Bourbaki

Appendix

Hardy Fields. (H) Functions

1. Hardy fields

Let F\mathfrak{F} be the filter base on R\mathbf{R} formed by the intervals of the form [x0,+∞[[x_0, +\infty[. Recall that we have defined an equivalence relation R∞R_\infty: on the set H(F,R)\mathscr{H}(\mathfrak{F}, \mathbf{R}) of real functions defined on sets belonging to F\mathfrak{F} “there exists a set M∈FM \in \mathfrak{F} such that f(x)=g(x)f(x) = g(x) on MM” (V, p. 211), and that the quotient set H(F,R)/R∞\mathscr{H}(\mathfrak{F}, \mathbf{R})/R_\infty is endowed with the structure of a ring with unit element.

Definition 1

Given a subset K\mathfrak{K} of H(F,R)\mathscr{H}(\mathfrak{F}, \mathbf{R}) one says that K/R∞\mathfrak{K}/R_\infty (the canonical image of K\mathfrak{K} in H(F,R)/R∞\mathscr{H}(\mathfrak{F}, \mathbf{R})/R_\infty) is a Hardy field, if K\mathfrak{K} satisfies the following conditions:

1. K/R∞\mathfrak{K}/R_\infty is a subfield of the ring H(F,R)/R∞\mathscr{H}(\mathfrak{F}, \mathbf{R})/R_\infty. 2. *Every function in K\mathfrak{K} is continuous and differentiable on an interval [a,+∞[[a, +\infty[ (depending on the function), and the class with respect to R∞R_\infty of its derivative belongs to K/R∞\mathfrak{K}/R_\infty.

The hypothesis that K/R∞\mathfrak{K}/R_\infty is a field is equivalent to the following conditions: if f∈Kf \in \mathfrak{K} and g∈Kg \in \mathfrak{K} then f+gf + g and fgfg are equal to functions in K\mathfrak{K} on some set in F\mathfrak{F}; further, if ff is not identically zero on a set in F\mathfrak{F} then there exists a set MM in F\mathfrak{F} on which ff does not vanish, and 1/f1/f is equal to a function from K\mathfrak{K} on MM; by condition 2' one can always assume that MM is taken so that ff is continuous on MM, and consequently has constant sign on this interval.

By an abuse of language, if K\mathfrak{K} is such that K/R∞\mathfrak{K}/R_\infty is a Hardy field we shall say in what follows that K\mathfrak{K} is itself a Hardy field.

Example 1

Every Hardy field contains the field of rational constants (the smallest field of characteristic 0, cf. Alg., V, §1), which one can identify with the field Q\mathbf{Q}; moreover, since two constants cannot be congruent modulo R∞R_\infty unless they are equal, Q/R∞\mathbf{Q}/R_\infty is identical to Q\mathbf{Q}. The real constants also form a Hardy field, which one can identify with R\mathbf{R}.

Cites 1

Cites A V § 1. Prime fields. Characteristic

Example 2

A very important example of a Hardy field is the set of rational functions with real coefficients, which we shall denote by R(x)\mathbf{R}(x), by an abuse of language; if f(x)=p(x)/q(x)f(x) = p(x)/q(x) is a rational function with real coefficients, not identically zero, it is continuous, differentiable, and ≠0\neq 0 on an interval [a,+∞[[a, +\infty[, where aa is strictly greater than the largest of the real roots of the polynomials p(x)p(x) and q(x)q(x); thus every element of R(x)/R∞\mathbf{R}(x)/R_\infty other than 0 is invertible. We note again that two rational functions cannot be congruent modulo R∞\mathbf{R}_\infty unless they are equal, so R(x)/R∞\mathbf{R}(x)/\mathbf{R}_\infty can again be identified with R(x)\mathbf{R}(x).

2. Extension of a hardy field

Given a Hardy field K\mathfrak{K} we shall see how one can form new Hardy fields K′⊃K\mathfrak{K}' \supset \mathfrak{K} such that K′/R∞\mathfrak{K}'/\mathbf{R}_\infty is obtained by adjoining (in the algebraic sense of the term, cf. Alg., V. §2) new elements, of a form which we shall make precise, to K/R∞\mathfrak{K}/\mathbf{R}_\infty.

Lemma 1

Let* a(x),b(x)a(x), b(x) be two continuous real functions that do not change sign on an interval [x0,+∞[[x_0, +\infty[. If, on this interval, the function y(x)y(x) *is continuous and differentiable, and satisfies the identity

y′=ay+b y' = ay + b

(1)

then there exists an interval [x1,+∞[[x_1, +\infty[ on which yy does not change sign.

Indeed, let us put z(x)=y(x)exp⁡(−∫x0xa(t) dt)z(x) = y(x) \exp \left( - \int_{x_0}^x a(t) \, dt \right) (cf. IV, p. 183); then, by (1), z′(x)=b(x)exp⁡(−∫x0xa(t) dt)z'(x) = b(x) \exp \left( - \int_{x_0}^x a(t) \, dt \right). If b(x)⩾0b(x) \geqslant 0 for x⩾x0x \geqslant x_0 then zz is increasing on this interval, so either is <0< 0 on all this interval, or is zero on an interval [x1,+∞[[x_1, +\infty[, or else is >0> 0 on an interval [x1,+∞[[x_1, +\infty[; since yy has the same sign as zz the proposition is proved in this case. The argument is similar if b(x)⩽0b(x) \leqslant 0 for x⩾x0x \geqslant x_0.

Remark

This very elementary property does not extend to linear differential equations of order >1> 1; for example, the function y=sin⁡xy = \sin x satisfies y′′+y=0y'' + y = 0, but changes sign on every neighbourhood of +∞+\infty.

Lemma 2

Let* a(x)a(x) and b(x)b(x) be two functions belonging to a given Hardy field K\mathfrak{K} and y(x)y(x) a function satisfying the identity (1) on an interval [x0,+∞[[x_0, +\infty[ where aa and bb are defined and continuous. If p(u)p(u) is a polynomial in uu whose coefficients are functions of xx belonging to K\mathfrak{K}, defined and differentiable on [x0,+∞[[x_0, +\infty[, then there exists an interval [x1,+∞[[x_1, +\infty[ on which the function p(y)p(y) does not change sign.

The proposition is trivial if p(u)p(u) has coefficients which are identically zero on [x0,+∞[[x_0, +\infty[, or if p(u)p(u) is of degree 0 in uu, since any function in K\mathfrak{K} is of constant sign on an interval [x1,+∞[[x_1, +\infty[. Suppose that p(u)p(u) is of degree n>0n > 0; the leading coefficient cc of p(u)p(u) is then ≠0\neq 0 on an interval [α,+∞[[\alpha, +\infty[; one can thus write p(u)=c(un+c1un−1+⋯+cn)p(u) = c(u^n + c_1 u^{n-1} + \cdots + c_n) where c,c1,c2,…,cnc, c_1, c_2, \ldots, c_n are functions belonging to K\mathfrak{K} and differentiable on [α,+∞[[\alpha, +\infty[; so it is enough to prove the lemma for c=1c = 1. We argue by induction on nn; one has

ddx(p(y))=(ay+b)(nyn−1+(n−1)c1yn−2+⋯+cn−1) \frac{d}{dx} (p(y)) = (ay + b)(ny^{n-1} + (n-1)c_1 y^{n-2} + \cdots + c_{n-1})
+c1′yn−1+⋯+cn′=na p(y)+q(y) + c'_1 y^{n-1} + \cdots + c'_n = na\ p(y) + q(y)

where q(u)q(u) is a polynomial of degree ≤n−1\leq n - 1, with coefficients in K\mathfrak{K}. By hypothesis the functions na(x)na(x) and q(y(x))q(y(x)) do not change sign on an interval [β,+∞[[ \beta, +\infty[; the lemma is thus a consequence of lemma 1.

Theorem 1

Let a(x)a(x) and b(x)b(x) be two functions belonging to a given Hardy field K\mathfrak{K} and y(x)y(x) a function satisfying (1) on an interval [x0,+∞[[ x_0, +\infty[. When r(u)=p(u)/q(u)r(u) = p(u)/q(u) runs through the set of rational fractions in uu with coefficients in K\mathfrak{K} such that q(y)q(y) is not identically zero on a neighbourhood of +∞+\infty the set K(y)\mathfrak{K}(y) of functions r(y)r(y) forms a Hardy field.

Cites 1

Cites App. 0. Hardy Fields. (H) Functions

Indeed, by lemma 2, there exists an interval [x1,+∞[[ x_1, +\infty[ on which r(y)r(y) is defined, continuous, and does not change sign, from which it follows immediately that K(y)/R∞\mathfrak{K}(y)/\mathbf{R}_\infty is a field; moreover, since

ddx(r(y))=r′(y)y′=r′(y)(ay+b) \frac{d}{dx}(r(y)) = r'(y)y' = r'(y)(ay + b)

(where r′(y)=(p′(y)q(y)−p(y)q′(y))/(q(y))2r'(y) = (p'(y)q(y) - p(y)q'(y))/(q(y))^2 is defined by hypothesis on a neighbourhood of +∞+\infty), the derivative of every function in K(y)\mathfrak{K}(y) belongs to K(y)\mathfrak{K}(y), which proves that K(y)\mathfrak{K}(y) satisfies the conditions of def. 1 of V, p. 247.

It is clear that K(y)/R∞\mathfrak{K}(y)/\mathbf{R}_\infty is obtained by the algebraic adjunction to K/R∞\mathfrak{K}/\mathbf{R}_\infty of the class of yy modulo R∞\mathbf{R}_\infty. Again one says that K(y)\mathfrak{K}(y) is obtained by adjoining yy to K\mathfrak{K}.

Corollary 1

If yy is a function in K\mathfrak{K}, not identically zero on a neighbourhood of +∞+\infty, then K(log⁡∣y∣)\mathfrak{K}(\log|y|) is a Hardy field.

Indeed, (log⁡∣y∣)′=y′/y(\log|y|)' = y'/y is equal to a function in K\mathfrak{K} on an interval [x0,+∞[[ x_0, +\infty[.

COROLLARY 2 If yy is any function in K\mathfrak{K}, then K(ey)\mathfrak{K}(e^y) is a Hardy field.

Indeed, (ey)′=eyy′(e^y)' = e^y y', and y′y' is equal to a function in K\mathfrak{K} on an interval [x0,+∞[[ x_0, +\infty[.

COROLLARY 3 If K\mathfrak{K} contains the real constants, and if yy is a function in K\mathfrak{K}, not identically zero on a neighbourhood of +∞+\infty, then K(∣y∣α)\mathfrak{K}(|y|^{\alpha}) is a Hardy field for every real number α\alpha.

Indeed, ddx(∣y∣α)=∣y∣α(αy′/y)\frac{d}{dx}(|y|^{\alpha}) = |y|^{\alpha} (\alpha y'/y), and αy′/y\alpha y'/y is equal to a function in K\mathfrak{K} on an interval [x0,+∞[[ x_0, +\infty[.

Finally, we note that if yy is the primitive of any function in K\mathfrak{K} then K(y)\mathfrak{K}(y) is again a Hardy field.

3. Comparison of functions in a hardy field

Proposition 1

Two functions in the same Hardy field are comparable to any order* (V, p. 232).

Cites 3

Cites V § 3. Asymptotic expansions of functions of a real variable, V § 1. Comparison of functions on a filtered set, Proposition 8, V § 3

Indeed, if ff belongs to a Hardy field K\mathscr{K}, then for every integer n>0n > 0 there exists an interval [x0,+∞[[x_0, +\infty[ on which ff is nn times differentiable, its nthn^{th} derivative being equal to a function in K\mathscr{K} on this interval. It is therefore enough to show that any two functions f,gf, g of K\mathscr{K} are comparable. This is evident if one of the functions is identically zero on a neighbourhood of +∞+\infty; one may therefore restrict oneself to the case where they are both strictly positive on a neighbourhood of +∞+\infty. But then, for every real number tt, f−tgf - tg is equal to a function in K\mathscr{K} on a neighbourhood of +∞+\infty, so has constant sign on a neighbourhood of +∞+\infty, which proves the proposition (V, p. 217, prop. 9).

One deduces immediately from this proposition that, if a Hardy field K\mathscr{K} contains the real constants (as we shall always assume in what follows), and if ff and gg are any two functions in K\mathscr{K} then any two of the functions ef,eg,log⁡∣f∣,log⁡∣g∣,∣f∣α,∣g∣αe^f, e^g, \log |f|, \log |g|, |f|^{\alpha}, |g|^{\alpha} (α\alpha an arbitrary real), ∫af,∫ag\int_a f, \int_a g (aa being any real number in an interval [x0,+∞[[x_0, +\infty[ where ff and gg are regulated) are comparable (when they are defined); indeed, any two of these functions belong to a given Hardy field obtained by adjoining them successively to K\mathscr{K}.

Similarly, every function f(x)f(x) in a Hardy field K\mathscr{K} is comparable to xx, since xx and f(x)f(x) belong to the Hardy field obtained by adjoining xx to K\mathscr{K}. One thus concludes (in particular) that ff is comparable to any order to every power xαx^{\alpha}, as well as to log⁡x\log x and to exe^x.

One also sees that if ff and gg belong to the same Hardy field K\mathscr{K}, if g(x)>0g(x) > 0 on an interval [x0,+∞[[x_0, +\infty[, and if g(x)g(x) tends to 0 or to +∞+\infty as xx tends to +∞+\infty, then the order of ff relative to gg (V, p. 219) is always defined.

Prop. 8 of V, p. 233, is therefore applicable to every function ff in a Hardy field, and proves that:

1) if ff is of order +∞+\infty relative to xx then ∫axf(t) dt∼(f(x))2/f′(x)\int_a^x f(t)\,dt \sim (f(x))^2/f'(x).

2) if ff is of order μ>−1\mu > -1 relative to xx then ∫axf(t) dt∼1μ+1xf(x)\int_a^x f(t)\,dt \sim \frac{1}{\mu+1} x f(x).

3) if ff is of order μ<−1\mu < -1 relative to xx then ∫a+∞f(t) dt∼−1μ+1xf(x)\int_a^{+\infty} f(t)\,dt \sim -\frac{1}{\mu+1} x f(x).

4) if ff is of order −∞-\infty relative to xx then ∫a+∞f(t) dt∼−(f(x))2/f′(x)\int_a^{+\infty} f(t)\,dt \sim -(f(x))^2/f'(x).

Further, we have the following proposition:

Proposition 2

Let ff be a function belonging to a Hardy field K\mathscr{K}.

Cites 1

Cites V § 3. Asymptotic expansions of functions of a real variable

1 If ff is of infinite order relative to xx, then, for every integer n>0n > 0,

f(n)(x)∼(f′(x))n(f(x))n−1. f^{(n)}(x) \sim \frac{(f'(x))^n}{(f(x))^{n-1}}.
(2)

2° If ff is of finite order μ\mu relative to xx then, for every n>0n > 0,

f(n)(x)∼μ(μ−1)…(μ−n+1)f(x)xn f^{(n)}(x) \sim \mu(\mu - 1) \ldots (\mu - n + 1) \frac{f(x)}{x^n}
∼(μ−1)…(μ−n+1)μn−1(f′(x))n(f(x))n−1 \sim \frac{(\mu - 1) \ldots (\mu - n + 1)}{\mu^{n-1}} \frac{(f'(x))^n}{(f(x))^{n-1}}
(3) except if μ\mu is an integer ⩾0\geqslant 0 and n>μn > \mu.

1 If ff is of infinite order relative to xx one has log⁡∣f∣≫log⁡x\log |f| \gg \log x, so, since log⁡∣f∣\log |f| and log⁡x\log x are comparable to any order, f′/f≫1/xf'/f \gg 1/x. Put g=f′/fg = f'/f; since gg is equal to a function in K\mathscr{K} on a neighbourhood of +∞+\infty one deduces from 1/g≪x1/g \ll x, that g′/g2≪1g'/g^2 \ll 1, and so g′/g≪g=f′/fg'/g \ll g = f'/f, or again fg′≪gf′fg' \ll gf'. From the relation f′=fgf' = fg one deduces by differentiating that

f′′=fg′+gf′∼gf′ f'' = fg' + gf' \sim gf'
or again f′′/f′∼f′/ff''/f' \sim f'/f. The same argument, applied to f(n)f^{(n)} instead of ff, shows, by induction on nn, that f(n)/f(n−1)∼f′/ff^{(n)}/f^{(n-1)} \sim f'/f; whence relation (2).

2 If ff is of finite order μ\mu relative to xx, and if μ≠0\mu \neq 0, one has log⁡∣f∣∼μlog⁡x\log |f| \sim \mu \log x, whence, on differentiating, f′(x)∼μf(x)xf'(x) \sim \mu \frac{f(x)}{x}; one deduces that f′f' is of order μ−1\mu - 1 relative to xx, which allows one to apply the same argument by induction on nn so long as μ≠n\mu \neq n, whence formula (3) when μ\mu is not an integer ⩾0\geqslant 0 and <n< n.

When ff is of integral order p⩾0p \geqslant 0 relative to xx one can write f(x)=xpf1(x)f(x) = x^p f_1(x), where f1f_1 is of order 0 relative to xx. By prop. 2 one has

f(p)∼p!f1. f^{(p)} \sim p! f_1.
To evaluate the derivatives of order n>pn > p one may restrict oneself to the case p=0p = 0. Then one has log⁡∣f∣≪log⁡x\log |f| \ll \log x, whence f′(x)/f(x)≪1/xf'(x)/f(x) \ll 1/x, in other words xf′(x)≪f(x)x f'(x) \ll f(x); if ff is not equivalent to a constant k≠0k \neq 0 one has, on differentiating this relation (V, p. 232, prop. 7), xf′′(x)+f′(x)≪f′(x)x f''(x) + f'(x) \ll f'(x), which implies that xf′′(x)∼−f′(x)x f''(x) \sim -f'(x). Taking account of this formula, one sees by induction on nn that f(n)f^{(n)} is of order ⩽−n\leqslant -n relative to xx, and that
f(n)∼(−1)n+1(n−1)!f′(x)xn−1. f^{(n)} \sim (-1)^{n+1} (n-1)! \frac{f'(x)}{x^{n-1}}.
(4)

If ff is equivalent to a constant k≠0k \neq 0 one has f(x)=k+f2(x)f(x) = k + f_2(x) with f2≪1f_2 \ll 1, and it reduces to studying the derivatives of f2f_2.

4. (H) functions

Proposition 3

If K0\mathscr{K}_0 is a Hardy field there exists a Hardy field K\mathscr{K} containing K0\mathscr{K}_0 and such that, for every function z∈Kz \in \mathscr{K}, not identically zero on a neighbourhood of +∞+\infty, both eze^z and log⁡∣z∣\log |z| belong to K\mathscr{K}.

Cites 1

Cites App. 0. Hardy Fields. (H) Functions

Denote by K\mathscr{K} the set of functions f∈H(F,R)f \in \mathscr{H}(\mathfrak{F}, \mathbf{R}) having the following properties: for each function f∈Kf \in \mathscr{K} there is a finite number of Hardy fields K1,K2,…,Kn\mathscr{K}_1, \mathscr{K}_2, \ldots, \mathscr{K}_n (the number nn and the fields Ki\mathscr{K}_i depending on ff) such that f∈Knf \in \mathscr{K}_n and, for 0≤i≤n−10 \leq i \leq n-1, one has Ki+1=Ki(ui+1)\mathscr{K}_{i+1} = \mathscr{K}_i(u_{i+1}) where ui+1u_{i+1} is equal either to ezie^{z_i} or to log⁡∣zi∣\log |z_i| with ziz_i belonging to Ki\mathscr{K}_i and not vanishing identically on a neighbourhood of +∞+\infty. One says that u1,u2,…,unu_1, u_2, \ldots, u_n form a definition sequence for the field Kn\mathscr{K}_n and of the function ff; the same function can naturally admit several definition sequences.

By def. 1 of V, p. 247, every function f∈Kf \in \mathscr{K}, not identically zero on a neighbourhood of +∞+\infty, has constant sign and is differentiable on an interval [x0,+∞[[x_0, +\infty[; if f∈Knf \in \mathscr{K}_n, then 1/f1/f and f′f' are equal to functions in Kn\mathscr{K}_n, thus to functions in K\mathscr{K}, on a neighbourhood of +∞+\infty. To see that K\mathscr{K} is a Hardy field it is enough to prove that if ff and gg are two functions in K\mathscr{K} then f−gf - g and fgfg are equal to functions in K\mathscr{K} on a neighbourhood of +∞+\infty. Now let u1,u2,…,umu_1, u_2, \ldots, u_m be a definition sequence for ff, and v1,v2,…,vnv_1, v_2, \ldots, v_n a definition sequence for gg. The sequence u1,u2,…,um,v1,v2,…,vnu_1, u_2, \ldots, u_m, v_1, v_2, \ldots, v_n obtained by concatenating the sequences (ui)(u_i) and (vi)(v_i) is again a definition sequence of a Hardy field Km+n\mathscr{K}_{m+n} and this field contains ff and gg, so f−gf - g and fgfg are equal to functions in Km+n\mathscr{K}_{m+n} on a neighbourhood of +∞+\infty.

One says that the Hardy field K\mathscr{K} defined in the proof of prop. 3 is the (H) extension of the Hardy field K0\mathscr{K}_0.

If K′\mathscr{K}' is another Hardy field possessing the properties stated in prop. 3, it follows from the construction of K\mathscr{K} that K/R∼\mathscr{K}/\mathbf{R}_\sim is contained in K′/R∼\mathscr{K}'/\mathbf{R}_\sim. By abuse of language, one says that the (H) extension of the Hardy field K0\mathscr{K}_0 is the smallest Hardy field K\mathscr{K} having these properties.

Definition 2

The field of (H) functions is called the (H) extension of the Hardy field R(x)\mathbf{R}(x) of rational functions with real coefficients. Any function belonging to this extension is called an (H) function.

Following this definition, if ff is an (H) function, not identically zero on a neighbourhood of +∞+\infty, then efe^f and log⁡∣f∣\log |f| are also (H) functions. More generally, if gg is a second (H) function, and u1,u2,…,unu_1, u_2, \ldots, u_n a definition sequence for gg, and if f(x)f(x) tends to +∞+\infty with xx, one sees by induction on nn that the composite functions u1∘f,u2∘f,…,un∘fu_1 \circ f, u_2 \circ f, \ldots, u_n \circ f and g∘fg \circ f are (H) functions.

5. Exponentials and iterated logarithms

We have already (V, p. 229) defined the iterated logarithms ln(x)l_n(x) by the conditions l0(x)=x,ln(x)=log⁡(ln−1(x))l_0(x) = x, l_n(x) = \log(l_{n-1}(x)) for n≥1n \geq 1. In the same way one defines the iterated exponentials en(x)e_n(x) by the conditions e0(x)=x,en(x)=exp⁡(en−1(x))e_0(x) = x, e_n(x) = \exp(e_{n-1}(x)) for n≥1n \geq 1. It is immediate, by induction on nn, that ln(x)l_n(x) is the inverse function of en(x)e_n(x), defined for x>en−1(0)x > e_{n-1}(0), and that em(en(x))=em+n(x),lm(ln(x))=lm+n(x)e_m(e_n(x)) = e_{m+n}(x), l_m(l_n(x)) = l_{m+n}(x). By the relations log⁡x≪xμ≪ex\log x \ll x^\mu \ll e^x for every μ>0\mu > 0, one has, for n≥1n \geq 1,

ln(x)≪(ln−1(x))μfor every μ>0(5) l_n(x) \ll (l_{n-1}(x))^\mu \quad \text{for every } \mu > 0 \tag{5}
en−1(x1+β)≪en(x1+δ)≪en((1−γ)x)≪(en(x))μ e_{n-1}(x^{1+\beta}) \ll e_n(x^{1+\delta}) \ll e_n((1-\gamma)x) \ll (e_n(x))^\mu
≪en((1+α)x)≪en(x1+β)(6) \ll e_n((1+\alpha)x) \ll e_n(x^{1+\beta}) \tag{6}

for all μ>0,α>0,β>0,0<γ<1,0<δ<1\mu > 0, \alpha > 0, \beta > 0, 0 < \gamma < 1, 0 < \delta < 1, these numbers being otherwise arbitrary (cf. V, p. 218, prop. 11).

We have already seen (V, p. 229) that, for n≥1n \geq 1, one has

ddx(ln(x))=∏i=0n−11li(x).(7) \frac{d}{dx}(l_n(x)) = \prod_{i=0}^{n-1} \frac{1}{l_i(x)}. \tag{7}

Similarly for n≥1n \geq 1

ddx(en(x))=∏i=1nei(x).(8) \frac{d}{dx}(e_n(x)) = \prod_{i=1}^n e_i(x). \tag{8}

whence, by prop. 8 of V, p. 233, for every μ>0\mu > 0

∫axen(tμ) dt∼xμen(xμ)∏i=0n−11ei(xμ)(9) \int_a^x e_n(t^\mu) \, dt \sim \frac{x}{\mu} e_n(x^\mu) \prod_{i=0}^{n-1} \frac{1}{e_i(x^\mu)} \tag{9}
∫x+∞dten(tμ)∼xμ∏i=0n1ei(xμ).(10) \int_x^{+\infty} \frac{dt}{e_n(t^\mu)} \sim \frac{x}{\mu} \prod_{i=0}^n \frac{1}{e_i(x^\mu)}. \tag{10}

One can show that if ff is any (H) function such that f≫1f \gg 1 then there exist two integers mm and nn such that

lm(x)≪f(x)≪en(x) l_m(x) \ll f(x) \ll e_n(x)

(V, p. 263, exerc. 1 and p. 264, exerc. 5) On the other hand, one can define increasing functions g(x)g(x) (which are no longer (H) functions) such that g(x)≫en(x)g(x) \gg e_n(x) for every n>0n > 0, or 1≪g(x)≪lm(x)1 \ll g(x) \ll l_m(x) for every m>0m > 0 (V, p. 265, exerc. 8, 9 and 10).

With the help of the iterated logarithms we shall show that one can define a comparison scale E\mathscr{E} (for xx tending to +∞+\infty) of (H) functions, which are >0> 0 on a neighbourhood of +∞+\infty and satisfy the following conditions:

a) the product of any two functions in E\mathscr{E} belongs to E\mathscr{E}; b) fμ∈Ef^\mu \in \mathscr{E} for every function f∈Ef \in \mathscr{E} and every real number μ\mu;

c) for every function f∈Ef \in \mathscr{E}, log⁡f\log f is a linear combination of a finite number of functions in E\mathscr{E}; d) for every function f∈Ef \in \mathscr{E}, apart from the constant 1, efe^f is equivalent to a function in E\mathscr{E}.

First we consider the set E0\mathscr{E}_0 of functions of the form ∏m=0∞(lm(x))αm\prod_{m=0}^{\infty} (l_m(x))^{\alpha_m}, where the αm\alpha_m are real numbers, zero apart from for a finite number of indices mm; it is immediate, from (5) (V, p. 253) that these functions form a comparison scale which satisfies conditions a),b)a), b) and c). Now we define, by recursion on nn, the set En\mathscr{E}_n (for n⩾1n \geqslant 1) formed by the constant 1 and by the functions of the form exp⁡(∑k=1pakfk)\exp \left( \sum_{k=1}^p a_k f_k \right), where pp is an arbitrary integer >0> 0, the functions fkf_k (1⩽k⩽p1 \leqslant k \leqslant p) are functions in En−1\mathscr{E}_{n-1} such that f1≫f2≫⋯≫fp≫1f_1 \gg f_2 \gg \cdots \gg f_p \gg 1, and the aka_k are real numbers ≠0\neq 0; we show by induction that En\mathscr{E}_n is a comparison scale satisfying a),b)a), b) and c) and containing En−1\mathscr{E}_{n-1}. In the first place, the relation En−1⊂En\mathscr{E}_{n-1} \subset \mathscr{E}_n holds for n=1n = 1, since the logarithm of any nonconstant function in E0\mathscr{E}_0 is of the form ∑k=1pakfk\sum_{k=1}^p a_k f_k, where the fkf_k are iterated logarithms, and so ≫1\gg 1; on the other hand, if En−2⊂En−1\mathscr{E}_{n-2} \subset \mathscr{E}_{n-1} one deduces from the definition of En\mathscr{E}_n that En−1⊂En\mathscr{E}_{n-1} \subset \mathscr{E}_n; this definition furthermore shows that En\mathscr{E}_n satisfies a),b)a), b) and c). It remains to see that En\mathscr{E}_n is a comparison scale: since the quotient of two functions in En\mathscr{E}_n again belongs to En\mathscr{E}_n it suffices to prove that every function ff of En\mathscr{E}_n, apart from the constant 1, cannot be equivalent to a constant ≠0\neq 0. Now one has log⁡f=∑k=1pakfk∼a1f1\log f = \sum_{k=1}^p a_k f_k \sim a_1 f_1 by construction, and since f1≫1f_1 \gg 1, log⁡f\log f tends to ±∞\pm \infty, so ff tends to 0 or to +∞+\infty as xx tends to +∞+\infty.

This being so, if E\mathscr{E} is the union of the En\mathscr{E}_n for n⩾0n \geqslant 0, then E\mathscr{E} is a comparison scale, for two functions in E\mathscr{E} belong to the same scale En\mathscr{E}_n; for the same reason, E\mathscr{E} satisfies a), and it is clear that it also satisfies b) and c). Finally, if f∈Ef \in \mathscr{E} there exists an nn such that f∈Enf \in \mathscr{E}_n; if ff is not the constant 1 then f(x)f(x) tends to 0 or to +∞+\infty as xx tends to +∞+\infty; in the first case ef∼1e^f \sim 1 and in the second, efe^f belongs to En+1\mathscr{E}_{n+1} by definition, and so to E\mathscr{E}.

Remark

Despite the practical usefulness of the scale E\mathscr{E} which we have just defined, it is easy to give examples of (H) functions which have no principal part with respect to E\mathscr{E}. Indeed, if ff is an (H) function such that f∼agf \sim a g, where aa is a constant >0> 0 and g∈Eg \in \mathscr{E}, then log⁡f−log⁡g−log⁡a\log f - \log g - \log a tends to 0 with 1/x1/x, so log⁡f\log f admits, relative to E\mathscr{E}, an asymptotic expansion whose remainder tends to 0, by property c). Now, if one considers for example the (H) function f(x)=e2(x+1x)f(x) = e_2 \left( x + \frac{1}{x} \right) one has log⁡f(x)=exp⁡(x+1x)\log f(x) = \exp \left( x + \frac{1}{x} \right), so the asymptotic expansions of log⁡f\log f relative to E\mathscr{E} are of the form

log⁡f(x)=ex+exx+12!exx2+⋯+1n!exxn+o(exxn)(n an integer >0). \log f(x) = e^x + \frac{e^x}{x} + \frac{1}{2!} \frac{e^x}{x^2} + \cdots + \frac{1}{n!} \frac{e^x}{x^n} + o \left( \frac{e^x}{x^n} \right) \quad (n \text{ an integer } > 0).

It is clear that the remainder in this expansion is equivalent to 1(n+1)!exxn+1\frac{1}{(n+1)!} \frac{e^x}{x^{n+1}}, so does not tend to 0. Hence ff does not have a principal part relative to E\mathscr{E}.

6. Inverse function of an (h) function

If ff is an (H) function, then ff is monotone and continuous on an interval [x0,+∞[[x_0, +\infty[, so the inverse function φ\varphi of the restriction of ff to this interval is monotone and continuous on a neighbourhood of the point a=lim⁡x→+∞f(x)a = \lim_{x \to +\infty} f(x); but, if aa is equal to +∞+\infty (resp. −∞-\infty, finite), one can show that φ(y)\varphi(y) (resp. φ(−y)\varphi(-y), φ(a+1y)\varphi \left( a + \frac{1}{y} \right) or φ(a−1y)\varphi \left( a - \frac{1}{y} \right)) is not in general equal to an (H) function on a neighbourhood of +∞+\infty. Nevertheless we shall see that in certain important cases one can obtain an (H) function equivalent to φ(y)\varphi(y) (resp. φ(−y)\varphi(-y), φ(a+1y)\varphi \left( a + \frac{1}{y} \right), φ(a−1y)\varphi \left( a - \frac{1}{y} \right)) and sometimes even an asymptotic expansion of this function relative to the scale E\mathscr{E} defined in V, p. 254.

We shall use the following proposition:

Proposition 4

Let pp and qq be two (H) functions which are strictly positive on an interval [x0,+∞[[x_0, +\infty[.

Cites 3 · Cited by 1

Cites I § 2. The mean value theorem, V § 3. Asymptotic expansions of functions of a real variable, App. 0. Hardy Fields. (H) Functions

Cited by Proposition 5

1° If q≪p/p′q \ll p/p' one has p(x+q(x))∼p(x)p(x + q(x)) \sim p(x). 2° If both q≪p/p′q \ll p/p' and q(x)≪xq(x) \ll x then p(x−q(x))∼p(x)p(x - q(x)) \sim p(x).

The two parts of this proposition are clear if p∼kp \sim k (constant ≠0\neq 0); one may therefore suppose that p(x)≪1p(x) \ll 1 (otherwise one applies the argument to 1/p1/p). One then deduces p′(x)≪1p'(x) \ll 1.

1° One can write p(x+q(x))=p(x)+q(x)p′(x+θq(x))p(x + q(x)) = p(x) + q(x)p'(x + \theta q(x)) with 0≤θ≤10 \leq \theta \leq 1 (I, p. 14, corollary). Since ∣p′(x)∣|p'(x)| tends to 0 as xx tends to +∞+\infty, and is equal to an (H) function on a neighbourhood of +∞+\infty, it is decreasing on an interval [x1,+∞[[x_1, +\infty[, so, for x≥x1x \geq x_1, one has ∣p′(x+θq(x))∣≤∣p′(x)∣|p'(x + \theta q(x))| \leq |p'(x)|; since qp′≪pqp' \ll p one has p(x+q(x))∼p(x)p(x + q(x)) \sim p(x).

2° The condition q(x)≪xq(x) \ll x ensures that x−q(x)x - q(x) tends to +∞+\infty with xx. Again one has p(x−q(x))=p(x)−q(x)p′(x−θp(x))p(x - q(x)) = p(x) - q(x)p'(x - \theta p(x)) with 0≤θ≤10 \leq \theta \leq 1. The same argument as in the first part of the proof shows that, for xx sufficiently large, ∣p′(x−θq(x))∣≤∣p′(x−q(x))∣|p'(x - \theta q(x))| \leq |p'(x - q(x))|. It reduces to showing that q(x)p′(x−q(x))p(x−q(x))q(x) \frac{p'(x - q(x))}{p(x - q(x))} tends to 0 as xx tends to +∞+\infty. The proposition is true if p′/p≫1p'/p \gg 1 since then ∣p′/p∣|p'/p| is an (H) function, increasing for xx large enough; thus q(x)∣p′(x−q(x))∣∣p(x−q(x))∣≤q(x)∣p′(x)∣∣p(x)∣q(x) \frac{|p'(x - q(x))|}{|p(x - q(x))|} \leq q(x) \frac{|p'(x)|}{|p(x)|}, and qp′≪pqp' \ll p by hypothesis. It is also true if p′/p∼kp'/p \sim k (kk constant ≠0\neq 0), for then

p′(x−q(x))p(x−q(x))∼p′(x)p(x) \frac{p'(x - q(x))}{p(x - q(x))} \sim \frac{p'(x)}{p(x)}

since x−q(x)x - q(x) tends to +∞+\infty. It only remains to examine the case where p′/p≪1p'/p \ll 1. First suppose that p(x)p(x) is of finite order relative to xx, and so (V, p. 232, prop. 7) p′(x)/p(x)≪1/xp'(x)/p(x) \ll 1/x. Then one has

p′(x−q(x))p(x−q(x))=1x−q(x)O1(1), \frac{p'(x-q(x))}{p(x-q(x))} = \frac{1}{x-q(x)} O_1(1),
so
q(x)p′(x−q(x))p(x−q(x))=q(x)x(1−q(x)x)−1O1(1)=q(x)xO2(1), q(x) \frac{p'(x-q(x))}{p(x-q(x))} = \frac{q(x)}{x} \left( 1 - \frac{q(x)}{x} \right)^{-1} O_1(1) = \frac{q(x)}{x} O_2(1),
and one sees in this case that the proposition is true subject only to the hypothesis that q(x)≪xq(x) \ll x. Finally we examine the case where 1/x≪p′(x)/p(x)≪11/x \ll p'(x)/p(x) \ll 1; the function r=p′/pr = p'/p is then of finite order relative to xx; since, by the preceding remark, prop. 4 of V, p. 255, is applicable to such a function, one has p′(x−q(x))/p(x−q(x))∼p′(x)/p(x)p'(x-q(x))/p(x-q(x)) \sim p'(x)/p(x), and the hypothesis qp′≪pqp' \ll p allows us to complete the proof.

Remark

The conditions imposed on q(x)q(x) may not be improved, as the following examples show:

a)

p(x)=ex,q(x)=1=p(x)p′(x),p(x+q(x))=ep(x) p(x) = e^x, \quad q(x) = 1 = \frac{p(x)}{p'(x)}, \quad p(x+q(x)) = ep(x)

b)

p(x)=log⁡x,q(x)=x−log⁡x≪p(x)p′(x)=xlog⁡x, p(x) = \log x, \quad q(x) = x - \log x \ll \frac{p(x)}{p'(x)} = x \log x,
p(x−q(x))=log⁡log⁡x≪p(x). p(x-q(x)) = \log \log x \ll p(x).

We shall first study the inverses of (H) functions of a particular sort:

Proposition 5

Let g be an* (H) function not equivalent to a constant ≠0\neq 0, and such that g(x)≪xg(x) \ll x, and let u(x)u(x) be the inverse function of x−g(x)x - g(x), defined on a neighbourhood of +∞+\infty. Let (un)(u_n) be the sequence of functions defined, by recursion on nn, by the conditions u0(x)=xu_0(x) = x, un(x)=x+g(un−1(x))u_n(x) = x + g(u_{n-1}(x)) for n⩾1n \geqslant 1; then un≫1u_n \gg 1 and

u(x)−un(x)∼g(x)(g′(x))n. u(x) - u_n(x) \sim g(x)(g'(x))^n.
(11)

Cites 2

Cites App. 0. Hardy Fields. (H) Functions, Proposition 4

Let us put y=u(x)y = u(x), yn=un(x)y_n = u_n(x); then x=y−g(y)x = y - g(y), y0=xy_0 = x and yn=x+g(yn+1)y_n = x + g(y_{n+1}). One first deduces that x/y=1−g(y)yx/y = 1 - \frac{g(y)}{y}; since yy tends to +∞+\infty with xx, the hypothesis g(x)≪xg(x) \ll x shows that y=u(x)∼x=y0y = u(x) \sim x = y_0; further,

y−x=g(y)=g(x)+(y−x)g′(z) y - x = g(y) = g(x) + (y - x)g'(z)
where zz belongs to the interval with endpoints x,yx, y; when xx tends to +∞+\infty so does zz, and since g(x)≪xg(x) \ll x, g′≪1g' \ll 1, thus g′(z)g'(z) tends to 0, and consequently
y−x=g(x)+o(y−x) y - x = g(x) + o(y - x)

whence

u(x)−x∼g(x). u(x) - x \sim g(x).
(12)

In the second place we show, by recursion on nn, that when xx tends to +∞+\infty one has un≫1u_n \gg 1, and

u(x)−un(x)≪u(x)−un−1(x). u(x) - u_n(x) \ll u(x) - u_{n-1}(x).
(13)

Indeed, y−yn=g(y)−g(yn−1)=(y−yn−1)g′(zn−1)y - y_n = g(y) - g(y_{n-1}) = (y - y_{n-1})g'(z_{n-1}), where zn−1z_{n-1} belongs to the interval with endpoints yy and yn−1y_{n-1}; by the induction hypothesis zn−1z_{n-1} tends to +∞+\infty with xx, and so g′(zn−1)g'(z_{n-1}) tends to 0, which proves (13). One deduces from this relation and from (12) that u(x)−un(x)≪u(x)−x∼g(x)≪x∼u(x)u(x) - u_n(x) \ll u(x) - x \sim g(x) \ll x \sim u(x), whence un(x)∼u(x)u_n(x) \sim u(x) and consequently un≫1u_n \gg 1. Finally, the relation u(x)−un(x)≪u(x)−xu(x) - u_n(x) \ll u(x) - x can also be written (u(x)−x)−(un(x)−x)≪u(x)−x(u(x) - x) - (u_n(x) - x) \ll u(x) - x, whence

un(x)−x∼u(x)−x∼g(x). u_n(x) - x \sim u(x) - x \sim g(x).
(14)

To prove (11), first remark that if t(x)t(x) is a function such that t(x)−x∼g(x)t(x) - x \sim g(x), then one has g′(t(x))∼g′(x)g'(t(x)) \sim g'(x). Indeed, for any ε>0\varepsilon > 0, for xx sufficiently large g′g' is monotone, so g′(t(x))g'(t(x)) lies between g′(x+(1+ε)g(x))g'(x + (1 + \varepsilon)g(x)) and g′(x+(1−ε)g(x))g'(x + (1 - \varepsilon)g(x)). Prop. 4 of V, p. 255, shows that g′(t(x))∼g′(x)g'(t(x)) \sim g'(x), provided one has established the relation g≪g′/g′′g \ll g'/g''. Now, if gg is of infinite order relative to xx one has (V, p. 251, prop. 2) g′′/g′∼g′/gg''/g' \sim g'/g, and, since g′≪1,g≪g/g′∼g′/g′′g' \ll 1, g \ll g/g' \sim g'/g''; if gg is of finite order μ\mu relative to xx one necessarily has μ≤1\mu \leq 1; if μ<1\mu < 1, since gg is not equivalent to a constant ≠0\neq 0, formulae (3) and (4) (V, p. 251) show that g′′/g′∼k/xg''/g' \sim k/x (kk constant ≠0\neq 0), whence again g≪g′/g′′g \ll g'/g''; finally, if μ=1\mu = 1 then g′g' is of order 0 relative to xx, so g′′/g′≪1/xg''/g' \ll 1/x, and so again g≪g′/g′′g \ll g'/g''.

This being so, since zn−1z_{n-1} lies between yy and yny_n, it follows from (14) that zn−1−x∼g(x)z_{n-1} - x \sim g(x), whence g′(zn−1)∼g′(x)g'(z_{n-1}) \sim g'(x) by the above; thus

y−yn∼(y−yn−1)g′(x), y - y_n \sim (y - y_{n-1})g'(x),
whence we obtain (11) by induction on nn.

Remark 1

If gg is of order <1< 1 relative to xx, the function u(x)−un(x)u(x) - u_n(x) tends to 0 with xx once nn is sufficiently large. Indeed, in the opposite case one would have gg′′n≫1g{g''}^n \gg 1 for every nn, so gg would be of infinite order relative to 1/g′1/g'; in other words, one would have log⁡∣g∣≫log⁡∣g′∣\log|g| \gg \log|g'|, whence, on differentiating, g′/g≫g′′/g′g'/g \gg g''/g'. But if gg is of order μ<1\mu < 1 relative to xx one has g′/g∼g′′/g′g'/g \sim g''/g' when μ=−∞,g′g∼μμ−1g′′g′\mu = -\infty, \frac{g'}{g} \sim \frac{\mu}{\mu - 1} \frac{g''}{g'} when μ≠0\mu \neq 0, and finally g′/g≪g′′/g′g'/g \ll g''/g' when μ=0\mu = 0 (V, p. 251, n 3).

In contrast, if gg is of order 1 relative to xx one can have gg′′n≫1g{g''}^n \gg 1 for every integer n>0n > 0, as the example g(x)=x/log⁡xg(x) = x/\log x shows.

Cites 1

Cites App. 0. Hardy Fields. (H) Functions

Remark 2

When g(x)g(x) is an (H) function equivalent to a constant k≠0k \neq 0 one has g(x)=k+g1(x)g(x) = k + g_1(x), with g1≪1g_1 \ll 1; the function u1(x)=u(x)−ku_1(x) = u(x) - k is the inverse of the function x−g1(x+k)x - g_1(x + k), and one is brought back to the case treated in prop. 5 of V, p. 256.

To have an asymptotic expansion for the function uu it thus suffices to have such an expansion for the function unu_n: if gg admits an asymptotic expansion relative to the scale under consideration one is thus led (by the definition of (H) functions) to the problems examined in V, p. 223 to 227.

The most general case, as follows, reduces to the case treated in prop. 5 of V, p. 256: the function y=u(x)y = u(x) is assumed to satisfy the relation

φ(x)=ψ(y)−g(y) \varphi(x) = \psi(y) - g(y)
where φ\varphi is an (H) function, ψ\psi is an (H) function such that ψ≫1\psi \gg 1 and such that the inverse function θ\theta of ψ\psi is also an (H) function, and gg is an (H) function such that g≪ψg \ll \psi. Let now v(x)v(x) be the inverse function of x−g(θ(x))x - g(\theta(x)); one has u=θ∘v∘φu = \theta \circ v \circ \varphi, and g(θ(x))≪xg(\theta(x)) \ll x; if one knows an asymptotic expansion for vv thanks to prop. 5 of V, p. 256, one can then deduce an asymptotic expansion for uu as in V, p. 223 to 227.

Cites 2

Cites App. 0. Hardy Fields. (H) Functions, V § 2. Asymptotic expansions

Example 1

Let us seek an asymptotic expansion for the inverse function v(x)v(x) of x5+xx^5 + x (for xx tending to +∞+\infty); putting x5=tx^5 = t one is led to seek an expansion of the inverse function u(t)u(t) of t+t1/5t + t^{1/5} (for tt tending to +∞+\infty), that is, to apply prop. 5 of V, p. 256, to the case where g(t)=−t1/5g(t) = -t^{1/5}. Let us, for example, calculate, u2(t)u_2(t); one has

u2(t)=t−(t−t1/5)1/5=t−t1/5+15t−3/5+225t−7/5+o1(t−7/5). u_2(t) = t - \left( t - t^{1/5} \right)^{1/5} = t - t^{1/5} + \frac{1}{5} t^{-3/5} + \frac{2}{25} t^{-7/5} + o_1\left( t^{-7/5} \right).
Moreover, by (11) (V, p. 256)
u(t)−u2(t)∼−125t−7/5 u(t) - u_2(t) \sim -\frac{1}{25} t^{-7/5}
whence
u(t)=t−t1/5+15t−3/5+125t−7/5+o2(t−7/5) u(t) = t - t^{1/5} + \frac{1}{5} t^{-3/5} + \frac{1}{25} t^{-7/5} + o_2\left( t^{-7/5} \right)
and one deduces the expansion
v(x)=u(x)1/5=x1/5−15x−3/5−125x−7/5+o3(x−7/5) v(x) = u(x)^{1/5} = x^{1/5} - \frac{1}{5} x^{-3/5} - \frac{1}{25} x^{-7/5} + o_3\left( x^{-7/5} \right)
sought.

Cites 1

Cites App. 0. Hardy Fields. (H) Functions

Example 2

Let us seek an asymptotic expansion for the inverse function of the function x/log⁡xx / \log x; from the identity x=y/log⁡yx = y / \log y, where y=v(x)y = v(x), one deduces log⁡x=log⁡y−log⁡log⁡y\log x = \log y - \log \log y; on putting z=log⁡y,t=log⁡xz = \log y, t = \log x, one has t=z−log⁡zt = z - \log z, and one is led to expand the inverse function u(t)u(t) of t−log⁡tt - \log t; for example, one has

u2(t)=t+log⁡(t+log⁡t)=t+log⁡t+log⁡tt−(log⁡t)22t2+o1(log⁡tt2) u_2(t) = t + \log(t + \log t) = t + \log t + \frac{\log t}{t} - \frac{(\log t)^2}{2t^2} + o_1\left( \frac{\log t}{t^2} \right)
and moreover, by (11) (V, p. 256)
u(t)−u2(t)∼log⁡tt2 u(t) - u_2(t) \sim \frac{\log t}{t^2}
whence
u(t)=t+log⁡t+log⁡tt−(log⁡t)22t2+log⁡tt2+o2(log⁡tt2) u(t) = t + \log t + \frac{\log t}{t} - \frac{(\log t)^2}{2t^2} + \frac{\log t}{t^2} + o_2\left( \frac{\log t}{t^2} \right)
and on returning to the original problem one obtains the asymptotic expansion
v(x)=xlog⁡x+xlog⁡log⁡x+xlog⁡log⁡xlog⁡x+o(xlog⁡log⁡xlog⁡x). v(x) = x \log x + x \log \log x + x \frac{\log \log x}{\log x} + o\left( x \frac{\log \log x}{\log x} \right).

Cites 1

Cites App. 0. Hardy Fields. (H) Functions

Remark

Note that two equivalent (H) functions can have non-equivalent inverses, as the example of the two functions log⁡x\log x and 1+log⁡x1 + \log x shows.

Exercises

See the exercises for Appendix 0.

Transcribed from the printed volume. Not translated.