Coding Theory 17: Elliptic Ruled Surface Codes

Coding theory / 17

Elliptic ruled surface codes

Ruled surfaces over elliptic curves are the point where vector-bundle classification enters the code parameter problem.

Let \(C/\mathbb F_q\) be an elliptic curve with \(\gamma=\#C(\mathbb F_q)\). For \(X=\mathbb P(E)\to C\), the natural evaluation length is

\[n=(q+1)\gamma.\]

Rank-two indecomposable vector bundles over elliptic curves are controlled by the classification of Atiyah, extended in the finite-field setting used here through Arason-Elman-Jacob. In the normalized indecomposable rank-two case, the degree is \(0\) or \(1\).

Example 1: degree-zero indecomposable case

For a normalized indecomposable degree-zero bundle \(E\) and integers \(0<b<\gamma\), \(0\le a<q+1\), one obtains codes with

\[n=(q+1)\gamma,\qquad k=(a+1)b,\qquad d\ge(q+1-a)(\gamma-b)\]

In this range, the dimension is exact because the symmetric powers decompose in a way controlled by the elliptic vector-bundle classification.

Proof audit

For a normalized indecomposable degree-zero bundle, the Atiyah-Arason-Elman-Jacob classification decomposes \(S^a(E)\) as a direct sum of indecomposable degree-zero bundles \(F_{r_i}\) with total rank \(\sum_i r_i=a+1\). After twisting by \(\mathcal O_C(bP_0)\) with \(b>0\), each summand is semistable of positive degree, so \(H^1=0\). Since an elliptic curve has genus \(1\), vector-bundle Riemann-Roch gives

\[h^0(F_{r_i}\otimes\mathcal O_C(bP_0))=\deg(F_{r_i}\otimes\mathcal O_C(bP_0))=r_i b.\]

Summing over \(i\) gives \(k=b\sum_i r_i=(a+1)b\).

Example 2: the product formula reappears with \(\gamma\)

Compare the product formula \((q+1)^2,(a+1)(b+1),(q+1-a)(q+1-b)\) with the elliptic degree-zero formula. The base line \(\mathbb P^1\) is replaced by an elliptic curve with \(\gamma\) rational points, and the genus-one cohomology changes \(b+1\) into \(b\).

Degree-one case

For the normalized indecomposable degree-one bundle, under the same range \(0<b<\gamma\), \(0\le a<q+1\), the distance lower bound has the shape

\[d\ge(q+1-a)(\gamma-a-b).\]

This is a useful positive distance estimate in the subrange \(a+b<\gamma\). In that subrange the same estimate also excludes a nonzero section vanishing on every evaluation point, so the germ map is injective and the section-space dimension is the code dimension. Outside that subrange the construction still gives an evaluation code, but this particular Hansen-Lomont lower bound is vacuous and injectivity has to be checked separately.

The dimension lower bound is

\[k\ge(a+1)\left(b+{a\over2}\right),\]

and it becomes exact under additional hypotheses such as \(b>a\) in the Chapter 6 discussion. The unresolved issue is the contribution of \(h^1(S^a(E)\otimes\mathcal O_C(bP_0))\) in the remaining range.

Proof audit

In degree one the ruled surface is ample, so the fiber-component parameter is \(m=b-ae=b+a\) because the invariant is \(e=-1\). Hansen-Lomont then gives

\[d\ge(q+1-a)(\gamma-a-b).\]

For the dimension, \(\deg E=1\) and

\[\operatorname{rank}S^a(E)=a+1,\qquad \deg S^a(E)={a(a+1)\over2}.\]

After twisting by \(\mathcal O_C(bP_0)\), Riemann-Roch on the elliptic base gives

\[h^0-h^1=(a+1)b+{a(a+1)\over2}=(a+1)\left(b+{a\over2}\right).\]

Thus the displayed formula is always a lower bound, and it is exact precisely in the ranges where the first cohomology term vanishes.

Example 3: why degree one can look worse numerically

If only a lower bound for \(k\) is known, then rate estimates may understate the true performance. The table comparing degree-zero and degree-one cases should be read with this asymmetry in mind.

Example 4: obstruction as missing cohomology

The problem is not the length or the fiber distance estimate. It is the inability to compute the decomposition of \(S^a(E)\) well enough to determine the first cohomology term for all \(a,b\). This is the Lomont gap addressed only partially by the available decomposition methods.