Consider the following integral:

$$I_k=\int_0^\infty \left(\text{sech}x\tanh\tfrac12x\right)^k\,dx$$

where $k\in\Bbb N$.

If we evaluate $I_1$, $I_2$, $I_3$, etc. we get the following pattern:

  • $I_1=\log(2)$
  • $I_2=-3+\pi$
  • $I_3=\frac 72-5\log(2)$
  • $I_4=22-7\pi$
  • $I_5=-\frac{341}{12}+41\log(2)$
  • $I_6=-\frac{968}{5}+\frac{493}{8}\pi$
  • ...

From this data we can see that:

$$\begin{align} \pi&=3+I_2=\frac{22}{7}-\frac17 I_4=\frac{7744}{2465}+\frac{8}{493}I_6 \\[4pt] \log(2)&=0+I_1=\frac{7}{10}-\frac15I_3=\frac{341}{492}+I_5 \end{align}$$

And because $I_k$ decreases very rapidly($I_{10}$ is in the order of $1e6$) we can set $I_k\approx 0$ for high $k$ and we get rational approximations of both, $\pi$ and $\log(2)$, for $I_{2k}$ and $I_{2k+1}$, respectively, that apparently go on forever.

I see some equations that somehow "encode" the information of a given number, but how is it that this integral has the information of both $\pi$ and $\log(2)$, apparently unrelated numbers?


  • $\begingroup$ Please write an informative title that will allow others to find the problem and solution later. Cutesy uninformative titles are annoying and unhelpful. $\endgroup$ – David G. Stork Aug 29 '19 at 2:04
  • $\begingroup$ Interesting exercise, for sure and $\to +1$ for the question. $\endgroup$ – Claude Leibovici Aug 29 '19 at 4:21
  • $\begingroup$ $\pi$ and $\ln 2$ are not actually very unrelated. The $\ln 2$ constant shows up naturally, for instance, in the study of integrals like $$\int_0^{\pi/2}\ln(\sin t)dt=-\frac{\pi}{2}\ln 2.$$ See also the relations between $\pi$, $\ln2$, $e$, and Catalan's constant $\mathrm{G}$ in this post. $\endgroup$ – clathratus Nov 19 '19 at 5:25

I'm not sure exactly what kind of an answer you're looking for---but the following might be a helpful heuristic:

  1. A standard substitution transforms the integrals $$I_k=\int_0^\infty \left(\operatorname{sech} x \tanh\frac{x}{2}\right)^k\,dx ,$$ (for $k$ a nonnegative integer, which we henceforth assume) into integrals of rational functions of a new variable.

  2. After applying the method of partial fractions, we can integrate term-by-term, and the only non-rational functions that occur in the antiderivatives have the form $\log q(u)$ and $\arctan r(u)$ for some affine functions $q, r$ with rational coefficients.

  3. At typical limits of integration the functions $q(u)$ usually take on values of rational number with small numerator and denominator, and so the contribution of these terms is a sum of logarithms of small numbers---of course, the smallest positive integer with nonzero logarithm is $2$, so it's no surprise that $\log 2$ occurs in the values of such integrals often.

  4. Likewise, if we can write $\arctan v$ without $\arctan$, typically the result is some rational multiple of $\pi$. In particular, we have $\arctan 1 = \frac{\pi}{4}$ and $\lim_{v \to \infty} \arctan v = \frac{\pi}{2}$, no it's no surprise that $\pi$ occurs often either.

More explicitly: Applying the hyperbolic analogue $$x = 2 \operatorname{artanh} t$$ of the Weierstrass substitution transforms $I_k$ into an integral with a rational integrand, $$I_k = 2 \int_0^1 \frac{(1 - t^2)^{k - 1} t^k \,dt}{(1 + t^2)^k}.$$ The next step for evaluating this integral in terms of elementary functions depends on the parity of $k$.

For $k$ odd, substituting $$u = t^2, \qquad du = 2 t \,dt$$ gives $$I_k = \int_0^1 \frac{(1 - u)^{k - 1} u^{(k - 1) / 2}}{(u + 1)^k} .$$ Expanding the integrand using partial fractions gives $$I_k = \int_0^1 \left(P(u) + \frac{A_t}{(u + 1)^k} + \cdots + \frac{A_2}{(u + 1)^2} + \frac{A_1}{u + 1} \right) du $$ for some rational polynomial $P(u)$ and rational coefficients $A_1, \ldots, A_k$. But the antiderivative of every term but $\frac{A_1}{u + 1}$ is a rational function, so each contributes some rational number, altogether contributing a rational total, call it $\alpha$. The value of the last term is $A_1 \int_0^1 \frac{du}{u + 1} = A_1 \log u\vert_0^1 = A_1 \log 2$, so $$\boxed{I_k = \alpha_k + \beta_k \log 2}$$ for some rational numbers $\alpha_k$ and $\beta_k := A_1$. A tedious but straightforward induction shows that $\beta_k \neq 0$.

For $k$ even, that substitution is not available (or more precisely, it makes the integrand worse), but applying the method of partial fractions again gives $$I_k = \int_0^1 \left(P(t) + \frac{A_k t + B_k}{(t^2 + 1)^k} + \cdots + \frac{A_2 t + B_2}{(t^2 + 1)^2} + \frac{A_1 t + B_1}{(t^2 + 1)} \right) dt $$ for some polynomial $P(t)$ and rational coefficients $A_1, \ldots, A_k, B_1, \ldots, B_k$. But our integrand in $u$ is even, and so $A_1 = \cdots = A_k = 0$. On the other hand, for $m > 1$, the integral of $\frac{1}{(t^2 + 1)^m}$ satisfies the reduction formula of the form $$\int \frac{dt}{(t^2 + 1)^m} = f_m(t) + \gamma_m \int \frac{dt}{(t^2 + 1)^{m - 1}}$$ for some rational function $f_m$ and rational constant $\gamma_m$ (both depending on $m$). In particular, induction gives $$\int \frac{dt}{(t^2 + 1)^m} = g_m(x) + \delta_m \int \frac{dx}{t^2 + 1} = g_m(t) + \delta_m \arctan t + C$$ for some rational function $g_m$ and rational constant $\delta_m$ (again both depending on $m$). Substituting back into our previous formula for $I_k$, we have that $$I_k = [h(t) + \zeta \arctan t]\vert_0^1 = h(1) - h(0) + \beta_k \pi$$ for some rational function $h$ and rational constant $\beta_k$, and so $$\boxed{I_k = \alpha_k + \beta_k \pi}$$ for rational numbers $\alpha_k := h(1) - h(0)$ and $\beta_k$. Again, an induction shows that $\beta_k \neq 0$ for $k > 0$.

Remark 1 Essentially the same phenomenon occurs for the similar family $$J_j := \int_0^1 \frac{x^{2 j} (1 - x)^{2 j}dx}{1 + x^2} ,$$ which generalizes the so-called Dalzell integral (the case $j = 2$, which gives $\frac{22}{7} - \pi$). If $j$ is odd, we get an expression of the form $\alpha + \beta \log 2$, and if $j$ is even we get $\alpha + \beta \pi$, with $\alpha, \beta \neq 0$ in both cases.

Remark 2 We can use the resulting integral expressions to derive rational approximations of $\pi$ and $\log 2$. On the interval $[0, 1]$, $\frac{1}{2^k} \leq \frac{1}{(1 + t^2)^k} \leq 1$, giving the bounds $$ \frac{1}{2^k} E_k < I_k < E_k, \\ \textrm{where} \quad E_k = 2 \int_0^1 u^k (1 - u^2)^k du = \frac{\Gamma(k) \Gamma\left(\frac{1}{2} k + \frac{1}{2}\right)}{\Gamma\left(\frac{3}{2} k + \frac{1}{2}\right)} \sim \frac{\sqrt{2 \pi}}{\sqrt{k}} \left(\frac{2}{3 \sqrt{3}}\right)^k ,$$ so for any particular $k$, rearranging gives rational bounds for $\pi$ or $\log 2$. (For odd $k = 2 l + 1$, by the way, we have $E_{2 l + 1} = \frac{(2 l)! l!}{(3 l + 1)!}$.)

For example, taking $k = 2$ gives $I_2 = \pi - 3$ and $L_2 = \frac{16}{105}$, and rearranging gives $$\frac{319}{105} < \pi < \frac{331}{105} .$$

| cite | improve this answer | |
  • 1
    $\begingroup$ Very nice work with the bounds. $\endgroup$ – Claude Leibovici Aug 29 '19 at 7:46
  • 1
    $\begingroup$ $K_j=\frac{\Gamma (2 j+1)^2}{\Gamma (4 j+2)}\sim \frac{\sqrt{\pi } 2^{-4 j-\frac{3}{2}}}{\sqrt{j}}$ seems to be interesting $\endgroup$ – Claude Leibovici Aug 29 '19 at 8:16

Probably not an answer but too long for the comment section.

You are considering $$I_k=\int_0^\infty \Big( \text{sech}(x) \tanh \left(\frac{x}{2}\right) \Big)^k \,dx$$ Let $x=2 \tanh ^{-1}(t)$ which makes $$J_k=\int \Big( \text{sech}(x) \tanh \left(\frac{x}{2}\right) \Big)^k \,dx=2 \int\left(1-t^2\right)^{k-1} \left(\frac{t}{1+t^2}\right)^k\,dt$$ that is to say $$J_k=2\frac{ t^{k+1}}{k+1}\, F_1\left(\frac{k+1}{2};1-k,k;\frac{k+3}{2};t^2,-t^2\right)$$ where appears the Appell hypergeometric function of two variables.

Integrating between $0$ and $1$, after simplification, this leads to $$I_k=\, _2\tilde{F}_1\left(k,\frac{k+1}{2};\frac{3 k+1}{2} ;-1\right)\, \Gamma (k)\, \Gamma \left(\frac{k+1}{2}\right)$$ where appears the regularized hypergeometric function.

As you noticed $I_{2k}=a_k-b_k \pi$ and $I_{2k+1}=c_k-d_k \log(2)$. So, for sure, if you make $I_k\sim \epsilon$, you have rational approximations of $\pi$ and $\log(2)$. The small problem I see is that they are not extremely accurate.

For example $$I_{20}=\frac{2357262305394688}{1119195}-\frac{21968591457761 \pi }{32768}\approx 1.7 \times 10^{-11}$$ would give as a rational approximation $$\pi \approx \frac{77242771223173136384}{24587137716568822395}=\color{red}{3.1415926535897932384}88023$$ while $$\pi \approx \frac{21053343141}{6701487259}=\color{red}{3.141592653589793238462}382$$ is better.

Similarly $$I_{21}= 4354393801 \log (2)-\frac{100374690765091043}{33256080}\approx 5.0 \times 10^{-12}$$ would give as a rational approximation $$\log(2)\approx \frac{100374690765091043}{144810068597560080}=\color{red}{0.69314718055994530941}60873$$ while $$\log(2)\approx \frac{34733068453}{50109225612}=\color{red}{0.693147180559945309417}8461$$ is better.

| cite | improve this answer | |

Too long for a comment.

We have the convenient relation $$\text{sech}(x)\tanh(x/2)=\frac{2}{e^x+1}-\frac{2}{e^{2x}+1},$$ so let $$L_n=I_{2n+1}=2^{2n+1}\sum_{k=0}^{2n+1}(-1)^k{2n+1\choose k}\int_0^{\infty}\frac{dx}{(e^x+1)^{2n-k+1}(e^{2x}+1)^k}$$ and $$P_n=I_{2n}=2^{2n}\sum_{k=0}^{2n}(-1)^k{2n\choose k}\int_0^{\infty}\frac{dx}{(e^x+1)^{2n-k}(e^{2x}+1)^k}.$$ These integrals may be easier to evaluate explicitly.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.