Strong faults move a break log's floor into the harness
Worth reading first: A harness reads a warp's tail best threaded at both ends · A break log reads the tail only above its fault floor · A warp breaks in a tail no tensile test reaches.
A loom’s break log, kept shaft by shaft, is an instrument for a question no tensile test can answer: how a warp yarn’s weakest places are distributed, out in the tail where a few breaks a beam come from. A harness reads a warp’s tail best threaded at both ends treated the threading as an experimental design and found the design that answers it fastest. Hang the warp on the front two shafts and the back two, about two fifths at the front, and a season’s log tells a lognormal yarn from a normal one by a hundred to one in six beams, against twenty-five for a straight draft.
The design worked because the log was a floor plus a cliff. The floor was discrete faults — slubs, bad piecings, fly woven in — that break on every shaft alike; the cliff was the yarn’s own spread of strength, rising toward the back where the tension is highest. The front shafts saw only the floor, so they measured it, and the back shafts saw the cliff on top of it.
Every fault was taken to be weaker than the front shaft. That was what made the floor flat. A fault population is not obliged to be so obliging, and the essay that built the design ended by asking what a spread of fault strengths would do to it. Its guess was that the best harness would gain a third group of shafts near the middle, where a gentle fault cliff and the yarn’s steep one differ most. The guess is half right, and the half that is wrong is the more useful.
The floor is a cliff when faults have strengths
A fault breaks on a shaft only if it is weaker than the shaft’s tension. The tensions in the shed’s own extension run from 14 per cent of the yarn’s 500-millimetre breaking load on the front shaft to 52 per cent on the back. A fault weaker than 14 per cent breaks everywhere, which is the flat floor; a fault at 30 per cent breaks only behind the shaft whose tension reaches 30, and passes unbroken through every shaft in front of it.
So with a spread of fault strengths the floor rises through the harness. The fault strengths are a chosen law, not a measured one: lognormal, with a median and a log-spread that are inputs here, because no fault classifier reports a fault’s strength — it counts faults by length and thickness, not by the load at which they part. At a median of a quarter of the breaking load and a log-spread of 0.4, the front shaft breaks 7 per cent of the faults that reach it, half of them break from the eighth shaft back, and the back shaft breaks 97 per cent.
The drawing shows what that does to the log. Both laws log 0.12 breaks a beam on the front shaft, almost all faults, and the floor climbs from there in a long gentle ramp. The normal law’s cliff leaves it from about the twentieth shaft and reaches 2.4 breaks on the back shaft; the lognormal’s barely leaves it, 1.6. What the front shafts were relied on to read — a flat floor, the same on every shaft — is not there to read.
The two-ended harness stops working
The design that read the tail in six beams read the floor at the front and assumed the floor at the front was the floor everywhere. Once it is not, the log has two things to learn about the faults rather than one — how many there are and how strong — and the normal law, given its own fault rate and its own fault median to fit with, has more room to imitate the lognormal.
With weak faults, the pairs need 6.2 beams — the earlier design’s six, give or take the cost of reading a median that turns out to be irrelevant. As the median rises past the front shaft’s tension the pairs fall apart quickly. From a median of about an eighth of the breaking load they need more beams than the straight draft: 30 at 15 per cent against the straight draft’s 27, and 33 at 25 per cent against 30. The design that was best by a factor of four is now worse than no design at all.
The reason is the same as the reason a single back shaft could never tell the laws apart in the account that first read the log by shaft: a log that sees too few tensions lets the alternative law move its nuisance parameters until it fits. The front pair used to pin the floor; now it sees a few per cent of the faults, a number the normal law can match with a stronger fault population, and once it can do that the back pair’s cliff is all that is left to argue from.
Knowing the faults’ strengths in advance would help — the dashed curve, fitting only the fault rate, needs 8.4 beams at a median of 15 per cent and 17 at 25 — but that is knowledge no mill has. A mill that knew its faults’ strength would not need a break log to learn about its yarn’s; both come from the same population of weak places.
The best threading moves its floor inward
The design was found the first time by searching a family of splits. With two nuisance parameters, the family is the wrong place to look, and the right method is older than any loom: the exchange that the optimal-design literature calls Fedorov–Wynn. Start with the warp spread evenly; at each round, find the normal log that best fits the current design’s lognormal one, see which shaft would add the most divergence per end against it, and move a small share of the warp there. The design that comes out is the one no allocation of ends can beat.
With faults all weaker than the front shaft the exchange finds the earlier answer, on single shafts rather than pairs: 42 per cent of the warp on the front shaft and 58 on the back, 4.2 beams. As the faults strengthen, the group that reads the floor walks into the harness. At a median of 10 per cent it sits on shaft 14, at 15 per cent on shaft 16, at 20 per cent on shafts 17 and 18, carrying the same 42 per cent throughout.
That is the half of the guess that was wrong, and it is the more useful half. The group that reads the floor is not reading it at the front because the front is the front; it reads it there because the front was the lowest tension at which every fault breaks and no yarn breaks. With strong faults that place is further back. The middle of the harness, at around two fifths of the breaking load, is where nine tenths of the faults have broken and neither law’s spread has yet risen far off the floor — so it is where the floor can be read nearly whole, and the design goes there.
A third group arrives, at the front
From a median of a quarter of the breaking load the design changes shape. Beside the floor group on shafts 18 and 19 and the cliff group on shaft 24, a third group appears on shaft 2, carrying 2.6 per cent of the warp. At a median of 30 per cent it is 4.9 per cent on shaft 3; at 35, 5.5 per cent on shaft 4. The beams needed rise only from 4.2 to 5.6 over the whole range.
That is the half of the guess that was right: Elfving’s rule, that a design needs as many distinct points of observation as its model has unknowns, comes true at three. But the third point is not in the middle where the two cliffs differ most. It is near the front, where only the weakest faults break, and it is small.
The log shows what each group reads. The front group reads how many faults break at a low tension, which is a reading of how strong they are: 0.13 breaks a beam on shaft 2 under the lognormal law with its true faults. The middle group reads how many there are, since nearly all of them break there: 14 breaks a beam on shaft 19. The back group reads the cliff: 21 breaks on shaft 24.
The normal law’s best answer to that log sets its fault rate at 73 per cent of the truth and its median at 20 per cent rather than 25, and it is wrong in all three places at once — 0.22 at the front, 11 in the middle, 26 at the back. Its own tail adds breaks in the middle and at the back that the lognormal’s does not, so it lowers its fault rate to compensate; to account for the front shaft’s breaks with fewer faults it makes them weaker; and the compromise leaves it over at the front, short in the middle and still over at the back. Three groups are what it takes to catch it in that contradiction.
The design is proved best, not found best
An exchange that stops is not thereby at an optimum. The test that it is comes from the same theory, and it can be drawn: at the best design, every shaft the design uses must add the same divergence per end against the best-fitting alternative, and no shaft it leaves empty may add more. If some empty shaft added more, moving an end there would improve the design.
The four shafts in use read 2.20 to 2.28 ten-thousandths per end. The front shaft, unused, reads 2.20, as much as shaft 2 — so the third group could sit on either and the design would barely notice, which is why the front group moved between shafts 1 and 4 as the median changed. Nothing reads more than the level.
The dead zones say where a log learns nothing. Shafts 8 and 9 add almost no divergence, because there both laws’ logs are the fault ramp and the fitted fault population matches it exactly. Shaft 23 adds a hundredth of the level, because it is where the fitted normal log’s cliff crosses the lognormal’s — too few breaks below and too many above, and none different at the crossing. A shaft placed at either is a shaft whose ends are logged for nothing.
How wide the faults spread decides whether the third group is needed
The median is one input; the spread is the other. A narrow spread is nearly the point population that made the floor flat, only moved to a stronger place; a wide one reaches down below the front shaft’s tension even when its median is well above it.
With a log-spread of 0.1, faults of median 25 per cent are a tight cluster that every shaft from about the fourteenth breaks and no shaft in front of it does; the design reads their floor on shaft 14 and the cliff on 24, two groups, 4.2 beams. Widening the spread pushes the floor group back — to 17 at 0.2 and 18 at 0.3 — because the last few strong faults break later and the floor is not complete until they have.
From a log-spread of 0.4 the third group appears, and from 0.5 it sits on the front shaft itself with about 3.4 per cent of the warp. A wide spread puts a measurable tail of weak faults under even the lowest tension, and a design that reads that tail learns the population’s shape from it. The beams needed are almost flat: 4.2 to 5.1 across the whole range. What a third group costs is a handful of heddles; what it saves, against the design that ignores fault strength, is more than twenty-five beams.
What the threading asks of the loom
Every design above is drawn on single shafts, which is what the theory treats as points. A weave needs a shaft for each of its lifting classes in every group, so a plain weave would use pairs and a 2/2 twill fours, as how many shafts a draft needs counts. Three groups of pairs is six shafts of the twenty-four; three groups of fours is twelve. The pair version was not searched here. The earlier essay’s experience — single back shaft never, back pair poorly, front and back pairs well — suggests that the move from single shafts to pairs costs beams through the spread of tensions inside each group, and the size of that cost here is not computed.
The group that reads the cliff carries more than half the warp on a single shaft, 2,214 ends; as pairs it is about 1,100 a shaft, against the even loading that where the heddles go treats as the default. That is the same heavy back frame the earlier design asked for, and what the shed costs in newtons prices lifting it. The new group in the middle carries 1,680 ends on shafts that a straight draft would give 167 each. The front group is a hundred ends.
And the empty shafts stay in the loom. The harness has a depth, and every shaft’s tension depends on how far behind the fell it hangs; a design that empties twenty frames needs them still in place, or the tensions the design was computed at are not the ones the warp sees.
How the logs and designs were computed
The yarn, the harness and the spread’s tail are the earlier log’s: a 20 tex ring-spun cotton, two strength laws fitted to the same expected breaking loads, a warp of 4,000 ends by 1,000 metres on twenty-four shafts whose tensions rise evenly from 14 to 52 per cent of the 500-millimetre breaking load, and each shaft’s expected spread breaks per end from the tail no tensile test reaches. The faults arrive at one per hundred kilometres of yarn and each has a strength drawn from a lognormal law with the stated median and log-spread; a fault breaks on a shaft if its strength is below the shaft’s tension, so a shaft’s expected fault breaks per end are the rate times the share of the population below its tension.
The divergence of a design is the expected log-likelihood ratio of one beam’s log under the lognormal law against the normal law with the fault rate and fault median that fit it best — Poisson counts, so a sum over shafts of , each weighted by the ends on that shaft. For a given median it is convex in the fault rate and the rate is found by a ternary search; the median is scanned on a logarithmic grid and the best cell refined by a golden-section search. Beams are the natural log of a hundred over the divergence. The fault log-spread is shared by the two laws and taken as known.
The best threading is found by Fedorov–Wynn exchange: 160 rounds over all twenty-four shafts, moving a share 1/(n + 2) of the warp at round n to the shaft adding the most divergence per end; shafts holding less than 2 per cent are then emptied and 640 further rounds run on those that remain. Required of it, and checked: that with faults all weaker than the front shaft and only the rate read, the divergence is the earlier log’s own to a part in a thousand; that reading the median as well never tells the laws apart better; that the weak-fault design has two groups and the median-25 one three; that the divergence per end is level within a tenth across the shafts used and no empty shaft exceeds it by more than 5 per cent; and that the best design beats both the straight draft and the front-and-back pairs. A fault population with no median or a negative spread, a negative fault rate, a threading of the wrong length and a log asked to read something other than its floor or its median are refused.
What the drawing cannot show, and what the model leaves out
The faults’ strengths. The law is lognormal because a strength law has to be something, and a median of a quarter of the breaking load is a guess at the faults that matter: the ones a winder’s clearer passes, which are therefore not grossly thin, and which still break on a loom. Every number above moves with that guess. What does not move is the shape: a floor group that walks inward as faults strengthen and a third group that appears once some faults are weaker than the front shaft and most are not.
Two laws, one alternative. The design discriminates lognormal from normal. A Weibull yarn, or a question about all three at once, is a different design with a different best-fitting alternative for each pair; the exchange runs the same way and was not run.
The fault spread as known. The log-spread is shared and fixed. Reading it from the log too would add a fourth unknown, and Elfving’s count says a fourth group might then be wanted; where it would go is not computed.
The scatter. Counts scatter only as chance does. Real logs carry more — beams differ, breaks are logged against the wrong shaft, and a periodic weak place repeats where a random one would not — and extra scatter raises every beam count here in proportion without changing which design is best.
Who worked what out
Logging warp breaks by shaft is long-standing weaving practice. The weakest-link reading of a yarn’s strength is Peirce’s, from 1926. Optimal design for discriminating between two models — the T-optimal criterion — is Atkinson and Fedorov’s, from 1975, and the exchange that finds such a design descends from Fedorov’s and Wynn’s work around 1970; Elfving’s 1952 rule on support points is the older ancestor. What is added here is the fault population with its own strengths in a loom’s break log; the finding that the front-and-back harness becomes worse than a straight draft once faults are stronger than the front shaft’s tension; that the best threading’s floor-reading group walks into the middle of the harness as faults strengthen, to the place where every fault has broken and no yarn yet has; that a small third group appears near the front once the faults are strong or widely spread; and that the design’s optimality can be checked shaft by shaft by the level of the divergence each one adds.
Still open: what a mill’s faults actually break at
Every design here waits on one measurement nobody makes. A fault’s strength is the load at which it parts, and a clearer’s cut log already holds the faults it removed — not the ones that passed, which are the ones that matter on the loom. Testing the passed faults is a tensile test of the yarn at the places a clearer marks as near its limit, which an electronic clearer can be set to flag without cutting.
A few hundred such tests would give a median and a spread for one mill’s faults, and the designs above would turn from a family of answers into one. The prediction is a check on itself: if the faults’ median is below the front shaft’s tension, the earlier two-ended harness is right and the third group never appears. If it is above, the mill’s front shafts have been logging a fraction of its faults all along, and a break log read the old way has been underestimating its floor by exactly the share the front shaft misses.
Shares its objects with
Essays naming at least two of the same things, that neither author linked.
- The quietest shafts belong at the back — both name harness, heddle, shed, threading
- A figure is harder on its warp — both name heddle, shed, threading
- The back shaft works hardest — both name heddle, shed, threading
- Three mistakes and the shape each one leaves — both name fault, harness, threading
- A jacquard harness needs three half-spans of height — both name harness, shed
- A jacquard is every end its own shaft — both name harness, threading
Named objects
A flat tag is an object no other essay names yet.