Mathematics authors titles recent submissions
arXiv:1709.06046v1 [math.NT] 18 Sep 2017
BY THE MULTISET OF ITS π -SUMS?
DMITRI V. FOMIN
This paper is dedicated to the memory of Oleg Izhboldin (1963-2000)
ABSTRACT. This is a survey of all available information on a remarkable problem in number theory proposed by Leo Moser in 1957. In general form the question is: can a collection ofπnumbers be uniquely restored given the collection of itsπ -sums? We describe results and techniques from sixty years of research in this area. Some new ο¬nds and open questions are presented.
1. INTRODUCTION
Sixty years ago, in 1957, American Mathematical Monthly has published the following relatively simple problem in number theory proposed by Leo Moser (see [2]):
Problem 1.1. (a) The ten numbers π 1 β©½ π 2 β©½ ... β©½ π 10 are the sums of the five unknown numbers
π₯1 β©½π₯2β©½ ...β©½π₯5taken two at a time. Determine theπ₯βs in terms of theπ βs.
(b) Show that if π 1 < π 2 < ... < π 6 are six distinct numbers formed by taking the sums of four numbers two at a time, then there exist four other numbers which give the same sums when added in pairs.
Naturally β after it was quickly solved β the problem was immediately generalized and reformu-lated. And then it turned out to be quite an interesting little question...
Problem 1.2. Let π΄be a collection (multiset) of π numbers π1 β©½ π2 β©½ ... β©½ ππ. Consider the multisetπ΄(2) of (π
2
)
2-sums of multisetπ΄, i.e., collection of all sums of the kind ππ
1 +ππ2, where
1β©½π1< π2 β©½π. Is it possible to restoreπ΄givenπ΄(2)?
The numbers here could be complex or even belong to an arbitrary ο¬eld of characteristic zero. That really doesnβt matter as we will learn shortly.
In this generalization the problem asks whether a multiset is uniquely determined by (or can be recovered from) the multiset of its2-sums.
It was later presented in the literature (e.g., see [8]) using somewhat diο¬erent terminology.
Problem 1.3. A malicious farmerβs apprentice was asked to provide the list of weights ofπbags of grain. Instead he weighed them two at a time and recorded allπ(πβ 1)β2combined weights written down in some random order. Is it possible to find the weights of bags (up to permutation of bags)? By the way, it seems that the apprentice was not only malicious but also somewhat dense β instead of performing onlyπweighings he did a whole lot more of them.
Again the problem is posed as a βrecoveryβ question β whether an unknown multiset can be uniquely restored from the multiset of its pairwise sums.
Date: September 19, 2017.
2010Mathematics Subject Classification. Primary: 11B75, 11P70; Secondary: 05A15.
Key words and phrases. integer multisets, multiset recovery, sumsets, symmetric polynomials. 1
(2)
Now, each interesting question, theorem or conjecture deserves a nice name that easily rolls of the tongue in lectures and discussions. βFermatβs Last Theoremβ, βRiemann Hypothesisβ, βP=NPβ, βCollatz 3k+1 Conjectureβ β all these names are short and to the point. I submit that βMultiset Recovery Problemβ sounds just as neat while describing the issue with decent precision.
The original problem1.1 was indeed quite easy. However, its generalization1.2 was not. Still, it did not present a serious obstacle; the answer was quickly discovered and so the problem was generalized even further.
Notation. For any pair of positive integers π and π such that π β©Ύ π we will denote by π΄(π ) the multiset ofπ -sums ofπ΄, i.e. collection of all sums of the kind
ππ
1 +ππ2 + β¦ +πππ ,
where1β©½π1< π2 <β¦< ππ β©½π.
Problem 1.4. Consider positive integersπandπ withπ > π . Do there exist two distinctπ-multisets
π΄andπ΅such thatπ΄(π )= π΅(π )?
Such a pair of multisets would represent a "recovery failure". Indeed, in this case, given multiset
π =π΄(π ) =π΅(π ) it is impossible to determine the original multiset.
Definition. If two multisetsπ΄and π΅ have the same collections ofπ -sums β that is, π΄(π ) = π΅(π ) β then we will call these multisets π -equivalent (or when this will not cause any confusion, simply equivalent) and we will denote that relation asπ΄βΌπ π΅ (or simplyπ΄βΌ π΅).
Definition. We will call a pair of natural numbers (π, π ) singular if it represents a non-trivial βmultiset recovery failureβ β i.e.,π > π and there exist two differentπ -equivalentπ-multisetsπ΄and
π΅ (π΄β π΅ & π΄βΌπ π΅).
So all the above problems can be reworded as questions about singular pairs. The ultimate goal is to describe those pairs in some easily βcomputableβ way.
Notation. For any natural number π byξΉπ we will denote the set of all natural numbersπ > π
such that pair(π, π )is singular.
For instance,ξΉ1is obviously empty. AlsoQuestion 1.1could be reformulated as follows: does ξΉ2contain numbers 10 and 4?
This article will present all currently known results on the Multiset Recovery Problem and the methods involved. We will also discuss some new facts and conjectures.
2. HISTORICAL TIMELINE
(1957). Just a few words about Moserβs possible motivation for this problem. A few years before that, in 1954, Leo Moser and Jim Lambek published article [1] about pairs of complementary subsets ofβwhere they have provedLambek-Moser Theoremabout partitions ofβand how they are related to sequences of numbers of the form{π(π) +π}whereπ βΆβββ€β©Ύ0is some arbitrary non-decreasing unbounded function.
This investigation seems to be quite close to questions about how ο¬nite or inο¬nite sets of natural numbers overlap or complement each other when being translated.
I suspect that this was how Leo Moser stumbled upon questions about multiset recovery for the case of π = 2 β but of course, this is pure speculation. However, in their later short article [4] Lambek and Moser mention both Multiset Recovery Problem and complementary sequences of integers literally on the same page.
(3)
(1958). In 1958, almost immediately after Moser has posed his original question, Selfridge and Straus in their article [3] provided solution for the βrealβ problem1.2as well as some other ques-tions. First, they have proved the following
Theorem 2.1. Multisetπ΄ofπnumbers is uniquely determined by multiset of its2-sumsπ΄(2)if and only ifπis not a power of 2.
In other words, pair(π,2)is singular if and only ifπis a power of 2. Or, using our notation, ξΉ2 = {4,8,16,32, ...} = {2π βΆ π >1}.
Second, they have explored case ofπ = 3of the more general problem 1.4. They have shown that forπ = 6 there are easily constructed examples of diο¬erent multisets ofπ numbersπ΄andπ΅
such thatπ΄βΌ3 π΅. For instance:
π΄= {15,β5}, π΅ = {(β1)5,5}βΉ π΄β π΅ π΄(3) = {310,(β3)10}, π΅(3) = {(β3)10,310}βΉ π΄(3)=π΅(3),
where15is not1as you might have thought. In standard multiset notationππmeans elementπwith
multiplicityπ(i.e.,π-multiple entry of numberπ), so we simply mean that
π΄= {1,1,1,1,1,β5}, π΅ = {β1,β1,β1,β1,β1,5}.
As for other values ofπ, the authors of [3] have proved that such example for caseπ = 3can exist only ifπ2β (2π+ 1)π+ 2β 3πβ1vanishes for some naturalπ < π. It is not very hard to prove that the only other non-trivial values ofπfor which that is possible areπ= 27(withπ= 5,9) andπ= 486
(π= 9).
Thus, they have showed that
{6}βξΉ3 β {6,27,486}.
Third, using the same technique for the case ofπ = 4it was shown that the only non-trivial values ofπwhen recovery might not be always possible areπ= 8,12.
Presenting an example forπ = 8is quite easy. Generally, one can always construct an example of βrecovery failureβ inProblem 1.4ifπ= 2π (we will do that later inSection 3). Again, this can be written as
{8}β ξΉ4β {8,12}.
Naturally, that suggested a few additional questions.
Question 2.2. Do pairs(π, π ) = (27,3)and(486,3)represent actual βrecovery failuresβ? In other words, do there exist forπ = 27andπ= 486examples of differentπ-multisetsπ΄andπ΅ such that
π΄(3) =π΅(3)?
Question 2.3. Same question about pair(π, π ) = (12,4). That is, do there exist two different12 -multisetsπ΄andπ΅such thatπ΄(4) =π΅(4)?
At that time both questions were left unsolved. Yet another important question from the same article:
Question 2.4. In cases when recovery is impossible, could there exist more than two π-multisets that generate identical multisets ofπ -sums?
(4)
(1959). Soon after the paper by Selfridge and Straus, Leo Moser and his coauthor Joachim (Jim) Lambek wrote a small article [4]. It started by acknowledging results of their colleagues from UCLA, and then they proceeded to develop the problem in a slightly diο¬erent direction.
Namely, they asked a question whether the set of non-negative integers can be split in two subsets
π΄= {π1, π2, ...}andπ΅ = {π1, π2, ...}such thatπ΄(2)andπ΅(2)coincide as multisets. They proved that the answer was positive and that there exists only one such decomposition ofβ€β©Ύ0.
They did that by using multiset generating functions.
Definition. For any finite multisetπ΄of positive integers of the form{ππ1
1 , π
π2
2 , ..., π
ππ
π }we define its
generating function (polynomial)ππ΄(π₯)by formula
ππ΄(π₯) = π β
π=1
πππ₯ππ.
Similarly this generating function can be defined for an infinite multisetπ΄ = {ππ1
1 , π
π2
2 , ...}as long as sequence{π1βππ
π }is bounded.
Authors proved that in their particular case generating functions satisο¬ed system of equations:
{
ππ΄(π₯) +ππ΅(π₯) = 1β(1 βπ₯) π2
π΄(π₯) βππ΄(π₯
2) =π2
π΅(π₯) βππ΅(π₯
2).
It was also proved that similar split ofππ = {0,1,2, ..., πβ1}is possible if and only ifπis a power of 2. That split is unique and is determined by the so-calledThue-Morse sequence{πΌπ}deο¬ned as πΌπ =π 2(π) (mod 2)whereπ 2(π)is thebinaryweight ofπ, i.e., sum of digits (or simply, the number of ones) in the binary representation ofπ. So if π = 2π and we deο¬ne setsπ΄ = {π
1, ..., ππ}and π΅ = {π1, ..., ππ}as follows
π΄= {πβππβΆ πΌπ = 0} π΅ = {πβππβΆ πΌπ = 1}
thenπ΄(2) =π΅(2). Indeed, if you setππ΄(π₯) = βππ=1π₯ππ andπ
π΅(π₯) = βπ
π=1π₯
ππthen we have
ππ΄(π₯) +ππ΅(π₯) = π’(π₯) = 1 βπ₯
2π
1 βπ₯ , ππ΄(π₯) βππ΅(π₯) =π£(π₯) = πβ1
β π=0
(1 βπ₯2π).
Thusππ΄= (π’+π£)β2andππ΅ = (π’βπ£)β2and from that it follows quite easily thatπ2
π΄(π₯) βππ΄(π₯2) = ππ΅2(π₯) βππ΅(π₯2). It is left to notice that the sides in the the last equality are generating functions for multisetsπ΄(2)andπ΅(2)respectively.
Somewhat similar βgenerating functionsβ approach was used later in [5], [8] and [9] in conjunc-tion with some other ideas.
(1962). Next paper on the subject appeared in 1962, when Fraenkel, Gordon and Straus published [5] proving that answer toQuestion 2.4was negative. This was not the last time when Multiset Recovery Problem deο¬ed the expectations.
Authors have found numerous multi-singularity examples for the simplest case ofπ = 2. More precisely, they have showed how to construct examples of three diο¬erent 8-multisetsπ΄, π΅ andπΆ
such thatπ΄(2) =π΅(2)=πΆ(2). Here is one of these examples:
(5)
After this, naturally, the original question was adjusted into asking how many diο¬erentπ-multisets could generate the same multiset ofπ -sums. The maximum possible number of such multisets was denoted by ξ²π (π). Of course, pair (π, π )must be singular to begin with β which is equivalent to inequalityξ²π (π)>1.
Then more inequalities forξ²π (π)were proved. For instance,
ξ²2(16)β©½3, 2β©½ ξ²3(6)β©½6, ξ²4(12)β©½2.
However, no other examples of triple (or greater) βmultiplicityβ were found, which led to another open question:
Question 2.5. a) Forπ = 2does there existπ = 2π > 8such that some three distinctπ-multisets
generate the same multisets ofπ -sums (i.e.,ξ²2(π)> 2)?
b) Generally, does there exist any singular pair (π, π )different from (8,2)with three pairwise distinctπ -equivalentπ-multisets (i.e.,ξ²π (π)>2)?
Two more results in the same article deserve mention. One was to prove that when dealing with any question about multiset recovery it was enough to work with ring of integers β€, instead of arbitrary ο¬elds of characteristic zero or torsion-free Abelian groups. The other result resolved one of the questions about number of elements inξΉπ . Namely, the authors proved thatξΉπ was ο¬nite for allπ >2.
(1962). The very same year the ο¬rst part of the originalProblem 1.1was used in one of the top math contests in the Soviet Union β namely, Moscow City Mathematical Olympiad. It is quite possible that some Soviet mathematician had seen article [5] and liked the original question well enough to submit it to the olympiad committee. It was given to high school juniors and proved to be one of the more diο¬cult problems of that year. An unpublished compilation of problems from that competition (translated into English) can be found in [13].
(1968). Among two questions about suspect pairs (Questions 2.2, 2.3) the latter βπ= 12, π = 4β seemed easier. So it was not surprising that βonlyβ ten years after original article [3], John Ewell published his paper [6] claiming that pair(12,4)was not singular and recovery was always possible for this case (it became a part of his Ph.D thesis). He also found a purely combinatorial and more direct proof of the very important formulaβ¨3β©(see below, inSection 4).
Many years later further investigation uncovered an error in calculations regarding pair(12,4). However, another result in the same article was clearly correct β namely, Ewell demonstrated that answer toQuestion 2.5 bwas positive. He has proved thatξ²3(6) = 4and then went on to provide complete characterization of all possible quartets of pairwise diο¬erent 3-equivalent 6-multisets.
We will give you one of these examples as a demonstration of Ewellβs discovery
π΄= {0,5,9,10,11,13};π΅= {1,5,8,9,10,15};πΆ= {1,6,7,8,11,15};π·= {3,5,6,7,11,16},
leaving the actual veriο¬cation as an easy exercise for the reader.
(1981). Richard Guy mentioned multiset recovery in his compendium of unsolved problems in number theory, see [7], Problem C5. He explained that it had been solved for π = 2 while the question for values ofξ²2 still had not been answered in full. Caseπ = 3forπ = 27andπ = 486
(6)
(1991). As far as we know, after Ewellβs thesis Multiset Recovery Problem slipped into relative obscurity until 1991, when Boman, Bolker and OβNeil have explored a slightly diο¬erent approach in their article [8].
More precisely, for point π₯ = (π₯1, π₯2, ..., π₯π)in π-dimensional euclidean spaceβπ and for any
π -subsetπ΄ = {π1, ..., ππ }ofπΌπ = {1,2, ..., π}let us deο¬neπ₯π΄ as the sumπ₯π
1 +...+π₯ππ . Then we can deο¬ne linear operator
π π,π βΆβπ ββ(ππ ), π
π,π (π₯1, π₯2, ..., π₯π) = (π₯π΄1, π₯π΄2, ..., π₯π΄(π
π )
),
whereπ΄1, ...,π΄(π π )
is the sequence of allπ -subsets inπΌπ. This is a βsortβ of discrete (combinatorial) version of Radon integral transform. Mappingπ π,π can obviously be transferred from euclidean spaces to their reductions modulo standard actions (permutations of coordinates) of symmetric groups ππ and π(π
π )
respectively so that we have π π,π βΆ ππ β π(π π )
where ππ = βπβππ. Then Multiset Recovery Problem can be posed as a question on whetherπ π,π is an injection. Or more generally, as a question on the size ofπ β1π,π (π₯), withπ₯βπ(π
π ) .
Using this notation and terminology they proved β among other things β that when recovering
π-multiset from the collection of its 2-sums one can not obtain more thanπβ 2diο¬erent multisets. Improving on that result they have also showed that for anyπβ 8this upper bound could actually be lowered to 2 and almost always to 1. Thus they solvedQuestion 2.5 a.
It is worth noting that judging by the list of the open questions, at that time the authors did not know about Ewellβs paper [6].
(1992). By some happy βaccidentβ in 1991 Question 1.2 was used in a studentsβ mathematical contest in St.Petersburg, USSR. Author of this survey was β as surely many other mathematicians before him β lured in by this seemingly simple problem, and started his own investigation. That resulted in article [9] by Fomin and Izhboldin submitted for Russian publication in 1992 (English translation was published in 1995).
Most of that article was about rediscovering the very same results already achieved in [3] and [5] β unfortunately, due to a rather poor access to international scientiο¬c magazines authors could not properly search for the papers already written on this issue. However, their article still contained one completely new result: singularity examples which positively answeredQuestion 2.2for both cases βunder suspicionβ. Pairs(27,3)and(486,3)were proved singular.
Thus, investigation of generalized Multiset Recovery Problem 1.4 for the case of π = 3 was closed.
(1996). Just a few short years later, Boman and Linusson have independently come up with sin-gularity examples for pairs (27,3) and (486,3)in [11]. Alas, they thought that case (12,4) was already resolved by Ewell β at the end of their paper they mentioned that they were told (apparently at the very last moment) about article [6]. They also made some inroads into ο¬nding all possible singularity examples forπ = 3.
(1997). Ross Honsberger dedicated a chapter called βA Gem from Combinatoricsβ of his book [12] to the caseπ = 2 of Multiset Recovery Problem. It is curious that he never mentions Leo Moser. Instead Honsberger stated that the results he had reproduced came from Paul ErdΕs and John Selfridge. This is the only time when ErdΕsβs name appears in this story. It is not clear whether he really has done something there or possibly it was just a mistake in attribution.
(7)
(2003). In chapter 46 of their engaging book [14], Savchev and Andreescu explained the solution for Question 1.2and also went over the results from Lambek and Moserβs article [4] concerning Thue-Morse sequence.
(2008). A slightly expanded version ofQuestion 1.3with extra items repeating parts of [5] and [8] was published as another problem in American Mathematical Monthly β submitted by Chen and Lagarias, [15]. Some of the results were subsequently posted and discussed on the Cut-The-Knot website, see [16].
(2016). Nothing signiο¬cant happened for quite some time until Isomurodov and Kokhas ([17]) have discovered that Ewell made a mistake in his lengthy polynomial computations for pair(12,4). We will never know how that happened, but nowadays mathematicians no longer have to do all these ex-hausting computations by hand β for instance, authors of [17] made use of symbolic computational package MAPLEβ’. After that the authors proved the existence of the βrecovery failureβ example and actually produced it, thus solvingQuestion 2.3and ο¬nalizing caseπ = 4ofProblem 1.4(see below inSection 4).
3. SOME SIMPLE EXAMPLES
This short section explains how construct some simple examples of singular pairs.
Case π = 2, π = 2π. The most obvious and trivial of all examples of βrecovery failureβ is pair (2,2): one cannot hope to restore a set of two numbers knowing only their sum. This is not a βrealβ singular pair (becauseπ= π ) but we can use it as a basis from which less trivial examples are built. Namely, if we have twoπ-multisetsπ΄andπ΅which are 2-equivalent, then for any numberπwe have
π΄βͺ (π΅+π)βΌ2 π΅βͺ (π΄+π) β¨1β©
whereπ +πis multiset obtained fromπ by addingπto all of its elements. So if we start with
π΄= {1,1}, π΅ = {0,2}, π΄βΌ2 π΅
then choosingπ = 1we get
π΄β²= {1,1,1,3}, π΅β² = {0,2,2,2}, π΄β² 2βΌπ΅β².
Proceeding in this manner, we can easily build examples of 2-equivalentπ-multisets for anyπwhich is a power of 2. Again, we will leave the proof ofβ¨1β©as an exercise for the reader.
Caseπ= 2π . Remember that singularity example forπ= 6,π = 3fromSection 2? It can be easily generalized for any pair(π, π )whereπ= 2π .
Namely, you can take some2π -multisetπ΄, ο¬nd its arithmetic meanπand reο¬ectπ΄with respect toπto obtain what we will call itsmirrormultisetπ΄Μ = 2πβπ΄. As long asπ΄is not symmetric,π΄Μ
will be diο¬erent fromπ΄andπ΄ βΌπ π΄Μ. To prove that it is suο¬cient to notice that for eachπ β π΄
with|π|=π the mirror image ofπ΄βπ (which is a sub-multiset ofπ΄Μconsisting ofπ numbers) has the same sum of elements.
For demonstration purposes we only need one example β let us considerπ΄= {12π β1,1 β 2π }and its mirrorπ΄Μ = {(β1)2π β1,2π β 1}.
All theπ -sums of numbers inπ΄β there are(2π
π )
of them β fall into two groups. One consists of the sums that include element(1 β 2π )β there are(2π β1
π β1
)
of those, and each one of these sums is equal to(π β 1)β 1 + (1 β 2π ) = βπ . The other one has(2π β1
π )
(8)
them, each one of them equal toπ . Thus we haveπ΄(π ) = {(βπ )π, π π}whereπ = (2π β1 π β1
)
= (2π β1 π
)
. You can see thatπ΄(π ) is symmetric (with respect to zero) and thusπ΄Μ(π ) =π΄(π ).
Duality (π, π ) β (π, πβ π ). If we have two π -equivalent π-multisets π΄ and π΅, then these same
multisets are(πβπ )-equivalent as well. To prove that, it is suο¬cient to demonstrate that sum of all elements ofπ΄equals to that ofπ΅. Quick computation shows that
β
1β©½π1<π2...<ππ β©½π
(ππ1 +ππ2 +...+πππ ) = (
πβ 1 π β 1
) β
1β©½πβ©½π ππ.
Thus sum of numbers inπ΄equals the sum of numbers inπ΅ and denoting that number byπwe have
π΄(πβπ ) =πβπ΄(π ), π΅(πβπ ) =πβπ΅(π ),
which proves the duality. This means we can always assume thatπβ©Ύ2π ; ifπ < π < 2π then we can switch to pair(π, π β²)whereπ β² = πβπ andπ > 2π β².
This duality allows us to generate more examples of singular pairs. For instance, since (8,2)
is singular then(8,6)is singular, too. As we will see soon, pairs (27,3), (486,3)and(12,4)are singular β therefore, pairs(27,24),(486,483)and(12,8)are singular as well.
Later in this article (seeSection 5) we will talk more about this duality and its partial expansion.
Linear transformations. Finally, one obvious but useful fact. Ifπ΄βΌπ π΅andπ(π₯) =ππ₯+πis some arbitrary linear function then multisetsπ(π΄) andπ(π΅)are also π -equivalent. That simply means we can translate and stretch/shrink singularity examples to obtain new ones.
For instance, if you consider π΄ = {0,5,9,10,11,13}, π΅ = {1,5,8,9,10,15}then π΄ βΌ3 π΅. Applyingπ(π₯) = 2π₯β 13we obtain new pair of 3-equivalent multisetsπ΄1= {β13,β3,5,7,9,13},
π΅1 = {β11,β3,3,5,7,17}.
Of course, all the singularity examples that can be obtained from each other by such operations will be considered identical for the purposes of this investigation.
4. SYMMETRIC POLYNOMIALS. COMPLETE SOLUTION FOR CASESπ = 2,3,4
Now let us delve into speciο¬c techniques used in multiset recovery. The main one is based on the following approach that utilizes symmetric polynomials.
Givenπ-multisetπ΄ = {π1, ..., ππ}we can produce sequence of sums of itsπ-th powers forπ = 1, ..., π. That is, we can apply power-sum symmetric polynomials ofπvariables
ππ(π₯1, ..., π₯π) = π β
π=1
π₯ππ
to multisetπ΄to obtain sequenceπ1(π΄), ..., ππ(π΄). It is well known thatπ΄can be restored from this
sequence β values of ππ(π΄) determine coeο¬cients of polynomial(π₯β π1)(π₯βπ2)...(π₯β ππ) and therefore determine the multiset of its roots.
Thus if values ofππ(π΄)forπ= 1, ..., πcan be deduced from values ofππ(π΄(π ))then multisetπ΄is uniquely determined by multisetπ΄(π ).
Let us start from small values ofπ. Forπ= 1we have already computed
π1(π΄(π )) =
( πβ 1 π β 1
) π1(π΄)
and therefore, ifπ΄(π ) =π΅(π ) thenπ
1(π΄) =π1(π΅). This means thatπ1(π΄)can always be found from
(9)
Now, ifπ= 2then
π2(π΄(π )) = β 1β©½π1<...<ππ β©½π
(ππ
1 +...+πππ )
2 =
( πβ 1 π β 1
) β
1β©½πβ©½π π2π +
( πβ 2 π β 2
) β
1β©½π<πβ©½π
2ππππ =
= (
πβ 1 π β 1
) β
1β©½πβ©½π π2π +
( πβ 2 π β 2
) β
1β©½πβ πβ©½π ππππ =
= ((
πβ 1 π β 1
) β
( πβ 2 π β 2
)) β
1β©½πβ©½π π2π +
( πβ 2 π β 2
) β
1β©½π,πβ©½π ππππ =
= (
πβ 2 π β 1
)
π2(π΄) + (
πβ 2 π β 2
)
π1(π΄)2.
We already knowπ1(π΄)and thus we can ο¬ndπ2(π΄)as long as coeο¬cient(πβ2
π β1
)
is not zero. But
π > π >0and thereforeπ2(π΄)is also always βrecoverableβ. Sinceππ(π΄(π ))is a symmetric polynomial ofπ
1, ... ππ then it can be expressed in the following
way:
ππ(π΄(π )) =πΌππ(π΄) +ξΌ(π1(π΄), π2(π΄), ..., ππβ1(π΄)), β¨2β© where πΌ is a constant (in terms of variablesππ) deο¬ned by three numbersπ, π , π, and ξΌ is some polynomial of πβ 1 variables whose coeο¬cients are fully deο¬ned by that triplet as well. For instance, as we have just shown, forπ= 2we haveπΌ =(πβ2
π β1
)
andξΌ(π₯) = (πβ2 π β2
) π₯2.
It follows that if we could show that coeο¬cientπΌ does not vanish then we will have proved that
ππ(π΄)is determined byπ1(π΄(π )),π 2(π΄(
π )), ...,π
π(π΄( π )).
Let us denote that coeο¬cientπΌasπΉπ ,π(π). Then the following equality is true:
Theorem 4.1.
πΉπ ,π(π) = π β
π=1
(β1)πβ1ππβ1
( π π βπ
)
. β¨3β©
It was proved in [5] by some neat manipulation of a formula from [3]. Later a purely combinato-rial and more direct proof ofβ¨3β©was given in [6]. And then it was again βrediscoveredβ and proved (in a somewhat diο¬erent manner, by making use of exponential generating functions) in [9].
We will leave it to the reader as a simple exercise to prove that forπ = 2formulaβ¨3β©is indeed equivalent toπΉπ ,2(π) =(πβ2
π β1
)
.
Incidentally, even without this formula caseπ = 2(Problem 1.2) can now be resolved in a very straightforward manner. Computingππ(π΄(2))we obtain (using notation fromβ¨2β©)
πΌ = πβ 2πβ1,
which means thatπ-multiset is always recoverable from the multiset of its 2-sums ifπis not a power of 2. As we already know, ifπis a power of 2 thenπ-multisetπ΄cannot always be recovered from
π΄(2).
Caseπ = 3. Equationβ¨3β©gives us
πΉ3,π(π) = (
π 2
) β 2πβ1
( π 1
)
+ 3πβ1, 2πΉ3,π(π) = π2βπ(2π+ 1) + 2β 3πβ1.
(10)
Investigation here is again relatively straightforward. First, we can prove thatπΉ3,π(π)cannot be zero for positive integerπ ifπ > 12. Second, we check all the cases withπ β©½ 12and verify that polynomialsπΉ3,πhave integer roots if and only ifπβ {1,2,3,5,9}. For these ο¬ve special cases we have
πΉ3,1(π) = 1
2(πβ 1)(πβ 2), πΉ3,2(π) = 1
2(πβ 2)(πβ 3), πΉ3,3(π) = 1
2(πβ 3)(πβ 6), πΉ3,5(π) = 1
2(πβ 6)(πβ 27), πΉ3,9(π) = 1
2(πβ 27)(πβ 486).
FromSection 3 we already know that pair(6,3)is singular. To prove the same for pairs(27,3)
and(486,3)we present the following examples:
π= 27 π΄β²27 = {0,116,210}, π΄β²β²27= {05,110,210,32}, π΄β²β²β²27 = {0,15,210,36,45}, π= 486 π΄486 = {022,1176,2231,356,4}.
We will leave it to the reader to verify that all these multisets are 3-equivalent to their mirrors. Nowadays, this can be done in minutes, using just a few lines of code in some decent computational package.
Summary: ξΉ3= {6,27,486}.
Caseπ = 4. Fromβ¨3β©we obtain
πΉ4,π(π) = (
π 3
) β 2πβ1
( π 2
) + 3πβ1
( π 1
)
β 4πβ1,
6πΉ4,π(π) =π
3βπ2(3 + 3β 2πβ1+ 1) +π(2 + 3β 2πβ1+ 2β 3π) β 6β 4πβ1.
Then using divisibility and other pretty routine number theory ideas we can prove that forπ >7
polynomialsπΉ3,πdo not have positive integer roots. And, ο¬nally,
πΉ4,1(π) = 1
6(πβ 1)(πβ 2)(πβ 3), πΉ4,2(π) = 1
6(πβ 2)(πβ 3)(πβ 4), πΉ4,3(π) = 1
6(πβ 3)(πβ 4)(πβ 8), πΉ4,4(π) = 1
6(πβ 4)(π
2β 23π+ 96),
πΉ4,5(π) = 1
6(πβ 8)(π
2β 43π+ 192),
πΉ4,6(π) = 1
6(πβ 12)(π
2β 87π+ 512),
πΉ4,7(π) = 1
6(πβ 8)(π
(11)
Caseπ= 8is βtrivialβ β this is the situationπ= 2π which is well known to us by now. The only other non-trivial root ofπΉ4,πpolynomials is 12. As we mentioned before, this case turned out to be tougher than the others β a pair of 4-equivalent 12-multisets was found only in 2016 by Isomurodov and Kokhas. Namely, if we consider the two following diο¬erent 12-multisets
π΄= {12,4,6,7,82,9,10,12,152}, π΅ = {0,3,4,5,6,7,9,10,11,12,13,16},
thenπ΄βΌ4 π΅.
In [17] the authors have actually proved that this is the only possible singularity example for case
(12,4)(considering pairs of multisets that diο¬er only by linear transformation as identical). Summary: ξΉ4= {8,12}.
5. DIGGING FOR ROOTS(OFπΉπ ,π)
Since we know that pair (π, π ) can be singular only if π is a root of πΉπ ,π then let us turn our attention to ο¬nding out more about those roots. (The rest of this section was inspired by the proof ofTheorem 7from [3]).
For this let us take another, closer, look at the polynomialsπΉπ ,πfor the ο¬rst few values ofπ.
Caseπ= 1. We already know that
πΉπ ,1(π) = (
πβ 1 π β 1
)
= 1
(π β 1)!(πβ 1)(πβ 2)...(πβπ + 1) = 1 (π β 1)!
π β1
β π=1
(πβπ).
Thus the roots are 1 throughπ β 1and of no interest to us βπhas to be greater thanπ to provide us with a possibly singular pair(π, π ).
Caseπ= 2. This one we have also computed before.
πΉπ ,2(π) = (
πβ 2 π β 1
)
= 1 (π β 1)!
π β
π=2
(πβπ).
Again, no roots of interest. Let us go on.
Caseπ= 3. It is still fairly easy to compute
πΉπ ,3(π) = 1
(π β 1)!(πβ 2π ) π β
π=3
(πβπ),
which (ο¬nally!) has a non-trivial rootπ = 2π . However, we already know about it β pair(2π , π )is always singular.
Caseπ= 4. This computation might take a little longer but eventually you will get the following formula (forπ > 2)
πΉπ ,4(π) = 1 (π β 1)!(π
2β (6π β 1)π+ 6π 2)
π β
π=4
(πβπ). β¨4β©
A-ha! We now have quadratic Diophantine equation for non-trivial rootsπ
π2β (6π β 1)π+ 6π 2 = 0, β¨5β©
which can be also rewritten as a quadratic equation forπ
(12)
Sum of the roots of equationβ¨6β©isπβ hence, if pair (π, π )withπ > π is a root of equation β¨4β©
then so is pair(π, πβπ ). This is clearly a direct analog of singular pairsβ duality we have described inSection 3. Let us call such pairsβ¨π,4β©-conjugatedor simplyπ-conjugated.
In the same manner from equationβ¨5β©we can conclude that if pair(π, π )is a root of equationβ¨4β©
then its other root is pair(6π β 1 βπ, π ). These two pairs will be calledβ¨π ,4β©-conjugated.
Obviously, both types of conjugation are symmetric. It is also clear from equationsβ¨5β©andβ¨6β©
that for any positive solution(π, π )we have6π β 1 > πandπ > π β otherwise left sides of these equations are positive. Thus, conjugate pair always consists of two positive integers as well.
Let us consider the smallest possible root of β¨5β©, namely π = (2,1) (since we are solving the equation in positive integers, we are allowed to talk about βsmallestβ solution). That pair is not something we can directly use becauseπ must be at least 3 for the formulaβ¨4β©to make sense. But it is still a root of our quadratic equationβ¨5β©and we will use it to produce others.
It is important to mention that pairπis self-π-conjugated (2 β 1 = 1) so the only way to produce a diο¬erent solution is viaπ -conjugation. So we jump to pair(3,1), then throughπ-conjugation to pair(3,2), then to(8,2), then to(8,6)and so on.
Proceeding like that we will obtain one inο¬nite chain of pairs-solutions of equationβ¨5β©:
(2,1)βπ (3,1)
π
β(3,2)
π
β(8,2)
π
β(8,6)
π
β(27,6)
π
β
(27,21)βπ (98,21)
π
β(98,77)
π
β(363,77)
π
β(363,286)
π
β... β¨7β©
We have marked the arrows with small letters "π" and "π " to show which type of conjugation was used in each case; also we have underlined pairs which are not βfully compliantβ β they are roots of equationβ¨5β© but they are not roots of corresponding polynomialπΉπ ,π withπ > 2(pair(8,2)is singular but we have discounted it because of π > 2requirement). So, starting from (8,6), pairs in the chain represent valid roots of polynomials πΉπ ,4. Thus, they are all βsuspectβ as possible singularities for Multiset Recovery Problem.
Caseπ= 5. In this case computation is also not terribly complicated. Forπ > 3we obtain
πΉπ ,5(π) = 1 (π β 1)!(π
2β (12π β 5)π+ 12π 2)(πβ 2π )
π β
π=5
(πβπ).
Again we have a quadratic Diophantine equation which can be written like this
π2β (12π β 5)π+ 12π 2 = 0. β¨8β©
or like this
π 2βπ π+π(π+ 5)β12 = 0. β¨9β©
As before, equationβ¨9β©gives us π-conjugation βdualityβ(π, π ) β (π, πβπ ). And equationβ¨8β©
provides us withβ¨π ,5β©-conjugation β namely,(π, π )β(12π β 5 βπ, π ).
Similarly to the previous subsection we obtain inο¬nite chain of pairs-solutions:
(3,1)βπ (4,1)
π
β(4,3)
π
β(27,3)
π
β(27,24)
π
β(256,24)
π
β(256,232)
π
β(2523,232)
π
β...
However this case diο¬ers somewhat from the previous one. The βminimum solutionβ pair(3,1)
hasπ-conjugate(3,2)that does not coincide with it β thus the chain can be extended in the other direction as well. Therefore we obtain more solutions:
(3,2)βπ (16,2)
π
β(16,14)
π
β(147,14)
π
β(147,133)
π
β(1444,133)
π
β(1444,1311)
π
β...
(13)
...(1444,1311)βπ (1444,133)
π
β(147,133)
π
β(147,14)
π
β(16,14)
π
β
π
β(16,2)
π
β(3,2)
π
β(3,1)
π
β(4,1)
π
β(4,3)
π
β(27,3)
π
β
π
β(27,24)
π
β(256,24)
π
β(256,232)
π
β(2523,232)... β¨10β©
In both cases π = 4 andπ = 5 it is easy to prove that all positive integer solutions of equations
β¨5β©andβ¨8β©belong to the chains 7 and 10 respectively. We will leave that to the reader.
Caseπ= 6. It would be great if same ideas could be applied for this and subsequent cases as well. However, computation ofπΉπ ,6shows that forπ >4we have the following formula:
πΉπ ,6(π) = 1
(π β 1)!ππ ,6(π) π β
π=6
(πβπ),
where
ππ ,6(π) =π4β (30π β 16)π3+ (150π 2β 90π + 11)π2β (240π 3β 90π 2+ 4)π+ 120π 4.
Some of the polynomialsππ ,6have integer roots. For instance,
π8,6(π) = (πβ 12)(π3β 212π2+ 6347πβ 40960), π10,6(π) = (πβ 32)(π3β 252π2+ 6047πβ 37500), π22,6(π) = (πβ 32)(π3β 612π2+ 51047πβ 878460), π30,6(π) = (πβ 32)(π3β 852π2+ 105047πβ 3037500).
But that is basically all we can get forπ = 6. Alas, no more quadratic Diophantine equations, no chains of conjugation. Also, it seems likely that polynomialsππ ,6do not have integer roots other than the ones shown above.
And, of course, same happens with cases of even greater values ofπ β and so this line of inves-tigation ends here.
6. COMPUTER TO THE RESCUE
Roots ofπΉπ ,π. Trying to ο¬nd more roots for cases π > 4in hope of some insight, I have written a short program in SAGE which was then run through SageMath web interface at CoCalc.com for π = 3,4,5,6,7, etc. until the server started to stumble (that happened somewhere aroundπ = 40). After that I have switched to local install of SAGEand proceeded until π = 200 when every new value ofπ started to require almost a day to process (and then my computer ran out of operational memory).
The program did the following. For every ο¬xed value ofπ it ran the loop forπfrom 1 to 1000, where at each step it computed polynomial πΉπ ,π, factorized it over β€ and in case of non-trivial factorization printed out the roots of the polynomial. At the end it also producedπmax(π )β the last value ofπfor which a non-trivial factorization ofπΉπ ,πoccurred.
Below (in Table 1) you can see the summary of all non-trivial roots (with pairs(2π , π )excluded) obtained from this experiment.
We have marked each entryπwith the ο¬rst value of πfor whichπΉπ ,π(π) = 0. So, for instance, mark[4]corresponds to chainβ¨7β©, and[5]β to chainβ¨10β©.
For all other values ofπ between 3 and200the only roots found were eitherπ= 2π (which would have been marked with[3]) or trivial (1 throughπ ) and therefore of no interest for us.
(14)
π π
3 27[5], 486[9] 4 12[6]
6 8[4],27[4] 8 12[6] 10 32[6]
14 16[5],147[5] 21 27[4],98[4]
22 32[6]
24 27[5],256[5] 30 32[6]
62 64[7]
77 98[4],363[4] 126 128[8]
133 147[5], 1444[5]
TABLE1. Non-trivial roots ofπΉπ ,πfor3β©½π β©½200
Roots ofπΉπ ,πwhich have not been yet veriο¬ed as multiset recovery singularities are emphasized inbold. They represent the current βsuspectβ cases.
In addition the experiment showed that for all 3 < π β©½ 200the value ofπmax(π ) was equal to
2π β 1. Claiming that to be always true is what we will callπmax-Conjectureβ seeConjecture 7.6
below, inSection 7.
The following proposition can be considered as a very easy βhalfβ of this conjecture.
Proposition 6.1. For anyπ >2,π= 2π and any oddπsuch that1< π < πwe have πΉπ ,π(π) = 0. Proof. We can rewrite this statement by usingβ¨3β©, adding summand withπ= 0, and reversing the summation indexπ. As a result we obtain
π β
π=0
(β1)π(π βπ)π (
π π
) = 0
for any even number0< π < π. Now, since
( π π
) =
( π πβπ
)
, (π βπ)π = (π β (πβπ))π,
the equation above is equivalent to
π β
π=0
(β1)π(π βπ)π (
π π
) = 0.
Any polynomial of π with degree less than π (such as (π β π)π) can be expressed as a linear
(15)
proposition then follows from well-known formula
π β
π=0
(β1)ππ[π]
( π π
) = 0,
which can be easily proved using generating functionπ(π₯) = (1 +π₯)π = βπ π=0
(π π )
π₯π. Polynomial π(π₯)has(β1)as a root of orderπ, thus itsπ-th derivativeπ(π)(π₯) =βπ
π=0
(π π )
π[π]π₯πβπ also must have (β1)as a root for any0< π < π.
β And then fromπΉπ ,2π β1(2π ) = 0follows
Corollary 6.2. For anyπ >2we haveπmax(π )β©Ύ2π β 1.
Singularity search. Well, since we already started using computer assistance, let us continue down this slippery slope. The next idea in automating our investigation is to hunt not for the roots of polynomialsπΉπ ,πbut for the singular multisets themselves.
The objective is to try and ο¬nd singularity examples for the smallest βsuspectβ pairs(27,6)and
(32,10). The other suspects, not π-conjugated to these two, are too large to hope for any βbrute forceβ computer search to succeed.
The main idea of this approach is to restrict the realm of theπ-multisets that we deal with. Con-sider all (π+πβ1
πβ1
)
weak compositionsofπ intoπparts, that is, representations of πas a sum ofπ
non-negative integersππ,(π= 1,2, ..., π):
ξΌ βΆ π=π1+π2+...+ππ, ππββ€β©Ύ0.
Each weak composition of this form can be treated as sequence of multiplicities β that is, from each compositionξΌ we will constructπ-multiset
π΄ξΌ = {1π1,2π2, ..., πππ}. β¨11β© Alas, in both cases(27,6)and(32,10)we cannot hope to ο¬ndπ-multisetπwhich isπ -equivalent to its own mirrorπΜ. Indeed, ifπ βΌπ πΜ then without loss of generality we can assume thatπ1(π) = 0. ThusπΜ = βπ and for any evenπ >0the sum ofπ-th powers of numbers inπwill be equal to that ofπΜ.
From the experiment we already know (and can easily verify this formally) that πΉ6,π(27) = 0
iο¬π = 4andπΉ10,π(32) = 0 iο¬π = 6. Therefore, for anyπ such that π(π ) = πΜ(π ) we will have
ππ(π) = ππ(π)Μ for all values of1β©½ π β©½ πβ the only exceptions we could have hoped for were 4 and 6 (forπ= 27andπ= 32respectively) and we have just eliminated them.
So, how can we proceed and what are the challenges?
First of all, we cannot aο¬ord to generate an array of all weak compositions (or multisets of type
β¨11β©) and then analyze the result β the computer would soon run out of memory. For example, if
π= 27andπ= 10then we get almost 100 million of such compositions (94,143,280 to be exact). Thus, the algorithms here have to be iterative. It is fairly easy to write an iterator function which generates next weak composition based on the previous one. Hint: ο¬nd the last non-zero part, increase the previous part by one and make all the following parts zero except for the last one. If necessary, same can be done when going through allπ -subsets in anπ-multiset.
Second, the check function that veriο¬es whether the two given multisetsπ΄andπ΅areπ -equivalent to its mirror π΄Μ must be written very carefully and very eο¬ciently because it will be called quite a few times. To make it work as fast as possible the function needs to implement some βquick
(16)
rejectionβ checks. For instance, the sum of the ο¬rstπ numbers fromπ΄(let us assume it is sorted) is always equal to the minimum number inπ΄(π ); hence these sums forπ΄andπ΅ must coincide. The more simple checks of this sort are employed the better.
Third, calling this check function for every pairπ΄ξΌ andπ΄ξΌβ²of constructed multisets is absolutely
out of the question. So, some sort of simpliο¬ed βsignatureβ has to be computed for each multiset
π΄(ξΌπ ) (alas, no quick rejections there) so we can compare these numbers instead of comparing very large multisetsπ΄(ξΌπ ). But even with that we cannot go much farther beyondπ= 10for the reason I already mentioned above β such huge arrays of data will exhaust the computer memory.
My own implementation of this approach did not ο¬nd any examples of 6-equivalent 27-multisets of typeβ¨11β© forπ < 10. As a sanity check I ran the same code for pair(12,4)withπ = 17and after a few hours of number crunching it resulted in the same unique example of two 4-equivalent 12-multisets already found in [17].
Clearly, absence of positive results in this computational experiment doesnβt mean much as there could be 6-equivalent 27-multisets that span longer stretches of integers. And it is always possible that singularity example for(27,6)simply doesnβt exist. For now, this remains an open question.
If some of the readers become interested in this line of investigation I will gladly send them my code β I am sure it can be made more eο¬ective while consuming less memory. And then, who knows, perhaps the next value ofπwill ο¬nally yield the desired singularity.
7. OPEN QUESTIONS
Here is a list of a few open questions which have come up during this surveyβs investigations.
Question 7.1. Is it true thatξΉ5 = {10}?
Question 7.2. Which of the new βsuspectβ pairs(π, π)found inSection 6are singular?
Of course, we are talking here about pairs highlighted in bold in Table 1. Perhaps some cleverly written computer program could answer this question at least for the smallest βsuspectβ pairs(27,6)
and(32,10).
Question 7.3. Does a non-trivial root of πΉπ ,π guarantee an existence of corresponding singular pair? That is, is it true that for any positive integersπ ,π,πsuch thatπ > π ,π β©ΎπandπΉπ ,π(π) = 0
there exists a pair of differentπ-multisetsπ΄andπ΅such thatπ΄(π ) =π΅(π )?
All the results accumulated over the last sixty years so far conο¬rm this hypothesis; however, it looks like an extremely diο¬cult nut to crack. The following question could perhaps serve as a small step in that direction.
Question 7.4. A singular pairπ = (π, π )is such thatπΉπ ,π(π) = 0forπ= 4or5. Is pairπβ² = (π, π )
obtained fromπ byβ¨π , πβ©-conjugation also singular?
We know that π-conjugation between roots of πΉπ ,π has its direct analog in (π, π ) β (π, πβ π )
duality between singular pairs. However, the similar question aboutπ -conjugation does not seem to be even remotely as simple.
Question 7.5. Is there a less βaccidentalβ explanation for at least one singular pair withπ β 2π ,
π >2?
So far, all such examples were constructed in a ratherad hocmanner by grinding through solu-tions for simultaneous equasolu-tionsππ(π΄(π )) =π
π(π΅
(17)
overwhelmingly complex. Perhaps for at least some singular pairs there exists a less βaccidentalβ construction, combinatorial or algebraic.
I think that the following hypothesis is the most important among reasonably hard open questions on multiset recovery.
Conjecture 7.6. (πmax-Conjecture)Is it true thatπmax(π ) = 2π β 1for allπ > 3? In other words, prove or disprove that polynomialπΉπ ,πcan have integer roots only ifπβ©½ 2π β 1.
Positive answer to this question would mean a considerable breakthrough in any β computer-aided or not β search for the roots of polynomialsπΉπ ,π.
For starters, we could immediately claim thatξΉπ = {2π }for many small values ofπ . Proof of that forπ = 5(that is, a solution toQuestion 7.1) would go like this:
Proof. Ifπ >5andπβ 10thenπΉ5,π(π)β 0for anyπ >0. Indeed,πmax-Conjectureimplies there is no need to check valuesπβ©Ύ10. FromSection 5we know the same is true forπβ©½ 5. So we only need to examineπΉ5,6,πΉ5,7,πΉ5,8andπΉ5,9:
πΉ5,6(π) =π4β 134π3+ 3311π2β 27754π+ 75000, πΉ5,7(π) =π4β 262π3+ 9527π2β 107570π+ 375000, πΉ5,8(π) =π4β 518π3+ 27791π2β 420490π+ 1875000, πΉ5,9(π) =π4β 1030π3+ 81815π2β 1653650π+ 9375000.
How do we do that? Again, we can use computer to help us produce a veriο¬able computer-independent proof. As an example, let us prove (quite formally) that πΉ5,6 has no integer roots in[5; β[⧡{10}. A couple of lines of code in MATLABβ’ or in SAGE will get us real roots of this polynomial
π₯1= 6.014875..., π₯2 = 7.745287..., π₯3= 15.348149..., π₯4 = 104.891687... ,
We cannot use that as a proof, but we can computeby handvaluesπΉ5,6(π₯)for π₯= 6, 7, 8, 15, 16, 104 and 105. The results β 24, (-600), 360, 2040, (-4776), (-745560) and 93480 β prove that there is a non-integer root inside each one of segments[6; 7],[7; 8],[15; 16]and[104; 105]. Those four non-integer numbers obviously constitute the set of all roots ofπΉ5,6.
In exactly the same way (but with longer computations) one can prove the same for polynomials
πΉ5,7,πΉ5,8andπΉ5,9(polynomialsπΉ5,7andπΉ5,9both have one integer root but it is equal to 10). Finally, sinceπΉ5,π(π)β 0, then fromTheorem 4.1it follows thatπ-multisetπ΄is always
recover-able fromπ΄(5). β
The following hypothesis proposes an update toQuestion 2.5.
Conjecture 7.7. (Magical Triplet Conjecture) If π > 2 and π > 2π then for any three distinct
π-multisets some two of them are notπ -equivalent to each other.
If we agree to exclude fully investigated cases ofπ = 2andπ= 2π then let us call three diο¬erent
π-multisets such that they are allπ -equivalent to each other, amagical triplet. I submit thatmagical triplets do not existβ in other words, when one tries to recover a multiset from its collection of
(1)
...(1444,1311)βπ (1444,133) π
β(147,133) π
β(147,14) π
β(16,14) π β π
β(16,2) π β(3,2)
π β(3,1)
π β(4,1)
π β(4,3)
π
β(27,3) π β π
β(27,24) π
β(256,24) π
β(256,232) π
β(2523,232)... β¨10β©
In both cases π = 4 andπ = 5 it is easy to prove that all positive integer solutions of equations
β¨5β©andβ¨8β©belong to the chains 7 and 10 respectively. We will leave that to the reader.
Caseπ= 6. It would be great if same ideas could be applied for this and subsequent cases as well. However, computation ofπΉπ ,6shows that forπ >4we have the following formula:
πΉπ ,6(π) = 1
(π β 1)!ππ ,6(π)
π β
π=6
(πβπ),
where
ππ ,6(π) =π4β (30π β 16)π3+ (150π 2β 90π + 11)π2β (240π 3β 90π 2+ 4)π+ 120π 4.
Some of the polynomialsππ ,6have integer roots. For instance,
π8,6(π) = (πβ 12)(π3β 212π2+ 6347πβ 40960), π10,6(π) = (πβ 32)(π3β 252π2+ 6047πβ 37500), π22,6(π) = (πβ 32)(π3β 612π2+ 51047πβ 878460), π30,6(π) = (πβ 32)(π3β 852π2+ 105047πβ 3037500).
But that is basically all we can get forπ = 6. Alas, no more quadratic Diophantine equations, no chains of conjugation. Also, it seems likely that polynomialsππ ,6do not have integer roots other than the ones shown above.
And, of course, same happens with cases of even greater values ofπ β and so this line of inves-tigation ends here.
6. COMPUTER TO THE RESCUE
Roots ofπΉπ ,π. Trying to ο¬nd more roots for cases π > 4in hope of some insight, I have written a short program in SAGE which was then run through SageMath web interface at CoCalc.com for π = 3,4,5,6,7, etc. until the server started to stumble (that happened somewhere aroundπ = 40). After that I have switched to local install of SAGEand proceeded until π = 200 when every new value ofπ started to require almost a day to process (and then my computer ran out of operational memory).
The program did the following. For every ο¬xed value ofπ it ran the loop forπfrom 1 to 1000, where at each step it computed polynomial πΉπ ,π, factorized it over β€ and in case of non-trivial factorization printed out the roots of the polynomial. At the end it also producedπmax(π )β the last value ofπfor which a non-trivial factorization ofπΉπ ,πoccurred.
Below (in Table 1) you can see the summary of all non-trivial roots (with pairs(2π , π )excluded) obtained from this experiment.
We have marked each entryπwith the ο¬rst value of πfor whichπΉπ ,π(π) = 0. So, for instance, mark[4]corresponds to chainβ¨7β©, and[5]β to chainβ¨10β©.
For all other values ofπ between 3 and200the only roots found were eitherπ= 2π (which would have been marked with[3]) or trivial (1 throughπ ) and therefore of no interest for us.
(2)
π π
3 27[5], 486[9]
4 12[6]
6 8[4],27[4]
8 12[6]
10 32[6]
14 16[5],147[5] 21 27[4],98[4] 22 32[6]
24 27[5],256[5] 30 32[6]
62 64[7]
77 98[4],363[4] 126 128[8]
133 147[5], 1444[5]
TABLE1. Non-trivial roots ofπΉπ ,πfor3β©½π β©½200
Roots ofπΉπ ,πwhich have not been yet veriο¬ed as multiset recovery singularities are emphasized inbold. They represent the current βsuspectβ cases.
In addition the experiment showed that for all 3 < π β©½ 200the value ofπmax(π ) was equal to
2π β 1. Claiming that to be always true is what we will callπmax-Conjectureβ seeConjecture 7.6 below, inSection 7.
The following proposition can be considered as a very easy βhalfβ of this conjecture.
Proposition 6.1. For anyπ >2,π= 2π and any oddπsuch that1< π < πwe have πΉπ ,π(π) = 0. Proof. We can rewrite this statement by usingβ¨3β©, adding summand withπ= 0, and reversing the summation indexπ. As a result we obtain
π β
π=0
(β1)π(π βπ)π (
π π
)
= 0
for any even number0< π < π. Now, since
( π π
)
=
( π πβπ
)
, (π βπ)π = (π β (πβπ))π,
the equation above is equivalent to
π β
π=0
(β1)π(π βπ)π (
π π
)
= 0.
Any polynomial of π with degree less than π (such as (π β π)π) can be expressed as a linear
(3)
proposition then follows from well-known formula
π β
π=0
(β1)ππ[π]
( π π
)
= 0,
which can be easily proved using generating functionπ(π₯) = (1 +π₯)π = βπ π=0
(π π )
π₯π. Polynomial π(π₯)has(β1)as a root of orderπ, thus itsπ-th derivativeπ(π)(π₯) =βπ
π=0
(π π )
π[π]π₯πβπ also must have
(β1)as a root for any0< π < π.
β
And then fromπΉπ ,2π β1(2π ) = 0follows
Corollary 6.2. For anyπ >2we haveπmax(π )β©Ύ2π β 1.
Singularity search. Well, since we already started using computer assistance, let us continue down this slippery slope. The next idea in automating our investigation is to hunt not for the roots of polynomialsπΉπ ,πbut for the singular multisets themselves.
The objective is to try and ο¬nd singularity examples for the smallest βsuspectβ pairs(27,6)and
(32,10). The other suspects, not π-conjugated to these two, are too large to hope for any βbrute forceβ computer search to succeed.
The main idea of this approach is to restrict the realm of theπ-multisets that we deal with. Con-sider all (π+πβ1
πβ1
)
weak compositionsofπ intoπparts, that is, representations of πas a sum ofπ
non-negative integersππ,(π= 1,2, ..., π):
ξΌ βΆ π=π1+π2+...+ππ, ππββ€β©Ύ0.
Each weak composition of this form can be treated as sequence of multiplicities β that is, from each compositionξΌ we will constructπ-multiset
π΄ξΌ = {1π1,2π2, ..., πππ}. β¨11β©
Alas, in both cases(27,6)and(32,10)we cannot hope to ο¬ndπ-multisetπwhich isπ -equivalent to its own mirrorπΜ. Indeed, ifπ βΌπ πΜ then without loss of generality we can assume thatπ1(π) = 0. ThusπΜ = βπ and for any evenπ >0the sum ofπ-th powers of numbers inπwill be equal to that ofπΜ.
From the experiment we already know (and can easily verify this formally) that πΉ6,π(27) = 0
iο¬π = 4andπΉ10,π(32) = 0 iο¬π = 6. Therefore, for anyπ such that π(π ) = πΜ(π ) we will have
ππ(π) = ππ(πΜ)for all values of1β©½ π β©½ πβ the only exceptions we could have hoped for were 4 and 6 (forπ= 27andπ= 32respectively) and we have just eliminated them.
So, how can we proceed and what are the challenges?
First of all, we cannot aο¬ord to generate an array of all weak compositions (or multisets of type
β¨11β©) and then analyze the result β the computer would soon run out of memory. For example, if
π= 27andπ= 10then we get almost 100 million of such compositions (94,143,280 to be exact). Thus, the algorithms here have to be iterative. It is fairly easy to write an iterator function which generates next weak composition based on the previous one. Hint: ο¬nd the last non-zero part, increase the previous part by one and make all the following parts zero except for the last one. If necessary, same can be done when going through allπ -subsets in anπ-multiset.
Second, the check function that veriο¬es whether the two given multisetsπ΄andπ΅areπ -equivalent to its mirror π΄Μ must be written very carefully and very eο¬ciently because it will be called quite a few times. To make it work as fast as possible the function needs to implement some βquick
(4)
rejectionβ checks. For instance, the sum of the ο¬rstπ numbers fromπ΄(let us assume it is sorted) is always equal to the minimum number inπ΄(π ); hence these sums forπ΄andπ΅ must coincide. The more simple checks of this sort are employed the better.
Third, calling this check function for every pairπ΄ξΌ andπ΄ξΌβ²of constructed multisets is absolutely out of the question. So, some sort of simpliο¬ed βsignatureβ has to be computed for each multiset
π΄(ξΌπ ) (alas, no quick rejections there) so we can compare these numbers instead of comparing very large multisetsπ΄(ξΌπ ). But even with that we cannot go much farther beyondπ= 10for the reason I already mentioned above β such huge arrays of data will exhaust the computer memory.
My own implementation of this approach did not ο¬nd any examples of 6-equivalent 27-multisets of typeβ¨11β© forπ < 10. As a sanity check I ran the same code for pair(12,4)withπ = 17and after a few hours of number crunching it resulted in the same unique example of two 4-equivalent 12-multisets already found in [17].
Clearly, absence of positive results in this computational experiment doesnβt mean much as there could be 6-equivalent 27-multisets that span longer stretches of integers. And it is always possible that singularity example for(27,6)simply doesnβt exist. For now, this remains an open question.
If some of the readers become interested in this line of investigation I will gladly send them my code β I am sure it can be made more eο¬ective while consuming less memory. And then, who knows, perhaps the next value ofπwill ο¬nally yield the desired singularity.
7. OPEN QUESTIONS
Here is a list of a few open questions which have come up during this surveyβs investigations. Question 7.1. Is it true thatξΉ5 = {10}?
Question 7.2. Which of the new βsuspectβ pairs(π, π)found inSection 6are singular?
Of course, we are talking here about pairs highlighted in bold in Table 1. Perhaps some cleverly written computer program could answer this question at least for the smallest βsuspectβ pairs(27,6)
and(32,10).
Question 7.3. Does a non-trivial root of πΉπ ,π guarantee an existence of corresponding singular pair? That is, is it true that for any positive integersπ ,π,πsuch thatπ > π ,π β©ΎπandπΉπ ,π(π) = 0 there exists a pair of differentπ-multisetsπ΄andπ΅such thatπ΄(π ) =π΅(π )?
All the results accumulated over the last sixty years so far conο¬rm this hypothesis; however, it looks like an extremely diο¬cult nut to crack. The following question could perhaps serve as a small step in that direction.
Question 7.4. A singular pairπ = (π, π )is such thatπΉπ ,π(π) = 0forπ= 4or5. Is pairπβ² = (π, π ) obtained fromπ byβ¨π , πβ©-conjugation also singular?
We know that π-conjugation between roots of πΉπ ,π has its direct analog in (π, π ) β (π, πβ π )
duality between singular pairs. However, the similar question aboutπ -conjugation does not seem to be even remotely as simple.
Question 7.5. Is there a less βaccidentalβ explanation for at least one singular pair withπ β 2π ,
π >2?
So far, all such examples were constructed in a ratherad hocmanner by grinding through solu-tions for simultaneous equasolu-tionsππ(π΄(π )) =π
π(π΅
(5)
overwhelmingly complex. Perhaps for at least some singular pairs there exists a less βaccidentalβ construction, combinatorial or algebraic.
I think that the following hypothesis is the most important among reasonably hard open questions on multiset recovery.
Conjecture 7.6. (πmax-Conjecture)Is it true thatπmax(π ) = 2π β 1for allπ > 3? In other words, prove or disprove that polynomialπΉπ ,πcan have integer roots only ifπβ©½ 2π β 1.
Positive answer to this question would mean a considerable breakthrough in any β computer-aided or not β search for the roots of polynomialsπΉπ ,π.
For starters, we could immediately claim thatξΉπ = {2π }for many small values ofπ . Proof of that forπ = 5(that is, a solution toQuestion 7.1) would go like this:
Proof. Ifπ >5andπβ 10thenπΉ5,π(π)β 0for anyπ >0. Indeed,πmax-Conjectureimplies there is no need to check valuesπβ©Ύ10. FromSection 5we know the same is true forπβ©½ 5. So we only need to examineπΉ5,6,πΉ5,7,πΉ5,8andπΉ5,9:
πΉ5,6(π) =π4β 134π3+ 3311π2β 27754π+ 75000, πΉ5,7(π) =π4β 262π3+ 9527π2β 107570π+ 375000, πΉ5,8(π) =π4β 518π3+ 27791π2β 420490π+ 1875000, πΉ5,9(π) =π4β 1030π3+ 81815π2β 1653650π+ 9375000.
How do we do that? Again, we can use computer to help us produce a veriο¬able computer-independent proof. As an example, let us prove (quite formally) that πΉ5,6 has no integer roots in[5; β[⧡{10}. A couple of lines of code in MATLABβ’ or in SAGE will get us real roots of this polynomial
π₯1= 6.014875..., π₯2 = 7.745287..., π₯3= 15.348149..., π₯4 = 104.891687... ,
We cannot use that as a proof, but we can computeby handvaluesπΉ5,6(π₯)for π₯= 6, 7, 8, 15, 16, 104 and 105. The results β 24, (-600), 360, 2040, (-4776), (-745560) and 93480 β prove that there is a non-integer root inside each one of segments[6; 7],[7; 8],[15; 16]and[104; 105]. Those four non-integer numbers obviously constitute the set of all roots ofπΉ5,6.
In exactly the same way (but with longer computations) one can prove the same for polynomials
πΉ5,7,πΉ5,8andπΉ5,9(polynomialsπΉ5,7andπΉ5,9both have one integer root but it is equal to 10). Finally, sinceπΉ5,π(π)β 0, then fromTheorem 4.1it follows thatπ-multisetπ΄is always
recover-able fromπ΄(5). β
The following hypothesis proposes an update toQuestion 2.5.
Conjecture 7.7. (Magical Triplet Conjecture) If π > 2 and π > 2π then for any three distinct
π-multisets some two of them are notπ -equivalent to each other.
If we agree to exclude fully investigated cases ofπ = 2andπ= 2π then let us call three diο¬erent
π-multisets such that they are allπ -equivalent to each other, amagical triplet. I submit thatmagical triplets do not existβ in other words, when one tries to recover a multiset from its collection of
(6)
REFERENCES
[1] J. Lambek, L. Moser, (1954),Inverse and Complementary Sequences of Natural Numbers, Amer. Math. Monthly.,
33, pp.454β458
[2] L. Moser, (1957)Problem E1248, Amer. Math. Monthly,64, p.507
[3] J.L. Selfridge, E.G. Straus, (1958),On the determination of numbers by their sums of a fixed order, Paciο¬c J. Math.,
8, pp.847β856
[4] J. Lambek, L. Moser, (1959),On some two way classifications of integers, Canad.J. Math.,2, pp.85β89
[5] A.S. Fraenkel, B. Gordon and E.G. Straus, (1962),On the determination of sets by the sets of sums of a certain
order, Paciο¬c J. Math.,12, pp.187β196
[6] J.A. Ewell, (1968),On the determination of sets by sets of sums of fixed order, Canad. J. Math.,20, pp.596β611 [7] R.K. Guy, (1981),Unsolved problems in Number Theory, Springer-Verlag New York Inc.
[8] J. Boman, E. Bolker, P. OβNeil, (1991),The Combinatorial Radon Transform modulo the Symmetric Group, Ad-vances in Applied Mathematics,12, pp.400-411
[9] D.V. Fomin, O.T. Izhboldin (1994), "Nabory Kratnyh Summ" (in Russian), Proceedings of the St. Petersburg
Mathematical Society, Volume III, St.Petersburg University Publishing, St.Petersburg, Russia
[10] D.V. Fomin, O.T. Izhboldin (1995), "Collections of Multiple Sums",Proceedings of the St. Petersburg
Mathe-matical Society Volume III, AMS Translations, Series 2, vol.166, American Mathematical Society, Providence,
R.I.
[11] J. Boman, S. Linusson, (1996),Examples of Non-Uniqueness for the Combinatorial Radon Transform modulo
the Symmetric Group, Mathematica Scandinavica,78, pp.207-212
[12] R. Honsberger (1997),In PΓ³lyaβs Footsteps. Miscellaneous Problems and Essays, Dolciani Mathematical Expo-sitions,19, The Mathematical Association of America
[13] G. Galperin, A. Tolpygo, D. Leites (transl.) (1997), 60-odd Years of Moscow Mathematical Olympiads, MMO 1935-1997
[14] S. Savchev, T.Andreescu (2003),Mathematical Miniatures, New Mathematical Library,43, The Mathematical Association of America
[15] E.R. Chen, J.C. Lagarias (2008),Problem E11389, Amer. Math. Monthly,115(8), p.758 [16] A. Bogomolny (2011),A Property of the Powers of 2, Cut-The-Knot
[17] J.E. Isomurodov, K.P. Kokhas (2017),A set of 12 numbers is not determined by its set of 4-sums, Journal of Math. Sciences,224, No.2, pp. 258-262
BOSTON, USA