Let $f_n(x)$ be recursively defined as $$f_0(x)=1,\ \ \ f_{n+1}(x)=\sqrt{x+f_n(x)},\tag1$$ i.e. $f_n(x)$ contains $n$ radicals and $n$ occurences of $x$: $$f_1(x)=\sqrt{x+1},\ \ \ f_2(x)=\sqrt{x+\sqrt{x+1}},\ \ \ f_3(x)=\sqrt{x+\sqrt{x+\sqrt{x+1}}},\ \dots\tag2$$

The functions $f_0(x)$, $f_1(x)$ and $f_2(x)$ are integrable in elementary functions, e.g.: $$\int\sqrt{x+\sqrt{x+1}}\,dx=\left(\frac{2\,x}3+\frac{\sqrt{x+1}}6-\frac14\right)\sqrt{x+\sqrt{x+1}}+\frac58\ln\left(2\,\sqrt{x+1}+2\,\sqrt{x+\sqrt{x+1}}+1\right).\tag3$$

Question: Is there an integer $n>2$ such that $f_n(x)$ is integrable in elementary functions?

Update: The question is reposted at MathOverflow as was suggested by moderator.

Vladimir Reshetnikov
  • 45,303
  • 7
  • 151
  • 282
  • Have you looked in such proofs with [differential Galois theory](http://en.wikipedia.org/wiki/Differential_Galois_theory)? – user64494 Sep 21 '13 at 20:44
  • 6
    I am very far from being an expert in that area. My major was theoretical physics, we were trained to solve quite complex intgration problems, but many theoreticals aspects of math were only incidentally mentioned. – Vladimir Reshetnikov Sep 22 '13 at 19:22
  • There is Liouville's theorem in differential Galois theory which says that if $f$ has an antiderivative, then $f$ has the form $f=\sum c_i g_i'/g_i+h'$, where $g_i$, $h$ belong to a differential field $F$ that contains $f$. So IIUC, because $f_n^2=x+f_{n-1}$ and $f_n'=(1+f_{n-1}')/(2f_n)$, the antiderivative of $f_n$ can only involve a rational functions of $\{x,f_1,\ldots,f_n\}$ and a sum of logarithms of some such functions. This limits the scope of what functions one needs to look at. [Link to a description](http://www.claymath.org/programs/outreach/academy/LectureNotes05/Conrad.pdf). – Kirill Sep 30 '13 at 05:08
  • @Kirill "Has the form" means "can be represented in the form", right? So, one would need to not only look at the apparent form of the integrand, but to also prove that it cannot be transformed to a certain form. – Vladimir Reshetnikov Sep 30 '13 at 16:21
  • @Kirill Anyways, feel free to publish a partial answer, if you feel it can help to make the progress towards the final result. – Vladimir Reshetnikov Sep 30 '13 at 16:23
  • I updated my answer to reflect partial progress (I don't know if you get notifications for answer updates). I believe $f_3$ is (provably) integrable in an elementary form that I haven't been able to find. – Kirill Oct 12 '13 at 17:41
  • @Kirill Thanks, it is very interesting! – Vladimir Reshetnikov Oct 12 '13 at 17:56
  • @VladimirReshetnikov: I'm actually wondering if posting a new question on MO would be a better course of action. While there are no complete solutions, I think there is some valuable content here, and migrating all of this to MO would be to the detriment of math.SE. If you do cross-post this question, make sure that each version contains a link to the other. (If you absolutely insist on migration, I can still do this.) – user642796 Nov 28 '13 at 09:22
  • @ArthurFischer Ok. I'll repost. – Vladimir Reshetnikov Nov 28 '13 at 17:51

4 Answers4


I posted an answer to this question as an answer to the related question on MathOverflow. See my (second) answer at https://mathoverflow.net/questions/171733. However, as Krokop pointed out, it would be better to have a copy here on Math.SE to the original question. Thanks to Krokop for copying my answer over to Math.SE.

I am going to show that there is no elementary antiderivative of $f_n$ when $n>2$.

Assume $n>2$ (NB: This is important, because the argument below will not work for $n\le2$; the reader may enjoy finding where it breaks down), and let $K_n = {\mathbb C}\bigl(x,f_n(x)\bigr)$ be the elementary differential field generated by $x$ and $f_n(x)$. Then $K_n$ is the field of meromorphic functions on the normalization $\hat C_n$ of the algebraic curve $C_n$ defined by the minimal degree $y$-monic polynomial $P_n(x,y)$ that satisfies $P_n\bigl(x,f_n(x)\bigr) \equiv 0$. This minimal degree is $2^n$; for example, $P_2(x,y) = (y^2-x)^2-x-1$ and $P_3(x,y) = \bigl((y^2-x)^2-x\bigr)^2-x-1$, etc.

Since $P_{n+1}(x,y) = (P_n(x,y)+1)^2-x-1$ for $n\ge 1$ with $P_1(x,y)=y^2-x-1$, one sees, by applying the Eisenstein Criterion to $P_n(x,y)$ regarded as an element of $D[y]$ with $D$ being the integral domain ${\mathbb C}[x]$, that $P_n(x,y)$ is irreducible for all $n\ge 1$. Hence, $\hat C_n$ is connected.

It will be important in what follows to observe that $K_n$ has an involution $\iota$ that fixes $x$ and sends $f_n(x)$ to $-f_n(x)$; this is because $P_n(x,y)$ is an even polynomial in $y$. The fixed field of $\iota$ is ${\mathbb C}\bigl(x,\,f_n(x)^2\bigr)$, and the $(-1)$-eigenspace of $\iota$ is ${\mathbb C}\bigl(x,\,f_n(x)^2\bigr)f_n(x) = K_{n-1}{\cdot}f_n(x)$.

Now, the curve $C_n\subset \mathbb{CP}^2$ has only one point on the line at infinity, namely $[1,0,0]$, but the normalization $\hat C_n$ has $2^{n-1}$ points lying over this point. They can be parametrized as follows: First, establish the convention that $\sqrt{u}$ means the unique analytic function on the complex $u$-plane minus its negative axis and $0$ that satisfies $\sqrt1 = 1$ and $\bigl(\sqrt{u}\bigr)^2 = u$. Let $\epsilon = (\epsilon_1,\ldots,\epsilon_{n-1})$ be any sequence with ${\epsilon_k}^2=1$ and consider the sequence of functions $g^\epsilon_k(t)$ defined by the criteria $g^\epsilon_1(t) = \sqrt{1+t^2}$ and $g^\epsilon_{k+1}(t) = \sqrt{1+\epsilon_{n-k}t g^\epsilon_k(t)}$ for $1\le k < n$. Choose, as one may, a $\delta_n>0$ sufficiently small so that, when $t$ is complex and satisfies $|t|<\delta_n$, all of the functions $g^\epsilon_k$ are analytic when $|t|<\delta_n$. In particular, one finds an expansion $$ g^\epsilon_n(t) = 1+\tfrac12\epsilon_1\,t + \tfrac18(2\epsilon_1\epsilon_2-1)t^2 + O(t^3). $$

Also, it is easy to verify that the disk in $\mathbb{CP}^2$ defined by $$ [x,y,1] = [1,\ t g^\epsilon_n(t),\ t^2]\qquad\text{for}\quad |t|<\delta_n $$ is a nonsingular parametrization of a branch of $C_n$ in a neighborhood of the point $[1,0,0]$. In the normalization $\hat C_n$, this is then a local parametrization of a neighborhood of a point $p_\epsilon\in \hat C_n$. Obviously, this describes $2^{n-1}$ distinct points on $\hat C_n$.

When $x$ and $f_n$ are regarded as meromorphic functions on $\hat C_n$, it follows that there is a unique local coordinate chart $t_\epsilon:D_\epsilon\to D(0,\delta_n)\subset \mathbb{C}$ of an open disk $D_\epsilon\subset \hat C_n$ about $p_\epsilon$ such that $t_\epsilon(p_\epsilon)=0$ and on which one has formulae $$ x = \frac1{{t_\epsilon}^2} \quad\text{and}\quad f_n(x) = \frac{g^\epsilon_n(t_\epsilon)}{t_\epsilon} = \frac{1+\tfrac12\epsilon_1\ t_\epsilon +\tfrac18(2\epsilon_1\epsilon_2-1)\ {t_\epsilon}^2} {t_\epsilon} + O({t_\epsilon}^2). $$ In particular, it follows that $f_n(x)$, as a meromorphic function on $\hat C_n$, has polar divisor equal to the sum of the $p_\epsilon$ and hence has degree $2^{n-1}$. Of course, this implies that the zero divisor of $f_n(x)$ on $\hat C_n$ must be of degree $2^{n-1}$ as well.

Note that the functions $g^\epsilon_k$ satisfy $g^{-\epsilon}_k(-t) = g^{\epsilon}_k(t)$, where $-\epsilon = (-\epsilon_1,\ldots,-\epsilon_{n-1})$. This implies that $\iota(p_\epsilon) = p_{-\epsilon}$ and that $t_\epsilon\circ\iota = -t_{-\epsilon}$.

Now, the $2^{n-1}$ zeroes of $f_n(x)$ on $\hat C_n$ are distinct, for they are the zeros of the polynomial $q_n(x) = P_n(x,0) = (q_{n-1}+1)^2-x-1$, and the discriminant of $q_n$, being the resultant of $q_n$ and $q_n'$, is clearly an odd integer, and hence is not zero. Thus, $C_n$ is a branched double cover of $C_{n-1}$, branched exactly where $f_{n}$ has its zeros. This induces a branched cover $\pi_n:\hat C_n\to \hat C_{n-1}$ that is exactly the quotient of $\hat C_n$ by the involution $\iota$ (whose fixed points are where $f_n$ has its zeros). Since one then has the Riemann-Hurwitz formula $$ \chi(\hat C_n) = 2\chi(\hat C_{n-1}) - B_n = 2\chi(\hat C_{n-1}) - 2^{n-1}, $$ and $\chi(\hat C_1) = \chi(\hat C_2) = 2$, induction gives $\chi(\hat C_n) = (3{-}n)2^{n-1}$, so the genus of $\hat C_n$ is $(n{-}3) 2^{n-2} + 1$. (This won't actually be needed below, but it is interesting.)

The only poles of $x$ and $f_n(x)$ on $\hat C_n$ are the points $p_\epsilon$, and computation using the above expansions shows that, in a neighborhood of $p_\epsilon$, one has an expansion of the form $$ f_n(x)\,\mathrm{d} x - \mathrm{d}\left(f_n(x)\bigl(\tfrac12\ x + \tfrac16\ f_n(x)^2\bigr) \right) = \left(\frac{ (1-\epsilon_1\epsilon_2) } {4{t_\epsilon}^2} + O({t_\epsilon}^{-1})\right)\ \mathrm{d} t_\epsilon\ . $$ Thus, the meromorphic differential $\eta$ on $\hat C_n$ defined by the left hand side of this equation has, at worst, double poles at the points $p_\epsilon$ and no other poles.

Now, by Liouville's Theorem, $f_n$ has an elementary antiderivative if and only if $f_n(x)\ \mathrm{d} x$ and, hence, the form $\eta$ are expressible as finite linear combinations of exact differentials and log-exact differentials. Thus, $f_n(x)$ has an elementary antiderivative if and only if $\eta$ is expressible in the form $$ \eta = \mathrm{d} h + \sum_{i=1}^m c_i\,\frac{\mathrm{d} g_i}{g_i} $$ for some $h,g_1,\cdots g_m\in K_n$ and some constants $c_1,\ldots,c_m$. Suppose that these exist. Since $\eta$ has, at worst, double poles at the $p_\epsilon$ and no other poles, it follows that $h$ must have, at worst, simple poles at the points $p_\epsilon$ and no other poles; in fact, $h$ is uniquely determined up to an additive constant because its expansion at $p_\epsilon$ in terms of $t_\epsilon$ must be of the form $$ h = \frac{\epsilon_1\epsilon_2-1}{4t_\epsilon} + O(1). $$
Moreover, because $\eta$ is odd with respect to $\iota$, it follows that $h$ (after adding a suitable constant if necessary) must also be odd with respect to $\iota$. This implies, in particular, that $h$ vanishes at each of the zeros of $f_n$ (which, by the argument above, are simple zeros). This implies that $h = r\,f_n$ for some $r\in K_{n-1}$ that has no poles and satisfies $r(p_\epsilon) = (\epsilon_1\epsilon_2-1)/4$ for each $\epsilon$. However, since $r$ has no poles and $\hat C_n$ is connected, it follows that $r$ is constant. Thus, it cannot take the two distinct values $0$ and $-1/2$, as the equation $r(p_\epsilon) = (\epsilon_1\epsilon_2-1)/4$ implies.

Thus, the desired $h$ does not exist, and $f_n$ cannot be integrated in elementary terms for any $n>2$.

Robert Bryant
  • 1,126
  • 10
  • 10
  • 1
    I copied your answer. I hope its okay. And +1 for your excellent answer. – Krokop Jun 18 '14 at 11:39
  • 2
    @Krokop: Yes, it's OK. In fact, thanks for doing this. However, I think that I will need to edit the opening sentences so that they will make sense. – Robert Bryant Jun 18 '14 at 11:47

This is only a partial attempt at answering the question. I read this description of differential Galois theory to find out about this. I'll try to present the relevant bits of those notes---this is not original, and all mistakes are mine.

First, define what is meant by an elementary function. Let $\mathbb{C}(x,g_1,\ldots,g_n)$ be the set of all rational functions of $x$ and $g_i$, where $g_i(x)$ are functions of $x$. An elementary field $K=\mathbb{C}(x,g_1,\ldots,g_n)$ is a field of functions such that each $g_j$ is an exponential, logarithm or algebraic function of an element of $\mathbb{C}(x,g_1,\ldots,g_{j-1})$. Such an elementary field is closed under differentiation.

Liouville's theorem then says that if $f$ is a function in $K$ that has an elementary antiderivative, that antiderivative cannot be just any arbitrary function of elements of $K$, but must be equal to $$ \sum_i c_i \log g_i + h, $$ for some functions $g_i$, $h$ in $K$, and constants $c_i$.

For the given functions $f_n$, $f_n^2=x+f_{n-1}$, we can choose elementary functions to mean $K=\mathbb{C}(x,f_1,\ldots,f_n)$ (i.e., all rational expressions involving $f_i$ and $x$), and deduce that if $\int f_n$ is elementary, we must have $$ f_n = \sum_i c_i \frac{g_i'}{g_i} + h', $$ where $g_i$ and $h$ are in $K$.

While this doesn't seem like all that much, it does mean that one (rigourously) knows one doesn't ever have to look at functions like $\sqrt{x+1}\log\sqrt{x+1}$ or $e^{\sqrt{x+1}}$ when integrating $\sqrt{x+\sqrt{x+1}}$, if "elementary" functions are as defined above. As to how you can actually find those functions $g_i$, $h$ or disprove their existence, I have no idea.

I tried to integrate $f_3$ and $f_4$ by guessing a general form and solving for coefficients, but it didn't work.

Update. (This is a work in progress, so maybe the claims and expressions will be somewhat of a mystery. The question is certainly interesting.)

There is an algorithm, the Risch-Hermite-Trager algorithm (see Computer Algebra by Geddes, Czapor, Labahn, and also Integrating Algebraic Functions by Trager), for integrating arbitrary algebraic expressions in elementary form or proving their integrals are not elementary. The function $f_3(x)$ is maybe integrable in elementary form, with $$\int f_3 = \frac{-3+8x+2f_2+3 f_2 f_3/x}{12}f_3 + \int q_3, $$ where $\int q_3$ is a sum of logarithmic terms as in Liouville's theorem (but I am unable to integrate it explicitly, even the case of $f_2$ is hard for me: I asked related questions here and here, but my knowledge of algebraic geometry is lacking). $$ q_3 = \left(\left(11 x^2+23 x^3+x^4-13 x^5-x^6+x^7+\left(-6 x-13 x^2-x^3+18 x^4+6 x^5-6 x^6\right) f_2+f_1 \left(-4 x-9 x^2-9 x^3+15 x^5+x^6-4 x^7+\left(8+17 x-5 x^3-2 x^4-23 x^5+13 x^6\right) f_2\right)\right) f_3\right)/\left(32 x^2 (1+x) \left(-1-x+x^2\right) \left(-1-x+x^2-2 x^3+x^4\right)\right). $$ I am assured by the algorithm that it is integrable in elementary terms if and only if it is possible to write $$ \int q_3 = \sum c_i \log v_i, $$ where $c_i$ are numbers that belong to the field $\mathbb{Q}$ extended by the roots of the polynomial $$\begin{gathered} R(Z) = \left(-288966001+75889061888 Z^2-15925264777216 Z^4+88562225643520 Z^6+439804651110400 Z^8\right)^2 \times\\ \left(-14508481-784902408192 Z^2-757665491845120 Z^4-103142060671792840704 Z^6+635008673359146778624 Z^8\right)^2 \times\\ \left(46732663382478474361-89569724082204845664256 Z^2+904900289526729208095571968 Z^4+1144371243549912082526161076224 Z^6+3200188454940133648592935688601600 Z^8+11158139832684515154655834451715555328 Z^{10}+9545270929062413812474597798367805308928 Z^{12}-124380415953981823700469862336514528116736 Z^{14}+50404501905167945967686052562499557916672 Z^{16}\right)^2. \end{gathered}, $$ and $v_i$ have certain prescribed roots and poles, but I don't know how to calculate $c_i$ and $v_i$ explicitly (in theory it is possible to either find them or establish they cannot exist).

The point is that there is a deterministic algorithm for figuring out the algebraic portion of the integral $\int f_3$, and the integral can be evaluated or proven to be non-elementary by constructing a certain polynomial $R(Z)$, whose roots form a basis over $\mathbb{Q}$ of the field to which the coefficients $c_i$ belong. If all the roots are zero (non-zero integrand that has only zero logarithmic terms), the integral is known to be non-elementary. Otherwise, the logarithmic term is given by $$\sum_{\beta,R(\beta)=0} c_\beta \log v_\beta, $$ where (roughly) $v_\beta$ is a function of $(x,y)$ that has a root/pole wherever $q_3\,dx$ (viewed as a function of both $x$ and $y$, where $y$ is a multivalued algebraic function of $x$) has a pole with residue $\beta$, and no other roots or poles anywhere else. This is complicated by a change of variables $x=z^{-2}$, $\hat y=z y$ necessary earlier, and the fact that the curve $(x,y)$ has singular points that make some things ill-defined. Such functions $v_\beta$ (they are generators of specific principal ideals) may not exist for some other reasons, but nothing about the function $q_3$ prevents their existence.

I also calculated $R(Z)$ for $f_4$, and it has non-zero roots, meaning integrability of $f_4$ in elementary terms is equivalent to finding a set of those functions $v_\beta$; I don't know if they exist or not.

  • 14,246
  • 1
  • 35
  • 72
  • 1
    Does there any inductive theorem for unintegrable in elementary functions? For example the one which is very useful for this question, if $f(x)$ is unintegrable in elementary functions, does $\sqrt{x+f(x)}$ also always unintegrable in elementary functions? If has, the proof can be much more simplified as we only need to prove $f_3(x)$ is unintegrable in elementary functions. For trying to prove $f_3(x)$ is unintegrable in elementary functions, you can try to prove the last integral in my answer is unintegrable in elementary functions. – Harry Peter Oct 02 '13 at 13:45
  • I think not @HarryPeter +1 though. – mick Oct 12 '13 at 00:09
  • @Kirill Strictly speaking, Risch method is a [semi-algorithm](http://en.wiktionary.org/wiki/semi-algorithm), because sometimes it needs to check if a given elementary expression is equal to zero, and this problem is known to be undecidable in general. – Vladimir Reshetnikov Oct 12 '13 at 17:50
  • @VladimirReshetnikov Interesting. But if the expression is a rational function in $\mathbb{Q}(x,y)$ with a polynomial relation $f(x,y)=0$, then I don't see what's undecidable about that. Same for elements of an algebraic extension of $\mathbb{Q}$. – Kirill Oct 12 '13 at 17:57
  • @Kirill Yes, I think in this case undecidability is not relevant. See [this](http://en.wikipedia.org/wiki/Risch_algorithm#Decidability) for details. – Vladimir Reshetnikov Oct 12 '13 at 18:04
  • The link is broken – 2016 Jun 15 '14 at 20:30

I think there is highly have no chance to have an integer $n>2$ such that $f_n(x)$ is integrable in elementary functions.

The reasons are mainly because of the processes of eliminating the radicals.

For $f_0(x)$ and $f_1(x)$ , they are trivially integrable in elementary functions.

For $\int f_2(x)~dx$ ,

Let $u=\sqrt{x+1}$ ,

Then $x=u^2-1$



$=\int2u\sqrt{u^2+u-1}~du$ , which can express in elementary functions.

Starting from $\int f_3(x)~dx$ ,

Let $u=\sqrt{x+1}$ ,

Then $x=u^2-1$




Introduce the Euler substitution:

Let $v=u+\sqrt{u^2+u-1}$ ,

Then $u=\dfrac{v^2+1}{2v+1}$




$=\int4\dfrac{(v^2+1)(v^2+v-1)}{(2v+1)^3}\sqrt{\dfrac{(v^2+1)^2-(v^2+1)(2v+1)+(v-1)(2v+1)^2}{(2v+1)^2}}~dv$ , which is highly have no chance to express in elementary functions.

Harry Peter
  • 7,551
  • 1
  • 20
  • 50

If you pick $f_0 = 1, K_0 = \Bbb C(X)$ and $f_{n+1} = \sqrt{X+f_n}, K_{n+1} = K_n(f_{n+1})$, the first three fields have genus $0$, so integrating $f_{n}$ is as easy as integrating a rational fraction.

For example, if we put $w = \frac {f_1-1}{f_2-1}$ then $K_2 = \Bbb C(w)$ and we have formulas

$$ x = \frac{3w^4-8w^3+10w^2-4w}{(w^2-1)^2}, f_1 = \frac{-2w^2+2w-1}{w^2-1}, f_2 = \frac{w^2-3w+1}{w^2-1} $$

Then integrating $f_2dx$ is integrating $f_2(dx/dw) dw$, and we obtain a result in terms of $w$ and $\log(w^2-1)$.

However, $K_3 = K_2(\sqrt{x+f_2}) = K_2(v = \sqrt{4w^4-11w^3+10w^2-w-1})$ is obtained by adjoining a square root of an element of degree $4$, so $K_3$ is an elliptic curve (the following fields' genus then grow exponentially). And you want to integrate the $1$-form $f_3 dx$ on this curve.

If I am not mistaken, the poles of the $1$-form are at the $4$ points at $x = \infty$ (or rather $w^2 = 1$). A computation shows that the residues are $+1/4$ around the points (in the coordinates $(v,w)$) $A =(+5,-1), B = (+1,+1)$ and $-1/4$ around the points $C = (-5,-1), D = (-1,+1)$. Sadly, $[A]+[B]-[C]-[D] \neq 0$. I think this is a point of infinite order on the elliptic curve so there is no rational function we can plug into a $\log$ to cancel the residues of the $1$-form.

So unless I'm mistaken, a primitive of $f_3dx$ can't even be expressed using logarithms and elliptic functions.

  • 48,894
  • 2
  • 77
  • 128