## L-functions

This section describes routines related to L-functions. We first introduce the basic concept and notations, then explain how to represent them in GP. Let Γ(s) = π-s/2Γ(s/2), where Γ is Euler's gamma function. Given d ≥ 1 and a d-tuple A = [α1,...,αd] of complex numbers, we let γA(s) = ∏α ∈ A Γ(s + α).

Given a sequence a = (an)n ≥ 1 of complex numbers (such that a1 = 1), a positive conductor N ∈ ℤ, and a gamma factor γA as above, we consider the Dirichlet series L(a,s) = ∑n ≥ 1 an n-s and the attached completed function Λ(a,s) = Ns/2γA(s).L(a,s).

Such a datum defines an L-function if it satisfies the three following assumptions:

* [Convergence] The an = O_ε(nk1) have polynomial growth, equivalently L(s) converges absolutely in some right half-plane Re(s) > k1 + 1.

* [Analytic continuation] L(s) has a meromorphic continuation to the whole complex plane with finitely many poles.

* [Functional equation] There exist an integer k, a complex number ε (usually of modulus 1), and an attached sequence a* defining both an L-function L(a*,s) satisfying the above two assumptions and a completed function Λ(a*,s) = Ns/2γA(s). L(a*,s), such that Λ(a,k-s) = ε Λ(a*,s) for all regular points.

More often than not in number theory we have a^ *= a (which forces |ε |= 1), but this needs not be the case. If a is a real sequence and a = a*, we say that L is self-dual. We do not assume that the an are multiplicative, nor equivalently that L(s) has an Euler product.

Remark. Of course, a determines the L-function, but the (redundant) datum a,a*, A, N, k, ε describes the situation in a form more suitable for fast computations; knowing the polar part r of Λ(s) (a rational function such that Λ-r is holomorphic) is also useful. A subset of these, including only finitely many an-values will still completely determine L (in suitable families), and we provide routines to try and compute missing invariants from whatever information is available.

Important Caveat. The implementation assumes that the implied constants in the O_ε are small. In our generic framework, it is impossible to return proven results without more detailed information about the L function. The intended use of the L-function package is not to prove theorems, but to experiment and formulate conjectures, so all numerical results should be taken with a grain of salt. One can always increase `realbitprecision` and recompute: the difference estimates the actual absolute error in the original output.

Note. The requested precision has a major impact on runtimes. Because of this, most L-function routines, in particular `lfun` itself, specify the requested precision in bits, not in decimal digits. This is transparent for the user once `realprecision` or `realbitprecision` are set. We advise to manipulate precision via `realbitprecision` as it allows finer granularity: `realprecision` increases by increments of 64 bits, i.e. 19 decimal digits at a time.

#### Theta functions

Given an L-function as above, we define an attached theta function via Mellin inversion: for any positive real t > 0, we let θ(a,t) := (1)/(2π i)∫Re(s) = c t-s Λ(s) ds where c is any positive real number c > k1+1 such that c + Re(a) > 0 for all a ∈ A. In fact, we have θ(a,t) = ∑n ≥ 1 an K(nt/N1/2)   where K(t) := (1)/(2π i)∫Re(s) = c t-s γA(s) ds. Note that this function is analytic and actually makes sense for complex t, such that Re(t2/d) > 0, i.e. in a cone containing the positive real half-line. The functional equation for Λ translates into θ(a,1/t) - ε t^kθ(a*,t) = P_Λ(t), where P_Λ is an explicit polynomial in t and log t given by the Taylor development of the polar part of Λ: there are no log's if all poles are simple, and P = 0 if Λ is entire. The values θ(t) are generally easier to compute than the L(s), and this functional equation provides a fast way to guess possible values for missing invariants in the L-function definition.

#### Data structures describing L and theta functions

We have 3 levels of description:

* an `Lmath` is an arbitrary description of the underlying mathematical situation (to which e.g., we associate the ap as traces of Frobenius elements); this is done via constructors to be described in the subsections below.

* an `Ldata` is a computational description of situation, containing the complete datum (a,a*,A,k,N,ε,r). Where a and a* describe the coefficients (given n,b we must be able to compute [a1,...,an] with bit accuracy b), A describes the Euler factor, the (classical) weight is k, N is the conductor, and r describes the polar part of L(s). This is obtained via the function `lfuncreate`. N.B. For motivic L-functions, the motivic weight w is w = k-1; but we also support non-motivic L-functions.

Design problem. All components of an `Ldata` should be given exactly since the accuracy to which they must be computed is not bounded a priori; but this is not always possible, in particular for ε and r.

* an `Linit` contains an `Ldata` and everything needed for fast numerical computations. It specifies the functions to be considered (either L(j)(s) or θ(j)(t) for derivatives of order j ≤ m, where m is now fixed) and specifies a domain which limits the range of arguments (t or s, respectively to certain cones and rectangular regions) and the output accuracy. This is obtained via the functions `lfuninit` or `lfunthetainit`.

All the functions which are specific to L or theta functions share the prefix `lfun`. They take as first argument either an `Lmath`, an `Ldata`, or an `Linit`. If a single value is to be computed, this makes no difference, but when many values are needed (e.g. for plots or when searching for zeros), one should first construct an `Linit` attached to the search range and use it in all subsequent calls. If you attempt to use an `Linit` outside the range for which it was initialized, a warning is issued, because the initialization is performed again, a major inefficiency:

```  ? Z = lfuncreate(1); \\ Riemann zeta
? L = lfuninit(Z, [1/2, 0, 100]); \\ zeta(1/2+it), |t| < 100
? lfun(L, 1/2)    \\ OK, within domain
%3 = -1.4603545088095868128894991525152980125
? lfun(L, 0)      \\ not on critical line !
*** lfun: Warning: lfuninit: insufficient initialization.
%4 = -0.50000000000000000000000000000000000000
? lfun(L, 1/2, 1) \\ attempt first derivative !
*** lfun: Warning: lfuninit: insufficient initialization.
%5 = -3.9226461392091517274715314467145995137
```

For many L-functions, passing from `Lmath` to an `Ldata` is inexpensive: in that case one may use `lfuninit` directly from the `Lmath` even when evaluations in different domains are needed. The above example could equally have skipped the `lfuncreate`:

```  ? L = lfuninit(1, [1/2, 0, 100]); \\ zeta(1/2+it), |t| < 100
```

In fact, when computing a single value, you can even skip `lfuninit`:

```  ? L = lfun(1, 1/2, 1); \\ zeta'(1/2)
? L = lfun(1, 1+x+O(x^5)); \\ first 5 terms of Taylor development at 1
```

Both give the desired results with no warning.

Complexity. The implementation requires O(N(|t|+1))1/2 coefficients an to evaluate L of conductor N at s = σ + i t.

We now describe the available high-level constructors, for built-in L functions.

#### Dirichlet L-functions

Given a Dirichlet character χ:(ℤ/Nℤ)* → ℂ, we let L(χ, s) = ∑n ≥ 1 χ(n) n-s. Only primitive characters are supported. Given a fundamental discriminant D, the function L((D/.), s), for the quadratic Kronecker symbol, is encoded by the `t_INT` D. This includes Riemann ζ function via the special case D = 1.

More general characters can be represented in a variety of ways:

* via Conrey notation (see `znconreychar`): χN(m,.) is given as the `t_INTMOD` `Mod(m,N)`.

* via a znstar structure describing the abelian group (ℤ/Nℤ)*, where the character is given in terms of the znstar generators:

```    ? G = znstar(100, 1); \\ (Z/100Z)*
? G.cyc \\ ~ Z/20 . g1  + Z/2 . g2 for some generators g1 and g2
%2 = [20, 2]
? G.gen
%3 = [77, 51]
? chi = [a, b]  \\ maps g1 to e(a/20) and g2 to e(b/2);  e(x) = exp(2ipi x)
```

More generally, let (ℤ/Nℤ)^ *= ⨁ (ℤ/diℤ) gi be given via a znstar structure G (`G.cyc` gives the di and `G.gen` the gi). A character χ on G is given by a row vector v = [a1,...,an] such that χ(∏ gini) = exp(2π i∑ ai ni / di). The pair [G, v] encodes the primitive character attached to χ.

* in fact, this construction [G, m] describing a character is more general: m is also allowed to be a Conrey index as seen above, or a Conrey logarithm (see `znconreylog`), and the latter format is actually the fastest one.

* it is also possible to view Dirichlet characters as Hecke characters over K = ℚ (see below), for a modulus [N, ] but this is both more complicated and less efficient.

In all cases, a non-primitive character is replaced by the attached primitive character.

#### Hecke L-functions

The Dedekind zeta function of a number field K = ℚ[X]/(T) is encoded either by the defining polynomial T, or any absolute number fields structure (preferably at least a bnf).

Given a finite order Hecke character χ: Clf(K) → ℂ, we let L(χ, s) = ∑A ⊂ OK χ(A) (NK/ℚA)-s.

Let Clf(K) = ⨁ (ℤ/diℤ) gi given by a bnr structure with generators: the di are given by `K.cyc` and the gi by `K.gen`. A character χ on the ray class group is given by a row vector v = [a1,...,an] such that χ(∏ gini) = exp(2π i∑ ai ni / di). The pair [bnr, v] encodes the primitive character attached to χ.

```  ? K  = bnfinit(x^2-60);
? Cf = bnrinit(K, [7, [1,1]], 1); \\ f = 7 oo1 oo2
? Cf.cyc
%3 = [6, 2, 2]
? Cf.gen
%4 = [[2, 1; 0, 1], [22, 9; 0, 1], [-6, 7]~]
? lfuncreate([Cf, [1,0,0]]); \\  χ(g1) = ζ6, χ(g2) = χ(g3) = 1
```

Dirichlet characters on (ℤ/Nℤ)* are a special case, where K = ℚ:

```  ? Q  = bnfinit(x);
? Cf = bnrinit(Q, [100, ]); \\ for odd characters on (Z/100Z)*
```

For even characters, replace by `bnrinit(K, N)`. Note that the simpler direct construction in the previous section will be more efficient.

#### Artin L functions

Given a Galois number field N/ℚ with group G = `galoisinit`(N), a representation ρ of G over the cyclotomic field ℚ(ζn) is specified by the matrices giving the images of `G.gen` by ρ. The corresponding Artin L function is created using `lfunartin`.

```     P = quadhilbert(-47); \\  degree 5, Galois group D5
N = nfinit(nfsplitting(P)); \\ Galois closure
G = galoisinit(N);
[s,t] = G.gen; \\ order 5 and 2
L = lfunartin(N,G, [[a,0;0,a^-1],[0,1;1,0]], 5); \\ irr. degree 2
```

In the above, the polynomial variable (here `a`) represents ζ5 := exp(2iπ/5) and the two matrices give the images of s and t. Here, priority of `a` must be lower than the priority of `x`.

#### L-functions of algebraic varieties

L-function of elliptic curves over number fields are supported.

```  ? E = ellinit([1,1]);
? L = lfuncreate(E);  \\ L-function of E/Q
? E2 = ellinit([1,a], nfinit(a^2-2));
? L2 = lfuncreate(E2);  \\ L-function of E/Q(sqrt(2))
```

L-function of hyperelliptic genus-2 curve can be created with `lfungenus2`. To create the L function of the curve y^2+(x^3+x^2+1)y = x^2+x:

```  ? L = lfungenus2([x^2+x, x^3+x^2+1]);
```

Currently, the model needs to be minimal at 2, and if the conductor is even, its valuation at 2 might be incorrect (a warning is issued).

#### Eta quotients / Modular forms

An eta quotient is created by applying `lfunetaquo` to a matrix with 2 columns [m, rm] representing f(τ) := ∏m η(mτ)rm. It is currently assumed that f is a self-dual cuspidal form on Γ0(N) for some N. For instance, the L-function ∑ τ(n) n-s attached to Ramanujan's Δ function is encoded as follows

```  ? L = lfunetaquo(Mat([1,24]));
? lfunan(L, 100)  \\ first 100 values of tau(n)
```

More general modular forms defined by modular symbols will be added later.

#### Low-level Ldata format

When no direct constructor is available, you can still input an L function directly by supplying [a, a*,A, k, N, ε, r] to `lfuncreate` (see `??lfuncreate` for details).

It is strongly suggested to first check consistency of the created L-function:

```  ? L = lfuncreate([a, as, A, k, N, eps, r]);
? lfuncheckfeq(L)  \\ check functional equation
```

#### lfun(L, s, {D = 0})

Compute the L-function value L(s), or if `D` is set, the derivative of order `D` at s. The parameter `L` is either an Lmath, an Ldata (created by `lfuncreate`, or an Linit (created by `lfuninit`), preferrably the latter if many values are to be computed.

The argument s is also allowed to be a power series; for instance, if s = α + x + O(x^n), the function returns the Taylor expansion of order n around α. The result is given with absolute error less than 2-B, where B = realbitprecision.

Caveat. The requested precision has a major impact on runtimes. It is advised to manipulate precision via `realbitprecision` as explained above instead of `realprecision` as the latter allows less granularity: `realprecision` increases by increments of 64 bits, i.e. 19 decimal digits at a time.

```  ? lfun(x^2+1, 2)  \\ Lmath: Dedekind zeta for Q(i) at 2
%1 = 1.5067030099229850308865650481820713960

? L = lfuncreate(ellinit("5077a1")); \\ Ldata: Hasse-Weil zeta function
? lfun(L, 1+x+O(x^4))  \\ zero of order 3 at the central point
%3 = 0.E-58 - 5.[...] E-40*x + 9.[...] E-40*x^2 + 1.7318[...]*x^3 + O(x^4)

\\ Linit: zeta(1/2+it), |t| < 100, and derivative
? L = lfuninit(1, , 1);
? T = lfunzeros(L, [1,25]);
%5 = [14.134725[...], 21.022039[...]]
? z = 1/2 + I*T;
? abs( lfun(L, z) )
%7 = 8.7066865533412207420780392991125136196 E-39
? abs( lfun(L, z, 1) )
%8 = 0.79316043335650611601389756527435211412  \\ simple zero
```

The library syntax is `GEN lfun0(GEN L, GEN s, long D, long bitprec)`.

#### lfunabelianrelinit(bnfL, bnfK, polrel, sdom, {der = 0})

Returns the `Linit` structure attached to the Dedekind zeta function of the number field L (see `lfuninit`), given a subfield K such that L/K is abelian. Here `polrel` defines L over K, as usual with the priority of the variable of `bnfK` lower than that of `polrel`. `sdom` and `der` are as in `lfuninit`.

```   ? D = -47; K = bnfinit(y^2-D);
? rel = quadhilbert(D); T = rnfequation(K.pol, rel); \\ degree 10
? L = lfunabelianrelinit(T,K,rel, [2,0,0]); \\ at 2
time = 84 ms.
? lfun(L, 2)
%4 = 1.0154213394402443929880666894468182650
? lfun(T, 2) \\ using parisize > 300MB
time = 652 ms.
%5 = 1.0154213394402443929880666894468182656
```

As the example shows, using the (abelian) relative structure is more efficient than a direct computation. The difference becomes drastic as the absolute degree increases while the subfield degree remains constant.

The library syntax is `GEN lfunabelianrelinit(GEN bnfL, GEN bnfK, GEN polrel, GEN sdom, long der, long bitprec)`.

#### lfunan(L, n)

Compute the first n terms of the Dirichlet series attached to the L-function given by `L` (`Lmath`, `Ldata` or `Linit`).

```   ? lfunan(1, 10)  \\ Riemann zeta
%1 = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
? lfunan(5, 10)  \\ Dirichlet L-function for kronecker(5,.)
%2 = [1, -1, -1, 1, 0, 1, -1, -1, 1, 0]
```

The library syntax is `GEN lfunan(GEN L, long n, long prec)`.

#### lfunartin(nf, gal, rho, n)

Returns the `Ldata` structure attached to the Artin L-function provided by the representation ρ of the Galois group of the extension K/ℚ, defined over the cyclotomic field ℚ(ζn), where nf is the nfinit structure attached to K, gal is the galoisinit structure attached to K/ℚ, and rho is given either

* by the values of its character on the conjugacy classes (see `galoisconjclasses` and `galoischartable`)

* or by the matrices that are the images of the generators `gal.gen`.

Cyclotomic numbers in `rho` are represented by polynomials, whose variable is understood as the complex number exp(2 i π/n).

In the following example we build the Artin L-functions attached to the two irreducible degree 2 representations of the dihedral group D10 defined over ℚ(ζ5), for the extension H/ℚ where H is the Hilbert class field of ℚ(sqrt{-47}). We show numerically some identities involving Dedekind ζ functions and Hecke L series.

```  ? P = quadhilbert(-47)
%1 = x^5 + 2*x^4 + 2*x^3 + x^2 - 1
? N = nfinit(nfsplitting(P));
? G = galoisinit(N); \\ D_10
? [T,n] = galoischartable(G);
? T  \\ columns give the irreducible characters
%5 =
[1  1              2              2]

[1 -1              0              0]

[1  1 -y^3 - y^2 - 1      y^3 + y^2]

[1  1      y^3 + y^2 -y^3 - y^2 - 1]
? n
%6 = 5
? L2 = lfunartin(N,G, T[,2], n);
? L3 = lfunartin(N,G, T[,3], n);
? L4 = lfunartin(N,G, T[,4], n);
? s = 1 + x + O(x^4);
? lfun(-47,s) - lfun(L2,s)
%11 ~ 0
? lfun(1,s)*lfun(-47,s)*lfun(L3,s)^2*lfun(L4,s)^2 - lfun(N,s)
%12 ~ 0
? lfun(1,s)*lfun(L3,s)*lfun(L4,s) - lfun(P,s)
%13 ~ 0
? bnr = bnrinit(bnfinit(x^2+47),1,1);
? bnr.cyc
%15 =  \\ Z/5Z: 4 non-trivial ray class characters
? lfun([bnr,], s) - lfun(L3, s)
%16 ~ 0
? lfun([bnr,], s) - lfun(L4, s)
%17 ~ 0
? lfun([bnr,], s) - lfun(L3, s)
%18 ~ 0
? lfun([bnr,], s) - lfun(L4, s)
%19 ~ 0
```

The first identity identifies the non-trivial abelian character with (-47,.); the second is the factorization of the regular representation of D10; the third is the factorization of the natural representation of D10 ⊂ S5; and the final four are the expressions of the degree 2 representations as induced from degree 1 representations.

The library syntax is `GEN lfunartin(GEN nf, GEN gal, GEN rho, long n, long bitprec)`.

#### lfuncheckfeq(L, {t})

Given the data attached to an L-function (`Lmath`, `Ldata` or `Linit`), check whether the functional equation is satisfied. This is most useful for an `Ldata` constructed "by hand", via `lfuncreate`, to detect mistakes.

If the function has poles, the polar part must be specified. The routine returns a bit accuracy b such that |w - w| < 2b, where w is the root number contained in `data`, and w is a computed value derived from θ(t) and θ(1/t) at some t != 0 and the assumed functional equation. Of course, a large negative value of the order of `realbitprecision` is expected.

If t is given, it should be close to the unit disc for efficiency and such that θ(t) != 0. We then check the functional equation at that t.

```  ? \pb 128       \\ 128 bits of accuracy
? default(realbitprecision)
%1 = 128
? L = lfuncreate(1);  \\ Riemann zeta
? lfuncheckfeq(L)
%3 = -124
```

i.e. the given data is consistent to within 4 bits for the particular check consisting of estimating the root number from all other given quantities. Checking away from the unit disc will either fail with a precision error, or give disappointing results (if θ(1/t) is large it will be computed with a large absolute error)

```  ? lfuncheckfeq(L, 2+I)
%4 = -115
? lfuncheckfeq(L,10)
***   at top-level: lfuncheckfeq(L,10)
***                 ^ —  —  —  —  —  —
*** lfuncheckfeq: precision too low in lfuncheckfeq.
```

The library syntax is `long lfuncheckfeq(GEN L, GEN t = NULL, long bitprec)`.

#### lfunconductor(L, {ab = [1, 10000]}, {flag = 0})

Compute the conductor of the given L-function (if the structure contains a conductor, it is ignored); `ab` = [a,b] is the interval where we expect to find the conductor; it may be given as a single scalar b, in which case we look in [1,b]. Increasing `ab` slows down the program but gives better accuracy for the result.

If `flag` is 0 (default), give either the conductor found as an integer, or a vector (possibly empty) of conductors found. If `flag` is 1, same but give the computed floating point approximations to the conductors found, without rounding to integers. It `flag` is 2, give all the conductors found, even those far from integers.

Caveat. This is a heuristic program and the result is not proven in any way:

```   ? L = lfuncreate(857); \\ Dirichlet L function for kronecker(857,.)
? \p19
realprecision = 19 significant digits
? lfunconductor(L)
%2 = [17, 857]
? lfunconductor(L,,1) \\ don't round
%3 = [16.99999999999999999, 857.0000000000000000]

? \p38
realprecision = 38 significant digits
? lfunconductor(L)
%4 = 857
```

Note. This program should only be used if the primes dividing the conductor are unknown, which is rare. If they are known, a direct search through possible prime exponents using `lfuncheckfeq` will be more efficient and rigorous:

```   ? E = ellinit([0,0,0,4,0]); /* Elliptic curve y^2 = x^3+4x */
? E.disc  \\ |disc E| = 2^12
%2 = -4096
\\ create Ldata by hand. Guess that root number is 1 and conductor N
? L(N) = lfuncreate([n->ellan(E,n), 0, [0,1], 2, N, 1]);
? fordiv(E.disc, d, print(d,": ",lfuncheckfeq(L(d))))
1: 0
2: 0
4: -1
8: -2
16: -3
32: -127
64: -3
128: -2
256: -2
512: -1
1024: -1
2048: 0
4096: 0
? lfunconductor(L(1)) \\ lfunconductor ignores conductor = 1 in Ldata !
%5 = 32
```

The above code assumed that root number was 1; had we set it to -1, none of the `lfuncheckfeq` values would have been acceptable:

```   ? L2(N) = lfuncreate([n->ellan(E,n), 0, [0,1], 2, N, -1]);
? [ lfuncheckfeq(L2(d)) | d<-divisors(E.disc) ]
%7 = [0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, -1, -1]
```

The library syntax is `GEN lfunconductor(GEN L, GEN ab = NULL, long 10000], long bitprec)`.

#### lfuncost(L, {sdom}, {der = 0})

Estimate the cost of running `lfuninit(L,sdom,der)` at current bit precision. Returns [t,b], to indicate that t coefficients an will be computed, as well as t values of `gammamellininv`, all at bit accuracy b. A subsequent call to `lfun` at s evaluates a polynomial of degree t at exp(h s) for some real parameter h, at the same bit accuracy b. If L is already an `Linit`, then sdom and der are ignored and are best left omitted; the bit accuracy is also inferred from L: in short we get an estimate of the cost of using that particular `Linit`.

```  ? \pb 128
? lfuncost(1, ) \\ for zeta(1/2+I*t), |t| < 100
%1 = [7, 242]  \\ 7 coefficients, 242 bits
? lfuncost(1, [1/2, 100]) \\ for zeta(s) in the critical strip, |Im s| < 100
%2 = [7, 246]  \\ now 246 bits
? lfuncost(1, , 10) \\ for zeta(1/2+I*t), |t| < 100
%3 = [8, 263]  \\ 10th derivative increases the cost by a small amount
? lfuncost(1, [10^5])
%3 = [158, 113438]  \\ larger imaginary part: huge accuracy increase

? L = lfuncreate(polcyclo(5)); \\ Dedekind zeta for Q(zeta5)
? lfuncost(L, ) \\ at s = 1/2+I*t), |t| < 100
%5 = [11457, 582]
? lfuncost(L, ) \\ twice higher
%6 = [36294, 1035]
? lfuncost(L, [10^4])  \\ much higher: very costly !
%7 = [70256473, 45452]
? \pb 256
? lfuncost(L, ); \\ doubling bit accuracy
%8 = [17080, 710]
```

In fact, some L functions can be factorized algebraically by the `lfuninit` call, e.g. the Dedekind zeta function of abelian fields, leading to much faster evaluations than the above upper bounds. In that case, the function returns a vector of costs as above for each individual function in the product actually evaluated:

```  ? L = lfuncreate(polcyclo(5)); \\ Dedekind zeta for Q(zeta5)
? lfuncost(L, )  \\ a priori cost
%2 = [11457, 582]
? L = lfuninit(L, ); \\ actually perform all initializations
? lfuncost(L)
%4 = [[16, 242], [16, 242], [7, 242]]
```

The Dedekind function of this abelian quartic field is the product of four Dirichlet L-functions attached to the trivial character, a non-trivial real character and two complex conjugate characters. The non-trivial characters happen to have the same conductor (hence same evaluation costs), and correspond to two evaluations only since the two conjugate characters are evaluated simultaneously. For a total of three L-functions evaluations, which explains the three components above. Note that the actual cost is much lower than the a priori cost in this case.

The library syntax is `GEN lfuncost0(GEN L, GEN sdom = NULL, long der, long bitprec)`. Also available is `GEN lfuncost(GEN L, GEN dom, long der, long bitprec)` when L is not an `Linit`; the return value is a `t_VECSMALL` in this case.

#### lfuncreate(obj)

This low-level routine creates `Ldata` structures, needed by lfun functions, describing an L-function and its functional equation. You are urged to use a high-level constructor when one is available, and this function accepts them, see `??lfun`:

```  ? L = lfuncreate(1); \\ Riemann zeta
? L = lfuncreate(5); \\ Dirichlet L-function for quadratic character (5/.)
? L = lfuncreate(x^2+1); \\ Dedekind zeta for Q(i)
? L = lfuncreate(ellinit([0,1])); \\ L-function of E/Q: y^2=x^3+1
```

One can then use, e.g., `Lfun(L,s)` to directly evaluate the respective L-functions at s, or `lfuninit(L, [c,w,h]` to initialize computations in the rectangular box Re(s-c) ≤ w, Im(s) ≤ h.

We now describe the low-level interface, used to input non-builtin L-functions. The input is now a 6 or 7 component vector V = [a, astar, Vga, k, N, eps, poles], whose components are as follows:

* `V = a` encodes the Dirichlet series coefficients (an). The preferred format is a closure of arity 1: `n- > vector(n,i,a(i))` giving the vector of the first n coefficients. The closure is allowed to return a vector of more than n coefficients (only the first n will be considered) or even less than n, in which case loss of accuracy will occur and a warning that `#an` is less than expected is issued. This allows to precompute and store a fixed large number of Dirichlet coefficients in a vector v and use the closure `n- > v`, which does not depend on n. As a shorthand for this latter case, you can input the vector v itself instead of the closure.

```  ? z = lfuncreate([n->vector(n,i,1), 1, , 1, 1, 1, 1]); \\ Riemann zeta
? lfun(z,2) - Pi^2/6
%2 = -5.877471754111437540 E-39
```

A second format is limited to L-functions affording an Euler product. It is a closure of arity 2 `(p,d)- > F(p)` giving the local factor Lp(X) at p as a rational function, to be evaluated at p-s as in `direuler`; d is set to `logint`(n,p) + 1, where n is the total number of Dirichlet coefficients (a1,...,an) that will be computed. In other words, the smallest integer d such that p^d > n. This parameter d allows to compute only part of Lp when p is large and Lp expensive to compute: any polynomial (or `t_SER`) congruent to Lp modulo X^d is acceptable since only the coefficients of X^0,..., Xd-1 are needed to expand the Dirichlet series. The closure can of course ignore this parameter:

```  ? z = lfuncreate([(p,d)->1/(1-x), 1, , 1, 1, 1, 1]); \\ Riemann zeta
? lfun(z,2) - Pi^2/6
%4 = -5.877471754111437540 E-39
```

One can describe separately the generic local factors coefficients and the bad local factors by setting `dir` = [F, Lbad], were Lbad = [[p1,Lp1],...,[pk,Lpk]], where F describes the generic local factors as above, except that when p = pi for some i ≤ k, the coefficient ap is directly set to Lpi instead of calling F.

```  N = 15;
E = ellinit([1, 1, 1, -10, -10]); \\ = "15a1"
F(p,d) = 1 / (1 - ellap(E,p)*'x + p*'x^2);
Lbad = [[3, 1/(1+'x)], [5, 1/(1-'x)]];
L = lfuncreate([[F,Lbad], 0, [0,1], 2, N, ellrootno(E)]);
```

Of course, in this case, `lfuncreate(E)` is preferable!

* `V = astar` is the Dirichlet series coefficients of the dual function, encoded as `a` above. The sentinel values 0 and 1 may be used for the special cases where a = a* and a = a*, respectively.

* `V = Vga` is the vector of αj such that the gamma factor of the L-function is equal to γA(s) = ∏1 ≤ j ≤ dΓ(s+αj), where Γ(s) = π-s/2Γ(s/2). This same syntax is used in the `gammamellininv` functions. In particular the length d of `Vga` is the degree of the L-function. In the present implementation, the αj are assumed to be exact rational numbers. However when calling theta functions with complex (as opposed to real) arguments, determination problems occur which may give wrong results when the αj are not integral.

* `V = k` is a positive integer k. The functional equation relates values at s and k-s. For instance, for an Artin L-series such as a Dedekind zeta function we have k = 1, for an elliptic curve k = 2, and for a modular form, k is its weight. For motivic L-functions, the motivic weight w is w = k-1.

By default we assume that an = O_ε(nk1), where k1 = w and even k1 = w/2 when the L function has no pole (Ramanujan-Petersson). If this is not the case, you can replace the k argument by a vector [k,k1], where k1 is the upper bound you can assume.

* `V = N` is the conductor, an integer N ≥ 1, such that Λ(s) = Ns/2γA(s)L(s) with γA(s) as above.

* `V = eps` is the root number ϵ, i.e., the complex number (usually of modulus 1) such that Λ(a, k-s) = ϵ Λ(a*, s).

* The last optional component `V = poles` encodes the poles of the L or Λ-functions, and is omitted if they have no poles. A polar part is given by a list of 2-component vectors [β,Pβ(x)], where β is a pole and the power series Pβ(x) describes the attached polar part, such that L(s) - P_β(s-β) is holomorphic in a neighbourhood of β. For instance P_β = r/x+O(1) for a simple pole at β or r1/x^2+r2/x+O(1) for a double pole. The type of the list describing the polar part allows to distinguish between L and Λ: a `t_VEC` is attached to L, and a `t_COL` is attached to Λ. Unless a = a* (coded by `astar` equal to 0 or 1), it is mandatory to specify the polar part of Λ rather than those of L since the poles of L* cannot be infered from the latter ! Whereas the functional equation allows to deduce the polar part of Λ* from the polar part of Λ.

Finally, if a = a*, we allow a shortcut to describe the frequent situation where L has at most simple pole, at s = k, with residue r a complex scalar: you may then input `poles` = r. This value r can be set to 0 if unknown and it will be computed.

The library syntax is `GEN lfuncreate(GEN obj)`.

#### lfundiv(L1, L2)

Creates the `Ldata` structure (without initialization) corresponding to the quotient of the Dirichlet series L1 and L2 given by `L1` and `L2`. Assume that vz(L1) ≥ vz(L2) at all complex numbers z: the construction may not create new poles, nor increase the order of existing ones.

The library syntax is `GEN lfundiv(GEN L1, GEN L2, long bitprec)`.

#### lfunetaquo(M)

Returns the `Ldata` structure attached to the L function attached to the modular form z`: — >`i = 1^n η(Mi,1 z)Mi,2 It is currently assumed that f is a self-dual cuspidal form on Γ0(N) for some N. For instance, the L-function ∑ τ(n) n-s attached to Ramanujan's Δ function is encoded as follows

```  ? L = lfunetaquo(Mat([1,24]));
? lfunan(L, 100)  \\ first 100 values of tau(n)
```

For convenience, a `t_VEC` is also accepted instead of a factorization matrix with a single row:

```  ? L = lfunetaquo([1,24]); \\ same as above
```

The library syntax is `GEN lfunetaquo(GEN M)`.

#### lfungenus2(F)

Returns the `Ldata` structure attached to the L function attached to the genus-2 curve defined by y^2 = F(x) or y^2+Q(x) y = P(x) if F = [P,Q]. Currently, the model needs to be minimal at 2, and if the conductor is even, its valuation at 2 might be incorrect (a warning is issued).

The library syntax is `GEN lfungenus2(GEN F)`.

#### lfunhardy(L, t)

Variant of the Hardy Z-function given by `L`, used for plotting or locating zeros of L(k/2+it) on the critical line. The precise definition is as follows: if as usual k/2 is the center of the critical strip, d is the degree, αj the entries of `Vga` giving the gamma factors, and ϵ the root number, then if we set s = k/2+it = ρ e and E = (d(k/2-1)+∑1 ≤ j ≤ dαj)/2, the computed function at t is equal to Z(t) = ϵ-1/2Λ(s).|s|-Eedtθ/2 , which is a real function of t for self-dual Λ, vanishing exactly when L(k/2+it) does on the critical line. The normalizing factor |s|-Eedtθ/2 compensates the exponential decrease of γA(s) as t → oo so that Z(t) ~ 1.

```  ? T = 100; \\ maximal height
? L = lfuninit(1, [T]); \\ initialize for zeta(1/2+it), |t|<T
? \p19 \\ no need for large accuracy
? ploth(t = 0, T, lfunhardy(L,t))
```

Using `lfuninit` is critical for this particular applications since thousands of values are computed. Make sure to initialize up to the maximal t needed: otherwise expect to see many warnings for unsufficient initialization and suffer major slowdowns.

The library syntax is `GEN lfunhardy(GEN L, GEN t, long bitprec)`.

#### lfuninit(L, sdom, {der = 0})

Initalization function for all functions linked to the computation of the L-function L(s) encoded by `L`, where s belongs to the rectangular domain `sdom` = [center,w,h] centered on the real axis, |Re(s)-center| ≤ w, |Im(s)| ≤ h, where all three components of `sdom` are real and w, h are non-negative. `der` is the maximum order of derivation that will be used. The subdomain [k/2, 0, h] on the critical line (up to height h) can be encoded as [h] for brevity. The subdomain [k/2, w, h] centered on the critical line can be encoded as [w, h] for brevity.

The argument `L` is an `Lmath`, an `Ldata` or an `Linit`. See `??Ldata` and `??lfuncreate` for how to create it.

The height h of the domain is a crucial parameter: if you only need L(s) for real s, set h to 0. The running time is roughly proportional to (B / d+π h/4)d/2+3N1/2, where B is the default bit accuracy, d is the degree of the L-function, and N is the conductor (the exponent d/2+3 is reduced to d/2+2 when d = 1 and d = 2). There is also a dependency on w, which is less crucial, but make sure to use the smallest rectangular domain that you need.

```  ? L0 = lfuncreate(1); \\ Riemann zeta
? L = lfuninit(L0, [1/2, 0, 100]); \\ for zeta(1/2+it), |t| < 100
? lfun(L, 1/2 + I)
? L = lfuninit(L0, ); \\ same as above !
```

The library syntax is `GEN lfuninit0(GEN L, GEN sdom, long der, long bitprec)`.

#### lfunlambda(L, s, {D = 0})

Compute the completed L-function Λ(s) = Ns/2γ(s)L(s), or if `D` is set, the derivative of order `D` at s. The parameter `L` is either an `Lmath`, an `Ldata` (created by `lfuncreate`, or an `Linit` (created by `lfuninit`), preferrably the latter if many values are to be computed.

The result is given with absolute error less than 2-B|γ(s)Ns/2|, where B = realbitprecision.

The library syntax is `GEN lfunlambda0(GEN L, GEN s, long D, long bitprec)`.

#### lfunmfspec(L)

Returns `[valeven,valodd,omminus,omplus]`, where `valeven` (resp., `valodd`) is the vector of even (resp., odd) periods of the modular form given by `L`, and `omminus` and `omplus` the corresponding real numbers ω^- and ω^+ normalized in a noncanonical way. For the moment, only for modular forms of even weight.

The library syntax is `GEN lfunmfspec(GEN L, long bitprec)`.

#### lfunmul(L1, L2)

Creates the `Ldata` structure (without initialization) corresponding to the product of the Dirichlet series given by `L1` and `L2`.

The library syntax is `GEN lfunmul(GEN L1, GEN L2, long bitprec)`.

#### lfunorderzero(L, {m = -1})

Computes the order of the possible zero of the L-function at the center k/2 of the critical strip; return 0 if L(k/2) does not vanish.

If m is given and has a non-negative value, assumes the order is at most m. Otherwise, the algorithm chooses a sensible default:

* if the L argument is an `Linit`, assume that a multiple zero at s = k / 2 has order less than or equal to the maximal allowed derivation order.

* else assume the order is less than 4.

You may explicitly increase this value using optional argument m; this overrides the default value above. (Possibly forcing a recomputation of the `Linit`.)

The library syntax is `long lfunorderzero(GEN L, long m, long bitprec)`.

#### lfunqf(Q)

Returns the `Ldata` structure attached to the Θ function of the lattice attached to the definite positive quadratic form Q.

```  ? L = lfunqf(matid(2));
? lfunqf(L,2)
%2 = 6.0268120396919401235462601927282855839
? lfun(x^2+1,2)*4
%3 = 6.0268120396919401235462601927282855839
```

The library syntax is `GEN lfunqf(GEN Q, long prec)`.

#### lfunrootres(data)

Given the `Ldata` attached to an L-function (or the output of `lfunthetainit`), compute the root number and the residues.

The output is a 3-component vector [[[a1,r1],...,[an, rn], [[b1, R1],...,[bm, Rm]], w], where ri is the polar part of L(s) at ai, Ri is is the polar part of Λ(s) at bi or [0,0,r] if there is no pole, and w is the root number. In the present implementation,

* either the polar part must be completely known (and is then arbitrary): the function determines the root number,

```  ? L = lfunmul(1,1); \\ zeta^2
? [r,R,w] = lfunrootres(L);
? r  \\ single pole at 1, double
%3 = [[1, 1.[...]*x^-2 + 1.1544[...]*x^-1 + O(x^0)]]
? w
%4 = 1
? R \\ double pole at 0 and 1
%5 = [[1,[...]], [0,[...]]
```

* or at most a single pole is allowed: the function computes both the root number and the residue (0 if no pole).

The library syntax is `GEN lfunrootres(GEN data, long bitprec)`.

#### lfunsympow(E, m)

Returns the `Ldata` structure attached to the L function attached to the m-th symmetric power of the elliptic curve E defined over the rationals.

The library syntax is `GEN lfunsympow(GEN E, ulong m)`.

#### lfuntheta(data, t, {m = 0})

Compute the value of the m-th derivative at t of the theta function attached to the L-function given by `data`. `data` can be either the standard L-function data, or the output of `lfunthetainit`. The result is given with absolute error less than 2-B, where B = realbitprecision.

The theta function is defined by the formula Θ(t) = ∑n ≥ 1a(n)K(nt/sqrt(N)), where a(n) are the coefficients of the Dirichlet series, N is the conductor, and K is the inverse Mellin transform of the gamma product defined by the `Vga` component. Its Mellin transform is equal to Λ(s)-P(s), where Λ(s) is the completed L-function and the rational function P(s) its polar part. In particular, if the L-function is the L-function of a modular form f(τ) = ∑n ≥ 0a(n)q^n with q = exp(2π iτ), we have Θ(t) = 2(f(it/sqrt{N})-a(0)). Note that a(0) = -L(f,0) in this case.

The library syntax is `GEN lfuntheta(GEN data, GEN t, long m, long bitprec)`.

#### lfunthetacost(L, {tdom}, {m = 0})

This function estimates the cost of running `lfunthetainit(L,tdom,m)` at current bit precision. Returns the number of coefficients an that would be computed. This also estimates the cost of a subsequent evaluation `lfuntheta`, which must compute that many values of `gammamellininv` at the current bit precision. If L is already an `Linit`, then tdom and m are ignored and are best left omitted: we get an estimate of the cost of using that particular `Linit`.

```  ? \pb 1000
? L = lfuncreate(1); \\ Riemann zeta
? lfunthetacost(L); \\ cost for theta(t), t real >= 1
%1 = 15
? lfunthetacost(L, 1 + I); \\ cost for theta(1+I). Domain error !
***   at top-level: lfunthetacost(1,1+I)
***                 ^ —  —  —  —  —  — --
*** lfunthetacost: domain error in lfunthetaneed: arg t > 0.785
? lfunthetacost(L, 1 + I/2) \\ for theta(1+I/2).
%2 = 23
? lfunthetacost(L, 1 + I/2, 10) \\ for theta^((10))(1+I/2).
%3 = 24
? lfunthetacost(L, [2, 1/10]) \\ cost for theta(t), |t| >= 2, |arg(t)| < 1/10
%4 = 8

? L = lfuncreate( ellinit([1,1]) );
? lfunthetacost(L)  \\ for t >= 1
%6 = 2471
```

The library syntax is `long lfunthetacost0(GEN L, GEN tdom = NULL, long m, long bitprec)`.

#### lfunthetainit(L, {tdom}, {m = 0})

Initalization function for evaluating the m-th derivative of theta functions with argument t in domain tdom. By default (tdom omitted), t is real, t ≥ 1. Otherwise, tdom may be

* a positive real scalar ρ: t is real, t ≥ ρ.

* a non-real complex number: compute at this particular t; this allows to compute θ(z) for any complex z satisfying |z| ≥ |t| and |arg z| ≤ |arg t|; we must have |2 arg z / d| < π/2, where d is the degree of the Γ factor.

* a pair [ρ,α]: assume that |t| ≥ ρ and |arg t| ≤ α; we must have |2α / d| < π/2, where d is the degree of the Γ factor.

```  ? \p500
? L = lfuncreate(1); \\ Riemann zeta
? t = 1+I/2;
? lfuntheta(L, t); \\ direct computation
time = 30 ms.
? T = lfunthetainit(L, 1+I/2);
time = 30 ms.
? lfuntheta(T, t); \\ instantaneous
```

The T structure would allow to quickly compute θ(z) for any z in the cone delimited by t as explained above. On the other hand

```  ? lfuntheta(T,I)
***   at top-level: lfuntheta(T,I)
***                 ^ —  —  —  — --
*** lfuntheta: domain error in lfunthetaneed: arg t > 0.785398163397448
```

The initialization is equivalent to

```  ? lfunthetainit(L, [abs(t), arg(t)])
```

The library syntax is `GEN lfunthetainit(GEN L, GEN tdom = NULL, long m, long bitprec)`.

#### lfuntwist(L, chi)

Creates the Ldata structure (without initialization) corresponding to the twist of L by the primitive character attached to the Dirichlet character `chi`. The conductor of the character must be coprime to the conductor of the L-function L.

The library syntax is `GEN lfuntwist(GEN L, GEN chi)`.

#### lfunzeros(L, lim, {divz = 8})

`lim` being either a positive upper limit or a non-empty real interval inside [0,+ oo [, computes an ordered list of zeros of L(s) on the critical line up to the given upper limit or in the given interval. Use a naive algorithm which may miss some zeros: it assumes that two consecutive zeros at height T ≥ 1 differ at least by 2π/ω, where ω := `divz`.(dlog(T/2π) +d+ 2log(N/(π/2)^d)). To use a finer search mesh, set divz to some integral value larger than the default ( = 8).

```  ? lfunzeros(1, 30) \\ zeros of Rieman zeta up to height 30
%1 = [14.134[...], 21.022[...], 25.010[...]]
? #lfunzeros(1, [100,110])  \\ count zeros with 100 <= Im(s) <= 110
%2 = 4
```

The algorithm also assumes that all zeros are simple except possibly on the real axis at s = k/2 and that there are no poles in the search interval. (The possible zero at s = k/2 is repeated according to its multiplicity.)

Should you pass an `Linit` argument to the function, beware that the algorithm needs at least

```     L = lfuninit(Ldata, [T+1])
```

where T is the upper bound of the interval defined by `lim`: this allows to detect zeros near T. Make sure that your `Linit` domain contains this one, i.e. a domain [1,T+1] is fine but [0, T] is not! The algorithm assumes that a multiple zero at s = k / 2 has order less than or equal to the maximal derivation order allowed by the `Linit`. You may increase that value in the `Linit` but this is costly: only do it for zeros of low height or in `lfunorderzero` instead.

The library syntax is `GEN lfunzeros(GEN L, GEN lim, long divz, long bitprec)`.