stillwater-sc / universal / 26726097404

84%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

feat(elreal): infinite summation (Phase 5, #929) (#1033)

* feat(elreal): infinite summation (Phase 5, dissertation 4.2.3)

Implement sum() over a series of ZBCL terms (the McCleeary "stream of streams"),
the workhorse for Taylor series and transcendentals in Phase 7.

New headers:
- series.hpp    -- series<FpType>, the lazy co-list of ZBCL<FpType> terms (the
                   issue's colist<ZBCL<FpType>>), mirroring zbcl.hpp's memoised
                   thunk idiom; generators series_from_vector / _from_generator.
- exceptions.hpp -- elreal_exception base + elreal_sum_budget_exceeded.
- sum.hpp       -- sum<FpType>(series, max_depth): collects up to max_depth terms
                   (the budget), guards divergence (term magnitudes must trend to
                   zero), and renormalises the exact total of all term blocks into
                   the unique 0-overlap (DBL_k) stream via priestRenorm.

Design note: the dissertation defines sum as foldr add(). A naive lazy foldr over
the Phase 4 add() builds a deep lazy tower that, when forced, can emit a
value-correct but non-0-overlap stream (add()'s lazy composition does not
preserve the invariant across many nested calls -- see PR notes; worth a Phase 4
follow-up). Because the ZBCL value of a finite block multiset is exactly the
two-sum-exact total of its blocks, we instead renormalise the pooled blocks with
priestRenorm -- the same exact renormalisation add() uses internally -- giving a
result that is bit-exact on the represented terms AND 0-overlap. This makes sum()
eager over the budgeted prefix, which is the intended behaviour for a depth-bounded
summation.

Tests (elastic/elreal/summation/, dyadic-exact oracle, hosts double/float/cfloat<32,8>):
- geometric.cpp   -- sum (1/2)^n, (1/4)^n, (-1/2)^n, (1/3)^n; exact-oracle contract
                     exact_value(sum) == sum_i exact_value(term_i), 0-overlap, closed form.
- alternating.cpp -- Leibniz pi/4 (slow convergence; many compositions).
- ex... (continued)

Coverage Stats

80 of 85 new or added lines in 3 files covered. (94.12%)

18 existing lines in 2 files now uncovered.

47067 of 55945 relevant lines covered (84.13%)

6022107.95 hits per line