Repository navigation
Run the Jacobian weight sweeps in ExpScaled - #68
Merged
Merged
Conversation
The sweeps that balance a Jacobian cover are seeded with reciprocal
weights, which spread as widely as the weights themselves. A node
partway along a path could then leave Float64's range even where the
weight the sweep returns lies inside it: on Brown's almost-linear
system with one coordinate at 7e297, the reverse sweep's seed is
multiplied by that coordinate before the remaining factors of 0.5
bring it back, and cover_jacobian threw although a representable
cover exists.
The b-step (_sum_merge_b), the a-step (_marginal_weights) and the
grouped sweep of cover_jacobian_hybrid now compute in
ExpScaled{magtype(T)}, seeds included, and _toweights converts their
results back. A weight outside the normal range of magtype(T) is an
error naming the weight and whether it overflows or underflows;
below that range the conversion would round away significant bits,
and a weight rounded down need not cover. Interval tapes keep their
upward-rounded sweeps in Upward{F}, and an overflow there is now
reported as one. _check_cover_factors names the actual fault
(underflow, overflow, NaN or a negative value) instead of always
reporting underflow.
ExpScaled gains hypot, which the sign-averaged sweeps accumulate
with; it rounds as F's hypot does in F's normal range.
The wider arithmetic costs about 1.65x end to end on a 2000-variable
brownal.
Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The sweeps that balance a Jacobian cover are seeded with reciprocal weights, which spread as widely as the weights themselves. A node partway along a path could then leave Float64's range even where the weight the sweep returns lies inside it: on Brown's almost-linear system with one coordinate at 7e297, the reverse sweep's seed is multiplied by that coordinate before the remaining factors of 0.5 bring it back, and cover_jacobian threw although a representable cover exists.
The b-step (_sum_merge_b), the a-step (_marginal_weights) and the grouped sweep of cover_jacobian_hybrid now compute in ExpScaled{magtype(T)}, seeds included, and _toweights converts their results back. A weight outside the normal range of magtype(T) is an error naming the weight and whether it overflows or underflows; below that range the conversion would round away significant bits, and a weight rounded down need not cover. Interval tapes keep their upward-rounded sweeps in Upward{F}, and an overflow there is now reported as one. _check_cover_factors names the actual fault (underflow, overflow, NaN or a negative value) instead of always reporting underflow.
ExpScaled gains hypot, which the sign-averaged sweeps accumulate with; it rounds as F's hypot does in F's normal range.
The wider arithmetic costs about 1.65x end to end on a 2000-variable brownal.
Assisted-by: Claude Opus 5.5 noreply@anthropic.com