A course is one thread and a warp is many

Barré — a horizontal shading in a knitted fabric, a band a few courses deep that is slightly darker or slightly wider than the fabric above and below it — is the defect that knitters fear most and weavers rarely see. It is usually attributed to the yarn, and that attribution is correct as far as it goes.

But the same yarn woven does not do it. The difference is not in the fibre, the dye, the twist or the count. It is in how many threads the fabric is made of at once.

A woven fabric’s warp is two thousand threads side by side, drawn from two thousand packages. A weft knit’s course is one thread, drawn from one.

The claim #

A woven fabric averages its yarn’s package-to-package variation across its own width, and a weft knit does not average it at all.

A warp is a parallel structure: each end comes from a different package, so a difference between packages is an independent error per end, and independent errors are divided by the square root of how many are in view. A knit is a serial structure: one thread makes a whole course, so a difference between packages is one error applied identically to every stitch in that course — coherence in its purest form.

That is the whole mechanism, and it explains the asymmetry without appealing to anything about knitting or about dye uptake.

The argument #

Consider a yarn whose count drifts slowly over the length of a package — a per cent or two, over hundreds of metres, which is entirely ordinary.

In a warp, that drift is a difference between one end and another. Each end runs the whole length of the piece at whatever count its own package supplies, so the fabric has two thousand slightly different ends running down it. A drift between packages produces a stripe down the cloth, one end wide, at a contrast of a per cent or two — which is invisible, because one end in two thousand is a twentieth of a per cent of the width.

In a weft knit, the same drift is a difference between one course and another. Every stitch across the whole width of that course comes from the same yarn at the same moment, so the entire course is a per cent or two different from its neighbours. That is a band the full width of the fabric, and it is visible.

The same variation, in the same yarn, at the same amplitude. What differs is whether the fabric’s structure runs it across the direction it is being looked at or along it.

The half of the asymmetry that is about weaving #

The comparison is not quite between knitting and weaving. It is between a warp and everything else, and a woven pick behaves exactly like a knitted course.

A woven fabric has this problem in one direction and not the other. Its weft comes from one package at a time, so a weft-way drift makes bars — and weft bars are a real and familiar woven defect. Its warp comes from thousands, so a warp-way drift is averaged away.

So the honest statement is: a fabric averages a yarn’s variation in whichever direction it uses many threads at once, and shows it in whichever direction it uses one. A woven cloth averages in one direction and a weft knit averages in neither, because a weft knit uses one thread in the course direction and re-uses that same thread, course after course, in the other.

That is what makes the knit’s case worse rather than merely different. A warp knit does not have the problem at all, because it is fed from a beam of many ends exactly as a woven warp is, which is a structural argument for warp knitting that has nothing to do with speed or with stability.

Why the same argument makes plating work #

There is a second remedy in the trade, older than multi-feeder rotation and more reliable, and the population argument says why it is the better one.

Plating knits two yarns together as one stitch, one lying on the face and one on the back — two yarns where the loop normally has one. Each stitch then carries the mean of two independent sources rather than one, so the between-package variation is divided by the square root of two — a modest gain — and, far more importantly, the gain is achieved without imposing any period at all. Both yarns are present in every course, so there is no arrangement for the eye to lock onto.

That is a genuinely different kind of remedy from rotation. Rotation reduces the amplitude by F and creates a period; plating reduces it by √2 and creates nothing. By the ratio above, a period is worth about thirteen at ordinary viewing widths, so plating’s √2 with no period beats rotation’s four with one whenever the packages are not extremely well matched.

The comparison is worth putting in one line, because it is the sort of thing that is decided by habit rather than by arithmetic:

The middle row’s entry is a range rather than a number because it depends on a quantity nobody measures: how different the four packages are from each other. That is the whole reason the row is contentious in practice — the remedy’s worth is decided by a number that does not appear on any certificate.

The remedy, and when it makes things worse #

The trade’s answer to barré is to feed the machine from several packages in rotation — four, six, eight feeders, each taking its own package, so that consecutive courses come from different yarn and no single package’s drift owns a band.

The averaging is real and the arithmetic is simple: a drift spread over F feeders appears at a fraction of its amplitude in any one course.

But it does not remove the error. It changes its arrangement, from a slow drift to a period of F courses — and by the arithmetic of the previous rung in this ladder, that is a bad trade unless the numbers are right.

A periodic error is √(2n/π) times as visible as a random one of the same size, where n is how much fabric is in view. Over two hundred and fifty-six courses that is 12.8. So feeding from four packages helps only if the differences between those four packages are more than about thirteen times smaller than the drift being suppressed. If they are not, the machine has converted an invisible gradual shading into a visible regular stripe.

And it explains why the number of feeders is chosen as it is. More feeders means better averaging of the drift and a finer period for whatever residual difference remains — and a finer period, at some point, stops being read as a stripe at all and becomes texture. So the remedy works at both ends of the feeder count and fails in the middle, which matches practice: knitters use either very few feeders with carefully matched packages, or a great many.

What the loop adds that the weave does not #

There is a second mechanism in a knit, and it works in the same direction.

A knit’s dimensions come from its loop length and from very little else: the width of the fabric per wale, the courses per centimetre, the whole geometry of the relaxed structure follow from one length of yarn per stitch. So a variation in the yarn does not merely change the appearance of a course. It changes that course’s dimensions.

A woven cloth has no equivalent. Its geometry is set by the reed and the take-up, which are machine settings and do not vary with the yarn — so a thick pick in a woven cloth is a thick pick in a fabric whose spacing is unchanged. A thick course in a knit is a course of different length, in a fabric that will find its own width.

The consequence is that a knit’s bands are not only optical. A band of longer loops is a band of looser fabric, wider per wale, and one of shorter loops is narrower — so a knitted tube with barré in it is not cylindrical. That is the same fault reported as a dimensional defect rather than an appearance one, and the two are not usually connected.

Counting the independent sources #

The comparison can be made exact, and the numbers are startling enough to be worth setting down.

Take a square metre of fabric, at ordinary constructions, and count how many independent yarn sources contribute to it.

The woven warp and the warp knit are parallel structures with thousands of sources. The weft of anything, and the whole of a weft knit, are serial structures with one.

A structure with two thousand sources divides a between-source variation by about forty-five before it becomes visible. A structure with one divides it by nothing. That factor of forty-five is the whole of why the same yarn behaves so differently in the two fabrics, and it is why the remedy for barré is always some way of manufacturing extra sources — rotating feeders, plating two yarns together, blending packages at winding — rather than improving the yarn.

And it says which fabrics are worth making from a yarn known to drift. A woven warp will bury it. A weft knit will display it, in the one direction a garment is looked at across. Nothing about the yarn’s certificate distinguishes these two outcomes, because the certificate reports a coefficient of variation and the certificate has no way of saying whether the variation is between packages or within one — which is exactly the distinction that decides everything above.

The averaging is in the wrong direction to be useful #

There is a detail in the parallel case that the square-root argument passes over, and it decides whether the averaging is worth anything at all.

A warp’s two thousand sources are arranged across the cloth, and the drift they carry runs along it. So each end is not a sample of a population that averages within itself; it is one source, running the whole length of the piece at its own value. What the fabric averages is the difference between neighbours, seen sideways.

That is why the argument works, and it is also why it works only for a slow drift. The claim that a warp buries variation is really the claim that an eye integrating across many ends cannot resolve a one-end stripe of a per cent — a statement about resolution, not about arithmetic. If the same variation were arranged as a band of ends rather than as one end at a time, nothing would average: a hundred adjacent ends all a per cent heavy is a hundred-end stripe, and it is as visible as any knitted band.

And that is exactly what a beam warped badly delivers. Ends are drawn onto a warping beam in sections, so a section wound from one creel setting shares whatever that creel was carrying — which turns the warp’s independent sources into a much smaller number of grouped ones. A warp warped in twenty sections has twenty sources, not two thousand, for anything that varies between creel loads.

So the factor of forty-five is a best case that assumes the ends were randomised, and nothing in a warping room randomises them. The woven warp’s advantage is real and is an advantage of the beaming practice as much as of the structure, which is worth separating because only one of the two can be improved.

The knitted case has no such caveat, in either direction. One thread per course is one thread per course however the packages were wound, so the knit’s exposure is a structural fact and the warp’s protection is a contingent one.

That asymmetry is worth carrying because it says where the effort belongs. Improving a knitted fabric’s barré means manufacturing extra sources, which is what every remedy above does. Improving a woven cloth’s warp-way uniformity means keeping the sources it already has independent — a warping practice rather than a fabric decision, and one that costs nothing when it is done and cannot be recovered afterwards.

Where the model stops #

The drift is treated as a single number per package. A real package varies along its own length in a complicated way — a drift, plus short-term variation, plus whatever the winding imposed — and only the slow part is what the argument above is about.

Dye uptake is not modelled at all, and it is the largest contributor to real barré. A yarn that differs slightly in draw ratio, twist or fibre blend dyes differently, and the resulting difference in shade is far more visible than a difference in count. The structural argument here says only why the difference is arranged in bands; how large the visible contrast is depends on the dye, which this collection has no instrument for.

The feeder arithmetic assumes independent packages, and packages from the same spinning frame position are not independent. If four feeders are supplied from four packages that were wound side by side, they share whatever their frame was doing, and the remedy delivers less than the count suggests.

And the loop-length mechanism assumes positive feed. A machine metering yarn by length gives every stitch the same length whatever the count; one running on tension does not, and the two respond to a thick place in opposite directions. Which of the two is in the machine decides the sign of the dimensional effect.

The generalisation #

Whether variation averages depends on whether the structure is parallel or serial in the direction being looked at, and that is a property of how a thing is made rather than of what it is made from.

The diagnostic question is worth stating plainly: how many independent sources contribute to what is being seen at one time? Many, and their variation averages down as the square root. One, and it comes through undiminished, applied identically to everything that source touched.

The second lesson is that averaging by rotation is not free. Spreading one source’s variation across several outputs in rotation reduces the amplitude and imposes a period, and a period is worth an order of magnitude in visibility. That trade is usually made without noticing that it is a trade — the amplitude is what gets measured and the coherence is not.

Who found it, and when #

Barré is as old as knitting and its association with yarn variation is universal; the multi-feeder remedy is standard practice and so is the feeder stripe it sometimes produces. What appears not to be written down is the exchange rate between them — that suppressing a drift by rotating packages is a good trade only when the package-to-package difference is more than an order of magnitude below the drift, and that the order of magnitude is √(2n/π) and therefore depends on how much fabric the customer sees at once.

The structural half — that a warp averages and a course does not — is implicit in every account of why warp knitting is dimensionally better behaved, and stated in none of them as a fact about populations.

Where the ladder goes next #

This is the last rung of the ladder about variation. What follows is the tail of the same distribution, where the deviation is not a per cent or two but a thread that is not there at all: a missing end is a fault the length of the piece, and the arithmetic changes from an average to an accident.

Sideways, the loop that carries a knit’s dimensions is also what makes a knit recover and what gives it no closure condition — so a variation in loop length propagates into every property a knit has, rather than into one of them.