Some bounds for the logarithmic function

Flemming Topsøe ∗
University of Copenhagen
[email protected]

Abstract
Bounds for the logarithmic function are studied. In particular, we
establish bounds with rational functions as approximants. The study
leads into the fascinating areas of Padé approximations ([2], [6]), con-
tinued fractions ([7], [11]) and orthogonal polynomials ([14], [4]) as
well as the somewhat frightening jungle of special functions and asso-
ciated identities ([5], [9]). Originally, the results aimed at establishing
certain inequalities for Shannon entropy but are here discussed in their
own right (the applications to entropy inequalities will be published
elsewhere).
The exposition is informal, a kind of essay, with only occasional
indications of proofs. The reader may take it as an invitation to further
studies. Enough details are provided to enable the reader to verify all
statements. To the expert in the fields pointed to there is little or
nothing new.

Keywords Logarithmic inequalities, Padé approximation, continued frac-

tions, Jacobi polynomials, Legendre polynomials.

1 Basic inequalities
Consider first the truly basic inequalities:
1
1−≤ ln x ≤ x − 1 for x > 0 . (1)
x
Here and in all inequalities below it is understood that the inequalities shown
are strict, except in easily recognizable cases of equality. One may prefer to
write (1) in the form
∗
This work was supported by the Danish Natural Science Research Council.

1
 x
≤ ln (1 + x) ≤ x for x > −1 . (2)
1+x
The bounds (1) and (2) involve rational functions of type [1, 1] (the left-
hand-side) and of type [1, 0] (the right-hand-side). Here, the type refers to the
degrees of numerator and denominator in the rational function concerned.
The simplicity, generality and usefulness of (1) and (2) is unbeatable.
Nevertheless, there are many other interesting inequalities for the logarithmic
function. For instance, if we separate positive and negative values of x, none
of the following two inequalities is much more complex:
2x x 2+x
≤ ln (1 + x) ≤ · for 0 ≤ x < ∞, (3)
2+x 2 1+x
2x x 2+x
≥ ln (1 + x) ≥ · for − 1 < x ≤ 0 (4)
2+x 2 1+x
–and these inequalities are sharper than (2). Also note that the first inequal-
ity of (3) is an improvement, for x ≥ 0, over the first inequality of (2) with
a rational function of the same type – whereas the second inequality of (2)
cannot, of course, be improved in a similar way.
The inequalities (3) and (4) may be written as a single double inequality,
e.g. by taking absolute values. Another possibility is to define the function
λ by
ln (1 + x)
λ(x) = (5)
x
(λ represents the slope of the chorde on the function x y ln (1 + x) which
connects (0, 0) with (x, ln (1 + x))). By continuity, λ(0) = 1. Then
2 2+x
≤ λ(x) ≤ for x > −1 . (6)
2+x 2 + 2x
For some possibilities to sharpen the bounds (3) and (4), see (22) and
(23) further on. However, the form chosen is convenient for a special reason.
Actually, (3) and (4) are equivalent as follows by writing ln (1 + x) in the
x
form − ln (1 − 1+x ). To further exploit this observation, we introduce a
notion of duality between functions defined on [0, ∞[ and functions defined
on ] − 1, 0]. The dual of one such function, say φ, is defined by the relation
 
∗ −x
φ (x) = −φ . (7)
1+x
Inequalities of the form φ(x) ≤ ln (1 + x) ≤ ψ(x) for x ∈ [0, ∞[ (respect-
ively, for x ∈] − 1, 0]) then translate into inequalities φ∗ (x) ≥ ln (1 + x) ≥
ψ ∗ (x) for x ∈] − 1, 0] (respectively, for x ∈ [0, ∞[).

2
 In (3) and (4) we have an instance with two self-dual functions in the
sense that for the two analytic functions – say φ and ψ – given on the full
interval ] − 1, ∞[ by, respectively the left-hand and the right-hand expression
in (3) and (4), we find that

(φ|[0,∞[)∗ = φ|]−1,0] (8)

and similarly for ψ (here, the subscripts refer to restrictions of the domain of
definition). As the functions we shall consider will be analytic in ] − 1, ∞[,
we do not find it necessary to indicate the restrictions to the appropriate
interval since the analytic form of the functions will be the same whether we
consider the restriction to [0, ∞[ or to ] − 1, 0]. For the functions above we
may thus write simply φ∗ = φ and ψ ∗ = ψ.
Let us return to (2). Also there, duality is relevant. Indeed, writing (2)
in the form φ(x) ≤ ln (1 + x) ≤ ψ(x), we find that φ∗ = ψ (and ψ ∗ = φ), i.e.
(φ, ψ) is a dual pair. This explains why there is no restriction on x in (2),
except the natural one, x > −1.
Duality takes a different – and somewhat simpler – form when having
bounds for ln x in mind rather than bounds for ln (1 + x). In order to
save on notation, we use only one symbol, a “star” , for duality and then
indicate by an overline “tilde” if the functions are intended as bounds for
ln (x) rather than for ln (1 + x). The duality definition for “tilde-functions”
is then given by
φ̃∗ (x) = −φ̃(x−1 ) . (9)
As an example of a basic inequality of ln x-type, we mention
1 1
| ln x| ≤ |x − | for x > 0 (10)
2 x
which is obtained from the right-hand inequalities of (3) and (4).
From (10), and after a further substitution x := xa with a a positive
parameter, we obtain useful approximations of ln x. In more detail, one
finds that
1 a
(x − x−a ) ↓ ln x as a ↓ 0 for 1 ≤ x < ∞ , (11)
2a
and that
1 a
(x − x−a ) ↑ ln x as a ↓ 0 for 0 < x ≤ 1 . (12)
2a
These results – clearly duals of each other – can also be derived directly
from the expansion
∞
1 a −a
X a2n
(x − x ) = ln x + ( ln x)2n+1 . (13)
2a n=1
(2n + 1)!

3
 Consider, for instance, the case a = 12 . This leads to the inequality
x−1
ln x ≤ √ for 1 ≤ x < ∞ (14)
x
with reversal of the inequality sign for 0 < x < 1. Expressed in terms of the
λ-function this shows that
1
λ(x) ≤ √ for x > −1 . (15)
1+x
This useful inequality – which one could just as well had been led to by
considering the product of the extreme terms in (6) – goes back at least to
Karamata [8], cf. also Mitrinović [12, Section 3.6.15].
Let us end this section with a more special inequality which is, again,
related to (6). It is our only result which involves two parameters (other such
inequalities can be found in [12]). The inequality states that for 0 ≤ x ≤ 1
and 0 ≤ y < ∞,
1−x
(2 − x)λ(y) − ≤ λ(xy) ≤ xλ(y) + (1 − x) . (16)
1+y
This follows by a standard analysis of the inequalities which result when
you keep y fixed (for the proof, consider the function equal to the difference
between terms of the inequality you wish to prove, and differentiate twice).
The indicated proof makes use of (6) (which is needed in order to discuss
endpoint behaviour). One may also note that, for 0 ≤ x ≤ 1, the second
inequality in (6) follows by comparison of the two extreme terms in (16).

2 Rational approximants for ln (1 + x)

We have demonstrated the usefulness of (6) and now turn to a more system-
atic study of similar bounds with more general rational functions. We choose
to focus on bounds for ln (1 + x) for x ≥ 0. We already have some examples
of lower and upper bounds of ln (1 + x) and realize that it is natural to seek
lower bounds of type [n, n] and upper bounds of type [n, n − 1]. We shall
find such bounds for each n ≥ 1.
To find criteria for the selection of good lower bounds, assume that φ(x) =
F
G
is such a bound with F and G polynomials, both of degree n, say. We
insist that the bound is exact for x = 0, hence the constant term in F is 0.
For the difference function δ(x) = ln (1 + x) − φ(x), we find that
 
2 0 0
G(x) − (1 + x) F (x)G(x) − F (x)G (x)
δ 0 (x) = . (17)
(1 + x)G(x)2

4
For a loose consideration, let us neglect the fact that the denominator here
varies with the choice of φ. Then we only need to assure that the numerator
is small and non-negative. To choose among the possible bounds, note that
the numerator is a polynomium of degree 2n. We aim at a bound which is
especially good for small values of x and realize that we should attempt to
make lower terms vanish. We achieve this by insisting that all terms in the
numerator vanish, except the leading term.
F
To be precise, we denote by φn the function of the form φn = G with F
and G two polynomials of degree n such that F (0) = 0 and G(0) > 0 and
such that the numerator in (17) is a positive constant times x2n . Similarly, ψn
F
is defined by ψn = G , where F is a polynomial of degree n with F (0) = 0 and
G a polynomial of degree n − 1 with G(0) > 0 and such that the numerator
in (17) is a negative constant times x2n−1 .
It turns out that, for each n ≥ 1, φn and ψn are uniquely determined by
these requirements. We introduce the following standard representations:

xPn−1 (x) xRn−1 (x)

φn (x) = , ψn (x) = (18)
Qn (x) Sn−1 (x)

with (Pn )n≥0 , (Qn )n≥0 , (Rn )n≥0 and (Sn )n≥0 four sets of polynomials of de-
grees as indicated by the subscript and such that

Qn,n = 1 , Rn−1,n−1 = 1 (19)

(with natural notation for the coefficients of the polynomials we consider).

For convenience, we put Q0 (x) ≡ 1. By (19), the leading term in Qn is xn and
the leading term in Rn−1 is xn−1 , i.e. Qn and Rn−1 are monic polynomials.
Observe that we normalize via the denominator in φn and via the numerator
in ψn . With the chosen normalization of the representations for the φ- and
ψ-functions, the P -, Q-, R- and S-polynomials are all uniquely determined.
To save on notation, we shall normally write Pn (respectively Qn , Rn , Sn and
also φn and ψn ) instead of Pn (x) (respectively Qn (x), Rn (x) etc.). We often
refer to the four sets of polynomials as the P QRS-polynomials.
To recapitulate, φn and ψn are defined by (18) and (19) and by the es-
sential requirements

Q2n − (1 + x) (xPn−1 )0 Qn − xPn−1 Q0n = x2n ,

(20)
0 0 2 2n−1


(1 + x) (xRn−1 ) Sn−1 − xRn−1 Sn−1 − Sn−1 = Sn−1,n−1 x . (21)

Clearly, φn is a lower- and ψn an upper bound of ln (1 + x) for x ∈ [0, ∞[.

This follows as the functions x y ln (1+x)−φn (x) and x y ψn (x)− ln (1+x)
both vanish at x = 0 and have positive derivatives in ]0, ∞[.

5
 The considerations leading to the above definitions express a key idea
in the theory of Padé approximation, cf. Baker and Graves-Morris [2], a
standard reference. In the terminology of that theory, φn is the [n, n]-Padé
approximant and ψn the [n, n − 1]-Padé approximant of ln (1 + x). We shall
refer to the φn ’s as simply the lower approximants (to ln (1 + x)) and to the
ψn ’s as the upper approximants.
Instead of just referring the reader to the literature on Padé approxima-
tion, we base our exposition on experiments facilitated by modern computing
tools. This will lead rather quickly to desired formulas and other insights.
Full proofs of relations found experimentally may not always be so obvious.
We shall include enough details to enable the reader to validate all state-
ments in the usual rigorous mathematical style. The interested reader will
find further results in the literature referred to.

3 Some experiments
In order to get a feel for the nature of the approximants defined in the
previous section, it is natural to work out a number of examples. This can be
done by equating coefficients of polynomials occuring in the defining relations
(20) and (21). With hinsight, this can be done much more conveniently by
recursion formulas developed later ((60)–(63)) or by simply asking MAPLE
to work out the relevant Padé approximants. Anyhow, in one way or another
we can get at the first few approximants and obtain a suitable table, cf. Table
1.
For instance, φ1 (x) ≤ ln (1 + x) for x ≥ 0 which we recognize as the first
inequality of (3). As φ∗1 = φ1 , the first inequality of (4) also follows. The
determination of ψ2 gives the inequality
x(6 + x)
ln (1 + x) ≤ for x ≥ 0 , (22)
2(3 + 2x)
a strengthening of the second inequality in (3) with a bound of the same type
([2, 1]). And, if we dualize the inequality for ψ2 , we get ln (1 + x) ≥ ψ2∗ (x)
for −1 < x ≤ 0, i.e.

x(6 + 5x)
ln (1 + x) ≥ for − 1 < x ≤ 0 , (23)
2(1 + x)(3 + x)
a streghtening of the second inequality of (4), though with a function of a
different type ([2, 2] rather than [2, 1]).
By (18) and (19), Table 1 reveals the identity of the first P QRS-polynomials,
cf. Table 2.

6
 n φn ψn
2x
1 2+x x
3x(2+x) x(6+x)
2 6+6x+x 2 2(3+2x)
x(60+60x+11x2 ) x(30+21x+x2 )
3 3(20+30x+12x2 +x3 ) 3(10+12x+3x2 )
5x(84+126x+52x2 +5x3 ) x(420+510x+140x2 +3x3 )
4 6(70+140x+90x2 +20x3 +x4 ) 12(35+60x+30x2 +4x3 )
x(7560+15120x+9870x2 +2310x3 +137x4 ) x(3780+6510x+3360x2 +505x3 +6x4 )
5 30(252+630x+560x2 +210x3 +30x4 +x5 ) 30(126+280x+210x2 +60x3 +5x4 )
7x(1320+3300x+2960x2 +1140x3 +174x4 +7x5 ) x(13860+30870x+23520x2 +7035x3 +672x4 +5x5 )
6 10(924+2772x+3150x2 +1680x3 +420x4 +42x5 +x6 ) 30(462+1260x+1260x2 +560x3 +105x4 +6x5 )

Table 1: Lower- and upper approximants to ln (1 + x)

n 0 1 2 3
Pn 2 6 + 3x 20 + 20x + 113
x2 70 + 105x + 130
3
x2 + 25
6
x3
2 2 3
Qn 1 2+x 6 + 6x + x 20 + 30x + 12x + x
Rn 1 6+x 30 + 21x + x2 140 + 170x + 140
3
x2 + x 3
Sn 1 6 + 4x 30 + 36x + 9x2 140 + 240x + 120x2 + 16x3

Table 2: P QRS-polynomials of degrees 0, 1, 2 and 3.

A graphical plot of the first four approximants is shown in Figure 1. After

more extensive plotting – if the reader is not yet convinced – one will realize
that the φn ’s increase with n and the ψn ’s decrease, both sequences with
ln (1 + x) as limit function. In order to test this, we look at differences of
the relevant functions. One finds that
x2n
ψn − φ n = , (24)
Qn Sn−1
x2n+1
ψn+1 − φn = , (25)
Q n Sn
2 x2n+1
φn+1 − φn = , (26)
n + 1 Qn+1 Qn
x2n
ψn − ψn+1 = (2n + 1) . (27)
Sn−1 Sn
This may be obtained experimentally but also follows from the definitions of
the approximants (except for the constants in (26) and (27) which depend
on the later relations (29) and (30)).

7
 y1

y2

ln(1+x)

f2

f1

Fig.1. Upper and lower bounds of ln (1 + x) for x ≥ 0.

In order to find good approximations of ln (1 + x) (for x ≥ 0), we could

proceed differently from the above approach which is tied to the defining
relations (20) and (21). As a start, take x as a reasonable approximation
of ln (1 + x). This is too much, thus in the next step we divide by 1 plus
some non-negative function. Not to increase the complexity too much, take
this function to be proportional to x, say cx. Then we get a type-[1, 1]
approximation of ln (1 + x). If this is set to φ1 – known to be a reasonable
bound – we get c = 12 . Now, φ1 is too small, hence in our next step we divide
x by 2 plus a non-negative function. Choosing again a function proportional
to x, we obtain a type-[2, 1] bound and it is reasonable to adjust this so that
it coincides with ψ2 . We now have the bound
x
ψ2 (x) =
1 + 2+x x
3

and may continue based on the identification of the first few good bounds
given in Table 1.
The approach just described is of course well known and nothing but an
attempt to represent ln (1 + x) by a continued fraction. The approach above
does not lead in a unique way to (the beginnings of) a continued fraction.
E.g., above we could choose to look, as we did, at x divided by 2 + something
or we could have looked at 21 x divided by 1 plus something. Experimenting
with the possibilities, you soon see that a very simple structure emerges if
you at each step divide by n plus something. Indeed, you then arrive at the
beautiful representation

x 12 x 12 x 22 x 22 x n2 x n2 x
ln (1 + x) = ··· ··· . (28)
1+ 2 + 3 + 4 + 5 + + 2n + 2n + 1 +

8
 This representation is not new. You find it in [3], [4], [2], [7] and [11].
In fact, the result is more than two hundred years old (!) and goes back to
Lambert, cf. [10]. Regarding historical comments – to this and other aspects
of the paper – we refer to [3] and to [7].
In a sense, the continued fraction (28) tells us all we need to know as
it provides us with easy access to the approximants φn and ψn . We shall
elaborate more on this in Section 5. For now, let us look at a few more
“experiments” .
Some coefficients for the P QRS-polynomials are easy to guess from Tables
1 and 2, e.g., we see that
Sn,n = (n + 1)2 . (29)
2
As another example, we realize that Pn,n − Pn−1,n−1 = n+1
, hence

n+1
X 1
Pn,n =2 . (30)
ν=1
ν

One may also go hunting for relationships among the coefficients by con-
sulting the “On-Line Encyclopedia of Integer Sequences” , cf. Sloane [13].
For instance, enquiring about the sequence 6, 30, 90, 210, the feed-back from
this source will reveal the fact that
1
Qn,n−2 = [n + 2]4 .1
4
Many other relations for the P QRS-coefficients can be discovered in this
way. Instead of pursuing this line of investigation, we shall have a look at the
approximants after a transformation so that they approximate ln x rather
than ln (1 + x). In other words, we are asking about the “tilde-functions”

φ̃n (x) = φn (x − 1) ; ψ̃n (x) = ψn (x − 1) . (31)

This leads to the functions in Table 3.
We end this section by an investigation of the zeroes of the P QRS-
polynomials. The result – found by numerical computation using MAPLE
– is reported in Table 4. We realize that the interlacing property holds for
each set of polynomials, e.g., between adjacent zeroes of Pn you find a zero
of Pn−1 . And also some “mixed interlacing” takes place, as we shall return
to at the very end of the paper.
1
The notation [a]k is used for the descending factorial a(a − 1) · · · (a − k + 1).

9
n φ̃n ψ̃n
1 2(x−1)
1+x
x−1
2 3(x−1)(1+x)
1+4x+x2
(x−1)(x+5)
2(1+2x)
(x−1)(11+38x+11x2 ) (x−1)(10+19x+x2 )
3 3(1+9x+9x2 +x3 ) 3(1+6x+3x2 )
5(x−1)(5+37x+37x2 +5x3 ) (x−1)(47+239x+131x2 +3x3 )
4 6(1+16x+36x2 +16x3 +x4 ) 12(1+12x+18x2 +4x3 )
(x−1)(137+1762x+3762x2 +1762x3 +137x4 ) (x−1)(131+1281x+1881x2 +481x3 +6x4
5 30(1+25x+100x2 +100x3 +25x4 +x5 ) 30(1+20x+60x2 +40x3 +5x4 )
7(x−1)(7+139x+514x2 +514x3 +139x4 +7x5 ) (x−1)(142+2272x+6397x2 +4397x3 +647x4 +5x5 )
6 10(1+36x+225x2 +400x3 +225x4 +36x5 +x6 ) 30(1+30x+150x2 +200x3 +75x4 +6x5 )

Table 3: Lower- and upper approximants to ln (x)

P1 -2.0000
P2 -4.1356 -1.3189
P3 -7.2397 -2.0000 -1.1603
P4 -11.3204 -2.9243 -1.5197 -1.0969
P5 -16.3907 -4.0764 -2.0000 -1.3251 -1.0650
Q1 -2.0000
Q2 -4.7321 -1.2679
Q3 -8.8730 -2.0000 -1.1270
Q4 -14.4026 -3.0302 -1.4926 -1.0746
Q5 -21.3174 -4.3334 -2.0000 -1.3000 -1.0492
R1 -6.0000
R2 -19.4582 -1.5418
R3 -42.7683 -2.6743 -1.2240
R4 -77.0830 -4.2481 -1.7114 -1.1242
R5 -123.2945 -6.2526 -2.3648 -1.4089 -1.0792
S1 -1.5000
S2 -2.8165 -1.1835
S3 -4.7094 -1.6934 -1.0992
S4 -7.1551 -2.4015 -1.3828 -1.0606
S5 -10.1487 -3.2837 -1.7793 -1.2469 -1.0415

Table 4: Zeroes of the P QRS-polynomials.

10
4 Some facts
In this section it is assumed that any occurring x is non-negative unless stated
otherwise explicitly.
The experiments of the previous section lead to important facts about
the φn ’s and ψn ’s. Let us start by looking at the quality of these functions
as bounds for ln (1 + x).
Apparently, all coefficients in the P QRS-polynomials are positive. From
(24) - (27) it therefore follows that

φ1 (x) ≤ φ2 (x) ≤ · · · ≤ ln (1 + x) ≤ · · · ≤ ψ2 (x) ≤ ψ1 (x) . (32)

Furthermore, by (25), recalling also (19) and (29), we see that

x
ψn+1 (x) − φn (x) ≤ . (33)
(n + 1)2
Though this bound is quite loose, it is sharp enough to imply that, for each
x ≥ 0, the sequences φn (x) and ψn (x) both converge to ln (1 + x) as n → ∞.
We also note that
x2n+1
ln (1 + x) − φn (x) ≤ , (34)
Qn (x)Sn (x)
x2n
ψn (x) − ln (1 + x) ≤ . (35)
Qn (x)Sn−1 (x)

These bounds are less explicit but typically much sharper than (33).
The behaviour of φn (x) and ψn (x) for large x is given by the relations
n
X 1 x
φn (x) ≈ 2 ; ψn (x) ≈ (36)
ν=1
ν n2

in the sense that the ratios involved converge to 1 for x → ∞. These relations
follow from (29) and (30).
We also note that the functions φn and ψn are increasing. As for φn this
follows from (20) (see also (17)) which tells us that

Qn (x)2 − x2n
φ0n (x) = ≥ 0. (37)
(1 + x)Qn (x)2

The approximants may be characterized in a natural way which is quite

different from the definitions via (20) and (21). Indeed, for each n ≥ 1, φn is
the unique type-[n, n] rational function with φn (0) = 0 and φn (x) ≤ ln (1+x)

11
for x ≥ 0 which dominates any other such function locally, i.e., for any such
function f , there exists ε > 0 such that φn (x) ≥ f (x) for 0 ≤ x < ε. In fact,
for this conclusion we may assume about type-[n, n] functions f considered
only that f ≤ ψn+1 , locally, instead of f (x) ≤ ln (1 + x) for x ≥ 0. A similar
characterization holds for ψn .
Note that we have to work locally with inequalities in the characteriza-
tions just pointed out (consider, for example, the function given by f (x) =
3x
4+x
; then f (x) ≤ ln (1 + x) for x ≥ 0 – as f ≤ φ2 – but f ≤ φ1 does not hold
globally for x ≥ 0, only locally). The proofs of the characterizations rest on
the facts that if a type-[n, n]-function is not identical to φn , then there is a
(ν)
ν ≤ 2n such that f (ν) (0) 6= φn (0) and if a type-[n, n − 1] function g is not
(ν)
identical to ψn , then there is a ν ≤ 2n − 1 such that g (ν) (0) 6= ψn (0).
The determination of the P QRS-polynomials in closed form is not that
obvious based on our experiments. Most strikingly, perhaps, is that if we
define Q̃n by
Q̃n (x) = Qn (x − 1) ,
then n  2
X n
Q̃n (x) = xk ,
k=0
k

cf. Table 3. As Qn (x) = Q̃n (1 + x), this leads to the formula

n−k 
 X  
n n−k n
Qn,k = .
k ν=0 ν n−k−ν

The sum here can be evaluated by using the binomial Chu-Vandermonde

formula:
N  
X N
[x]j [y]N −j = [x + y]N ,
j=0
j

or, more directly, by the equivalent Chu-Vandermonde convolution formula:

N     
X x y x+y
= .
j=0
j N −j N

For these identities, see [1]. What we find is that

    
n 2n − k 2n − k
Qn,k = = . (38)
k n−k k, n − k, n − k

12
 Thus, the coefficients in Qn have been identified as certain trinomial coef-
ficients.
Having determined the Qn,k ’s we may use the defining relation (20) to
determine the Pn−1,k ’s. With some effort this leads to the formula
k
X (−1)ν
Pn−1,k = Qn,k−ν . (39)
ν=0
ν+1

Combined with (38) this gives a reasonable determination of the
coefficients

of the P -polynomials, e.g. it shows that Pn−1,0 = Qn,0 = 2n n
. However,
the formulas P
can sometimes be simplified (consider the case k = n − 1 when
Pn−1,n−1 = 2 n1 ν1 according to (30)).
A striking feature of the φ̃n -functions is that they are self-dual. This is
seen from Table 3. Indeed, writing

(x − 1)P̃n−1 (x)
φ̃n (x) =
Q̃n (x)

with P̃n−1 (x) = Pn−1 (x − 1) and Q̃n (x) = Qn (x − 1), we realize that these
polynomials are self-reflected in the sense that
1
xn−1 P̃n−1 ( ) = P̃n−1 (x) ,
x
1
xn Q̃n ( ) = Q̃n (x) .
x

By simple calculation, this implies that the φ̃n ’s are self-dual. Hence also
the φn ’s are self-dual: φ∗n = φn . By duality, this implies that φn is a best
type-[n, n] upper bound of ln (1 + x) for −1 < x ≤ 0 (with “ best” being
understood in much the same way as discussed for the bounds found for
x ≥ 0).
The functions ψn are not self-dual. The duals ψn∗ are even of a different
type, viz. of type [n, n]. These functions then, are the best type-[n, n] lower
bounds of ln (1 + x) for −1 < x ≤ 0.

5 Exploiting Lamberts continued fraction

It is natural to seek procedures which allow recursive determination of the
P QRS-polynomials. This can be achieved based on the defining relations
and the derived relations (24)–(27). However, this will, typically, involve

13
products of polynomials and simpler procedures are desirable. Here Lamberts
expansion (28) provides the proper tool.
Following usual terminology of continued fraction theory, let An and Bn
denote the approximants related to (28). These are polynomials associated
in the natural manner to the corresponding finite continued fractions, e.g.

A1 x A2 x A3 x
= , = 2
, = .
B1 1 B2 1 x B3 12 x
1+ 1+
2 12 x
2+
3
The first polynomials are

A0 = 0 , A 1 = x , B 0 = 1 , B 1 = 1 (40)

and, exploiting the special form of (28), the following recursive relations hold
for n ≥ 1:

A2n = 2nA2n−1 + n2 xA2n−2 , (41)

A2n+1 = (2n + 1)A2n + n2 xA2n−1 , (42)
B2n = 2nB2n−1 + n2 xB2n−2 , (43)
B2n+1 = (2n + 1)B2n + n2 xB2n−1 . (44)

We realize that, for n ≥ 1,

A2n A2n−1
φn = , ψn = . (45)
B2n B2n−1
The polynomials A2n−1 , A2n , B2n , and B2n+1 are all of degree n and for
basic coefficients we find

An,0 = 0 , An,1 = Bn,0 = n! , (46)

2
A2n+1,n+1 = B2n,n = n! . (47)
Together with (45), (47) tells us that

A2n B2n
xPn−1 = 2
, Qn = , (48)
(n!) (n!)2
A2n−1 B2n−1
xRn−1 = , S n−1 = . (49)
((n − 1)!)2 ((n − 1)!)2

14
 Now, Qn is known, cf. (38), hence also B2n can be found. And then, from
(43), we can determine B2n−1 , hence, by (49), also Sn−1 . This leads to the
formula
n−1  
X 2n − ν − 1
Sn−1 = n xν . (50)
ν=0
ν, n − ν, n − ν − 1
Similarly, a formula for Rn−1 can be found based on (39), (41) and (49).
Going through the details you find that
n−1 X
ν+1  
X 2n − ν + µ − 2
Rn−1 = n (−1) µ−1
µ−1 xν .
ν=0 µ=1
ν − µ + 1, n − ν + µ − 1, n − ν + µ − 2
(51)
The expressions (38), (39), (50) and (51) thus provide formulas for the
P QRS-polynomials in closed form with very satisfactory expressions for the
denominator polynomials (Q, S) and somewhat less satisfactory formulas for
the numerator polynomials (P , R).
For practical calculations, however, the recursive relations (41)–(44) are
more expedient and simple to programme. By (48) and (49) these relations
may be written directly in terms of the P QRS-polynomials:
2
Pn−1 = Rn−1 + xPn−2 , (52)
n
Rn = (2n + 1)Pn−1 + xRn−1 , (53)
2
Qn = Sn−1 + xQn−1 , (54)
n
Sn = (2n + 1)Qn + xSn−1 . (55)

Here, n ≥ 1 and for (52) we have put P−1 = 0.

The special structure of these recurrence relations may be emphasized by
introducing the notation ∆Pn for the polynomial

(∆Pn )(x) = Pn (x) − xPn−1 (x)

and similarly for ∆Qn etc. Then,

2
∆Pn−1 = Rn−1 ; ∆Rn = (2n + 1)Pn−1 , (56)
n
2
∆Qn = Sn−1 ; ∆Sn = (2n + 1)Qn . (57)
n
Here, n ≥ 1 (with ∆P0 = 2).

15
 Another form of (52)–(55) is obtained for the reflected polynomials. Fol-
lowing [6], the reflected polynomial of a polynomial A is the polynomial
denoted A which is given by
A(x) = xN A(x−1 )
with N the degree of A. We find:
2
P n−1 − P n−2 = Rn−1 ; Rn − Rn−1 = (2n + 1)xP n−1 , (58)
n
2
Qn − Qn−1 = xS n−1 ; S n − S n−1 = (2n + 1)Qn . (59)
n

6 Relations to orthogonal polynomials

The recurrence relations (52)–(55) are ideal for recursive computation of
the P QRS-polynomials. In (52) and (53), numerator polynomials are “en-
tangled” , and in (54) and (55) so are denominator polynomials. It is, however
easy to “unentangle” the formulas so as to only involve one set of polynomials
at a time. In this way you find, for n ≥ 2:
(n + 1)Pn − (2n + 1)(x + 2)Pn−1 + nx2 Pn−2 = 0, (60)
nQn − (2n − 1)(x + 2)Qn−1 + (n − 1)x2 Qn−2 = 0, (61)
2
(2n − 1)Rn − 4nx + (4n2 − 1) Rn−1 + (2n + 1)x2 Rn−2


= 0, (62)
n
2
(2n − 1)Sn − 4nx + (4n2 − 1) Sn−1 + (2n + 1)x2 Sn−2


= 0. (63)
n
In passing we note that it is easy to prove from (60) and (61) that the
φn ’s are self-dual.
It is well known that orthogonal polynomials satisfy three-term recur-
rence relations. With the above formulas in mind it therefore lies nearby
to investigate the possible relations of the P QRS-polynomials to orthogonal
polynomials. However, the recurrence relations (60)-(63) are not in standard
form as known from the theory of orthogonal polynomials because of the
appearence of x in the last term. This points to the desirability of rewriting
the formulas by introducing a suitable transformation.
Such a transformation may be suggested by looking at the zeroes of the
P QRS-polynomials, recalling that these are dense in the support of any
underlying measure. Here, the information provided by Table 4 is useful.
It shows that the zeroes lie in ] − ∞, −1[ and also, that the “interlacing”
behaviour which is characteristic for orthogonal polynomials applies to our
polynomials.

16
 n Pn Qn Rn Sn
0 2 1 1 1
1 3x x 3x − 2 3x + 1

Table 5: PQRS-polynomials of degrees 0 and 1.

Possibly after some experimentation, it is found that a transformation

which places the zeroes in the interval ] − 1, 1[ have a simplifying effect. This
may be achieved by first making a transformation to the reflected polynomi-
als, which brings the zeroes into ] − 1, 0[, and then consider the natural linear
transformation into ] − 1, 1[. We therefore define new polynomials, indicated
by the letters P, Q, R, S. For the P-polynomials, definition and notation is as
follows:
1  2 
Pn (x) = P n ( (x − 1)) = 2−n (x − 1)n Pn .
2 x−1
The other polynomials are defined similarly. We find:

2
Pn − Pn−1 = Rn , (64)
n+1
x−1
Qn − Qn−1 = Sn−1 , (65)
n
1
Rn − Rn−1 = (n + )(x − 1)Pn−1 , (66)
2
Sn − Sn−1 = (2n + 1)Qn , (67)

and the three-term recurrence relations become

(n + 1)Pn − (2n + 1)xPn−1 + nPn−2 = 0, (68)

nQn − (2n − 1)xQn−1 + (n − 1)Qn−2 = 0, (69)
n(2n − 1)Rn − (4n2 − 1)x + 1 Rn−1 + n(2n + 1)Rn−2


= 0, (70)
n(2n − 1)Sn − (4n2 − 1)x + 1 Sn−1 + n(2n + 1)Sn−2


= 0. (71)

Note the disappearence of any x’s in the last terms. We may apply (68)–
(71) with n ≥ 2 in connection with the start polynomials given in Table
5.
We shall relate the PQRS-polynomials to the classical Jacobi polynomials,
here denoted Pnα,β , which are associated with the measures on [−1, 1] with
densities (1 − x)α (1 + x)β . From standard sources, e.g. [14], we see that these

17
polynomials may be determined from the recurrence relations
(2n − 1 + α + β) (2n + α + β)(2n + α + β − 2)x + (α2 − β 2 ) α,β


Pnα,β − Pn−1
2n(n + α + β)(2n − 2 + α + β)
(n + α − 1)(n + β − 1)(2n + α + β) α,β
+ Pn−2 = 0
n(n + α + β)(2n − 2 + α + β)
(72)
for n ≥ 2 in conjunction with the first polynomials which are given by
1
P0α,β = 1 ; P1α,β =


(α + β + 2)x + (α − β) . (73)
2
The Jacobi polynomials have a simple expression in closed form:
n   
α,β 1 X n+α n+β
Pn = n (x − 1)n−ν (x + 1)ν . (74)
2 ν=0 ν n−ν

We realize that
Qn = Pn0,0 , (75)
i.e. the Qn ’s are nothing but the classical Legendre polynomials. By (65)
the S-polynomials are closely related to these polynomials. However, apart
from a constant factor, they may also be identified directly as orthogonal
polynomials. Indeed, by Table 5, (71), (72) and (73) it is easy to check that

Sn = (n + 1)Pn1,0 . (76)

The P- and R-polynomials are harder to identify. We appeal to [4] for

a systematic table of orthogonal polynomials. The table works with monic
polynomials. Let the monic polynomials corresponding to the Pn ’s be de-
noted [Pn ], i.e.
2n−1
[Pn ] = 2n+1
Pn . (77)
n
Then you find the recursion formula
n2
[Pn ] − x[Pn−1 ] + [Pn−2 ] = 0 . (78)
4n2 − 1
By a look-up in [4] (case 5, page 219), you realize that the [Pn ]’s are
orthogonal polynomials associated with the somewhat bizarre measure on
] − 1, 1[ with density function
4
.
1+x 2


π 2 + ln 1−x

18
 We have not identified the R-polynomials in a similar manner. Possibly,
it is more reasonable to look at the polynomials Sn − Rn when searching for
properties related to orthogonal polynomials.
Lastly we note two interlacing properties with mixed polynomials. For
this, we may as well return to the original P QRS-polynomials. In fact, from
the defining relation (20) and an investigation of the sign of Pn−1 at zeroes of
Qn it follows that between any two neighbouring zeroes of Qn there is a zero
of Pn−1 (and of course only one zero). And from (21) a similar investigation
shows that between any two neighbouring zeroes of Sn−1 there is a zero of
Rn−1 . This accounts for n − 2 zeroes of Rn−1 . As is easily seen, the last zero
of Rn−1 is smaller than the smallest zero of Sn−1 .

Acknowledgement
Thanks are due to Jacob Stordal Christiansen for some of the references to
the literature and for identification of the associated orthogonal polynomials
related to the Pn ’s.

References
[1] E. S. Andersen and M. E. Larsen. Combinatorial Summation Identities.
In Analysis, algebra, and computers in mathematical research, volume
156 of Lecture Notes in Pure and Applied Mathematics, pages 1–23.
Marcel Dekker, 1994.

[2] G. A. Baker and P. Graves-Morris. Padé Approximants, volume 59 of

Encyclopedia of Mathematics and its Applications. Cambridge Univer-
sity Press, 2.nd edition, 1996.

[3] C. Brezinski. History of Continued Fractions and Padé Approximants.

Springer, Berlin Heidelberg, 1991.

[4] T. S. Chihara. An Introduction to Orthogonal Polynomials. Gordon and

Breach, New York London Paris, 1978.

[5] R. Askey G. E. Andrews and R. Roy. Special Functions. Cambridge

University Press, Cambridge, 1999.

[6] J. Gilewicz. Approximants de Padé, volume 667 of Lecture Notes in

Mathematics. Springer, 1978.

19
 [7] W. B. Jones and W. J. Thron. Continued Fractions - Analytic Theory
and Applications, volume 11 of Encyclopedia of Mathematics and its
Applications. Addison-Wesley, 1980.

[8] J. Karamata. Sur quelques problème posés par Ramanujan. J. Indian

Math. Soc. (N.S.), 24:343–365, 1960.

[9] W. Koepf. Computer Algebra Algorithms for Orthogonal Polynomials

and Special Functions, Leuven 2002. In Orthogonal Polynomials and
Special Functions, volume 1817 of Lecture Notes in Mathematics, pages
1–24. Springer, 2003.

[10] J. H. Lambert. Memoire sur quelques propri’et’e remarquables des

quantit’es transcendantes circulaires et logarithmiques. Memoires de
l’Acad. de Berlin, Ann’ee 1761, pages 265–322, 1768.

[11] L. Lorentzen and H. Waadeland. Continued Fractions with Applications,

volume 3 of Studies in Computational Mathematics. North-Holland,
1992.

[12] D. S. Mitrinović. Analytic Inequalities. Springer, Berlin, 1970.

[13] N. J. A. Sloane. The On-Line Encyclopedia of Integer Sequences. Notices

of the AMS, 50:912–915, 2003. (The encyclopedia itself is available at
http://www.research.att.com/ njas/sequences/).

[14] F. G. Tricomi. Vorlesungen über Ortogonalreihen. Springer, Berlin

Heidelberg, 1955.

20