Two readings separate the square only where they do not fold
Worth reading first: A constant ties the polarisation · The worst system in the square was the solver's.
A composite method runs a mean field on the system it wants and adds a correlation correction computed somewhere else, and the transferability square has spent a long sequence of essays asking whether anything the mean field produces for nothing can say when that correction will be wrong. The most recent of them found a way to ask the question with no reference in it at all. If two systems give nearly the same reading of a diagnostic and miss correlation energies differing by , every function of that diagnostic misses one of them by at least , however it was fitted and to however many references. That is a floor, and the polarisation’s floor equalled what a single well-placed constant achieves while looking at nothing.
That essay ended on an opening. A function of two diagnostics is caught only by systems that tie in both, and requiring both collapsed the floor: 0.022 for the highest occupied level with the double occupancy, against 0.173 and 0.212 for each alone, and 0.052 for the polarisation with the mean-field energy. It also named the obvious suspicion. Eighty-one points in two dimensions leave near-ties scarce, so a low floor could be an empty cell rather than a separation, and the systems are cheap enough to settle it.
They are. Four of the six falls were the grid’s, and two are real. What decides which is not the number of systems but whether the two readings fold the square.
Refined twice, the floors mostly climb back
The square is a ring of four sites at half filling, with nine repulsions from 1 to 20 and nine staggered site energies from 0 to 10. It is refined by putting the geometric midpoint between each pair of consecutive repulsions and the arithmetic midpoint between each pair of site energies, which gives 17 by 17, then again for 33 by 33. Each grid contains the one before, so every number on the coarse square is still there on the fine one and every tie found there is still a tie.
That containment fixes the direction anything can move. A floor is the largest gap in correlation energy over tied pairs, and refining only adds pairs. A floor can stay or rise and cannot fall.
Four of the six rise a long way. The polarisation with the mean-field energy goes from 0.052 to 0.202 and then 0.209. The polarisation with the highest occupied level goes from 0.025 to 0.101 and then 0.136, and the double occupancy with the mean-field energy from 0.025 to 0.118 and then 0.123. The polarisation with the double occupancy was never low and climbs from 0.212 to 0.294, which is all but the finest grid’s constant. Each of the large jumps happens at the first refinement, and the second adds little, which is the shape of a coarse grid that had simply missed the pairs that set the floor rather than one that is still converging.
Two do not move much. The double occupancy with the highest occupied level goes from 0.022 to 0.025 and then 0.029, and the highest occupied level with the mean-field energy from 0.031 to 0.034 and then 0.035. On the finest grid there are 1,690 and 2,763 pairs tied in both readings to within a per cent of each range, against six and thirteen on the coarse square. The evidence behind those two numbers has gone from a handful of pairs to thousands, and the numbers barely changed.
So the suspicion was right about four pairs and wrong about two, and the question becomes what distinguishes them.
One reading alone was not converged either
Before that, the refinement says something about the single readings that the earlier floors depended on.
The mean-field energy was reported to have the most room under a constant, with a floor of 0.136 against the constant’s 0.216, and the note beneath it warned that finer sampling could only raise the floor. It raises it to 0.206 on 289 systems and 0.209 on 1,089. That room was a gap in the coarse square. The highest occupied level holds better, from 0.173 to 0.192. The polarisation and the double occupancy climb to 0.298 and 0.294.
The constant moves too, and for a reason worth stating separately. On the coarse square the largest correlation energy any system misses is 0.432, at a repulsion of eight and a site energy of eight. On the finest grid it is 0.596, at a repulsion of 9.80 and a site energy of 9.5, which lies between coarse points. The ridge along which repulsion and site energy compete is narrow enough for its crest to fall between the coarse square’s rows, and that crest is 38 per cent higher than anything the coarse square ever measured. Half the range is the constant’s floor, so the constant goes from 0.216 to 0.298 — and the polarisation’s floor goes to exactly the same number, 0.298. The tie with a constant survives the refinement, still set by two systems with no polarisation at all, one missing almost nothing and one at the crest.
Halving the tolerance
A floor is defined at a tolerance: two systems count as tied when their readings agree to within a per cent of each diagnostic’s range. That per cent is a choice, and it is the choice that turns out to tell a fold from a gap.
Suppose the two readings, taken together, are a smooth one-to-one map of the square. Then two systems that tie closely must be close on the square, and two systems close on the square have nearly the same correlation energy. Halve the tolerance and the tied pairs move closer together and the floor shrinks with them, towards zero, at a rate set by how steep the correlation energy is in the readings. Suppose instead that the map folds, so that two distant regions of the square land on the same readings. Then there are pairs that tie exactly in the readings while lying far apart on the square, and no tolerance excludes them. Halve the tolerance and the floor stays.
The two pairs that held up under refinement are the two that collapse here. The double occupancy with the highest occupied level falls from 0.029 to 0.016, 0.007 and 0.005, and the highest occupied level with the mean-field energy from 0.035 to 0.018, 0.008 and 0.005. At the tightest tolerance the same two systems set both floors, a repulsion of 18.91 and one of 20, at a site energy of 1.5 — adjacent points on the grid. There is no fold, only a slope.
The polarisation with the mean-field energy goes from 0.209 to 0.192 and stays there. The two systems that set it are 23 grid steps apart, a repulsion of 1.68 with a site energy of 1.25 and a repulsion of 9.80 with a site energy of 10, missing 0.08 and 0.47 of correlation energy. The polarisation with the double occupancy barely moves and is set by systems 17 steps apart. The polarisation with the highest occupied level falls only from 0.136 to 0.106. The double occupancy with the mean-field energy is the one intermediate case: it falls from 0.123 to 0.039 and its last pair is three steps apart, so it is a gap, but a steep one that the coarse square sampled too thinly to see.
Where the square folds
The pairs that fold all contain the polarisation, and the reason is the vertical line the reference-free score found.
A third of the square, 367 of the 1,089 systems, reads a polarisation of exactly zero: every system where the site energy is enough to hold the lowest mean field’s two spins together. All of them fall on one vertical line, and a second diagnostic has to do the whole work of separating them. Along that line the mean-field energy falls steadily as the site energy rises at every repulsion, so it orders the systems at one repulsion perfectly well. It does not order them across repulsions. A system at a repulsion of 1.68 and a site energy of 1.25, missing 0.08, and one at a repulsion of 9.80 and a site energy of 10, missing 0.47 near the ridge, reach mean-field energies of −3.165 and −3.160, and those are the two ringed. That is a fold in the strict sense, a pair of distant systems mapped to one pair of readings, and it is built into the polarisation’s zero rather than into anything a finer grid could reach.
The double occupancy with the highest occupied level draws the square as two sheets with no overlap. Every system whose lowest mean field is polarised has a double occupancy between 0.02 and 0.68, and every system whose mean field is not has one between 1.26 and 1.92, so the double occupancy alone already does what the polarisation does and keeps a scale on each side. The highest occupied level then spreads each sheet out. The ringed pair is two neighbours at the polarised sheet’s far edge, with double occupancies of 0.022 and 0.020. Nothing in this picture ties two distant systems, which is the whole of what the floor was measuring.
That is the answer to the question the coarse floor raised. A tenfold fall in a two-reading floor on eighty-one systems is evidence of two quite different things — a pair of readings that separates the systems and a pair whose fold the coarse grid happened to step over — and only the tolerance test tells them apart, because only a separating map gets better as the comparison gets finer.
A floor is not a forecast
A low floor says that no function of the two readings is forced to be wrong by much. It does not say that any function anyone would build is right. The cheapest test of the difference is to build one: give each of the 1,008 systems on the fine grid that the coarse square does not contain the correlation energy of the coarse system nearest to it in the readings, with each reading scaled by its range on the coarse square, and see how far off it is.
The best constant on the held-out systems misses by a median of 0.080 and at worst by 0.298. The separating pair does far better in the typical case, a median of 0.013 and nine in ten under 0.039, and it has the smallest worst error of any set tried, 0.166. The folding pairs do well in the median too — the polarisation with the mean-field energy is at 0.007 — and much worse at the worst, 0.321, because a fold sends some held-out systems to a coarse neighbour on the wrong side of it.
Two things about the separating pair’s worst error matter. Eleven of the held-out systems lie beyond the coarse square’s largest correlation energy, on the crest it never sampled, and no composite trained on it could have reached them; that alone guarantees a miss of 0.164. But the separating pair’s worst miss, 0.166, survives removing all eleven. It sits on a system inside the coarse square’s own range, where the readings separate the systems perfectly well and eighty-one references are too few to follow how fast the correlation energy changes. Adding the mean-field energy as a third reading changes nothing much, 0.172.
So the separating pair does what the floor allows and no more. It is the only reading in the whole sequence of these scores that beats a constant at the worst as well as in the median, and it does so by a factor of 1.8 rather than ten.
Every pair in one table
The table carries one more lesson about the earlier scores. The reference-dependent ratios ranked the highest occupied level the clear worst of the single diagnostics, because its medians stayed high. Here it is the reading every surviving pair contains. A diagnostic can be poor alone and indispensable in combination, because what makes a pair work is that its two readings never tie the same two distant systems at once.
How the grids and the tests were made
Every system’s exact ground state is found by diagonalising the ring of four with two electrons of each spin, and its lowest unrestricted mean field by the same search over starts that fixed the solver’s worst system, so the correlation energy is the exact energy less the lowest mean-field energy. The five readings are those of the reference-free score: the polarisation difference, the double occupancy, the highest occupied level, the mean-field energy and the product of the first two. Each diagnostic’s range is taken on the coarse square at every grid, so that a one per cent tolerance means the same reading gap on all three.
The arithmetic that has to hold is checked where the figures are drawn. The three grids nest at 81, 289 and 1,089 systems. The coarse floor of the double occupancy with the highest occupied level is the published 0.022. Every floor, single or paired, is no lower on a finer grid, and none is higher at a tighter tolerance. And the refusal is the correlation energy read against itself: given as one of its own readings, the floor must come out no larger than half the tolerance times its range on every grid, and it does. If it did not, the floor would belong to the method rather than to the diagnostics.
What a ring of four cannot settle
The square is a model square. Its two axes are a repulsion and a staggered site energy on a ring of four, which is the setting where each diagnostic was found to fail in its own region. A molecule has many more axes than two, and a pair of readings that separates a two-parameter family cannot separate a family with more parameters than it has readings, unless the extra parameters happen not to matter. On a real benchmark the right number of readings to ask for is the number of independent things the systems vary in, and a two-reading floor near zero there would be a much stronger statement than it is here.
The verdict of fold or gap is drawn from a grid. Two systems three steps apart are called a gap and seventeen steps apart a fold. The tolerance test is what justifies the line between them, not the step count, and the intermediate pair, the double occupancy with the mean-field energy, shows how much a single coarse grid can mislead in either direction.
The composite tested is the crudest one. A nearest neighbour is a fair test of what eighty-one references can do with no assumptions at all. An interpolant that used the shape of the sheets could do better inside them; nothing can do better at the crest the coarse square never saw.
The tie that the ratios disguised
The polarisation’s zero has now been met three times, in three disguises. Scored as a worst ratio it appeared as near-zero denominators. Scored against the correlation energy itself it appeared as twenty-seven identical readings with every correlation energy on the grid behind them. Read beside a second diagnostic it appears as a fold that no grid removes. Each was correct and each was partial, and the last is the one that says why the polarisation cannot be rescued by a partner: a partner would have to separate, on its own, every system that the polarisation sends to zero, which is a third of the square.
The same pattern runs through the warning a cheap calculation gives and its successors. A diagnostic is a map from systems to readings, and what decides whether it can warn is not how well its readings correlate with an error, nor how small a floor looks on a convenient sample, but whether the map folds. That is a geometric property of the diagnostic, and it can be asked of any proposed diagnostic on any benchmark at the cost of halving a tolerance.
Still open: a third axis, and the steep side of a sheet
The obvious open question is a square with a third axis. Every system here is a ring of four at half filling, and the two readings that separate it do so because two readings can separate a two-parameter family. A second hopping, or a filling away from half, adds an axis, and the double occupancy with the highest occupied level would then have to separate a three-parameter family with two numbers; it must fold somewhere, and where its fold appears would say which physical change it is blind to.
The nearer question is the steep stretch the separating pair could not follow. Its worst held-out miss is inside the coarse range, on a system whose readings are well separated from every other and whose correlation energy changes quickly with them. Where on the square those systems lie, and whether a handful of extra references placed there rather than on a uniform grid would bring the worst error down to the median’s scale, is a question about where references should go, and it is the first time in this sequence that the answer could be a placement rather than a better diagnostic.
What links here
Computed from the collection rather than written here: the essays that point at this one.
Reads more easily once this is understood
Essays that name this one as worth reading first.
Shares its objects with
Essays naming at least two of the same things, that neither author linked.
- The arms were the kindest part of the square — both name composite method, correlation energy, mean-field approximation, model limit, transferability
- The second number is the error, rearranged — both name composite method, correlation energy, mean-field approximation, transferability
- A mean field cannot get out of the way — both name correlation energy, model limit
- Half of it is given back at one bond — both name correlation energy, model limit
- The exponent was the window's — both name mean-field approximation, model limit
- The half of the square a ring of four cannot show — both name correlation energy, model limit
Named objects
A dashed tag is an object no other essay names yet.
Composite methodCorrelation energyMean-field approximationModel limitTransferability