and B L Rothschild Ramsey Theory Wiley New York

T Hiraguchi On the dimension of partially ordered sets Sci Rep Kanazawa Univ

R Kimble Extremal problems in dimension theory for partially ordered sets Ph D thesis M I T

J Qin Some Problems Involving the Dimension Theory for Ordered Sets and the FirstFit Coloring

Algorithm Ph D thesis Arizona State University

I Rabinovitch and I Rival The rank of distributive lattices Discrete Math

W T Trotter Inequalities in dimension theory for p osets Proc Amer Math Soc

W T Trotter On the construction of irreducible p osets University of South Carolina Technical

Memorandum

W T Trotter A forbidden subp oset characterization of an order dimension inequality Math Systems

Theory

W T Trotter Combinatorics and Partial ly Ordered Sets Dimension Theory The Johns Hopkins

University Press Baltimore Maryland