# Learning to SQI

Maria Corte-Real Santos, Giacomo Pope

# Between Kernels and Ideals

Central to constructive applications of the Deuring correspondence is efficiently converting between equivalent ideals and isogenies. The core of this work is converting between a cyclic ideal $I$ of norm $n(I) = D$ and a kernel generator $K$ of order $D$. This page is dedicated to discussing how to perform these conversions and follows the description of $\textsf{IdealToKernel}$ and $\textsf{KernelToIdeal}$, described by Algorithms 19 and 20 of Leroux’s Thesis.

Although we have attempted to make this page self-contained, we point an interested reader to both the first explicit formulation in [GPS17], and the clear and well motivated discussions from Leroux’s Thesis: Section 4.2, as well as the recent publication [EPSV23].

## Evaluating endomorphisms

Central to converting between kernel generators and ideals is being able to compute the isomorphism $\xi : \End(E) \to \OO$ and subsequently using this to evaluate the action of some algebra element $\alpha \in \BB$ on points $P \in E$ via the mapping $\xi^{-1}(\alpha)(P)$.

The construction of the isomorphism comes from our knowledge of the basis of the maximal order $\OO_0$ and the endomorphism ring of the special curve $E_0$:

\begin{aligned} \OO_0 &= \left\langle 1, i, \frac{i + j}{2}, \frac{1 + k}{2} \right\rangle, \\ \End(E_0) &= \left\langle 1, \iota, \frac{\iota + \pi}{2}, \frac{1 + \iota \pi}{2} \right\rangle, \\ \end{aligned}

where $\BB = \langle 1, i, j, k \rangle$, and $\iota$ and $\pi$ are the twisting and Frobenius endomorphisms respectively:

$$\iota : (x,y) \mapsto (-x, \sqrt{-1} y), \qquad \pi : (x,y) \mapsto (x^p,y^p).$$

Concretely, these endomorphisms can be evaluated on any point $P \in E_0$ with the following functions:

# Global values
p = E0.base().characteristic()
sqrt_minus_one = E0.base()(-1).sqrt(all=True)[0] # Deterministic

def E01(P):
"""
Identity map, does nothing
"""
return P

def E0ι(P):
"""
Returns ι(P) = (-x, √-1 y)
"""
return E0(-P[0], sqrt_minus_one * P[1])

def E0π(P):
"""
Returns π(P) = (X^p, Y^p, Z^p)
"""
return E0(P[0] ** p, P[1] ** p, P[2] ** p)

def E0ιπ(P):
"""
Returns ιπ(P) = (-X^p, √-1 Y^p, Z^p)
"""
return E0(-P[0] ** p, sqrt_minus_one * P[1] ** p, P[2] ** p)


With the endomorphisms implemented, we now need to look at how we can map algebra elements to endomorphisms. Generically, any element $\alpha \in \BB$ can be written in the following way:

$$\alpha = a_0 + a_1 i + a_2 j + a_3 k, \quad a_i \in \QQ.$$

However, we will be working with integral, cyclic ideals $I$ with left order $\OO_0$, and we will be evaluating either basis elements of $\OO_0$ or the action of generators of the ideal $I = \OO_0\langle \alpha, D \rangle$. For these cases, our elements can always be written as:

$$\alpha = \frac{a_0 + a_1 i + a_2 j + a_3 k}{d}, \quad a_i \in \ZZ, \quad d \in \{1, 2\}.$$

We can then compute the isomorphism $\xi : (1, i, j, k) \mapsto ([1], \iota, \pi, \iota \circ \pi)$, allowing us to express the endomorphism $\theta_\alpha = \xi^{-1}(\alpha) \in \End(E_0)$:

$$[d] \circ \theta_\alpha = [a_0] + [a_1] \circ \iota + [a_2] \circ \pi + [a_3] \circ \iota \circ \pi ,$$

where $[m]$ denotes the endomorphism of scalar multiplication by $m$.

When $d = 1$, we can simply evaluate the right-hand side to obtain the action of $\theta_\alpha$ on some point $P \in E_0$. For the case of $d = 2$, we must first compute some $Q$ such that $P= [2]Q$ such that $\theta_\alpha(P) = [2] \circ \theta_\alpha (Q)$.

SageMath allows us to easily compute all $Q$ such that $P = [2]Q$ with the function P.division_points(2). We only need one such point, so we can simply compute Q = P.division_points(2)[0].

Note: for SQISign, it is important that we can compute such a $Q$ without needing to perform a field extension to allow the division. This requires that $P$ cannot have maximal even torsion. Understanding this is particularly important when we are taking step sizes in IdealToIsogenyFromKLPT() by computing isogenies from ideals of norm $2^\text{step}$. The fact we may need to divide by two during this conversion means that if the maximal available torsion is $2^f$, then we are restricted to pick $\text{step} \leq f - 1$.

### Putting this all together

With all the above discussion, we have a fairly friendly looking implementation of evaluating the action of $\xi^{-1}(\alpha)$ on a point $P$ of order $D$:

def eval_endomorphism(α, P, D):
"""
Evaluates the action of an endomorphism
f ∈ End(E0) on a point P ∈ E.
"""
# Basis for the endomorphisms, functions as above
EndE0 = [E01, E0ι, E0π, E0ιπ]

# Unpack the coefficients of the generator α such that
# α = (w + x*i + y*j + z*k) / d for d,w,x,y,z in ZZ
d, *α_coeffs = α.denominator_and_integer_coefficient_tuple()

# For generators of integral ideals, we expect the denominator
# to be at most 2
assert d in (1, 2), "Something is wrong with the input ideal"
if d == 2:
# Divide out by two before evaluation if needed
P = P.division_points(d)[0]

# Compute the image of α(P)
# α(P) = [α0] * P + [α1] * ι(P) + [α2] * π(P) + [α3] * ι(π(P))
αP = sum(ai * θ(P) for ai, θ in zip(α_coeffs, EndE0))

return αP


### Pick a curve

Technically, we can work with any curve $E$ providing that we know the connecting isogeny $\phi : E_0 \to E$, and in our implementation deuring.py, we have allowed for this with an optional parameter connecting_isogenies which is the tuple $(\phi, \widehat\phi)$. However, in SQISign, everything is mapped back to $E_0$ or $\OO_0$ before converting, so to allow for a clearer discussion we have simply assumed that our ideals $I$ have left order $\OO_0$ and our kernel generators are points $P \in E_0$.

Essentially, the trick for working with generic curves is to take $P \in E$ and compute the image $\widehat\phi(P) \in E_0$ to use in the above function. Then, after we find $P_\alpha = \theta_\alpha \circ \widehat\phi(P) \in E_0$, we return $\phi(P_\alpha) \in E$.

Note: The composition $\phi \circ \widehat\phi = [D]$ where $D = \deg(\phi)$. So in the above, we have computed:

$$[D] \circ \xi^{-1}(\alpha) = \phi \circ \theta_\alpha \circ \widehat\phi,$$

So, to recover the true action of $\xi^{-1}(\alpha)(P)$, we must divide by $[D]$.

## Computing the kernels from ideals

With the ability to evaluate $\xi^{-1}(\alpha)(P)$ on a point $P \in E_0$, obtaining a kernel generator from an ideal is now straightforward. For the ease of notation, we will write $\alpha(P)$ as the action of some $\alpha \in \BB$ on a point $P \in E_0$, dropping the explicit mention of the isomorphism $\xi$.

For an ideal $I$, let us denote the $I$-torsion subgroup

$$E[I] = \{ P \in E \mathrel{|} \alpha(P) = 0 \;\; \forall \alpha \in I \} ,$$

which should be understood as the kernel $E_0[I] = \ker(\phi_I)$ for the isogeny $\phi_I : E_0 \to E_0 / \langle I \rangle$. For an ideal of norm $D$, the group $E_0[I]$ will have order $D$.

We will only be considering cyclic ideals as these correspond to cyclic isogenies. For clarification, see the page Working with Cyclic Ideals. When $I$ is a cyclic ideal, we have that $E_0[I]$ is a cyclic group and there is some generator point $K \in E_0[D]$ such that $E_0[I] = \langle K \rangle$.

Additionally, for any cyclic ideal with left order $\OO_0$, we can always find an element $\alpha \in I$ with reduced norm $\gcd(\Nrd(\alpha), D^2) = D$. We call this element the generator of the ideal, and we write:

$$I = \OO_0\langle \alpha, D \rangle = \OO_0 \alpha + \OO_0 D.$$

Given the generator $\alpha$, we can then recover $E_0[I] = E_0[\alpha] \cap E_0[D]$ by first computing the torsion basis $E_0[D] = \langle P, Q \rangle$ and then computing the kernel generator $K = [a]P + [b]Q \in E_0[D]$.

This is achieved by solving for $a,b$ such that the action of $\alpha(K)$ gives the identity point on $E_0$. We can efficiently solve this by exploiting the fact that $\xi^{-1} (\alpha)$ is an endomorphism, and hence a group homomorphism, to reduce the problem to finding $a,b$ such that $[a]\alpha(P) = -[b]\alpha(Q)$. We do this by solving a discrete logarithm problem, which is efficient as our points all have smooth order.

### Prime power norms

When $D$ is a prime power, we have that at least one of $\alpha(P)$ or $\alpha(Q)$ will have order $D$. Without loss of generality, assume $\alpha(Q)$ has order $D$. We can then use $\alpha(Q)$ as the base point for solving the discrete log and find the integer $a$ such that $\alpha(P) = [a] \alpha(Q)$. The kernel generator $K = P - [a]Q$ can then be returned.

We can solve this in SageMath with the following snippet:

αP, αQ = [eval_endomorphism(α, X, D) for X in (P, Q)]
if not has_order_D(αQ, D):
αP, αQ = αQ, αP
P, Q = Q, P
a = αQ.discrete_log(αP)
assert αP == a*αQ
return P - a*Q


### Allowing $D$ to be composite

Although working with prime powers allows an easier discussion, by allowing composite $D$ we can perform this algorithm only once, rather than for all primes factoring $D$. However, to do this, we must first slightly modify the discrete logarithm steps.

For generic $D$, we might have that neither $\alpha(P)$ nor $\alpha(Q)$ have order $D$, so we cannot solve the discrete logarithm as above to recover $a$. Instead, we compute $\alpha(P), \alpha(Q)$ as before and write them in linear combinations of the torsion basis $E_0[D]$: $$\alpha(P) = [a_1] P + [b_1] Q, \qquad \alpha(Q) = [a_2] P + [b_2] Q.$$ The integers $a_i, b_i$ can be efficiently recovered by solving two bi-dimensional discrete logarithm problem. In out implementation, this is the function BiDLP(R, P, Q, D) in the utilities.py file.

Given the integers $a_i, b_i$, we can recover $E_0[I]$ by first computing the kernel basis of the matrix:

$$\mathbf{A} = \begin{bmatrix} a_1 & b_1 \\ a_2 & b_2 \end{bmatrix} \in \mathbb{M}_2(\mathbb{Z} / D \mathbb{Z}), \qquad \ker(\mathbf{A}) = \begin{bmatrix} u_1 & v_1 \\ u_2 & v_2 \end{bmatrix}$$

and from this, we recover

$$E_0[I] = \langle [u_1] P + [v_1] Q, [u_2] P + [v_2] Q\rangle.$$

We can perform this computation in SageMath with the following snippet:

# Evaluate α(P) and α(Q)
R, S = [EvalEndomorphism(α, X) for X in (P,Q)]

# Find x,y such that R = xP + yQ
Rx, Ry = BiDLP(D, R, P, Q)

# Find x,y such that R = xP + yQ
Sx, Sy = BiDLP(D, S, P, Q)

# matrix of α on D-torsion w.r.t. basis (P,Q)
mat = matrix([(Rx, Ry), (Sx, Sy)])

# Thanks to Lorenz Panny for the following Sage
# Hack to find kernel of a matrix over ZZ/DZZ
# See: https://trac.sagemath.org/ticket/34862
ker = mat.stack(D*identity_matrix(2)).left_kernel().basis_matrix()[:,:2].change_ring(Zmod(D))

return [u*P + v*Q for u,v in ker if u or v]


For our implementation, we want to not just have the kernel $E_0[I] = \langle K_1, K_2\rangle$ but rather a single generating point $K$. Although there are deterministic ways to solve for this, it seems to require computing the order of both $K_1$ and $K_2$, which for large characteristic can be slow.

Instead, we find that a fairly sensible method is to just try linear combinations of $K_i$ until a point of order $D$ is found.

If we have that either $K_1$ or $K_2$ has order exactly $D$, we can return this as our kernel generator immediately. In the case where neither has full order, we can construct $K = K_1 + [x] K_2$ and see whether this has full order. Checking if an element has order exactly $D$ is about ten times faster than computing its order, so this check is almost always faster.

We implemented this as a small function, which we copy below:

def derive_cyclic_generator(P, Q, D):
"""
Given generators <P,Q> of a cyclic group
of order D, find K such that G = <K>

Heuristically, it seems easy to randomly
find a K this way, and is about 10x faster
than the deterministic method as we do not
need to compute the order of P or Q.
"""
K = P + Q
for _ in range(1000):
if has_order_D(K, D):
return K
K += Q
raise ValueError(f"Never found a cyclic generator!")


We acknowledge there may be a significantly better way to do this, but this works for now.

### Avoiding discrete logs

The description above is how our code was implemented while developing this project. However, there was a recent publication: Deuring for the People: Supersingular Elliptic Curves with Prescribed Endomorphism Ring in General Characteristic by Jonathan Komada Eriksen, Lorenz Panny, Jana Sotáková, and Mattia Veroni, which had some tricks to remove the discrete logarithms.

The authors noticed that instead of computing the action of $\alpha$ on the generators of $E_0[D] = \langle P, Q \rangle$ and then solving a discrete logarithm to recover the kernel, one could directly recover $E_0[I] = \langle \bar\alpha(P), \bar\alpha(Q) \rangle$ by computing the action of the conjugate of the generator. For more details, please see Section 4.1 of the above reference.

We now have a very simply algorithm for composite $D$. Given an ideal $I$ of norm $D$:

• Compute the generator of the ideal $\alpha$ and take the conjugate $\bar\alpha$
• Compute the torsion basis $E_0[D] = \langle P, Q \rangle$.
• Evaluate the action of $\bar\alpha$ on this basis to obtain $\langle \bar\alpha(P), \bar\alpha(Q) \rangle$.
• (Optionally) take a linear combination of $\langle \bar\alpha(P), \bar\alpha(Q) \rangle$ to find a point $K$ of order $D$.

### Putting it all together

Putting everything together, we get the following clean implementation

def ideal_to_kernel(E, I):
"""
Given a supersingular elliptic curve E0
and a cyclic ideal I with left order O0
produces a generator K_I ∈ E[n(I)] such that
ker(ϕ) = ⟨K_I⟩ = ⟨α_bar(P), α_bar(Q)⟩
"""
assert is_cyclic(I), "Input ideal is not cyclic"

# Degree of the isogeny we will to compute
D = ZZ(I.norm())

# Compute a generator such that I = O<α, D>
α = ideal_generator(I)

# Compute the torsion basis of E[D]
P, Q = torsion_basis(E, D)

# Evaluate R = α_bar(P)
α_bar = α.conjugate()
R = eval_endomorphism(α_bar, P, D)

# If this has full order, we can stop here
# as R = α_bar(P) generates the kernel
if has_order_D(R, D):
return R

# Same again for S = α_bar(Q)
S = eval_endomorphism(α_bar, Q, D)
if has_order_D(S, D):
return S

# Neither R or S had full order, so we to find a
# linear combination of R, S which has order D
return derive_cyclic_generator(R, S, D)


## Computing ideals from kernels

Unlike ideal_to_kernel(), which is used throughout SQISign, converting from a kernel generator to an ideal is only performed during the commitment stage. In this final section, we overview how this algorithm is implemented. We follow Leroux’s Thesis: Algorithm 20 without modification, so the summary is shorter than the section above.

We know some point $K \in E[D]$ which generates the kernel of a cyclic isogeny $\phi$ of degree $D$, and we wish to compute a generator $\alpha$ such that $I_\phi = \OO \alpha + \OO D$ for $\OO \cong \End(E)$.

To do this, we start with a basis for $\OO$ which we denote $\langle \beta_0, \beta_1, \beta_2, \beta_3 \rangle$, and we know that our generator can be written as some integral linear combination of the basis: $\alpha = a_0 \beta_0 + a_1 \beta_1 + a_2 \beta_2 + a_3 \beta_3$ for $a_i \in \ZZ$.

As $\ker(\phi) = \langle K \rangle$, if $\alpha$ the generator of the ideal $I_\phi$, we must have that $\alpha(K)$ is the identity point on $E$. This is what allows us to determine the coefficients $a_i$. Using our isomorphism $\xi$, we map the basis for $\OO$ to a basis of $\End(E)$ such that $\xi(\beta_i) = \theta_i$. We then map the above $\alpha$ to some $\theta \in \End(E)$:

$$\theta = [a_0] \theta_0 + [a_1] \theta_1 + [a_2] \theta_2 + [a_3] \theta_3.$$ Then, we look for $a_i$ such that $\theta(K)$ is the identity point on $E$.

The method described in Algorithm 20 is to find two basis elements $\theta_i$ and $\theta_j$ which through their action on $K$ generate the torsion basis, i.e., $E[D] = \langle \theta_i(K), \theta_j(K) \rangle$.

Given this basis, we can compute any point $P \in E[D]$ as a linear combination of $\theta_i(K)$ and $\theta_j(K)$. In particular, for some $\theta_k$ not equal to $\theta_i$ or $\theta_j$, we can represent the point $\theta_k(K) = [a] \theta_i(K) + [b] \theta_j(K)$ by solving the bi-dimensional discrete logarithm problem.

As we have $\theta_k(K) = [a] \theta_i(K) + [b] \theta_j(K)$, we have derived an endomorphism

$$\theta = \theta_k - [a] \theta_i - [b] \theta_j,$$

which maps $K$ to the identity point on $E$. This is precisely what we were looking for, and we can finish our algorithm by computing

$$\alpha = \beta_k - a \beta_i - b \beta_j, \quad I_\phi = \OO\langle \alpha, D \rangle.$$

### Putting it all together

Putting this together into one function, we have the following implementation:

def kernel_to_ideal(P, D):
"""
Given a point P ∈ E[D] compute the
ideal I(<P>))
"""
# Compute a basis β1,β2,β3,β4 of O0
# with norm coprime to D
βs = compute_coprime_basis(D)

# Compute the image of all the points
# β(P) by acting with θ ≅ β
θs = [eval_endomorphism(β, P, D) for β in βs]

# Find θi, θj which generates E[D]
i, j = find_torsion_basis_EndE(θs, D)
θi, θj = θs[i], θs[j]

# Pick k ≠ i,j such that
k = set([0, 1, 2, 3]).difference([i, j]).pop()
θk = θs[k]

# Solve the discrete log
a, b = BiDLP(θk, θi, θj, D)
assert a * θi + b * θj == θk

# Create the Quaternion Algebra element
α = βs[k] - a * βs[i] - b * βs[j]
return O0 * α + O0 * D


### Finding the torsion basis

To find $E[D] = \langle \theta_i(K), \theta_j(K) \rangle$ it is enough to just check all pairs $\theta_i(K)$, $\theta_j(K)$ by computing the Weil pairing $e(\theta_i(K), \theta_j(K))$. If the multiplicative order of this element is $D$, then $\theta_i(K)$, $\theta_j(K)$ must be linearly independent.

def find_torsion_basis_EndE(θPs, D):
"""
Looks for θi, θj such that θi(P), θj(P) generates E[D]
"""
for i in range(4):
for j in range(i + 1, 4):
eθiθj = θPs[i].weil_pairing(θPs[j], D, algorithm="pari")
if has_order_D(eθiθj, D, multiplicative=True):
return i, j
raise ValueError(f"No basis for E[D] found with given point")


### Computing a good basis

Another thing we must be careful of is making sure we compute a basis for $\OO$ such that the reduced norm of all the basis elements are coprime to $D$. This is done to ensure that we find the correct generator $\alpha$ and not something which only corresponds to some subgroup of the kernel.

Currently, we find a good basis with the following (fairly hacky) function:

def compute_coprime_basis(D):
"""
Start with basis <1, i, (i + j) / 2, (1 + k) / 2>
and find a new basis such that the norm of each basis
element is coprime to D.
"""
O0_basis = O0.basis()
θs = []
for f in O0_basis:
while True:
if gcd(f.reduced_norm(), D) == 1:
θs.append(f)
break
f += O0_basis[0] + O0_basis[1]
return θs


Potentially there is much nicer solution than this. If you have a suggestion, please let us know! For now, this is working, so hopefully that is good enough.