The offset is measured in floors
Worth reading first: The offset fades at the window's floor · The singlet the doublets never had.
A four-parameter fit to a magnetic susceptibility curve carries two nuisance parameters, an impurity fraction and a temperature-independent term, and at particular exchange couplings the two stop being correlated. The upper of those couplings, in units of a population coupling fitted on twenty five-centre doublets, puts every connected cluster of six spin-½ centres in a narrow band — except five. The singlet the doublets never had found what those five share: a ground level holding a triplet and one or more singlets at exactly the same energy, a kind of ground no five-centre cluster can have.
The offset fades at the window’s floor then broke the degeneracy. Changing the coupling on one edge splits the singlet from the triplet, and on the side that keeps a triplet ground the excess falls smoothly as the splitting grows. On twelve edges of the three clusters with a single singlet beside the triplet it halved at splittings of 11.1 to 20.4 K, median 14.9. The data ran from 20 to 300 K, and 14.9 K is near the bottom of that window and a twentieth of the top. The essay read that as the offset living on the scale of the lowest temperature in the data, and said what would test the reading: the same clusters, re-read on windows with a different floor.
That is a clean experiment, because the reading makes a prediction the alternative cannot. If the floor sets the scale, the half-way splitting moves with the floor. If something about the top or the model’s own couplings sets it, the half-way splitting stays where it was while the floor moves underneath it.
Four floors, four frames
A cluster is never read alone. Its upper separation becomes a ratio only against a population coupling, and the population coupling depends on a fraction fitted on the doublets; the excess is that ratio less the median of the ordinary triplets. Every one of those quantities was measured on the 20–300 K window, so a different floor needs a different frame, and the frame must be refitted from the census at that floor before any split cluster is read in it.
The census of all forty-one clusters has already been read with the top at 300 K and the floor at 10, 20 and 40 K, in the essay that moved both ends of the window. A fourth census, with the floor at 5 K, is read here the same way. At each floor the fraction is refitted on the doublets, the ordinary triplets’ median is taken over the triplets whose ground holds no singlet, and each split cluster’s spectrum — the same changed edges, the same twenty changes from −50 to +50 per cent — is fitted on that floor’s window and read in that floor’s frame.
The frames differ, and in an orderly way. The fitted fraction falls as the floor rises, 0.0160 at 5 K, 0.0128 at 10, 0.0076 at 20 and 0.0044 at 40; the ordinary triplets’ median falls with it, 0.330, 0.320, 0.299, 0.292. At 5 K one of the sixteen ordinary triplets is set aside, because every reading of this census sets aside a cluster whose product falls below half the median, and at this floor one ordinary triplet does.
Before any edge is changed, the five clusters keep an excess under every floor. The ordering among them does not move — cluster 98, with three singlets in its ground, sits highest under every floor, cluster 46 lowest — but every excess shrinks as the floor rises: cluster 35’s from 0.093 at 5 K to 0.063 at 40 K, cluster 98’s from 0.209 to 0.175. A fit that reads the bottom of the curve more closely sees more of the degenerate ground. That is already a small confirmation of the reading, though not the test it was built for.
It is worth being careful about what that fall is. Each floor’s excess is measured against that floor’s own ordinary median and its own fitted fraction, and both of those move as the floor moves; the excess is a difference between two quantities each of which depends on the window. What the figure shows is that the difference stays positive and ordered under every frame, and narrows by about a third over a factor of eight in the floor. A fit is a statement about how many parameters a curve is worth over a particular range, and the range is part of the statement; the moment a fit invents was the first place in this sequence where one sample read over four ranges gave four answers, and the unsplit excess is one more number of that kind.
How one edge’s offset falls under each floor
The test is the fall. Take one edge of cluster 35 — edge 0–2, whose change splits the ground at a middling rate — and follow its excess as the splitting grows, once for each floor.
The four curves start at different heights, as the unsplit excesses said they would, and fall at different splittings. Under the 5 K floor the excess has halved by 5.4 K of splitting; under 10 K, by 8.4 K; under 20 K, by 15.1; under 40 K, by 33.4. The lower the floor, the earlier the fall, and the curves cross: a splitting of 30 K leaves the 40 K window about half its offset and the 5 K window almost none.
That crossing is the whole of the argument in one picture. A splitting is a fixed property of the split cluster, the same number in every window. What changes from curve to curve is only the lowest temperature the fit reads, and that alone decides how large a splitting must be before the singlet stops counting as part of the ground.
Nearly in proportion to the floor
Every edge whose change leaves a pure triplet ground can be read the same way. On the three one-singlet clusters there are twelve such edges, and at each floor each edge has a splitting at which its excess halves.
The medians are 5.4 K at a 5 K floor, 8.4 at 10, 14.9 at 20 and 32.8 at 40. On logarithmic axes they lie close to a straight line of slope 0.87. A scale set by the window’s top would have given a slope of zero, since the top never moved; a scale set by the floor alone would give one. The measured slope is much nearer the second.
In units of the floor itself, the half-way splitting is 1.08, 0.84, 0.75 and 0.82 floors. That is the sense in which the offset is measured in floors: across a factor of eight in the window’s lowest temperature, the splitting that removes half the offset stays within about a third of the floor. In wavenumbers the floor at 5 K is 3.5 cm⁻¹ and the median half-way splitting 3.8.
The edges spread more as the floor rises. Under a 5 K floor the twelve half-way splittings run from 4.3 to 5.7 K, a range of a third; under 40 K they run from 20.7 to 51.9, a factor of two and a half, and on three edges the largest change tried — fifty per cent of one coupling — does not split the ground far enough to halve the excess at all. A low floor makes every edge behave alike because the floor, not the edge, is setting the scale; a high one lets the cluster’s own structure through.
The fall in units of the floor
If the floor sets the scale, then measuring each splitting in floors should bring the four windows’ falls together. It does, and not perfectly.
The forty-eight curves — twelve edges at four floors — start together near one, pass through a half between about 0.7 and 1.1 floors, and fan out below a quarter. Some cross zero and go negative, which means the split cluster’s ratio has fallen below the ordinary triplets’ median; a ground split by several floors is an ordinary triplet with an excited singlet, and nothing guarantees it will sit exactly on the median of other ordinary triplets.
The departures from a single curve are in the direction the design predicts. The census reads forty temperatures evenly spaced from the floor to the top. Under a 5 K floor that is a spacing of 7.6 K, larger than the floor itself, so the lowest two readings are at 5 and 12.6 K, and the region within a floor of the bottom holds one point. Under a 40 K floor the spacing is 6.7 K, a sixth of the floor, and the bottom is densely read. The spacing essay found the census’s separations depending on how densely the cold end is read rather than only on where it stops, and a precision is a spacing found that a better reading at low temperature is a denser reading there. A floor sampled by one point behaves as a floor somewhat higher than its nominal value, which is where the 5 K window’s 1.08 floors sits: above proportion, where an effectively higher floor would put it.
That is a reading of the departure, not a measurement of it. Separating the floor from the spacing at the floor would need the same four floors re-read with the spacing held, which this does not do.
Why the bottom of the window
The reason the floor matters is the job the two nuisance parameters do. An impurity fraction contributes a Curie tail, a susceptibility rising as the inverse of temperature, and it is visible only where the cluster’s own susceptibility has flattened — at the bottom of the window. A temperature-independent term is a constant, and it too is told apart from the cluster by the shape of the curve at the low end. The upper separation is the coupling at which those two parameters’ correlation changes sign, and the fit decides that from how the curve bends at its lowest temperatures.
A singlet degenerate with the triplet in the ground dilutes the ground’s moment: at temperatures where only the ground is populated, a quarter of its states carry none. Split the singlet off by much less than the floor’s thermal energy and every temperature the fit reads still sees it as part of the ground; split it by much more and the bottom of the window sees a pure triplet, with the singlet an excited state like any other. The excess is the fit’s response to a ground whose low-temperature moment is diluted, and the scale on which a dilution becomes visible is set by the coldest reading.
The top of the window plays no part in that argument, which is why it is a good thing the top was held fixed: a test that moved both ends would have measured the floor and the top together. The window sets the constant found the census’s own constant following kT at the top; this offset follows the floor. The two numbers this sequence has measured on one census belong to opposite ends of one window.
What this means for a measured cluster
Real susceptibility measurements usually start at 2 K, the bottom of a standard magnetometer’s range, and sometimes at 1.8. On the reading measured here, a mixed-ground offset would survive in such data only for clusters whose singlet and triplet lie within roughly 2 K of each other, about 1.4 cm⁻¹. That is a strong condition. On the split clusters, a one per cent change in one coupling splits the ground by 0.3 to 3.6 K depending on the edge, so even a cluster whose couplings are equal by symmetry to a per cent would lose much of its offset on the sensitive edges under a 2 K floor — and a cluster whose bridges differ by several per cent would lose it everywhere.
The direction of that conclusion runs against intuition. A colder measurement reads the ground more finely, and that is exactly why it is less forgiving of a near-degeneracy: the finer the reading, the smaller the splitting needed to make two levels distinguishable. A fit on data that start at 20 K will report a mixed-ground offset for clusters that a fit on data starting at 2 K reports as ordinary triplets.
How the floors were read
Each window is forty temperatures evenly spaced from its floor to 300 K, with flat weights, as in the census. The fit is the four-parameter model the census uses, a coupling, an impurity fraction, a temperature-independent term and a -factor, and the correlation between the two nuisance parameters, from the fit’s own covariance, is followed on 160 couplings logarithmically spaced from 3 to 1,500 cm⁻¹, each change of sign bisected thirty times. The upper separation is the second change, kept below 1,000 cm⁻¹.
At each floor the population fraction is refitted on the doublets kept by that floor’s census, by a grid in its logarithm and a golden-section refinement, and the ordinary triplets are those whose ground holds no singlet. The split clusters are the same five, with the same classes of equivalent edges, the same representative edge of each class and the same twenty changes as at 20 K; the half-way splitting is interpolated linearly between the two changes that straddle half the unsplit excess, in kelvin at the split cluster’s own upper separation.
At 20 K every reading reproduces the published one, half-way splittings of 11.1 to 20.4 K. At 10 and 40 K each mixed cluster’s unsplit upper separation agrees with the window census’s to a part in ten thousand, though the census followed its grid to 4,000 cm⁻¹ and these follow theirs to 1,500. The refusal is the one the reading itself proposed: had the top set the scale, the half-way splitting would not have moved with the floor, and a slope through the medians below one half would not distinguish the two readings. It is 0.87.
What four floors on six-centre clusters cannot settle
The fit is always to the model it was built for. Every curve here is computed from an isotropic exchange Hamiltonian, with no zero-field splitting, no anisotropy in and no intermolecular coupling, and every one of those acts at the bottom of the window — precisely where this offset lives. A real measurement down to 2 K reads zero-field splitting of a triplet ground as readily as it reads a split singlet, and a fit with four parameters would have to absorb both.
The floor is also not the only thing that moved. Holding forty points while moving the floor changes the spacing at the bottom, and the 5 K window’s departure from proportion is plausibly that; it was not separated here. Nor were the two clusters with more than one singlet in their ground followed through the floors: one edge cannot split them, so they have no half-way splitting to move.
And the census has one size. Six centres is the smallest cluster with singlets to mix into a triplet ground, and the edge-by-edge sensitivities found here belong to six-vertex graphs.
Still open: the floor’s spacing, and every coupling at once
The obvious open question is the separation this left undone. The four floors were read with forty points each, so the spacing at the bottom of the window shrank as the floor rose; re-reading the 5 K and 40 K floors with the spacing held at its 20 K value would say whether the departure from proportion at the lowest floor belongs to the floor or to the points near it, and the designs essays already know how to make that change without touching anything else.
The nearer question is disorder in every coupling. A real cluster’s couplings all differ a little, and a ground split by many small changes at once may behave differently from one split by a single edge. Drawing each of a cluster’s couplings from a distribution a per cent or two wide, and reading the excess against the resulting ground splitting under a 2 K floor — the floor a measurement would actually have — would say how many real clusters near a mixed degeneracy could ever show the offset at all, and what a ground with two or three singlets does when every edge can move one of them.
Shares its objects with
Essays naming at least two of the same things, that neither author linked.
- The model is what is fitted — both name exchange coupling, least-squares, spin state, temperature
- The sign a frustrated ring changes — both name degeneracy, exchange coupling, least-squares, magnetic susceptibility
- A moment between two integers — both name degeneracy, spin state, temperature
- Five more clusters break the band — both name exchange coupling, magnetic susceptibility, spin state
- It was the count, not the frustration — both name degeneracy, exchange coupling, magnetic susceptibility
- Purity renames the poles — both name exchange coupling, magnetic susceptibility, spin state
Named objects
A dashed tag is an object no other essay names yet.
DegeneracyExchange couplingLeast-squaresMagnetic susceptibilitySpin stateTemperature