It goes back to Lagrange that a real quadratic irrational has always a periodic continued fraction. Starting from decades ago, several authors proposed different definitions of a $p$-adic continued fraction, and the definition depends on the chosen system of residues mod $p$. It turns out that the theory of p-adic continued fractions has many differences with respect to the real case; in particular, no analogue of Lagranges theorem holds, and the problem of deciding whether the continued fraction is periodic or not seemed to be not known until now. In recent work with F. Veneziano and U. Zannier we investigated the expansion of quadratic irrationals, for the $p$-adic continued fractions introduced by Ruban, giving an effective criterion to establish the possible periodicity of the expansion. This criterion, somewhat surprisingly, depends on the ‘real' value of the $p$-adic continued fraction.[-]

Let $E$ be an elliptic curve over a number field $K$. For each integer $n > 1$ the action of the absolute Galois group $G_K := Gal(\overline{K}/K)$ on the $n$-torsion subgroup $E [n]$ induces a Galois representation $\rho_{E,n}:G_K \rightarrow$ Aut$(E[n]) \backsimeq GL_2(\mathbb{Z} /n\mathbb{Z})$. The representations $\rho_{E,n}$ form a compatible system, and after taking inverse limits one obtains an adelic representation $\rho_E:G_K \rightarrow GL_2(\hat{\mathbb{Z}})$. If $E/K$ does not have $CM$, then Serre's open image theorem implies that the image of $\rho_E$ has finite index in $GL_2(\hat{\mathbb{Z}})$; in particular, $\rho_{E,\ell}$ is surjective for all but finitely many primes $\ell$.
I will present an algorithm that, given an elliptic curve $E/K$ without $CM$, determines the image of $\rho_{E,\ell}$ in $GL_2(\mathbb{Z} /\ell\mathbb{Z})$ up to local conjugacy for every prime $\ell$ for which $\rho_{E,\ell}$ is non-surjective. Assuming the generalized Riemann hypothesis, the algorithm runs in time that is polynomial in the bit-size of the coefficients of an integral Weierstrass model for $E$. I will then describe a probabilistic algorithm that uses this information to compute the index of $\rho_E$ in $GL_2(\hat{\mathbb{Z}})$.[-]

This talk will focus on the last step of the number field sive algorithm used to compute discrete logarithms in finite fields. We consider here non-prime finite fields of very small extension degree: $1 \le n \le 6$. These cases are interesting in pairing-based cryptography: the pairing output is an element in such a finite field. The discrete logarithm in that finite field must be hard enough to prevent from attacks in a given time (e.g. $10$ years). Within the CATREL project we aim to compute DL records in finite fields of moderate size (e.g. in $GF(p^n$) of global size from $600$ to $800$ bits) to estimate more tightly the hardness of DL in fields of cryptographic size ($2048$ bits at the moment). The best algorithm known to compute discrete logarithms in large finite fields (with small $n$) is the number field sieve (NFS):

(1) polynomial selection: select two distinct polynomials $f,g$ defining two number fields, such that they share modulo $p$ an irreducible degree $n$ factor, and have additional properties to improve the next two steps.
(2) sieving: sieve over elements that satisfy relations, to build the factor basis made of prime ideals of small norm.
(3) linear algebra: compute the kernel of a large matrix computed the step before. Then the logarithm of each element in the factor basis is known.
(4) individual logarithm: for a given element $s \in GF(p^n)$, decompose it over the factor basis to finally compute its discrete logarithm.

The most time consuming steps are the second and third: sieving and linear algerbra. After the sieve and the linear algebra, the logarithms of the prime ideals of small norm are known. To finally compute the discrete logarithm of the given element $s$, we lift $s$ in one of the number fields and factor it in prime ideals as with “small” elements in the sieve step. However here, $s$ does not have a small norm (bounded by $B \ll Q$). Its norm is very large, in particular, larger than $Q$. The usual way is to test for many $s' = s \cdot g^e$ with $g$ the given generator of $GF(p^n)$ until the norm of $s'$ is smooth enough. The time spent to find a good $e$ is asymptotically less than the sieving time. In practice, another modification of $s'$ is computed to reduce its norm. In [?], the authors write $s' = a(x) / b(x)$ with $a, b$ of coefficients of size $\sim p^{1/2}$ instead of $p$. With $n = 4$ the norm of $s$ is $O(p^{11/2})$. Their method compute $a,b$ of norm $O(p^{7/2})$. One need to factor into small prime ideals two elements $a,b$ instead of one $s'$.
for our record computations of discrete logarithms in $\mathbb{F}_pn$ with $2 \leqslant n \leqslant 6$, we improve the preparation of $s$, so that its norm in the number field is less than $Q$. This improves its smoothness property. Assume that we want to compute the discrete logarithm of $s$ in the larger subgroup of prime order $\ell$ of $GF(p^n)$, with $\ell$ $|$ $\Phi_np$. We decompose $s$ in $\epsilon \cdot s'$ with $\epsilon$ in a subfield or in a subgroup of order prime to $\ell$ and $s?$ with reduced coefficient size. We still have $log_g s = log_g s'$ mod $\ell$. We use a tower representation of $GF(p^n)$ with subfields for our purpose. We reduce the norm of $s \in \mathbb{F}_{p4}$ from $O(p^{11/2})$ to $O(p^{7/2}), s \in GF(p^3)$ from $O(p^6)$ to $O(p^2)$ and $s \in \mathbb{F}_{p2}$ from $O(p^4)$ to $O(p)$. This does not change the asymptotic complexity of this last step but this improves a lot its running time for small $n$.[-]