On the derivative of the Minkowki question-mark

function.
Dmitry Gayfulin1
Abstract
We prove new non-improvable results on the derivative of the Minkowski question mark funtion.
arXiv:2107.00461v1 [math.NT] 1 Jul 2021

1 Introduction
1.1 The Minkowski function ?(x).
For an arbitrary x ∈ [0, 1] we consider its continued fraction expansion

1
x = [0; a1 , a2 , . . . , an , . . .] = , aj ∈ Z+
1
a1 +
a2 + · · ·
with natural partial quotients at . This representation is infinite when x 6∈ Q and finite for rational
x. For irrational numbers the continued fraction representation is unique, however each rational x
has two different representations

x = [0; a1 , a2 , . . . , an−1 , an ] and x = [0; a1 , a2 , . . . , an−1 , an − 1, 1], where an > 2.

By
pk
:= [0; a1 , . . . , ak ]
qk
we denote the kth convergent fraction to x. By Bn we denote the nth level of the Stern-Brocot tree,
that is
Bn := {x = [0; a1 , . . . , ak ] : a1 + . . . + ak = n + 1}.
In [12] Minkowski introduced the function ?(x) which may be defined as the limit distribution func-
tion of sets Bn . This function was rediscovered several times and studied by many authors (see
[11],[10],[1],[4],[13]). For irrational x = [0; a1 , a2 , . . . , an , . . .] the formula
∞
X (−1)k+1
?(x) = (1)
k=1
2a1 +...+ak −1

introduced by Denjoy [2, 3] and Salem [15] may be considered as one of the equivalent definitions of
the function ?(x). If x is rational, then the infinite series in (1) is replaced by a finite sum. Note that
?([0; a1 , . . . , at + 1]) =?([0; a1 , . . . , at , 1]) and hence ?(x) is well-defined for rational numbers too. It
is known that ?(x) is a continuous strictly increasing function, also its derivative ?′ (x), if exists, can
take only two values – 0 and +∞. Almost everywhere in [0; 1] in the sense of Lebesgue measure the
derivative exists and equals 0. If x is rational, then ?′ (x) also exists and equals zero.
2010 Mathematics Subject Classification:11J06.
Key words and phrases: Minkowski question mark function, Continued fractions.
1
Research is supported by RNF grant No. 19-11-00001.

1
1.2 Notation and parameters
We will denote the sequences by capital letters A, B, C and their elements by the corresponding small
letters ai , bj , ck . All sequences of the present paper contain positive integers unless otherwise stated.
For an arbitrary finite sequence B = (b1 , b2 , . . . , bn ) we denote
n
←
− X
B = (bn , bn−1 , . . . , b1 ), S(B) = bi .
i=1

By hAi we denote the continuant of (possibly empty) finite sequence A = (a1 , . . . , at ). It is defined
as follows: the continuant of the empty sequence equals 1, ha1 i = a1 , if t > 2 one has

ha1 , a2 , . . . , at i = at ha1 , a2 , . . . , at−1 i + ha1 , a2 , . . . , at−2 i. (2)

Note that the finite continued fraction [0; a1 , . . . , at ] can be expressed using continuants:

ha2 , . . . , at i
[0; a1 , . . . , at ] = . (3)
ha1 , a2 , . . . , at i

Rule (2) can be generalized as follows:

ha1 , a2 , . . . , at , at+1 , . . . , as i = ha1 , a2 , . . . , at ihat+1 , . . . , as i + ha1 , a2 , . . . , at−1 ihat+2 , . . . , as i =

(4)
= ha1 , a2 , . . . , at ihat+1 , at+2 , . . . , as i(1 + [0; at−1 , at−2 , . . . , a1 ][0; at+2 , at+3 , . . . , as ]).

One can find more about the properites of continuants in [9].

For an irrational x = [0; a1 , a2 , . . . , an , . . .] we consider the sum Sx (t) of its partial quotients up
to t−th:
Sx (t) = a1 + a2 + . . . + at .
For an arbitrary finite sequence B we will denote by S(B) and Π(B) the sum and the product of its
elements respectively.
Throughout the paper we always denote the sequence of partial quotients of x by a1 , a2 , . . . , at , . . .
unless otherwise stated. We will also denote the sequence of t first elements of this infinite sequence
by At . Thus, Sx (t) = S(At ).
We also need the following constants
√
1+ 5 2 log Φ
Φ= , κ1 = = 1.3884 . . . , (5)
2 log 2
√
n + n2 + 4
λn = , (6)
2
4 log λ5 − 5 log λ4
κ2 = √ = 4.401 . . . , (7)
log λ5 − log λ4 − log 2
r
κ1 − 1
κ4 = = 0.7486 . . . . (8)
log 2
For an arbitrary sequence A of length t we denote S(A) − κ1 t by ϕ(1) (A). For x = [0; a1 , . . . , at , . . .]
(1)
we denote ϕ(1) (At ) by ϕx (t).

2
1.3 Critical values
In the paper [14] it was shown by by J. Paradis, P. Viader, L. Bibiloni that the value of the derivative
of the function ?(x) is connected with the limit behaviour of Sxt(t) . They showed that if Sxt(t) < κ1
and ?′ (x) exists, then ?′ (x) = +∞. On the other side, denote by z ≈ 5.319 is the root of the equation
2 log(1 + z) = z log 2. If Sxt(t) > z and ?′ (x) exists, then ?′ (x) = 0.
In the paper [4] A. Dushistova and N. Moshchevitin improved the results of [14] and formulated
the following two theorems.
Theorem A.
(i) Let for real irrational x ∈ (0, 1) the inequality Sx (t) < κ1 t hold for all t large enough. Then the
derivative ?′ (x) exists and equals +∞.
(ii) For any positive ε there exists an irrational number x ∈ (0, 1), such that ?′ (x) = 0 and the
inequality Sx (t) < (κ1 + ε)t holds for all t large enough.
Theorem B.
(i) Let for real irrational x ∈ (0, 1) the inequality Sx (t) > κ2 t hold for all t large enough. Then the
derivative ?′ (x) exists and equals 0.
(ii) For any positive ε there exists an irrational number x ∈ (0, 1), such that ?′ (x) = +∞ and the
inequality Sx (t) > (κ2 − ε)t holds for all t large enough.
One can see that the constants κ1 and κ2 in theorems A and B are non-improvable.

1.4 The dual problem

In the paper [5] the dual problem was considered. Suppose that ?′ (x) = 0. How small can be the
(1) (1)
difference ϕx (t) = Sx (t) − κ1 t? Statement (ii) of Theorem A implies that ϕx (t) can be less then
εt for any positive ε. The first non-trivial estimate of ϕ1x (t) was obtained in [5] by A. Dushistova, I.
Kan and N. Moshchevitin.
Theorem C.
(i) Let for irrational x ∈ (0, 1) the derivative ?′ (x) exists and ?′ (x) = 0. Then for any ε > 0 for all t
large enough one has
p
max ϕ(1)
x (u) = max (Sx (u) − κ1 u) > (κ4 − ε) t log t. (9)
u6t u6t

(ii) There exists an irrational x ∈ (0, 1), such that ?′ (x) = 0 and or any ε > 0 for all t large enough
one has √ p
ϕ(1)
x (t) = Sx (t) − κ1 t 6 (2 2 + ε)κ4 t log t. (10)
In the paper [8] a strengthened version of the inequality (9) was obtained within the same condition
 
(1) 2 p
max ϕx (u) > √ + ε κ4 t log t. (11)
u6t 3
Of course, one can ask the same question when ?′ (x) = +∞. How small can be the difference
(2)
ϕx (t) := κ2 t − Sx (t)? The first result in this area was also obtained in [5]. It was improved several
times and for now the best known estimates are the following:
Theorem D.[7]
(i) Let for real irrational x ∈ (0, 1) the derivative ?′ (x) exists and ?′ (x) = +∞. Then for all t large
enough one has √
max (κ2 u − Sx (u)) > 0.06222 t. (12)
u6t

3
(ii) There exists an irrational x ∈ (0, 1), such that ?′ (x) = +∞ and for all t large enough one has
√
κ2 t − Sx (t) 6 0.26489 t. (13)

2 Main results
Note that there are different quantities on the left-hand sides of the inequalities (9) and (10). One
(1)
can say that the inequality (9) considers the uniform behavior of ϕx (t), whereas the inequality (10)
deals with the local behaviour of this quantity. In the present paper we consider upper and lower
estimates in both cases. Our first theorem states that statement (ii) of Theorem C is non-improvable.

Theorem 1. Let for irrational x ∈ (0, 1) the derivative ?′ (x) exists and ?′ (x) = 0. Then for any
ε > 0 for infinitely many t one has
√ p
ϕ(1)
x (t) > (2 2 − ε)κ4 t log t. (14)
(1)
Our second theorem provides optimal estimate of the uniform behavior of ϕx (t).

Theorem 2. (i)Let for irrational x ∈ (0, 1) the derivative ?′ (x) exists and ?′ (x) = 0. Then for any
ε > 0 for all t large enough one has
√ p
max ϕ(1)
x (u) > ( 2 − ε)κ4 t log t. (15)
u6t

(ii) For any ε > 0 there exists an irrational x ∈ (0, 1) such that ?′ (x) = 0 and for infinitely many t
one has √ p
max ϕ(1)
x (u) 6 ( 2 + ε)κ4 t log t. (16)
u6t

3 Auxiliary lemmas
3.1 Increment lemmas and their corollaries
The following two lemmas shed light on the connection between the behavior of the sum of partial
quotients of x and the value of the derivative ?′ (x).

Lemma 3.1. ([5], Lemma 1) For an irrational x = [0; a1 , a2 , . . . , at , . . .] and for δ small in absolute
value, there exists a natural t = t(x, δ) such that

?(x + δ)−?(x) hAt ihAt−1 i

> . (17)
δ 2Sx (t)+4
Lemma 3.2. ([5], Lemma 2) For an irrational x = [0; a1 , a2 , . . . , at , . . .] and for δ small in absolute
value, there exists a natural t = t(x, δ) such that

?(x + δ)−?(x) hAt i2

6 Sx (t)−2 . (18)
δ 2
It is not convenient for us that the numerators of the right-hand sides of (17) and (18) do not
coincide. That is why we prove a ”symmetric” corollary of Lemmas 3.1 and 3.2

4
Lemma 3.3. For an irrational x = [0; a1 , a2 , . . . , at , . . .] the derivative ?′ (x) equals zero if and only
if
hAt i
lim √ Sx (t) = 0. (19)
t→∞
2
Proof. (⇐) Follows immedeately from Lemma 3.2.
(⇒) It follows from Lemma 3.1 that
hAt−1 i
lim √ Sx (t) = 0, (20)
t→∞
2
hAt i
Suppose that √ Sx (t) does not tend to zero as t → ∞. That is, there exists a positive constant c such
2
that
hAt i
√ Sx (t) > c (21)
2
for infinitely many t. Consider N ∈ N such that for any n > N

hAn−1 i c
√ Sx (n) < 100 .
2
Consider an arbitrary integer t > N + aN such that the inequality (21) holds. Then

(at + 1)c (at + 1)hAt−1 i hAt i

> √ S (t)
> √ Sx (t)
>c
100 2
x
2
Thus, at > 100. On the other hand,
√ at √ at
2 c 2 hAt−1 i hAt−2 i c
< √ < √ <
(at−1 + 1)(at + 1) at−1 + 1 2Sx (t) 2
Sx (t−1) 100
√ at
Therefore, we have 100 2 < (at−1 + 1)(at + 1). As at > 100, one can easily see that at−1 > at + 1.
From (21) one can also derive that

hAt−1 i at + 1 hAt−1 i hAt i

√ Sx (t−1) > √ at √ Sx (t−1) > √ Sx (t) > c.
2 2 2 2
Repeating the same argument t − N times we obtain that aN > at + t − N > at + aN and therefore
we come to a contradiction.
(1)
Lemma 3.4. Suppose that ?′ (x) = 0. Then ϕx (t) > 0 for all t large enough.

Proof. From Lemma 3.3 one has

hAt i 1
√ Sx (t) > 2 .
2
Φt
Using an obvious estimate hAt i > 2
from (22), one has

1 hAt i 1 Φt
< √ Sx (t) < √ (1)
.
2 2 2 κ1 t+ϕx (t)
2
√ κ1
As 2 = Φ, we obtain the statement of the lemma.

5
3.2 Continuant lower estimate
The following lemma is a useful tool to estimate the values of continuants, most of whose elements
are equal to 1. We introduce some notation first. Having a sequence At , denote by w(At ) the
number of its elements greater than 1. The set of such elements forms the sequence that we denote
by dAt (1), dAt (2), . . . , dAt (w(At )).
Lemma 3.5. For an arbitrary continuant hAt i = ha1 , a2 , . . . , at i one has
w(At ) 
1 t Y dAt (i)
ha1 , a2 , . . . , at i > Φ . (22)
2 i=1
4

Proof. We prove by induction on w(At ). If w(At ) equals zero, then hAt i = Ft+1 – (t+ 1)-th Fibonacci
number and one can easlily verify (22) using Binet’s formula.
Now suppose that the inequality (22) holds for w(At ) = n. Consider an arbitrary continuant hAt i
having w(At ) = n + 1. Let us also consider the continuant hA′t i, which is obtained from hAt i by
replacing one of its elements greater than one by 1. Of course, w(A′t ) = n and one has
n  
hAt i ′ hAt i 1 t Y dA′t (i)
hAt i = ′ hAt i > Φ . (23)
hAt i hA′t i 2 i=1 4

Applying (4), one can easily see that for arbitrary finite sequences A and B one has
←
− ←
− ←
−
hA, x, Bi = hA, xihBi(1 + [0; x, A ][0; B]) = xhAihBi(1 + [0; A ][0; x])(1 + [0; x, A ][0; B]).

Hence,
←
− ←
−
hA, x, Bi 1 + [0; x, A ][0; B] 1 + [0; A ][0; x] 1 1 x
=x ←
− ←− >x · = . (24)
hA, 1, Bi 1 + [0; 1, A ][0; B] 1 + [0; A ] 2 2 4
hAt i
Applying the lower estimate of hA′t i
from (24) to (23), we prove the induction step. The lemma is
proved.
Lemma 3.6. Consider two arbitrary real numbers β > α > 3 and a real number s > β. Let R(s, α, β)
be the set of all finite sequences R = (r1 , . . . , rk ) of real numbers such that α 6 ri 6 β ∀1 6 i 6 k
and S(R) = s. Then  
s
min Π(R) > β β (25)
R∈R(s,α,β)

Proof. Note that the number k is not fixed in the definition of R(s, α, β). Denote

f (s, α, β) = min Π(R).

R∈R(s,α,β)

It is clear that the function f (s, α, β) is monotonic in the first argument. Hence, without loss of
generality one can say that βs ∈ Z. Consider the sequence Rβ = (β, β, . . . , β ). Using the compactness
| {z }
s/β times
argument, one can easily see that there exists a sequence R0 = (r10 , . . . , rm
∈ R(s, α, β) such that 0
)
Π(R0 ) = f (s, α, β). Suppose that R0 6= Rβ . One can easily show that R0 cannot contain more than
two elements not equal to α or β. Indeed, if α < ri0 6 rj0 < β, then there exists δ > 0 such that
ri0 − δ > α and rj0 + δ < β. As (ri0 − δ)(rj0 + δ) < ri0 rj0 , one can see that

Π(r10 , . . . , ri−1
0
, ri0 −δ, ri+1
0 0
, . . . , rj−1 , rj0 +δ, rj+1
0 0
, . . . , rm ) < Π(r10 , . . . , ri−1
0
, ri0 , ri+1
0 0
, . . . , rj−1 , rj0 , rj+1
0 0
, . . . , rm )

6
and we obtain a contradiction with the definition of R0 .
On the other hand, it follows from the definition of R0 that

Π(R0 ) 6 Π(Rβ ). (26)

Without loss of generality one can assume that R0 does not contain elements equal to β. Indeed, if
we remove all such elements from R0 and the same number of elements from Rβ , the inequality (26)
will be still satisfied. Thus, R0 has the form R0 = (x, α, α, . . . , α) up to transposition of elements.
| {z }
(s−x)/α times
Here α 6 x < β.
Denote n = s−x
α
. Inequality (26) can be written as
αn+x
αn x 6 β β . (27)

But n fact, the opposite inequality is true for all α 6 x 6 β. Taking into account the fact that
αβ > β α for β > α > 3, one can easily verify that
αn+x
αn x > β β (28)
αn+x
for x = α and x = β. As β β is a convex downward function of x, one can deduce that it lies below
the linear function αn x for all α 6 x 6 β. Thus, we obtain a contradiction with (26) and the lemma
is proved.

3.3 Elimination of small elements greater than 1

Lemma 3.7. Let A, C be arbitrary (possibly empty) sequences of positive integers. Let B be a
symmetric sequence. Consider two arbitrary integers p > m > 1. Then

hA, 1, B, p + m − 1, Ci 6 hA, m, B, p, Ci. (29)

p+m
Proof. Denote q = 2
. Consider the function

f (x) = hA, q + x, B, q − x, Ci,

where x run through the set of integers if p + m is even and runs through the set of half-integers
otherwise. One can see that f (x) is a quadratic polynomial with negative leading coefficient. The
maximum of f (x) is attained at the point1
←
− ←
−
[0; B ] − [0; B] + [0; C] − [0; A ]
xm = .
2
←−
As B is a symmetric sequence, [0; B ] = [0; B] and therefore |xm | 6 12 . Hence, as p > m > 1, one can
easily see that
   
p+m−2 p−m
hA, 1, B, p + m − 1, Ci = f − 6f − = hA, m, B, p, Ci.
2 2

1
see [6], Lemma 5.4 for computational details

7
Lemma 3.8. Suppose that ?′ (x) = 0. Then there exists an irrational number y = [0; b1 , b2 , . . . , bt , . . .]
such that:

1. ?′ (y) = 0.

2. For all i ∈ N either bi = 1 or bi > 12.

(1) (1)
3. For all i ∈ N one has ϕy (i) 6 ϕx (i).

Proof. First, let us eliminate all elements equal to 2 from the sequence a1 , a2 , . . .. Denote by s1 the
smallest index such that as1 = 2. Denote by t1 the smallest index greater than s1 such that at1 > 1.
Now the procedure is repeated recursively:

si = min{n : n > ti−1 , an = 2}, ti = min{n : n > si , an > 1}

Thus, we obtain the two (possibly infinite) growing sequences s1 < t1 < s2 < t2 < . . .. Note that
if sj < n < tj for some j, then an = 1. Define the irrational number x′ = [0; a′1 , a′2 , . . . , a′t , . . .] as
follows: 
an − 1 = 1 if n = si for some i ∈ N

′
an = an + 1 if n = ti for some i ∈ N


an otherwise
(1) (1)
One can easily see from the definition of x′ that ϕx′ (t) 6 ϕx (t) for all t > 1. Let us now show that
?′ (x′ ) = 0.
It follows from Lemma 3.7 that ha′1 , a′2 , . . . , a′t i := hA′t i 6 hAt i for all t > 1. On the other hand
|Sx′ (t) − Sx (t)| 6 1. Hence, as
hAt i
lim √ Sx (t) = 0
t→∞
2
by Lemma 3.3, we obtain that
hA′ i
lim √ S t′ (t) = 0
t→∞
2 x
and therefore ?′ (x′ ) = 0. Using the same argument, one can eliminate the elements equal to
3, 4, . . . , 12 from the sequence a′1 , a′2 , . . .. The lemma is proved.

Lemma 3.9. Suppose that ?′ (x) = 0 and all partial quotients of x are either equal to 1 or greater
(1)
than 11. There exists T ∈ N such that for all t > T the inequality ϕx (t) > 3w(At) holds.

Proof. Lemma 3.3 implies that there exists T such that ∀t > T one has

hAt i 1
√ Sx (t) < 2 .
2
x
√ κ1
Hence, as 4
> 3 for x > 12 and 2 = Φ, one has by Lemma 3.5

1 hAt i 1 Φt 3w(At ) 1 3w(At )

> √ Sx (t) > √ = . (30)
2 2 2 (1)
κ1 t+ϕx (t) 2 √ ϕ(1)
x (t)
2 2
Statement of the lemma immediately follows from (30).

8
4 Blocks structure
4.1 Parameters introduction
Let x = [0; a1 , a2 , . . . , at , . . .] be an irrational number such that ?′ (x) = 0. By Lemma 3.8, without
loss of generality one can say that either ai = 1 or ai > 6 for all i ∈ N. Throughout the remaining
part of the paper we consider ε as a fixed positive real number from the statements of Theorems 1
and 2. Let λ = λ(ε) be an arbitrary rational number such that 1 > λ > 1 − ε6 . Define the following
integer constants:
 
10 log ε log 6
M= , P = + 1, N = 2M(P + 2). (31)
log λ log (1 + ε2 )
t0
Now we select an integer parameter t0 large enough such that (1 − λ)λN t0 > log t0
and

λN t0 ∈ Z. (32)

Denote ti = λi t0 , where 1 6 i 6 N. We define Bi = (ati +1 , ati +2 , . . . , ati−1 ) – the i-th block, where
1 6 i 6 N. Denote tN +1 = 0 and BN +1 = (atN+1 +1 , . . . , atN ). Thus, we have

[0; At0 ] = [0; a1 , . . . , at0 ] = [0; BN +1 , BN , . . . , B1 ].

NP
+1 NP
+1
(1) (1)
One can easily see that Sx (t0 ) = S(Bi ) and ϕx (t0 ) = ϕx (Bi ).
i=1 i=1
For each block Bi denote its greatest element by Mi and the index of such element by mi (if the
greatest element is not unique, we take the rightmost one). Thus, ami = Mi . Denote ck = √ Mk .
tk−1 log t0
Let us also consider for each 1 6 i 6 N + 1 the short block Bi′ = (ati +1 , . . . , ami −1 ). Note that

ϕ(1) (1) (1) (1) (1) ′

x (mi ) = ϕ (BN +1 ) + ϕ (BN ) + . . . + ϕ (Bi+1 ) + ϕ (Bi ) + (Mi − κ1 ). (33)

For each 1 6 i 6 N + 1 define the real numbers fi and fi′ from the following identities:
√ S(Bi )+fi √ti−1 log t0 ′
√ S(Bi′ )+fi′ √ti−1 log t0
hBi i = 2 , hBi i = 2 . (34)

4.2 Lower estimate of ϕ(1) (Bk )

Lemma 4.1. Suppose that ?′ (x) = 0. Then for all 1 6 i 6 N one has
N
X +1
p p
fi′ ti−1 log t0 + fk tk−1 log t0 < 0. (35)
k=i+1

Proof. As ?′ (x) = 0, by Lemma 3.3, whithout loss of generality one can say that for all 1 6 i 6 N
the inequality
hBN +1 , BN , . . . , Bi+1 , Bi′ i
√ S(BN+1 )+S(BN )+...+S(Bi+1 )+S(Bi′ ) < 1
2
is satisfied. Using (4) we obtain
hBN +1 ihBN i . . . hBi+1 ihBi′ i
√ S(BN+1 )+S(BN )+...+S(Bi+1 )+S(Bi′ )
< 1. (36)
2
Substituting (34) to (36) and taking logarithm of both parts, we get the statement of the lemma.

9
Lemma 4.2. Suppose that ?′ (x) = 0. If the inequality

|ϕ(1) (Bk )| > κ1 (tk−1 − tk )ε5 , (37)

holds for some 1 6 k 6 N + 1, then one has:

max ϕ(1) 0.9

x (u) > t0 . (38)
tN 6u6t0

Proof. Suppose that

ϕ(1) (Bk ) > κ1 (tk−1 − tk )ε5 .
As,
ϕ(1) (1) (1)
x (tk−1 ) = ϕx (tk ) + ϕ (Bk ),
(1)
using the fact that ϕx (tk ) > 0 by Lemma 3.4, we have

max ϕ(1) (1) (1) 5

x (u) > ϕx (tk−1 ) > ϕ (Bk ) > κ1 (tk−1 − tk )ε = κ1 λ
k−1
(1 − λ)t0 > t00.9 .
tN 6u6t0

On the other hand, if

ϕ(1) (Bk ) 6 −κ1 (tk−1 − tk )ε5 ,
(1)
we again use the fact that ϕx (tk−1 ) > 0 and obtain

max ϕ(1) (1) (1) (1) 5 0.9

x (u) > ϕx (tk ) = ϕx (tk−1 ) − ϕ (Bk ) > κ1 (tk−1 − tk )ε > t0 .
tN 6u6t0

One can easily deduce from Lemma 4.2 that if the inequality (37) holds, then the inequalities (14)
and (15) are also satisfied. Therefore, throughout the remaining part of the paper we will assume
that
S(Bk ) = κ1 (tk−1 − tk )(1 + o(ε4 )) (39)
for all 1 6 k 6 N + 1.

Lemma 4.3. Suppose that ?′ (x) = 0. Then for all 1 6 k 6 N + 1 one has
√
(1) (κ1 − 1)(tk−1 − tk ) log t0 (1 + o(ε4)) p
ϕ (Bk ) > √ − fk tk−1 log t0 . (40)
ck tk−1 log 2

Proof. We recall that w(Bk ) is the number of elements of the block Bk which are greater than 1 and
dBk (1), . . . , dBk (w(Bk )) is the sequence of such elements. It follows from Lemma 4.2 and Lemma 3.9
that
w(Bk )
X
dBk (i) = (κ1 − 1)(tk−1 − tk )(1 + o(ε4 )).
i=1
dBk (i)
Therefore, as dBk (i) > 12 for all i, one has 4
> 3. Now we can obtain a lower estimate of

w(Bk ) 
Y dBk (i)
i=1
4

10
 p
applying Lemma 3.6 for s = 14 (κ1 − 1)(tk−1 − tk )(1 + o(ε4 )), α = 3, and β = ck tk−1 log t0 . We have
w(Bk )   p  (κ1 −1)(t√k−1 −tk )(1+o(ε4 ))
Y dBk (i) ck tk−1 log t0 ck tk−1 log t0
> . (41)
i=1
4 4

Substituting the estimate (41) to (22) we obtain

 p  (κ1 −1)(t√k−1 −tk )(1+o(ε4 ))
c k tk−1 log t0 ck tk−1 log t0
hBk i > Φtk−1 −tk . (42)
4
√ κ1
Taking into account that 2 = Φ, from (34), (42) and Lemma 3.3 we get:
 p  (κ1 −1)(t√k−1 −tk )(1+o(ε4 )) √
9ci tk−1 log t0 ci tk−1 log t0 √ ϕ(1) (Bk )+fi tk−1 log t0
6 2 . (43)
16
Taking logarithms of both parts of (43), after some transformations we obtain the inequality
(κ1 − 1)(tk−1 − tk )(1 + o(ε4 )) log t0 p √
p 6 (ϕ(1) (Bk ) + fk tk−1 log t0 ) log 2 (44)
ci tk−1 log t0 2
which is equivalent to (40). Lemma is proved.

4.3 Main estimate

For k 6 N one can write (40) as
 
(1) (κ1 − 1)(1 − λ)(1 + o(ε4 )) p
ϕ (Bk ) > − fk tk−1 log t0 . (45)
ck log 2
For k = N + 1 we have
 
(1) (κ1 − 1)(1 + o(ε4 )) p
ϕ (BN +1 ) > − fN +1 tN log t0 . (46)
cN +1 log 2
p
Let the greatest element of the short block Bk′ be equal to c′k tk−1 log t. Using the argument from
Lemma 4.3, one can deduce the following lower estimate of ϕ(1) (Bk′ ):
√
(1) ′ S(Bk′ ) log t0 (1 + o(ε4 )) ′
p
ϕ (Bk ) > √ − fk tk−1 log t0 . (47)
c′k tk−1 log 2
Lemma 4.4. Suppose that ?′ (x) = 0. Then for all 1 6 i 6 N one has
X N √ k−i  !
(κ1 − 1)(1 − λ) ( λ) p

ϕ(1)
x (mi ) > + ci ti−1 log t0 1 + o(ε4 ) . (48)
log 2 ck
k=i+1

Proof. Substituting the estimates (40) for k = N, N − 1, . . . , i + 1, (46), and (47) to (33) and taking
into account the inequality (35) we obtain:
 √
(1) 4 κ1 − 1 p S(Bi ) log t0
ϕx (mi ) > (1 + o(ε )) tN log t0 + √ +
log 2cN +1 log 2c′i ti−1
√ N  (49)
(κ1 − 1) log t0 X tk−1 − tk p
√ + (ci ti−1 log t0 − κ1 ).
log 2 k=i+1
ck tk−1

11
Using the trivial estimates
√
S(Bi ) log t0 κ1 − 1 p
√ > 0, tN log t0 > 0
log 2c′i ti−1 log 2cN +1
we obtain
 √ N 
(κ1 − 1) log t0 X tk−1 − tk p
ϕ(1)
x (mi )
4
> (1 + o(ε )) √ + (ci ti−1 log t0 − κ1 ). (50)
log 2 ck tk−1
k=i+1

Taking into account the fact that tk = λk t0 , we immediately obtain (48). The lemma is proved.
Inequality (48) is the key tool that we will use in proofs of Theorem 1 and the first statement of
Theorem 2. Let us simplify it using a new notation. Denote
(κ1 − 1) 1
α= , η= . (51)
log 2 α
Then, one can rewrite (48) as follows:
 N √ 
X ( λ)k−i p

ϕ(1)
x (mi ) > (1 − λ) + ηci α ti−1 log t0 . 1 + o(ε4 ) . (52)
k=i+1
ck

5 Key lemmas
(1)
The inequality (52) reduces the estimation of max ϕx (mi ) to the problem of finding maximum of
16i6N
the following quantity:
N √
X ( λ)k−i
(1 − λ) + ηci , i = 1, 2, . . . , N.
k=i+1
c k

Note that this problem does not deal with the Minkowski function, continued fractions etc. It is
purely combinatorial. The following lemma allows us to estimate the desired maximum.
Lemma 5.1. Let η, c1 , c2 , . . . , cN be arbitrary positive real numbers. Define the real numbers ϕi as
follows:
N √ k−i
X λ
ϕi = (1 − λ) + ηci . (53)
k=i+1
ck
Then the following inequality holds:
p
max ϕi > 8η(1 + o(ε)). (54)
16i6N

Proof. The proof of the lemma will be splitted into several steps. We also recall that the constants
M, N, and P used in our argument are defined in (31).
Lemma 5.2. Suppose that the inequality (54) is not satisfied. Then there exists a natural number
i1 6 M such that ci1 > 2√1 η . Moreover, for all i 6 N one has

3
ci < √ . (55)
η

12
Proof. Suppose the contrary. Let us estimate ϕ1 from below:
N √ k M √ k M
X λ X λ √ X √ k
ϕ1 > (1 − λ) + ηc1 > (1 − λ) > 2 η(1 − λ) λ =
k=1
c k+1
k=1
c k+1
k=1
√ M (56)
√ √ 1− λ √ √ p
= 2 η(1 − λ) λ √ = 2 η(1 + λ)(1 + o(ε4 )) > 8η(1 + o(ε4 )).
1− λ
We obtain a contradiction with (54). The estimate (55) comes from the trivial inequality ϕi > ηci .
Lemma is proved.
Lemma 5.3. Suppose that the inequality (54) is not satisfied. Then for all im < N − M, m > 1
there exists a number im < im+1 < im + M such that cim+1 > (1 + ε2 )cim .
Proof. Suppose the contrary. Let the inequality cim +j < cim (1 + ε2 ) be satisfied for all 1 6 j 6 M.
Then, using the argument from Lemma 5.2 we obtain
M √ j M √ j
X λ 1−λ X λ
ϕim > (1 − λ) + ηcim > + ηcim >
c
j=1 im +j
1 + ε2 j=1 cim
  r (57)
2 8η

2
+ ηcim (1 + o(ε4 )) > 2
1 + o(ε4 ) .
(1 + ε )cim 1+ε
In the last ”>” of (57) we use the Cauchy-Shwartz inequality. We come to a contradiction with (54).
Lemma is proved.
Now we are ready to prove Lemma 5.1. Suppose that the inequality (54) is not satisfied. By
Lemma 5.2 there exists 1 6 i1 < M such that ci > 2√1 η . Applying Lemma 5.3
 
log 6
P = log (1+ε2 ) + 1 times, we obtain the number ciP +1 such that ciP +1 > 2√1 η (1 + ε2 )P η > √3η . We
come to a contradiction with (55). Lemma is proved.
One can rewrite (52) as
 N √ 
√ i X ( λ)k−i p
ϕ(1)
x (mi ) > λ (1 − λ) + ηci α t0 log t0 .(1 + o(ε4 )). (58)
ck
k=i+1
√ (1)
Note that λ = 1 + o(ε5 ). As Theorem 2 requires us to show that max ϕx (mi ) is greater than
16i6N
√
t0 log t0 multiplied by some constant factor, we need to estimate the maximum of the following
quantity
 N √ 
i
X ( λ)k−i
λ (1 − λ) + ηci , i = 1, 2, . . . , N.
ck
k=i+1
This estimate is provided by the following lemma.
Lemma 5.4. Let C = (c1 , c2 , . . . , cN ) be an arbitrary sequence of non-negative real numbers. Let η
be an arbitrary positive real number. Define the numbers ϕi using (53). Define the numbers ϕ′i as
follows:
N √ k
′
√ i X λ √ i
ϕi (C) = λ ϕi = (1 − λ) + λ ηci . (59)
k=i+1
c k

Then one has: p

max ϕ′i (C) > 2η(1 + o(ε)). (60)
16i6N

13
Proof. The proof of the lemma will be also splitted into several steps. First, using the substitution
√ j
dj = λ cj , we write (59) as
N
X λk
ϕ̃′i (D) := ϕ′i (C) = (1 − λ) + ηdi . (61)
k=i+1
d k

Here D = (d1 , d2 , . . . , dN ). Denote ϕ̃′max (D) = max ϕ̃′k (D). In order to prove the lemma it is enough
16k6N
for us to show that p
min ϕ̃max (D) = 2η(1 + o(ε)). (62)
D∈RN
+

Denote the minimum of (62) by ymin .

Lemma 5.5. Suppose that ϕ̃′max (D) = ymin for some D ∈ RN
+ . Then for all 1 6 k 6 N one has
ϕ̃′k (D) = ymin .
Proof. Suppose that ϕ̃′max (D) = ymin , but for some i one has ϕ̃′i (D) < ymin . We call the index n
minimizing if ϕ′n (D) = ymin . Let k be the largest non-minimizing index. Without loss of generality
one can say that all indices less than k are not minimizing too. Indeed, consider the sequence

D ′ = (d1 , . . . , dk−1, dk + δ, dk+1 , . . . , dN ),

where δ > 0 is some small parameter. One can easily see that ϕ̃′i (D ′ ) < ϕ̃′i (D) for i < k, ϕ̃′i (D ′ ) >
ϕ̃′i (D) for i= k and ϕ̃′i (D ′ ) = ϕ̃′i (D) for i > k. As k is a non-minimizing index, there exists δ > 0
such that ϕ̃′i (D ′ ) < ymin for all i 6 k.
Thus, k + 1 is the smallest minimizing index. As all indices less than k + 1 are not minimizing,
there exists δ > 0 such that for the sequence

D ′′ = (d1 , . . . , dk , dk+1 − δ, dk+2 , . . . , dN )

one has ymin > ϕ̃′i (D ′′ ) > ϕ̃′i (D) for i 6 k and ϕ̃′i (D ′′ ) < ϕ̃′i (D) = ymin for i = k + 1. Thus, we
obtained the sequence whose smallest minimizing index is at least k + 2. Repeating this argument
(N ) (N ) (N )
we obtain the sequence D (N ) = (d1 , d2 , . . . , dN ) with the smallest minimizing index equal to N.
One can easily see that for the sequence
(N ) (N ) (N )
D ′(N ) = (d1 , d2 , . . . , dN − δ)

for δ > 0 small enough one has ϕ̃′max (D ′(N ) ) < ϕ̃′max (D (N ) ) = ymin and we obtain a contradiction
with the definition of ymin . Lemma is proved.
Lemma 5.6. There exists a unique sequence D = (d1 , d2 , . . . , dN ) such that ϕ̃′max (D) = ymin . The
elements of this sequence satisfy the recurrent equation
q
k+1
dk + d2k + 4(1−λ)λ
η
dk+1 = (63)
2
with the initial condition d1 = 0.
Proof. Lemma 5.5 implies that ϕ̃′N (D) = ymin . This fact yields a linear equation ηdN = ymin from
which dN is uniquely defined. Then we substitute dN to the equation ϕ̃′N −1 (D) = ymin and find dN −1
etc. Therefore the sequence D is uniquely defined.

14
 Suppose that d1 > 0. But if we decrease d1 , we would also decrease ϕ̃′1 (D), but ϕ̃′max (D) would be
still equal to ymin . We obtain a contradiction with the uniqueness of D. Finally, from the equation
ϕ̃′k+1 (D) = ϕ̃′k (D) one can derive
λk+1
η(dk+1 − dk ) = (1 − λ) . (64)
dk+1
Considering (64) as the quadratic equation on dk+1 and choosing the positive root, we obtain (63).
Lemma is proved.
Thus, ymin = dηN where dN can be evaluated from the recurrent equations (63). Multiplying both
√
sides of (63) by η, we obtain
√ p√
√ ηdk + ( ηdk )2 + 4(1 − λ)λk+1
ηdk+1 = . (65)
2
√
Put ek = ηdk . Using the introduced notation, one can write (65) as follows:
p
ek + e2k + 4(1 − λ)λk+1
ek+1 = . (66)
2
Pn
Denote2 δk = (1 − λ)λk и Xn = δn .
k=1
√ √ √
Lemma 5.7. For n > 1 one has en < 2Xn and en+1 − en > 2Xn+2 − 2Xn+1 .
Proof.
√ The first statement is proved by induction. For n = 1 one can easily verify the inequality. As
x is a convex function, one can easily see that
p √ p
en + e2n + 4δn+1 2Xn + 2Xn + 4δn+1 p p
en+1 = < < 2Xn + 2δn+1 = 2Xn+1 . (67)
2 2
Now we prove the second statement. Note that
p
e2n + 4δn+1 − en 2δ δn+1
en+1 − en = = p n+1 = . (68)
2 en + e2n + 4δn+1 en+1
√ √
On the other hand, as we already showed, 2Xn+2 > 2Xn+1 > en+1 and therefore
p p 2δn+2 δn+1
2Xn+2 − 2Xn+1 = √ √ < = en+1 − en . (69)
2Xn+2 + 2Xn+1 en+1
Lemma is proved.
√
Now we are ready to prove Lemma 5.4. By Lemma 5.7, 2Xn+1 − en forms a decreasing sequence
of positive real numbers. Hence
p p p p √
0 < 2XN − eN −1 < 2X2 − e1 < 2 X2 = 2(1 − λ)(λ + λ2 ) < 2 1 − λ = o(ε2 ).
N
P √
On the other hand, 2XN = 2 (1 − λ)λk = 2(1 − λN +1 ) = 2 + o(ε5 ). Thus, eN −1 = 2 + o(ε2 ). It
k=1 √
also follows from (66) that eN = 2 + o(ε2 ). Hence,
√ p
ymin = ηdN = ηeN = 2η + o(ε2 ).
Lemma is proved.
2 e −e 1
Equation (66) is equivalent to k+1
δk+1
k
= ek+1 . Thus, (66) might be considered as numerical integration of the
′ 1
differential equation y = y on the non-uniform grid Xi . I am thankful to I. Mitrofanov who drew my attention to
this fact.

15
6 Proof of Theorem 1 and the first statement of Theorem 2
Now we are ready to prove Theorem 1.
Proof. From (52) and Lemma 5.1 one can deduce that for all t0 large enough there exist integer
numbers i and mi satisfying 1 6 i 6 N and logt0t0 < mi 6 t0 such that
p p √ p
ϕ(1)
x (mi ) > 8ηα ti−1 log t0 (1 + o(ε)) = 2 2κ4 ti−1 log t0 (1 + o(ε)). (70)
ti
As ti < mi 6 ti−1 and ti−1
= λ = 1 + o(ε5 ), from (70) one has
√ p
ϕ(1)
x (mi ) > 2 2κ4 mi log mi (1 + o(ε)).
Theorem is proved.
Let us now prove the first statement of Theorem 2.
Proof. It follows from (58) and Lemma 5.4 that for all t0 large enough there exist integer numbers i
and mi satisfying 1 6 i 6 N and logt0t0 < mi 6 t0 such that
√ i (1) p p √ p
λ ϕx (mi ) > 2ηα ti log t0 (1 + o(ε)) = 2κ4 ti log t0 (1 + o(ε)).
In other words, √ p
max ϕ(1) (1)
x (u) > ϕx (mi ) > 2κ4 t0 log t0 (1 + o(ε)).
u6t0
The first statement of Theorem 2 is proved.

7 Proof of the second statement of Theorem 2

7.1 Superblocks definition
Proof. In this chapter we will construct an irrational number x = [0; a1 , . . . , an , . . .] that will satisfy
the conditions of the second statement of Theorem 2. The continued fraction of x will have the form
x = [0; B(0) , B(1) , B(2) , . . . , B(n) , . . .], (71)
where the segments B(i) will be defined later. We will call these segments superblocks. Recall that the
sequence dk is defined from the equations (63) with the initial condition
qd1 = 0. It will be convenient
for us to modify the first element of this sequence. We set d1 = d2 = (1−λ)λ η
= o(ε5 ).
Let us now define the sequence Ti that plays the key role in our construction. We choose  Ti−1an

N T1
arbitrary integer T1 , satisfying λ T1 > log T1 . Then, if the number Ti−1 is defined, we put Ti = λN .
(i)
For each i > 1 put t0 = Ti .
Now we describe the construction of the superblock B(i) for an arbitrary positive integer i. For
(i) (i) (i)
each k such that 1 6 k 6 N we select three natural numbers mk , nk и tk from the following
conditions:
p (i)
p
dk Ti log Ti 6 mk 6 dk Ti log Ti (1 + ε4 )
dk p (i) dk p
Ti log Ti 6 nk 6 Ti log Ti (1 + ε4 )
κ1 − 1 κ1 − 1
log Ti ε (i) (i) (i) log Ti ε (72)
(1 + − ε3 ) 6 mk + nk − κ1 (nk + 1) 6 (1 + + ε3 )
log 2 8 log 2 8
(i) (i) (i) 1 (i) (i) (i)
tk = λk t0 + θ(nk + 1), where |θ| 6 , (nk + 1) | (tk−1 − tk ).
2
16
 (i) (i) (i)
One can easily see that the numbers mk , nk , tk , satisfying conditions (72) always exists. Consider
(i)
the block Bk = (at(i) +1 , . . . , at(i) ) having the following structure:
k k−1

(i) (i) (i) (i)

Bk = (mk , 1, 1, . . . , 1, mk , 1, 1, . . . , 1, . . . , mk , 1, 1, . . . , 1). (73)
| {z } | {z } | {z }
(i) (i) (i)
nk numbers nk numbers nk numbers

(i) (i) (i)

We denote the sequence of N blocks (BN , BN −1 , . . . , B1 ) by B(i) . For the initial superblock B(0) =
(a1 , a2 , . . . , at(1) ) we set all its elements to be equal to 1. Thus, the construction of the continued
N
fraction (71) is fully described. Note that the initial superblock B(0) has fixed length and therefore
(1)
does not affect neither the value of ?′ (x) nor the behavior of ϕx (t) as t grows.

7.2 ?′ (x) = 0
As we already mentioned it is enough to show that for x′ = [0; B(1) , B(2) , . . . , B(n) , . . .]) one has
?′ (x′ ) = 0. Denote the elements of continued fraction expansion of x′ by a′1 , a′2 , . . .. By Lemma 3.3 it
is enough to show that the function

ha′ , a′2 , . . . , a′t i

fx′ (t) = √1 a′ +a ′ +...+a′ (74)
21 2 t

tends to 0 as t → ∞. One can easily see that the function fx′ (t) grows if a′t = 1 and decreases if
(i)
a′t > 5. Of course, all mk are greater than 5. Thus it is enough to consider only the case when
a′t = 1, but a′t+1 > 5. In this case the continuant ha′1 , a′2 , . . . , a′t i consists of the sequences of the form
(i)
(ml , 1, 1, . . . , 1). From the formula hA, Bi 6 2hAihBi and Lemma 3.3 one can easily see that if we
| {z }
(i)
nl numbers
show that
(i)
2hml , 1, 1, . . . , 1i
| {z }
(i)
nl numbers 1
√ (i) (i)
< , (75)
ml +nl 2
2
(i) (i) (i)
we will obtain the fact that ?′ (x) = 0. Note that 4hml , 1, 1, . . . , 1i < 8ml Φnl . From the definition
| {z }
(i)
nl
numbers
(i) (i) (i) log Ti
(72) one can easily see that ml + nl = κ1 nl + log 2
(1 + 8ε + o(ε2 )). Thus, (75) is equivalent to the
following:
(i) (i) (i)
8ml Φnl 8ml
√ (i) log Ti
=√ log Ti < 1. (76)
κ 1 nl + (1+ 8ε +o(ε2 )) log 2
(1+ 8ε +o(ε2 ))
2 log 2
2
(i)
Taking logarithm of both sides of (76) and substituting ml from (72) , we obtain that

log Ti  2
 log T 
i ε 2
 √ log Ti  ε 2

1 + o(ε ) < 1 + + o(ε ) log 2 = 1 + + o(ε ) . (77)
2 log 2 8 2 8

Therefore, ?′ (x) = 0.

17
7.2.1 The inequality (16) is satisfied
We will show that for x that we built in Section 7.1 for all t = Ti , i > 1 the inequality (16) is satisfied.
This inequality is equivalent to the following:
√ p
ϕ(1)
x (ν) = S x (ν) − κ1 ν 6 ( 2 + ε)κ4 t log t ∀ν 6 t. (78)
(1)
We will prove (78) by induction
on i. Suppose that t = T1 = t0 . As the segment B(0) =
(a1 , a2 , . . . , at(1) ) has length o(ε6 ) t, we will not take it into account in our further argument.
N
One can easily see that ϕ1 (ν) > ϕ1 (ν − 1) if and only if aν > 1. Thus, it is enough to verify
(1) (1)
the inequality (78) only in the case when aν > 1. Suppose that ti < ν 6 ti−1 It follows from the
definition (72) that
(1) (1) (1)
mi + ni − κ1 (ni + 1) > 0.
Hence for any finite sequence B and for any 1 6 j 6 N one has
(1)
ϕ(1) (B, 1, . . . , 1 , mj ) > ϕ(1) (B).
| {z }
(1)
nj numbers

(1) (1)
Thus, it is enough to verify (78) only for the largest tk < ν 6 tk−1 such that aν > 1. In this case
we have
N N  
(1)
X
(1) (1) (1)
X T1 (λi−1 − λi ) log T1  ε
(1)
ϕx (ν) 6 ϕ (Bi ) + mk = (1)
1+ 1 + o(ε2 ) + mk . (79)
i=k i=k ni + 1 log 2 8

This can be written as follows:

N
!
(1 − λ)(κ1 − 1)  ε  X λi−k p

ϕ(1)
x (ν) 6 1+ + dk T1 log T1 1 + o(ε2 ) . (80)
log 2 8 i=k+1 di

κ1 −1
Denote α = log 2
(1 + 8ε ), η = α1 . Using the introduced notation we have

N
!
X λi−k p

ϕ(1)
x (ν) 6 (1 − λ) + ηdk α T1 log T1 1 + o(ε2) . (81)
i=k+1
di

We recall that d1 , d2 , . . . , dN is the minimizing sequence for (61). Thus from Lemma 5.5 we obtain
that
 
(1)
p p √ p √ ε p
ϕx (ν) 6 α 2η T1 log T1 (1 + o(ε)) = 2α T1 log T1 (1 + o(ε)) 6 2+ κ4 T1 log T1 . (82)
2

Thus the inequality (78) is satisfied for t = T1 . Using the same argument one can derive that for all
n ∈ N one has  
(1) (n)
√ ε p
ϕ (B ) 6 2+ κ4 Tn log Tn (83)
2
and for all 1 6 k 6 N, n ∈ N
N  
X
(1) (n) (n)
√ ε p
ϕ (Bi ) + mk 6 2+ κ4 Tn log Tn . (84)
i=k
2

18
Now we show that (78) is satisfied for t = Tn when n > 2. Using the same argument, one can deduce
n−1
X N
X (n) (n)
ϕ(1)
x (ν) 6 (i)
ϕ1 (B ) + ϕ(1) (Bi ) + mk (85)
i=1 i=k

for some 1 6 k 6 N. From (83) it follows that

n−1   X n−1   n−1 p
X
(i)
√
ε p √ ε p X
ϕ1 (B ) 6 2+ κ4 Ti log Ti 6 2+ κ4 log Tn Ti =
i=1
2 i=1
2 i=1
  n−1   √ N  
√ ε p X √ Ni √ ε p λ p
= 2+ κ4 Tn log Tn λ 6 2+ κ4 Tn log Tn √ N = Tn log Tn o(ε4 ).
2 i=1
2 1− λ
(86)

Substituting the estimates (84) and (86) to (85) we obtain that

 
(1)
√ 2ε p
ϕx (ν) 6 2+ κ4 Tn log Tn (87)
3

and the second statement of Theorem 2 is proved.

Acknowledgements: I would like to thank Igor Kan for fruitful discussions. Most results of
this paper are based on his ideas.

References
[1] G. Alkauskas, Integral transforms of the Minkowski question mark function, PhD thesis, Not-
tingham 2008.

[2] A. Denjoy, Sur une fonction reelle de Minkowski, C. R. Acad. Sci. Paris 194 (1932), 44-46.

[3] A. Denjoy, Sur une fonction de Minkowski, J. Math. Pures Appl. 17 (1938), 105-151.

[4] A.A. Dushistova, N.G. Moshchevitin, On the derivative of the Minkowski question mark function
?(x) (Russian), Fundam. Prikl. Mat. 16 (2010), 33-44; translation in J. Math. Sci. (N.Y.) 182
(2012), 463-471.

[5] A.A. Dushistova, I.D. Kan, N.G. Moshchevitin, Differentiability of the Minkowski question mark
function, J. Math. Anal. Appl. 401 (2) (2013), 774-794.

[6] D Gayfulin, Derivatives of two functions belonging to the Denjoy–Tichy–Uits family St. Peters-
burg Mathematical Journal, 2016, 27:1, 51–85.

[7] Д. Р. Гайфулин, И. Д. Кан, Производная функции Минковского., Изв. РАН. Сер. матем.,
85:4 (2021) .

[8] Д. Р. Гайфулин, И. Д. Кан, Стационарные точки функции Минковского., Матем. сб. (to
appear in 2021).

[9] I. D. Kan, Methods for estimating continuants, J. Math. Sci. (NY), 182 (4) (2012), 508-517.

19
[10] M. Kesseböhmer, B. O. Stratmann, Fractal analysis for sets of non-differentiability of
Minkowski’s question mark function, J. Number Theory 128 (2008), 2663-2686.

[11] J. R. Kinney, Note on a singular function of Minkowski, Proc. Amer. Math. Soc. 11 (5) (1960),
788-794.

[12] H. Minkowski, Zur Geometrie der Zahlen, Verhandlungen des III Internationalen Mathematiker-
Kongresous, Heidelberg 1904, 164-173.

[13] J. Paradís, P. Viader, L. Bibiloni, A new light on Minkowski’s ?(x) function, J. Number Theory
73 (2) (1998), 212-227.

[14] J. Paradís, P. Viader, L. Bibiloni, The derivative of Minkowski’s ?(x) function. J. Math. Anal.
and Appl. 253 (2001), 107 - 125.

[15] R. Salem, On some singular monotonic functions which are strictly increasing, Trans. Amer.
Math. Soc. 53 (3) (1943), 427-439.

Dmitry Gayfulin,
Steklov Mathematical Institute of Russian Academy of Sciences
ul. Gubkina, 8, Moscow, Russia, 119991
gayfulin@rambler.ru

20