40 Years with Sloane’s Integer Sequences

Jeffrey Shallit
School of Computer Science, University of Waterloo
Waterloo, Ontario N2L 3G1, Canada
[email protected]
http://www.cs.uwaterloo.ca/~shallit

1/1
My Christmas Present in 1974

2/1
Sloane’s First Letter to Me

3/1
Modified Engel expansions
Kalpazidou, Knopfmacher, and Knopfmacher (1990) studied
expansions of the form
1 1 1 1 1 1
α = a0 + − · + · · − ···
a1 a1 + 1 a2 a1 + 1 a2 + 1 a3
for real numbers α, where the ai are non-negative integers with ai
positive for i ≥ 1.
This expansion is essentially unique provided ai+1 ≥ ai .
I tried expanding α = 2/5 and found doubly-exponential growth
(a0 , a1 , . . .) = (0, 2, 3, 7, 13, 97, 193, 18817, 37633, 708158977, . . .).
Although this particular sequence was not in the OEIS at the time,
its even-indexed subsequence
(2, 7, 97, 18817, 708158977, . . .)
certainly was (as A002812) and led to the discovery that ...
4/1
Modified Engel expansions

1 √ i √ i
a2i+1 = (2 + 3)2 + (2 − 3)2
2
and √ √
i i
a2i+2 = (2 + 3)2 + (2 − 3)2 − 1.
From this it was easy to deduce the general form of the expansion
for 2/(2r + 1).

Compare Aho and Sloane’s 1973 paper, “Some Doubly Exponential

Sequences”.

Open Problem: find a simple closed form for the expansion of

3/7.

5/1
Some confusing quotients

Suppose we set c0 = a, c1 = b, and for n ≥ 2 define cn to be the

smallest integer such that
cn cn−1
> .
cn−1 cn−2

The resulting sequence (cna,b )n≥0 = (cn )n≥0 is often easy to

describe:

If a = 1, b = 2, then cn = F2n+1 , the (2n + 1)’th Fibonacci

number.

If a = 2, b = 3, then cn = 2n + 1.

But for other choices, some weird behavior results...

6/1
Some confusing quotients

For example, for a = 8 and b = 55, the resulting sequence

8, 55, 379, 2612, 18002, 124071, 855106, 5893451, 40618081, 279942687, . .

appears to satisfy the linear recurrence

cn = 6cn−1 + 7cn−2 − 5cn−3 − 6cn−4 .

In fact, this is true up to c11055 , but fails for c11056 !

7/1
Some confusing quotients
The reason why was explained by David Boyd.

The reciprocal of the roots of the characteristic polynomial in the

denominator are the real numbers
6.892 · · · , .95484560059 · · ·
and two complex numbers with magnitude
.9548478767 · · · .

Note that the quotient of the the smaller real root with the
magnitude of the imaginary roots is 1.00000238378 · · · , very close
to 1.

This is A010918 in the OEIS.

Open Problem: It is still not known exactly which of these

sequences satisfy linear recurrences.
8/1
Counting monomials

Question:
how many distinct monomials are there in the (expanded) product

X1 (X1 + X2 )(X1 + X2 + X3 ) · · · (X1 + X2 + · · · + Xn )

X1 : 1
X1 (X1 + X2 ) : 2
X1 (X1 + X2 )(X1 + X2 + X3 ) : 5
X1 (X1 + X2 )(X1 + X2 + X3 )(X1 + X2 + X3 + X4 ) : 14

(2nn)
Answer: Cn = n+1 , the n’th Catalan number.

9/1
Lexicographically least sequences avoiding patterns

Consider generating a sequence avoiding “squares” (two

consecutive identical blocks, like the word hotshots) in a greedy
way, starting with 0.

We first write down 0.

Now we can’t follow it with 0, because that would form 00, a

square.

So the next symbol is 1, giving 01.

Now the next symbol can be 0, giving 010.

But the next one can’t be 0 or 1, so choose 2, giving 0102.

10 / 1
Lexicographically least sequences avoiding patterns

It is easy to see, and not hard to prove, that the resulting infinite
sequence over N is

01020103 · · · = (ν2 (n))n≥1 ,

where ν2 (n) is the exponent of the largest power of 2 dividing n.

This is OEIS sequence A007814.

It is also the fixed point of the map that sends each integer i to
the pair (0, i + 1).

11 / 1
Lexicographically least sequences avoiding patterns

Now change the problem very slightly. Instead of trying to avoid

squares, try to avoid overlaps. An overlap is a word of the form
axaxa, where a is a single letter and x is a (possibly empty) word,
like the English word alfalfa.

The resulting sequence is M =

0010011001002001001100100210010020010011001002001001200100110010020010011001003 ···
which is OEIS A161371.

12 / 1
Lexicographically least sequences avoiding patterns

Mathieu Guay-Paquet and I studied this sequence and proved a

number of properties of it:

For example,

- the first occurrence of i = 0, 1, 2 is at position

√
Pi = 1, 3, 13, 79, . . . where Pn = ⌊2n · n! · e⌋. This is OEIS
A010844. Once we knew the formula for Pn , we were able to prove
it rather easily.

13 / 1
Lexicographically least sequences avoiding patterns

- M is the fixed point of the map where 0 → 001, 1 → 1001002,

2 → 200100110010020010011001003, etc., where the length of the
image of i is Qi = 3, 7, 27, 159 . . . This is OEIS A161370 and
Qi = 2Pi + 1.

Open problem: describe the lexicographically least sequence over

N avoiding 52 -powers. Is it over the alphabet {0, 1, 2}?

14 / 1
Counting abelian squares
A word x is an abelian square if the first half is a permutation of
the second half, like the English word reappear.
Consider counting the number Ak (n) of length-2n abelian squares
over an alphabet of size k, say Σ = {1, 2, . . . , k}
Suppose there are i 1’s in the first half of the string. Choose the
position of the 1’s
in the first and last halves of the string. This
2
can be done in ni ways. Now fill in the remaining n − 2i positions
of the string with k − 1 symbols in Ak−1 (n − i) ways. Thus

X n2 X  n 2
Ak (n) = Ak−1 (n − i) = Ak−1 (n − i)
i n−i
0≤i≤n 0≤i≤n
X n2
= Ak−1 (j).
j
0≤j≤n

15 / 1
Counting abelian squares
Since A1 (j) = 1 we immediately get
X n2
A2 (n) = .
i
0≤i≤n

From this we can find an easy proof of the famous identity

X n2 2n
= :
i n
0≤i≤n

Consider a string of length 2n, and choose n positions in it. If a

position falls in the first half of the string, make it 1; if a position
falls in the last half of the string, make it 2. Of the remaining
unchosen positions, make them 2 if they fall in the first half and 1
if they fall in the last half.
It is easy to see that this gives a bijection with the set of abelian
squares.
16 / 1
k-regular sequences

Given a sequence s = (s(n))n≥0 , we can divide it into its odd- and

even-indexed subsequences (s(2n))n≥0 and (s(2n + 1))n≥0 , then
dividing these sequences up again into their odd- and even-indexed
subsequences

(s(4n))n≥0 , (s(4n + 1))n≥0 , (s(4n + 2))n≥0 , (s(4n + 3))n≥0 ,

and so forth. The resulting set of sequences

D2 (s) = {(s(2i n + j))n≥0 : i ≥ 0 and 0 ≤ j < 2i }

is called the 2-decimation of s.

In a similar fashion we can form the k-decimation of s.
If the k-decimation of s is finitely generated (that is, there is a
finite set S of sequences so that Dk (s) is a subset of the Z-span of
S), then s is called k-regular.

17 / 1
An example

Consider the divide-and-conquer linear recurrence counting the

number of comparisons in mergesort:
(
0, if n = 0, 1;
T (n) = n n
T (⌊ 2 ⌋) + T (⌈ 2 ⌉) + n − 1, if n > 1.

Then

T (4n + 2) = −4T (n) + T (2n) + 3T (2n + 1)

T (4n + 3) = −8T (n) + 2T (2n) + 6T (2n + 1) − T (4n + 1)
T (8n) = 4T (n) − 7T (2n) − T (2n + 1) + 4T (4n) + T (4n + 1)
T (8n + 1) = 4T (n) − 6T (2n) − 2T (2n + 1) + 2T (4n) + 3T (4n + 1)
T (8n + 4) = −12T (n) + T (2n) + 7T (2n + 1) + T (4n + 1)
T (8n + 5) = −20T (n) + 4T (2n) + 12T (2n + 1) − T (4n + 1)

18 / 1
An example

So any sequence in the 2-decimation of T is a linear combination

of the subsequences (T (n))n≥0 , (T (2n))n≥0 , (T (2n + 1))n≥0 ,
(T (4n))n≥0 , and (T (4n + 1))n≥0 .

There is a “closed form” for T, as follows:

T (n) = n⌈log2 n⌉ − 2⌈log2 n⌉ + 1.

19 / 1
k-regular sequences in the OEIS

Some more examples:

The Mallows sequence: the unique monotone sequence (a(n))n≥0

of non-negative integers such that a(a(n)) = 2n for n 6= 1. It is
A007378 and is 2-regular.
The Propp sequence: the unique monotone sequence (b(n))n≥0 of
non-negative integers such that a(a(n)) = 3n for n ≥ 0. It is
A007378 and is 3-regular.
We showed that the lexicographically least monotone solution to
a(a(n)) = dn is d-regular.
Number of binary overlap-free words of length n: Cassaigne
showed that this sequence is 2-regular.

20 / 1
Strauss’s sequence

Define
X 2i 
r (n) = ,
i
0≤i<n

and let f (n) = ν3 (r (n + 1)) be the exponent of the highest power

of 3 dividing r (n + 1). The following table gives the first few
values of f :

n 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
f (n) 0 1 2 0 2 3 1 2 4 0 1 2 0 3 4

21 / 1
Strauss’s sequence
A 3-regular sequence recognizer easily produces the following
conjectured relations:

f (3n + 2) = f (n) + 2;
f (9n) = f (9n + 3) = f (3n);
f (9n + 1) = f (3n) + 1;
f (9n + 4) = f (9n + 7) = f (3n + 1) + 1;
f (9n + 6) = f (3n + 1).

With a little more work, one arrives at the conjecture

 
2 2n
ν3 (r (n)) = ν3 (n ),
n

which we proved in 1989.

22 / 1
Strauss’s sequence

A beautiful proof of this identity using 3-adic analysis was later

given by Don Zagier. Zagier showed that if we set
2k
P

0≤k<n k
F (n) = ,
n2 2n


n

then F (n) extends to a 3-adic analytic function from Z3 to

−1 + 3Z3 , and can be evaluated at the negative integers as follows:

(2n − 1)! X (k!)2

F (−n) = −
(n!)2 (k − 1)!
0≤k<n

for n ≥ 0.

23 / 1
Strauss’s sequence

A heuristic k-regular sequence recognizer can produce many

interesting conjectures. For example, let
X n n + k 
a(n) = .
k k
0≤k≤n

Let b(n) = ν3 (a(n)). Then computer experiments strongly suggest:

(
b(⌊n/3⌋) + (⌊n/3⌋ mod 2), if n ≡ 0, 2 (mod 3);
b(n) =
b(⌊n/9⌋) + 1, if n ≡ 1 (mod 3).

Open problem: prove or disprove these relations. (verified for

0 ≤ n ≤ 10, 000).

24 / 1
k-regular sequences in the OEIS

It turns out that k-regular sequences form a large class of

commonly-occurring sequences that are usually easy to recognize.

Back in the early 90’s, when Allouche and I were first looking at
this class of sequences, Neil Sloane provided an electronic copy of
the EIS and we tested it. Out of about 5000 sequences at the
time, approximately 300 (or 6%) of them were k-regular or
“probably” k-regular.

It would be nice to add a k-regular recognizer to the OEIS.

25 / 1
Nørgård’s rhythmic infinity series
Start with two consecutive Fibonacci numbers, such as 55 and 89.
Replace each one with the previous two Fibonacci numbers,
switching order if position is even.
Keep going until 3 is reached.
The Danish composer Per Nørgård used the resulting sequence to
design a series of rhythms in his musical compositions.

26 / 1
Nørgård’s rhythmic infinity series

Write out the indices instead of the Fibonacci numbers:

4, 5, 6, 5, 6, 7, 6, 5, 6, 7, 8, 7, 6, 7, 6, 5, . . .
Subtract 4 from each number to get:

0, 1, 2, 1, 2, 3, 2, 1, 2, 3, 4, 3, 2, 3, . . .
which the OEIS gives immediately as A005811: number of 1’s in
Gray code for n.

27 / 1
Counting squares in finite Fibonacci words

The finite Fibonacci words are given by A061107: X0 = 0, X1 = 1,

and Xn = Xn−1 Xn−2 . Here are the first few:

0, 1, 10, 101, 10110, 10110101, . . .

Fraenkel and Simpson (1999) came up with an exact formula for

R(n), the number of occurrences of squares in the Fibonacci word
Xn :
4 2
R(n) = nF (n + 1) − (n + 6)F (n) − 4F (n − 1) + n + 1,
5 5
for n ≥ 3, where F (n) is the n’th Fibonacci number. (Small error
in their paper corrected here.) This is A248425, just added!

28 / 1
A decision procedure

Recently we implemented a procedure that can automatically

compute formulas of this sort.

Here is a sketch of the idea.

We start with an infinite word generated by iterated morphism, like

the infinite Fibonacci word f = 01001010 · · · , which is a fixed
point of the map sending

0 → 01; 1 → 0.

Next, we associate a numeration system with the word. In this

case the proper numeration system is the Zeckendorf or Fibonacci
numeration Psystem: every positive integer can be written uniquely
in the form i≥2 ǫi F (i), where ǫi ∈ {0, 1} and ǫi ǫi+1 = 0 for all i.

29 / 1
A decision procedure
Now we need two finite automata that read inputs expressed in
this numeration system: one that computes f[n] and one that adds
two Fibonacci representations.

Given all this, we can construct a decision procedure that can

decide all first-order queries about the sequence (in this case, f)
using operations like +, <, =, and indexing into the sequence.

Even more, we can find a linear representation for the number of n

satisfying some predicate.

Here, by a linear representation we mean square matrices M0 , M1

and vectors u, v such that

f (n) = uMa1 Ma2 · · · Mai v ,

where a1 a2 · · · ai is the Fibonacci representation of n.

30 / 1
A decision procedure

For example, here is the predicate that there exists a square of

length 2m beginning at position i in the Fibonacci word:

∀j < m f[i + j] = f[i + m + j].

We can also write a predicate for the assertion that the block
f[i..i + 2m − 1] never occurred previously:

∀ℓ(∀k < 2m f[i + k] = f[ℓ + k]) =⇒ ℓ ≥ i.

Finally, we can also write a predicate for the assertion that the
square occurs entirely within a prefix of length n of the Fibonacci
word: i + 2m < n.

31 / 1
A decision procedure

From the automaton we can compute a linear representation.

From the linear representation we can compute the minimal

polynomial of the matrix.

From the minimal polynomial we can get the roots.

From the fundamental theorem of linear recurrences with constant

coefficients we can get an expression in terms of the roots.

This gives a formula like the one above, in a purely mechanical

fashion!

32 / 1
Long live Neil Sloane!

Neil Sloane at the Fields Institute, Toronto, April 2013

And long live the Encyclopedia of Integer Sequences!

33 / 1
For Further Reading

J.-P. Allouche and J. Shallit, The ring of k-regular sequences, Theoret.

Comput. Sci. 98 (1992), 163–197.
David Boyd, Linear recurrence relations for some generalized Pisot sequences,
in Advances in Number Theory, Oxford, 1993, pp. 333–340.
J. Shallit, Rational numbers with non-terminating, non-periodic modified
Engel-type expansions, Fibonacci Quart. 31 (1993), 37–40.
Mathieu Guay-Paquet and Jeffrey Shallit, Avoiding squares and overlaps over
the natural numbers, Discrete Math. 309 (2009), 6245–6254.
L. B. Richmond and J. Shallit, Counting abelian squares, Electronic J.
Combinatorics 16 (1) (2009), Paper #R72.

C. F. Du, H. Mousavi, L. Schaeffer, and J. Shallit, Decision algorithms for

Fibonacci-automatic words, with applications to pattern avoidance, preprint,
June 2014, http://arxiv.org/abs/1406.0670.

34 / 1