Closed

A finite algebra over a DVR

Lemma. (A finite algebra over a DVR). Let (A,(π),k)(A,(\pi),k) be a DVR, and let BB be a commutative finite free AA-algebra of rank n≥1n\ge1. If b∈Bb\in B is a non-zero-divisor, then D=det⁡A(mb)D=\det_A(m_b) is nonzero, B/bBB/bB has finite AA-length, and ord⁡AD=length⁡A(B/bB)=∑q∈Max⁡B[κ(q):k] length⁡Bq(Bq/bBq).\begin{equation} \operatorname{ord}_A D =\operatorname{length}_A(B/bB) =\sum_{\mathfrak q\in\operatorname{Max}B} [\kappa(\mathfrak q):k]\, \operatorname{length}_{B_{\mathfrak q}}(B_{\mathfrak q}/bB_{\mathfrak q}). \end{equation} All maximal ideals of BB lie above (π)(\pi). Terms with bb a unit are zero. Full residue-field degrees and unreduced quotient rings are used. The assertion does not require BB to be a domain or reduced, or any residue extension to be separable.

Proof

Proof. Choose an AA-basis of BB, and let TT be the matrix of mbm_b. Its kernel is zero. Over Frac⁡A\operatorname{Frac}A, a nonzero kernel vector could be multiplied by one nonzero common denominator to give a nonzero kernel vector in AnA^n. Thus TT is invertible over that field and D≠0D\ne0.

Here is the required matrix reduction over AA. Choose a nonzero entry of smallest order and move it to the upper-left corner by row and column permutations. Multiply its row by a unit to make this entry πa\pi^a. It divides every matrix entry. Elementary row operations therefore clear its column; elementary column operations then clear its row. These operations are invertible over AA. The remaining square block has nonzero determinant, so induction reduces TT to a diagonal matrix diag⁡(πa1,…,πan)\operatorname{diag}(\pi^{a_1},\ldots,\pi^{a_n}) with ai≥0a_i\ge0. It follows that B/bB≃⨁iA/(πai),ord⁡AD=∑iai.B/bB\simeq\bigoplus_i A/(\pi^{a_i}),\qquad \operatorname{ord}_A D=\sum_i a_i. The successive powers of π\pi give a composition series of length aia_i for A/(πai)A/(\pi^{a_i}). Length is additive in a short exact sequence: intersect a chain with the submodule and project it to the quotient for one inequality, and concatenate a submodule chain with inverse images of a quotient chain for the other. This proves the first equality and finite length.

A finite algebra is integral: on its finite generating module, the adjugate of the matrix of multiplication gives a monic equation for each element. If q\mathfrak q is maximal in BB, integrality makes its contraction maximal in AA. Indeed, if a field is integral over a subdomain, then that subdomain is a field: a monic equation for the inverse of a nonzero element, multiplied by a suitable power of the element, puts its inverse in the subdomain. Therefore q∩A=(π)\mathfrak q\cap A=(\pi), and κ(q)/k\kappa(\mathfrak q)/k is finite.

The ring C=B/bBC=B/bB has finite length over itself, since every chain of its ideals is a chain of its AA-submodules. Tag 00JB decomposes CC into its finitely many Artinian local factors. In a BqB_{\mathfrak q}-composition series for CqC_{\mathfrak q}, every simple factor is κ(q)\kappa(\mathfrak q). Its AA-length is its full kk-dimension: the action factors through kk, so its AA-submodules are precisely its kk-subspaces. Adding the lengths of all factors gives the second equality in (C.1.42). If C=0C=0, the sum is zero and the same equality holds. Neither step removes nilpotents or counts embeddings instead of field dimensions. ◻

Citation

Non-Archimedean Geometry Foundations, Tag NAG-002552.

Proof in chapter · Fixed-edition record