Orbitals

The contrast that fails least is a middling one

With a diffuse third kind of fragment at nuclear charge 0.6, the consecutive triples of a counterpoise assembly failed for two chains in five, worst on a light pair inside a diffuse chain. The open question was the contrast: does the failure grow as the diffuse fragment spreads out and vanish as it merges with the light one? Swept from 0.3 to 1 across every composition of six fragments, it does neither. The worst chain loses 73 points at 0.3, falls to 36 at 0.8, rises to 38 at 0.9 and settles at 31.5 when the kinds merge. Three different compositions take the worst place along the way. A three-fragment table, computed with no chain at all, tracks the worst loss down to 0.8 with a log correlation of 0.975 and cannot see the return.

Worth reading first: The far triples need a diffuse ghost · The uniform chain was the worst one to start from.

When fragments are described in each other’s basis functions, a cluster is treated better than its separated pieces and the binding comes out too large; the counterpoise correction removes that error by recomputing each piece with its partners’ functions present as ghosts. For a cluster of fragments the correction can be assembled from pieces: each fragment’s correction with every other fragment’s basis functions present as ghosts is approximated by its corrections from one ghost at a time, plus the three-body increments that two ghosts add together, and so on. Pairs alone count one share twice and leave the three-body terms out, and adding every triple costs more than the full calculation it stands in for, so the practical question has been which triples to keep. Over light and heavy fragments the census of every composition found that keeping only consecutive triples came within a point of the best cutoff for nine chains in ten, and that the uniform light chain was the worst of all, because far triple increments exist only among light fragments.

The far triples need a diffuse ghost then added a third kind of fragment, more diffuse than the light one, with nuclear charge 0.6 and its basis optimised for that charge as the others are. The consecutive rule failed far more often, for two chains in five at seven fragments, and the worst chains were not the all-diffuse one but chains mixing light and diffuse fragments. The far triples turned out to be largest on a light fragment corrected by two diffuse ghosts, so it was the mixture that failed.

One number made that result: 0.6. That essay’s continuation asked what happens to the failure as the contrast changes. With the diffuse charge near one the third kind merges into the light one and the two-kind result returns; much smaller, and the diffuse fragment’s own correction might dominate. It also asked whether the failure’s dependence on the contrast could be read from two or three fragments before any chain is built. Swept across the whole range, the failure is least at a middling contrast and returns before the merge, and the table that tracks it so well over most of the range cannot see why.

One charge, swept

The census is the same calculation as before with the third kind’s nuclear charge made a variable. Every composition of six fragments drawn from light (charge 1), heavy (charge 2) and diffuse is solved at diffuse charges 0.3, 0.4 and so on to 0.9, and at 1, where the diffuse kind is the light kind. That is 378 chains at each charge, each solved at six basis sizes and four separations. Seven fragments, 1,134 chains, are solved at 0.4, 0.6 and 0.8 to check that the trend is not an artefact of length. For each chain, keeping triples up to each span gives a covered share of the line on which the corrected energy can be trusted, and the consecutive rule’s loss is how far span two falls short of the best span, in points of that line.

The line is the range of accuracy targets, a decade and a half either side of a kilocalorie a mole, over which a practitioner might ask whether an assembled correction is good enough. A point of it is a hundredth of that range in the logarithm. The measure exists because there is one basis size where the correction is worth doing at any single target, and the line was holding that answer up: read at one target, small changes look like jumps between basis sizes. Read as a share of the whole line, a loss of ten points means the consecutive triples are usable over a tenth less of the targets anyone might choose than the best span is. It is the same measure the chains of eight and the distance cutoff were read with, so the numbers here continue theirs.

Two checks fix the dial’s ends. At 0.6 the swept census reproduces the published one cell for cell, every full correction, pair sum and triple sum, to 10−1210^{-12}, so the sweep varies one number and nothing else. At 1, every chain’s loss equals the loss of the two-kind chain with each diffuse fragment read as light, to 10−910^{-9}. So the merge the question was posed against is computed rather than assumed, and the light chain’s 31.5 points are where the dial ends.

Not monotone

The worst chain's loss falls as the diffuse fragment's charge rises, to a minimum near 0.8, and rises again before the kinds merge. The covered share lost by the consecutive triples against the best triple span, for the worst composition of 6 light, heavy and diffuse fragments, at each diffuse charge: 0.3: DLLDDD loses 72.6 points; 0.4: DLLDDD loses 71.8 points; 0.5: DLLDDD loses 71.5 points; 0.6: DLLDDD loses 49.1 points; 0.7: HHDDDD loses 42.0 points; 0.8: HHDDDD loses 35.7 points; 0.9: DDDDDD loses 37.7 points; 1: LLLLLL loses 31.5 points. At 7 fragments: 0.4: DLLDLLD loses 87.1; 0.6: DDLLDDD loses 57.7; 0.8: HHDDDDD loses 38.6. At charge one every chain is a chain of light and heavy fragments, and the worst is the uniform light chain.
Fig. 1 The consecutive triples’ loss on the worst composition of six fragments at each diffuse charge, with the worst chain named, and the worst of seven at three charges.

The worst chain’s loss falls from 72.6 points at charge 0.3 to 71.8 and 71.5 at 0.4 and 0.5, then drops to 49.1 at the published 0.6, 42.0 at 0.7 and 35.7 at 0.8. Then it rises, to 37.7 at 0.9, before settling at 31.5 at the merge. Seven fragments follow the same course one step up: 87.1 at 0.4, 57.7 at 0.6, 38.6 at 0.8. Neither expectation the question offered survives. The failure does not grow steadily as the third kind spreads out, since the curve is flat from 0.3 to 0.5. It does not fade steadily toward the merge either, since 0.9 is worse than 0.8.

The share of chains the consecutive triples serve to within a point climbs to three in four at 0.7, slips back at 0.8 and 0.9, and recovers only at the merge. For every composition of 6 fragments at each diffuse charge, the share whose consecutive triples are within one point of the best span, and the share losing more than two: 0.3: 54% within one, 44% over two; 0.4: 53% within one, 43% over two; 0.5: 55% within one, 42% over two; 0.6: 65% within one, 31% over two; 0.7: 75% within one, 23% over two; 0.8: 73% within one, 25% over two; 0.9: 70% within one, 28% over two; 1: 80% within one, 20% over two. At 7 fragments: 0.4: 48% and 49%; 0.6: 59% and 38%; 0.8: 78% and 20%.
Fig. 2 For every composition of six at each charge, the share within one point of the best span and the share losing more than two, with seven fragments at three charges.

The shares tell the same story across all 378 chains, not only the worst. The consecutive rule is within a point of the best span for 54 per cent of chains at 0.3, 65 per cent at the published charge, 75 per cent at 0.7, then 73 and 70 per cent at 0.8 and 0.9, and 80 per cent when the kinds merge. The share losing more than two points is the mirror image: 44, 31, 23, then 25 and 28, then 20. At seven fragments the within-a-point share runs 48, 59 and 78 per cent at 0.4, 0.6 and 0.8. The mean loss over all chains has its minimum between 0.7 and 0.8, 3.08 and 3.10 points, and is 4.21 at 0.9, higher than at the merge.

Three compositions take the worst place

No composition is worst across the dial: a light pair in a diffuse chain fails at strong contrast, a heavy pair at 0.7, the uniform diffuse chain near the merge. The points each of four compositions of 6 loses by keeping only consecutive triples, at each diffuse charge: DLLDDD (light pair in a diffuse chain): 0.3 72.6, 0.4 71.8, 0.5 71.5, 0.6 49.1, 0.7 21.5, 0.8 19.0, 0.9 31.0, 1 31.5; HHDDDD (heavy pair, diffuse rest): 0.3 11.8, 0.4 14.9, 0.5 30.9, 0.6 39.6, 0.7 42.0, 0.8 35.7, 0.9 19.4, 1 0.0; DDDDDD (all diffuse): 0.3 42.1, 0.4 46.4, 0.5 34.7, 0.6 25.8, 0.7 28.0, 0.8 34.5, 0.9 37.7, 1 31.5; LLLLLL (all light): 0.3 31.5, 0.4 31.5, 0.5 31.5, 0.6 31.5, 0.7 31.5, 0.8 31.5, 0.9 31.5, 1 31.5.
Fig. 3 The loss of four compositions of six across the dial: a light pair inside a diffuse chain, a heavy pair beside four diffuse fragments, the uniform diffuse chain and the uniform light chain.

Following named chains across the dial explains the shape. The chain that was worst at the published charge, a light pair inside a diffuse chain (DLLDDD in the figures’ lettering), loses 72.6 points at 0.3 and keeps losing over 70 to 0.5. It drops to 49.1 at 0.6 and to 21.5 and 19.0 at 0.7 and 0.8, where it is unremarkable, and then climbs back to 31.0 at 0.9, almost the merged light chain’s loss. A heavy pair beside four diffuse fragments loses 11.8 at 0.3 and rises steadily to 42.0 at 0.7, where it is the worst chain of all, and falls away to 19.4 at 0.9 and to nothing at the merge. Two heavy fragments beside light ones are a chain the two-kind census serves exactly. The uniform diffuse chain loses 42.1 and 46.4 at 0.3 and 0.4, dips to 25.8 at 0.6 and climbs to 37.7 at 0.9, where it is the worst.

The cure changes with the mechanism too. Every worst chain from 0.3 to 0.8 has its best span at three or four, one or two separations beyond the consecutive triples, which is the cheap repair the published census found. The uniform diffuse chain’s best span is five at every charge, the whole length of a chain of six, so near the merge the repair is to keep every triple, the expensive answer the uniform light chain also needed. A practitioner who set a span from the worst case at one charge would be one or two separations short at another.

So the worst place is held by three different compositions on one dial — the light–diffuse mixture up to 0.6, the heavy–diffuse mixture at 0.7 and 0.8, and the uniform diffuse chain at 0.9 — before the uniform light chain takes it at the merge. Each has its own charge where it fails most, and the minimum of the worst-chain curve is the charge where the first mechanism has faded and the last has not yet arrived.

Among the chains made of light and diffuse fragments alone, the failures fall to thirteen in thirty-four at 0.7 and return to all of them by 0.9. Of the 34 compositions of 6 light and diffuse fragments that use both kinds, how many lose more than two points by keeping only consecutive triples, at each diffuse charge: 0.3: 27; 0.4: 23; 0.5: 21; 0.6: 18; 0.7: 13; 0.8: 26; 0.9: 34; 1: 34. At charge one they are all the uniform light chain, which loses 31.5.
Fig. 4 Among the 34 compositions of six light and diffuse fragments that use both kinds, how many lose more than two points at each diffuse charge.

The chains of light and diffuse fragments alone show the return most sharply, since they are the chains the merge is about. Of the 34 that use both kinds, 27 lose more than two points at 0.3, 23 at 0.4, 21 at 0.5, 18 at 0.6 and 13 at 0.7. Then 26 at 0.8 and all 34 at 0.9. Near the merge a light and a diffuse fragment are almost the same kind, so every such chain behaves almost like the uniform light chain, which loses 31.5. The alternating chain LDLDLD shows the whole course on its own: it loses 35.3 at 0.3, nothing at 0.8, 22.7 at 0.9 and 31.5 at the merge. The contrast that best protects a mixed chain is the one large enough to stop it looking uniform and small enough not to make its light members borrow from their diffuse neighbours.

What three fragments can see

The explanation the published charge offered was a table, and the table can be read at every charge without building a chain. A triple’s three-body increment is a solve with only its three members present, so the far increments — those spanning three or four separations, the ones a wide cutoff is needed to catch — can be tabulated by the kinds of the corrected fragment and its two ghosts.

Three fragments and no chain: the far triple increment of a light fragment between two diffuse ghosts falls eighteenfold across the dial. The median magnitude of a three-body counterpoise increment over basis sizes, separations and far arrangements (spans three and four), in accuracy lines, for four patterns of corrected fragment and ghosts, at each diffuse charge: light, two diffuse ghosts: 0.3 1.21e-2, 0.4 1.11e-2, 0.5 7.85e-3, 0.6 3.46e-3, 0.7 1.71e-3, 0.8 1.10e-3, 0.9 6.89e-4; diffuse, two diffuse ghosts: 0.3 9.92e-5, 0.4 2.25e-4, 0.5 4.64e-4, 0.6 6.02e-4, 0.7 6.27e-4, 0.8 4.72e-4, 0.9 3.88e-4; light, two light ghosts: 0.3 2.70e-4, 0.4 2.70e-4, 0.5 2.70e-4, 0.6 2.70e-4, 0.7 2.70e-4, 0.8 2.70e-4, 0.9 2.70e-4; diffuse, two light ghosts: 0.3 5.95e-6, 0.4 1.01e-5, 0.5 1.84e-5, 0.6 2.81e-5, 0.7 4.48e-5, 0.8 8.65e-5, 0.9 1.41e-4.
Fig. 5 The median far three-body increment, in accuracy lines, for four patterns of corrected fragment and ghosts, at each diffuse charge.

The light fragment corrected by two diffuse ghosts carries the largest far increment at every charge, and it falls monotonically and steeply, from 1.21×10−21.21 \times 10^{-2} accuracy lines at 0.3 to 6.9×10−46.9 \times 10^{-4} at 0.9, eighteen times smaller. Its shape is the flat-then-falling shape of the worst loss: nearly level from 0.3 to 0.4, then falling. The all-diffuse triple does something else. It is small at 0.3, 9.9×10−59.9 \times 10^{-5}, rises to a peak of 6.3×10−46.3 \times 10^{-4} at 0.7 and falls to 3.9×10−43.9 \times 10^{-4} at 0.9, still above the all-light triple’s constant 2.7×10−42.7 \times 10^{-4}. The likely reason is that a very diffuse fragment’s functions are so broad at these separations that a second ghost adds little to what the first already lent. A diffuse fragment corrected by two light ghosts stays an order of magnitude or more below the others until 0.8, and reaches half the all-light value only at 0.9.

Read off three fragments, the far increment orders the worst chain's loss down to a charge of 0.8, and misses the one after it. The worst chain's loss at 6 fragments against the far triple increment of a light fragment with two diffuse ghosts, on a logarithmic axis, one point per charge: 0.3: increment 1.21e-2, loss 72.6; 0.4: increment 1.11e-2, loss 71.8; 0.5: increment 7.85e-3, loss 71.5; 0.6: increment 3.46e-3, loss 49.1; 0.7: increment 1.71e-3, loss 42.0; 0.8: increment 1.10e-3, loss 35.7; 0.9: increment 6.89e-4, loss 37.7. Correlation in the logarithm 0.975 for the worst loss and 0.889 for the mean.
Fig. 6 The worst chain’s loss at six fragments against the far increment of a light fragment between two diffuse ghosts, one point per charge.

Read against the census, the light-between-diffuse increment does well. Over the seven charges below the merge its logarithm correlates with the worst loss’s at 0.975, and with the mean loss’s at 0.889, and it orders the charges correctly from 0.3 to 0.8. A three-solve table predicts which contrast will be dangerous for the consecutive rule, and roughly by how much, over the range where a light fragment borrowing from diffuse ghosts is the mechanism.

It cannot see 0.9. There the increment is at its smallest and the worst loss rises, because the worst chain at 0.9 is the uniform diffuse one, whose failure is the uniform light chain’s failure, with far increments among near-identical fragments at a slightly larger size. No single entry of the table carries both mechanisms. The all-diffuse entry has the right shape near the merge, but its correlation with the mean loss over the whole dial is −0.73. The worst loss is set by whichever mechanism is larger, and that depends on how many far triples of each kind a chain contains, which is a property of chains rather than of three fragments.

The census at every charge, 6 fragments: which chain fails worst, by how much, and where the uniform chains rank. For each diffuse charge, of 378 compositions: 0.3: worst DLLDDD losing 72.6 with best span 3, 54% within one point, 44% over two, mean loss 8.18, all-diffuse chain ranked 18 and all-light 37; 0.4: worst DLLDDD losing 71.8 with best span 3, 53% within one point, 43% over two, mean loss 8.46, all-diffuse chain ranked 15 and all-light 38; 0.5: worst DLLDDD losing 71.5 with best span 4, 55% within one point, 42% over two, mean loss 7.06, all-diffuse chain ranked 19 and all-light 23; 0.6: worst DLLDDD losing 49.1 with best span 3, 65% within one point, 31% over two, mean loss 4.49, all-diffuse chain ranked 19 and all-light 10; 0.7: worst HHDDDD losing 42.0 with best span 3, 75% within one point, 23% over two, mean loss 3.08, all-diffuse chain ranked 5 and all-light 3; 0.8: worst HHDDDD losing 35.7 with best span 3, 73% within one point, 25% over two, mean loss 3.10, all-diffuse chain ranked 2 and all-light 4; 0.9: worst DDDDDD losing 37.7 with best span 5, 70% within one point, 28% over two, mean loss 4.21, all-diffuse chain ranked 1 and all-light 7; 1: worst LLLLLL losing 31.5 with best span 4, 80% within one point, 20% over two, mean loss 3.96, all-diffuse chain ranked 36 and all-light 1. At charge one the light and diffuse kinds are one kind, so every chain of them ties with the light chain.
Fig. 7 At each diffuse charge, of the 378 compositions of six: the worst chain, its loss and best span, the shares within one and over two points, the mean loss, and where the two uniform chains rank.

The table’s last two columns are where the uniform chains sit, and they move in opposite directions along the dial. The uniform diffuse chain ranks 18th of 378 at 0.3, 19th at the published charge, fifth at 0.7, second at 0.8 and first at 0.9. The uniform light chain, whose loss never changes because it contains no diffuse fragment, ranks 37th at 0.3, tenth at 0.6 and third at 0.7, rising as the rest of the census improves around it and falling back to seventh at 0.9, when diffuse chains overtake it. Below the merge, the chain the whole sequence began with is never the worst of three kinds. At every charge something that contains a fragment it does not loses more.

How the census was solved at every charge

The solve is the published one, with the diffuse charge as a variable. For each basis size and separation, a fragment alone, each pair and each triple is solved once with only its members’ functions present on the line, and stored by kinds and offsets. Each chain’s pair and triple corrections are sums over those tables, and only each fragment’s full correction, with every other fragment’s ghost present, is a solve per chain. The covered share is the union, in the logarithm of the accuracy line, of the windows in which each basis size’s corrected error lies below its correction. The far table is the median magnitude over all basis sizes, separations and six far arrangements of offsets.

The checks are the two ends of the dial, already described, and a refusal. A light–diffuse alternating chain’s full correction at one basis size and separation must differ between charges 0.4 and 0.8 by more than a thousandth of itself. A census that varied the label of the charge without the charge reaching the solve would reproduce the published census and the merge perfectly and measure nothing.

What one dial cannot say

One dimension, one kind of function. The fragments are atoms on a line with one-dimensional Gaussians, and “diffuse” means a smaller nuclear charge with its basis optimised to it. A real diffuse fragment — an anion, a Rydberg-rich atom — differs from its neighbours in the shape and count of its functions as well as their width, and the same dial in three dimensions would move more than one thing at once.

Six fragments, and seven at three charges. The sweep is at six because a full census of seven takes seven minutes a charge. Seven follows six wherever both were solved, and the two-kind census found the light chain’s loss growing with length, so the minimum’s position may drift with length even if its existence does not.

A median is not a sum. The far table reports medians over cells and arrangements, and a chain’s loss is set by the sum of its own far increments in the cells where its window edges lie. That is why a table can rank the strong-contrast charges and miss the return: the return is about how many far triples a chain has, not how large a typical one is.

Charges are a proxy for contrast, not a measure of it. The quantity that matters to a ghost is how far a fragment’s functions reach at the separations used. The charge sets that through the basis optimisation, but two fragments with the same charge ratio at different separations would sit at different points on this dial.

A contrast in the middle

The census’s first answer was that the uniform chain was the worst one to start from. Its second was that a mixture is worse than either pure chain. This one says that how much worse depends on the contrast in a way with a minimum in it. A diffuse fragment very unlike the light one makes its light neighbours borrow from far ghosts, and the consecutive rule misses that borrowing. One very like the light one makes every chain of the two behave as a uniform chain, which the consecutive rule never served well. Between them, at a charge around 0.7 to 0.8 in this model, both effects are weak, three chains in four are served to within a point, and the worst place goes to a heavy–diffuse mixture that neither the light–diffuse table nor the uniform-chain argument predicted. A rule of thumb that says “the more different the fragments, the more far triples matter” is right at one end of the dial and wrong at the other.

Still open: the cells that decide, and a table with chains in it

The obvious open question is the one this sequence has deferred twice: which cell sets each chain’s loss. Every loss is set at the basis size and separation where a window edge moves, and the three mechanisms found here may each live in a different cell — the light–diffuse borrowing at the shortest separation, the uniform failure at the largest basis. If they do, the cutoff could be chosen per cell rather than per chain, and the composition, and with it the contrast, would stop mattering.

The nearer question is a predictor that counts. The far table failed at 0.9 because it reports one typical increment, and a chain’s loss depends on how many far triples of each pattern it contains. A sum over a chain’s own patterns — the table entries weighted by how often each occurs at span three and four — is cheap, needs no full-correction solve, and would say whether the return near the merge is predictable from three-fragment pieces after all or genuinely needs the whole chain.

Shares its objects with

Essays naming at least two of the same things, that neither author linked.

Named objects

A dashed tag is an object no other essay names yet.

Basis set superposition errorCounterpoiseEffective nuclear chargeFragment methodGaussian basisMany-body expansion