Setting and geometry

A harness reads a warp's tail best threaded at both ends

A break log could only tell a lognormal yarn from a normal one in twenty-five beams, and the obvious remedy was to thread more ends at the back, where the tail breaks. It cannot work the way it was expected to: a shaft's spread breaks and its fault breaks both scale with its ends, so the cliff never rises over the floor. What works is a harness threaded at its two ends and nowhere between — six beams for a plain weave, nine for a twill.

Worth reading first: A break log reads the tail only above its fault floor · A warp breaks in a tail no tensile test reaches · Where the heddles go.

A break log reads the tail only above its fault floor turned a loom’s record of warp breaks into a reading of a yarn’s far tail — the tail that no tensile test reaches, because a test samples the middle. A harness of twenty-four shafts holds the warp at twenty-four tensions, from 14 per cent of the breaking load at the front to 52 at the back, and each shaft logs the weak places that fall below its own. Discrete faults — slubs, bad piecings, fly woven in — break wherever they arrive, so they lay a flat floor under the log, and the tail’s breaks stand on it as a cliff toward the back. Read that way, a season’s log can say which strength law a yarn follows, and the hard question, a lognormal tail against a normal one, takes twenty-two beams at one fault in a hundred kilometres.

That essay ended on a remedy. Every shaft had carried the same number of ends, because a straight draft threads a harness that way, and nothing requires it. Put more ends at the back, it predicted, and the cliff would stand higher above the floor exactly where the log reads it.

The prediction is wrong about its mechanism and right that the threading is a lever. This essay takes the threading as the thing to be designed and finds that the best harness for reading a tail is not the one heavy at the back. It is the one threaded at its two ends and empty between.

A break log from a harness threaded on shafts 1, 2, 23, 24 only, with 40 per cent of the warp at the frontThe breaks a 24-shaft loom would log over one beam when its 4,000 ends hang on shafts 1, 2, 23, 24 only, 40 per cent of them on the front 2 and the rest on the back 2, at 1 per 100 km of discrete faults. Under the lognormal law the front shafts log 8.0 each and the back shaft 12; under the normal law the front shafts log 8.0 each and the back shaft 18. The front shafts log faults and nothing else, so they read the floor; the back shafts log the floor scaled by their own ends plus the spread's cliff. Read with the floor taken from the log itself, this threading tells a lognormal yarn from a normal one by a hundred to one in 6.1 beams, against 25 for the straight draft. The empty shafts are still in the loom and still set the tensions of the ones in use. The picture holds every fault weaker than the front shaft's tension.the front shafts read the floor and the back shafts the cliff, and nothing between is neededone beam · 1 per 100 km · 40% of 4,000 ends on shafts 1, 2, the rest on 23, 24 14812162024051015202530shaft, front to back (empty shafts log nothing)breaks logged on the shaft, per beamshafts 3 to 22 carry no endsbut stay in the loom, and set the tensions of the restlognormalnormalfaults alone6.1 beams to 100:1, floor read from the logstraight draft 25
Fig. 1 A plain weave’s two classes of end hung on shafts 1, 2, 23 and 24 of a twenty-four-shaft harness, forty per cent of the warp at the front, and the breaks each shaft would log over one beam at one fault in a hundred kilometres. The front shafts log eight breaks each under both laws, every one a fault. The back shafts log twelve under the lognormal law and eighteen under the normal. Read with the floor taken from the log itself, this harness tells the two laws apart by a hundred to one in 6.1 beams, against 25 for a straight draft. The slider moves the share of the warp at the front.

The cliff and the floor rise together

Start with the remedy as stated, because the reason it fails is one line long.

A shaft’s spread breaks are the number of ends it carries times the chance that one end’s thousand metres hold a place weaker than that shaft’s tension — a yarn breaks at its thinnest place, and a shaft is a tensile test with a whole beam between its clamps. Its fault breaks are the number of ends it carries times the fault rate per end. Both are proportional to the ends. Double the ends on a shaft and both double; so the ratio of one to the other — how far the cliff stands above the floor on that shaft — is fixed by the shaft’s tension and the yarn, and no threading can touch it.

A normal yarn's break log under a straight draft and under one threaded heavier toward the back. The breaks a 24-shaft loom would log on each shaft over one beam of a normal yarn at 1 per 100 km, with the warp threaded straight (167 ends a shaft) and with each shaft carrying ends in proportion to its number, so the back shaft holds 320 and the front 13. Solid lines are the whole log and dashed lines the fault floor under each. On the back shaft the straight draft logs 2.5 over a floor of 1.7 and the heavy-backed one 4.7 over 3.2: on a logarithmic scale each curve stands exactly as far above its own floor as the other does, on every shaft, because the spread's breaks and the faults both scale with the ends a shaft carries. The picture cannot show what the heavier back does buy, which is counts rather than contrast.
Fig. 2 A normal yarn’s log under the straight draft, 167 ends a shaft, and under a draft heavier toward the back, with ends in proportion to the shaft’s number — 13 at the front and 320 at the back. Solid lines are everything logged and dashed lines the faults alone. The heavy-backed log is higher at the back and lower at the front, and on a logarithmic scale each solid line stands exactly as far above its own dashed floor as the other does, shaft by shaft.

The numbers make it concrete. On the back shaft, at one fault per hundred kilometres, the straight draft’s normal yarn logs 2.5 breaks over a floor of 1.7; the heavy-backed draft logs 4.7 over a floor of 3.2. Both are a cliff of 0.47 breaks for every fault break, which is the back shaft’s tension speaking and nothing else. The front shaft of the heavy-backed draft logs less of everything, a tenth of a break a beam, and the cliff there is still the same tiny fraction of the floor it always was.

So a heavier back does not make the cliff more visible above the floor. It makes the log bigger at the back. Whether that is worth anything is a different question, and the answer is not obvious, because the same ends had to come from somewhere.

What a heavier back does buy

It buys counts. A log is a set of counts, each scattered as chance scatters a count of rare events, and what decides between two laws is how far apart the counts each law predicts are, measured against that scatter. Twice the ends on a shaft is twice the expected count under both laws, and the scatter of a count grows only as its square root. So the shafts where the two laws disagree most are worth more ends, and the shafts where they agree are worth fewer.

Under the normal and lognormal laws, those shafts are the last two or three, the ones that work hardest because their lift is largest. On a straight draft at one fault in a hundred kilometres, the back shaft alone carries 68 per cent of the evidence between the two laws, the twenty-third another 22, and the first eighteen shafts together less than one part in a thousand. The front of a straight harness is spending two thirds of the warp on shafts where the laws make identical predictions.

The heavy-backed draft moves ends from there to where the evidence is, and the log improves: 11.8 beams to a hundred to one, against 22.2 for the straight draft, with the fault rate known. Not because the cliff stands higher — it does not — but because the shafts on which the two laws differ now hold more of the warp.

That suggests pushing to the limit: every end on the back shaft. With the fault rate known in advance, that log decides in 1.36 beams, sixteen times faster than the straight draft. It is the best any threading of this warp can do, and it is not available, for a reason that is the centre of the whole problem.

Nobody knows the floor in advance

The fault rate is a property of the yarn lot, of how well the winder’s clearer was set, of the fly in the shed that week. A mill does not know it before the beam is woven. It reads it from the log — from the front shafts, where the spread makes no breaks under any law and every break is a fault.

So the honest comparison lets the alternative law have whatever floor fits the log best. A log that cannot pin the floor down cannot use it. Measured that way the straight draft needs 24.9 beams rather than 22.2: the front shafts read the floor well enough that not knowing it costs only an eighth more beams.

A harness with every end on its back shaft has no front shafts. Its log is one count, the back shaft’s breaks, and either law can produce any count at all. A lognormal yarn at one fault per hundred kilometres logs exactly what a normal yarn at 0.53 does, because on that shaft the normal yarn’s cliff supplies the other 0.47 — on a single shaft a floor and a cliff are the same thing. The best possible design with the floor known becomes the worst possible design with the floor read — it cannot tell the laws apart in any number of beams.

Six ways to thread one warp, and the beams each needs to tell a lognormal yarn from a normal one. Beams of break log to a hundred to one between the lognormal and normal laws at 1 per 100 km, for six threadings of the same 4,000 ends on a 24-shaft loom, each with the fault rate known in advance and with it read from the log: straight, 22 known and 25 read; heavy back, 12 known and 15 read; back pair, 2.0 known and 20 read; split pair, 3.4 known and 6.1 read; back four, 3.7 known and 11 read; split four, 5.0 known and 9.1 read. Knowing the floor, the back shafts alone are best by far. Reading it, the back pair loses most of that and a single back shaft would lose all of it, because a higher fault floor imitates the normal law's cliff exactly when every end sits at one tension; a split keeps most of the gain. A real mill reads its floor. The picture assumes counts that scatter only as chance does.
Fig. 3 Six threadings of the same 4,000 ends and the beams of log each needs to separate a lognormal yarn from a normal one, at one fault per hundred kilometres. Squares assume the fault rate is known in advance; circles read it from the log. Straight, 22 and 25; heavier toward the back, 12 and 15; every end on the back pair, 2.0 and 20; the pair split with the front pair, 3.4 and 6.1; every end on the back four, 3.7 and 11; the four split with the front four, 5.0 and 9.1.

The back pair shows the same thing less starkly. Two shafts at slightly different tensions do carry a little information about the floor — under a normal yarn the twenty-fourth logs more than the twenty-third by a ratio a pure floor cannot imitate — and that is worth something, but not much: two beams with the floor known, twenty with it read. The difference between the square and the circle on that column is the price of having no front.

Two ends of the harness, and nothing between

Add the front back in, but only the front. Hang a plain weave’s two classes of end on shafts 1 and 2 and on shafts 23 and 24 — each class split between a front shaft and a back shaft carrying the same lifting plan, which where the heddles go established is always allowed, since two shafts lifted together lift their ends together. The twenty shafts between stay in the loom and set the tensions of the four in use, and carry nothing.

The front pair reads the floor and nothing else; the back pair reads the floor plus the cliff, scaled by its own ends. Two readings of one floor at two known scales is exactly what is needed to separate it from the cliff, and nothing between adds to either reading.

Beams of break log needed to tell a lognormal yarn from a normal one, against the share of the warp threaded at the front. For a harness that uses only its front k shafts and its back k, with a share of the 4,000 ends on the front ones, the number of beams a log must cover to favour the lognormal law over the normal by a hundred to one on average, with the fault rate (1 per 100 km) read from the log itself. For plain weave, front and back pairs the best share is 40 per cent, at 6.1 beams; for 2/2 twill, front and back fours the best share is 25 per cent, at 9.1 beams; for an eight-class weave, front and back eights the best share is next to nothing at the front, at 11 beams. The straight draft needs 25 and a draft threaded in proportion to the shaft number 15. With no ends at the front a single back shaft cannot tell the laws apart at all, and a block of several does so only through the spread of tensions inside it. The picture assumes counts that scatter only as chance does.
Fig. 4 The beams of log needed against the share of the warp on the front shafts, for harnesses using their front and back pairs, fours and eights, with the floor read from the log. The pair is best at 40 per cent at the front, 6.1 beams; the four at 25 per cent, 9.1 beams; the eight at next to nothing at the front, 11 beams. The straight draft’s 25 and the heavier-back draft’s 15 are ruled across.

The pair’s curve has its minimum at forty per cent of the warp at the front, 6.1 beams — a quarter of the straight draft’s twenty-five, and with the floor genuinely read rather than assumed. Its two ends are why the curve is a valley. With too little at the front the floor reading is noisy and the back pair’s cliff cannot be told from a raised floor; with too much, the back pair’s counts are too small for the cliff to show. Between, there is a best share, and it is not a half.

That this is a plain weave matters. Plain weave has two classes of end, so two shafts at the front and two at the back is the smallest harness that can weave it this way, and a dobby with spare shafts weaving a plain cloth is not an exotic arrangement: heavy plain warps are commonly spread over four or more shafts precisely to spread the heddles. The split puts 800 heddles on each front shaft and 1,200 on each back one, against 167 a shaft straight, and a two-shaft loom weaving the same warp would hang 2,000 on each. The heddle loads are those of an ordinary small harness.

A twill, and a block that reads its own floor

A 2/2 twill has four classes — a twill’s rule gives four distinct ends a repeat — so its version of the design is four shafts at the front and four at the back. Its best share is a quarter at the front, and it needs 9.1 beams.

A break log from a harness threaded on shafts 1, 2, 3, 4, 21, 22, 23, 24 only, with 25 per cent of the warp at the front. The breaks a 24-shaft loom would log over one beam when its 4,000 ends hang on shafts 1, 2, 3, 4, 21, 22, 23, 24 only, 25 per cent of them on the front 4 and the rest on the back 4, at 1 per 100 km of discrete faults. Under the lognormal law the front shafts log 2.5 each and the back shaft 7.5; under the normal law the front shafts log 2.5 each and the back shaft 11. The front shafts log faults and nothing else, so they read the floor; the back shafts log the floor scaled by their own ends plus the spread's cliff. Read with the floor taken from the log itself, this threading tells a lognormal yarn from a normal one by a hundred to one in 9.1 beams, against 25 for the straight draft. The empty shafts are still in the loom and still set the tensions of the ones in use. The picture holds every fault weaker than the front shaft's tension.
Fig. 5 A 2/2 twill on shafts 1 to 4 and 21 to 24, a quarter of the warp at the front. The front shafts log 2.5 breaks each, all faults; the back four log the floor scaled to their 750 ends each, plus the cliff, reaching 7.5 on the back shaft under the lognormal law and 11 under the normal. This threading separates the two laws in 9.1 beams with the floor read from the log.

The twill teaches something the plain weave hid. With no ends at the front at all, the twill’s back four still separate the laws in 10.6 beams — barely worse than the best split. The reason is that four shafts at the back are four tensions, from 47 to 52 per cent of the breaking load, and across even that short range a normal yarn’s breaks per end rise more than sixfold while a floor stays flat. The back block reads its own floor from its own slope. The front four add only the last seventh.

Push to eight classes and the effect completes itself: the best eight-shaft design puts next to nothing at the front, because the back eight span 40 to 52 per cent of the breaking load and that is span enough to read a floor from. What the front shafts supply is not “the front” as such. It is a second tension far from the first, and a wide enough back block can supply its own.

The rule is older than any loom

That pattern — the evidence lives at the ends of the range, and the best experiment puts everything there — is not a fact about textiles. It is the oldest result in the design of experiments. To measure the slope of a straight line through points whose height is uncertain, the best allocation of a fixed number of observations is half at each end of the available range and none in the middle; Gustav Elfving gave the geometric account of why, for any linear model, in 1952.

The break log is that problem wearing a harness. Its two unknowns are a floor and a cliff, the one flat and the other rising steeply with tension, and every shaft is an observation at a tension the shed geometry fixes. A straight draft is the textbook’s bad experiment, observations spread evenly across the whole range; the split is the textbook’s good one. That the best share is forty per cent rather than half is because the two “ends” are not symmetric: the front measures one unknown purely, and the back measures both.

It is also the reason the quietest shafts belong at the back does not settle anything here. That essay ordered shafts by how much each end flexes and wanted the ends that change face least on the back shafts, where the lift is largest. Here the order is fixed — front is low tension and back is high — and what is chosen is how much warp each end of the order carries. The two are compatible: within the back block, the quiet classes can still go furthest back.

Where the best share moves as faults grow

A mill does not know its fault rate before it threads either, so it matters how sensitive the design is to it.

Beams of break log needed under a straight draft and two split threadings, against the rate of discrete faults. Beams to a hundred to one between the lognormal and normal laws, with the fault rate read from the log, against the rate of discrete faults, for the straight draft and for the best split of the warp between front and back pairs and between front and back fours: at 0.01 per 100 km, 3.3 straight, 0.37 on pairs (1 per cent at the front) and 0.60 on fours (1 per cent); at 1 per 100 km, 25 straight, 6.1 on pairs (40 per cent at the front) and 9.1 on fours (25 per cent); at 100 per 100 km, 2,060 straight, 592 on pairs (46 per cent at the front) and 835 on fours (29 per cent). All three grow in proportion to the floor once it is high, so the split's advantage settles to a constant factor rather than vanishing, and the share it puts at the front rises as the floor comes to matter. The picture assumes one fault rate, constant along each beam and the same from beam to beam.
Fig. 6 Beams of log needed against the fault rate, from a hundredth to a hundred per hundred kilometres, for the straight draft and for the best splits on pairs and on fours, each at its own best share. At a hundredth: 3.3 beams straight, 0.37 on pairs, 0.60 on fours. At one: 25, 6.1 and 9.1. At a hundred: 2,060, 592 and 835.

At very low fault rates the front hardly matters, because there is almost no floor to read, and the best share falls to a per cent or so: the split becomes the back block, and it beats the straight draft by a factor of five to nine. As faults grow the floor matters more, the best share rises — to 40 per cent for the pair at one fault per hundred kilometres and 46 per cent at a hundred — and the advantage settles. It does not vanish. Once the floor dominates, every design’s beams grow in proportion to it, and the split pair stays three and a half times faster than the straight draft, the split four two and a half.

The design is forgiving where it needs to be. The pair’s valley is broad: anywhere from a fifth to three fifths of the warp at the front stays within a fifth of the best, so a share chosen for one fault per hundred kilometres serves a mill whose rate turns out to be half or double that.

What the instrument costs in stops

Every end at the back is an end held at half its breaking load, and an end that breaks stops the loom. The instrument is not free, and its price depends on the yarn.

Warp breaks a beam under four threadings, for a yarn of each of the three strength laws. Expected warp breaks over one beam at 1 per 100 km for four threadings of the same warp, under each of the three laws: straight, 42 normal, 40 lognormal, 103 Weibull; heavy back, 43 normal, 40 lognormal, 153 Weibull; split pair, 49 normal, 40 lognormal, 321 Weibull; split four, 47 normal, 40 lognormal, 285 Weibull. A lognormal yarn breaks only at its faults and costs the same however it is threaded. A normal one costs a split an eighth to a sixth more stops. A Weibull yarn costs a split nearly three times the straight draft's — but a straight log tells a Weibull from either other law inside a tenth of a beam, so the yarn that makes the split expensive is the one that never needed it. The picture holds the fault rate fixed.
Fig. 7 Expected warp breaks over one beam at one fault per hundred kilometres, for four threadings, under each law. A lognormal yarn logs 40 whatever the threading, all faults. A normal yarn logs 42 straight, 43 heavier at the back, 49 on the split pair and 47 on the split four. A Weibull yarn logs 103 straight, 153, 321 and 285.

A lognormal yarn costs nothing extra: its tail is too far out for the back shafts to find, and its forty breaks a beam are all faults whichever shaft they land on. A normal yarn costs an eighth to a sixth more stops on a split — forty-seven or forty-nine against forty-two. A Weibull yarn costs three times as many, since a Weibull tail is heavy enough that piling the warp at 52 per cent of its breaking load finds a great many weak places.

That last number looks like a fatal objection and turns out to be the opposite. A straight log tells a Weibull yarn from either other law in 0.12 of a beam — an eighth of one warp’s length. The yarn that makes the split expensive is the one yarn the split was never needed for.

Thread the first warp straight

So the instrument has an order of use, and it falls out of the stops rather than being imposed. Weave the first warp of a new yarn straight. A fraction of a beam settles whether the tail is Weibull-heavy, and if it is, the question is answered and the straight draft is the cheaper one to keep. If it is not, the remaining question is normal against lognormal, the yarn’s tail is light, and the split costs at most a sixth more stops for a quarter of the beams.

The threading is fixed when the warp is drawn in, and a following warp tied on to the old ends keeps it, so the choice is made once a warp rather than once a beam. A mill could run its instrument warps on one loom and its production on the rest, which is how a test is usually run.

How the log was computed

The yarn and the harness are the earlier log’s: a 20 tex ring-spun cotton at a spinning index of 1.35, three strength laws (normal, lognormal and Weibull) each fitted to the same expected breaking loads at 100 and 500 millimetres, and 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. Discrete faults break on whatever shaft they reach, at a rate per hundred kilometres of yarn.

A threading is a number of ends on each shaft summing to 4,000. Four kinds were run: straight, 166.7 a shaft; heavier toward the back, in proportion to the shaft’s number; every end on the back k shafts; and a split putting a share on the front k and the rest on the back k, equally within each block, for k of two, four and eight. Each shaft’s expected breaks are its ends times the sum of the fault rate per end and the law’s weak places per end below that shaft’s tension.

The beams needed are the natural log of a hundred over the expected log-likelihood ratio of one beam’s log under the lognormal law against the normal, which for independent Poisson counts is the sum over shafts of λAln⁡(λA/λB)−λA+λB\lambda_A\ln(\lambda_A/\lambda_B)-\lambda_A+\lambda_B. With the floor read rather than known, that sum is minimised over the normal law’s fault rate; it is convex in that rate, because each term is convex in its mean and the mean is linear in the rate, so a one-dimensional search finds the minimum. The best share was searched over steps of a quarter of a per cent.

Required of it, and not shown: that a straight threading reproduces the earlier log to nine figures; that the cliff over the floor on every shaft is identical under the straight and heavy-backed threadings, for all three laws; that doubling every shaft’s ends doubles the log; that reading the floor never tells the laws apart better than knowing it; that every end on the one back shaft cannot tell them apart once the floor is read; and that the best four-shaft split beats the straight draft. Seven malformed threadings — a harness of one shaft, negative ends, an empty harness, a split with everything at the front, a block wider than half the harness, an unnamed kind, a floor that is neither known nor read — are refused.

What the drawing cannot show, and what the model leaves out

The shafts in between. The figures draw the empty shafts as gaps and say they still set the tensions of the rest. A shaft’s tension in the shed comes from its distance from the fell and the lift that distance requires, which the harness has a depth computed; an empty frame still occupies its place, and a loom whose middle frames are removed altogether would bring the back shafts forward and lower their tension. The design needs the frames, not the ends on them.

The tensions themselves. They are taken from the shed’s own extension, shaft by shaft, with a common let-off tension and nothing else. A back shaft carrying 1,200 heddles is heavier to lift than one carrying 167, and a heavier frame may be driven differently; nothing here says whether that changes the strain an end sees, and what the shed costs in newtons prices the force rather than its effect on the frame’s motion.

The faults’ own strengths. Every fault is taken to be weaker than the front shaft, which is what makes the floor flat and the front a pure fault counter. A fault population with a spread of strengths near the working tensions would put a second, gentler cliff under the first, and the front shafts would read only the weakest faults.

The counts’ scatter. Each count is taken to scatter only as chance does. Real logs carry more — beams differ in sizing and lot, breaks are logged against the wrong shaft — and a periodic weak place from a drafting roller breaks the same ends repeatedly, a pattern a random error hides and a periodic one shows and no Poisson model has. Extra scatter raises every beam count here in proportion, the straight draft’s as much as the split’s, and leaves the ranking alone.

Who measured what

Logging warp breaks by shaft, and knowing that the back shafts break more, are long-standing weaving practice. The weakest-link reading of a yarn’s strength is Peirce’s, from 1926, and the optimal allocation of observations at the ends of a range is Elfving’s, from 1952, with Kiefer’s general theory of optimal designs following through the 1950s. What is added here is the observation that a harness’s threading is an experimental design; that the cliff over the floor on a shaft is fixed by its tension and cannot be raised by threading; that the threading which reads a yarn’s tail fastest, with the floor known, is the worst possible once the floor must be read; and the split threadings — a plain weave on its front and back pairs, a twill on its front and back fours — that bring a season’s question within six to nine beams, with their costs in stops under each law.

Still open: whether a third group is needed once faults have strengths

Elfving’s rule carries a count with it: an optimal design needs at least as many distinct points of observation as the model has unknowns. The log here has two — a floor and a cliff — and the best harness uses two groups of shafts, or one group wide enough to span two tensions.

A fault population with its own spread of strengths would add a third unknown, the faults’ own cliff, rising gently through the middle of the harness where the spread’s cliff is still flat. The prediction is that the best harness would then acquire a third group, somewhere near the middle of the shed’s tension range where the two cliffs differ most, and that its share would measure how strong a mill’s faults are — a property no fault classifier reports, since it counts faults by size and not by strength. Whether that third group is worth its shafts depends on how wide a real fault population’s strengths are, and on how many shafts a draft needs to leave free; both are measurable, and neither is measured here.

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.

Named objects

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

FaultHarnessHeddleShedThreadingWeakest-link