One-ninth constant and Varga's constant for rational approximation on $[0,\infty)$

edit · history · discussion · files · short url · analysis special values
Numbers
$form$ 
$\Lambda$ or $V$
$\Lambda$:
0.1076539192264845766153234450909471905879765632901150866985681469819243414626426434127761990409158732
comment: The one-ninth constant, also called the Halphen constant.
$V$:
9.289025491920818918755449435951745061031694867750124408239700614217293752472865070705241587061424714
comment: Varga's constant, the reciprocal $1/\Lambda$.
Definition
The one-ninth constant $\Lambda$ is the limiting $n$th root of the least uniform error in real rational approximation of $e^{-x}$ on $[0,\infty)$ with numerator and denominator degrees at most $n$ [3] [2]. Varga's constant is $V=1/\Lambda$.
Parameters
$form$
—   reciprocal convention ($\Lambda$ or $V=1/\Lambda$)
Formulas
(1)
$\displaystyle E_n=\inf_{\substack{p,q\in\mathbb R[x],\ \deg p,\deg q\leq n\\q(x)\ne0\ (x\geq0)}}\sup_{x\geq0}|e^{-x}-p(x)/q(x)|$, and $\displaystyle \Lambda=\lim_{n\to\infty}E_n^{1/n}$ [1].
(2)
Let $K(m)$ and $E(m)$ denote the complete elliptic integral of the first kind and the complete elliptic integral of the second kind in the parameter convention $m=k^2$. There is a unique $m_0\in(0,1)$ satisfying $K(m_0)=2E(m_0)$, and $\Lambda=\exp\!\left(-\pi\,K(1-m_0)/K(m_0)\right)$ is the elliptic nome at $m_0$ [1] [2].
(3)
$\Lambda$ is the unique root in $(0,1)$ of $\displaystyle G(x)=\sum_{k=0}^{\infty}(2k+1)^2(-x)^{k(k+1)/2}=0$ [1].
(4)
$\displaystyle \sum_{j=1}^{\infty}a_j\Lambda^j=\frac18$, where $\displaystyle a_j=\left|\sum_{d\mid j}(-1)^d d\right|$ and the inner sum is over the positive divisors of $j$ [1].
(5)
$\Lambda V=1$.
Comments
(6)
The two entries are reciprocal conventions of one constant: either determines the other. The notation $V$ follows [4]; $\Lambda$ is also called the Halphen constant [2].
(7)
The name refers to the conjecture $\Lambda=1/9$; in fact, $1/10<\Lambda<1/9$, so $9<V<10$ [3].
(8)
The series characterisations appear in equations (30)–(32) of [7]; its section 4 relates them to the elliptic-integral characterisation.
Programs
(P1)
Python
from mpmath import mp
with mp.workdps(140):
    m = mp.findroot(lambda t: mp.ellipk(t)-2*mp.ellipe(t),
                    (mp.mpf(4)/5, mp.mpf(9)/10))
    lam = mp.exp(-mp.pi*mp.ellipk(1-m)/mp.ellipk(m))
    print(mp.nstr(lam, 100))
    print(mp.nstr(1/lam, 100))
References
[1]
Alphonse P. Magnus and Jean Meinguet, The elliptic functions and integrals of the '1/9' problem, Numerical Algorithms 24 (2000), pp. 117–139. (doi)
[2]
Steven R. Finch, Mathematical Constants, Cambridge University Press, 2003, section 4.5, pp. 259–262, especially p. 261.
Links
Similar tables
Values of the elliptic nome —   $\Lambda$ is the nome at the unique parameter satisfying $K(m)=2E(m)$
Complete elliptic integral of the first kind —   $K(m_0)$ and $K(1-m_0)$ determine the exponential convergence rate
Complete elliptic integral of the second kind —   $E(m_0)$ determines the parameter through $K(m_0)=2E(m_0)$
Lebesgue constants of the Fourier partial sums —   operator norms in periodic approximation, another classical approximation problem
Data properties
Entries are of type: real number
digits: 100 (Unknown key)
Table is complete: yes (it holds the constant and its reciprocal, the two conventions specified here)
How they were obtained:

Halphen's series is evaluated with Arb at 429 bits, including 96 guard bits beyond the requested 100 significant digits. Rational bisection starts in $[1/10,1/9]$ and retains opposite certified endpoint signs until the bracket width is less than $2^{-365}$. If the terms with $k\geq N$ are omitted, their absolute sum is bounded by $\frac98(2N+1)^2 9^{-N(N+1)/2}$ on this bracket. The final bracket is returned as a real ball, and ball division gives its reciprocal. The computed radii are less than $4.8\times10^{-111}$ for $\Lambda$ and $4.1\times10^{-109}$ for $V$.

more

Independent checks use Arb's complete elliptic integrals at 640 bits to bracket the root of $K(m)-2E(m)$ and evaluate its nome. Controls use the values at $m=0$ and Legendre's relation at $m=1/2$. The divisor-sum series is evaluated independently through $j=240$, with the omitted terms bounded by $\sum_{j\geq241}j^2(1/8)^j$; it gives opposite signs relative to $1/8$ at both endpoints of each stored decimal interval, after inversion for $V$. These checks also test the reciprocal relation and the stated inequalities. The results match all 105 displayed significant digits in [5] and all 101 in [6], and a 150-digit rerun agrees. The attached checks repeat the interval tests on the values read from the database.