
Read a wave seven times in one turn, at equal intervals. Every reading is zero. Now go back halfway between the first two marks and read it there. The height is . The seven readings were exact. Silence would have given you the same seven numbers.
The wave is easy to make. Take six sawtooths. In one turn of the circle the first rises once, the second twice, and so on through the sixth. Each rises from minus one toward plus one, and takes the value zero at the instant it starts over. Add their heights. At one seventh of a turn the slowest wave reads and the fastest reads . The second cancels the fifth, and the third cancels the fourth. The same three pairs cancel at every seventh.
You can find that missing wave with a pencil. Just after zero all six sawtooths are close to minus one. Until the first reading at , their sum climbs along the straight line . Put and out comes . Square the sum and average it over the whole turn and its energy is . The seven samples miss all of it.
In The Resets That Meet, I changed the weights on the waves to pick out the places where particular resets coincide. One such choice keeps the total jump at zero exactly as it was, yet the reading at goes from zero to . That left me with a practical question. Suppose I have to set the marks on the circle before I know the weights. How close together must I put them?
Give the six waves their original weights and divide each gap into seven. On the very first new mark, at , the pencil calculation gives . One reading settles whether there is a wave. Measuring its energy asks more of the grid.
In The Geometry of Arithmetic Sampling I bound how far the grid can read above or below the wave’s energy, as a fraction of that energy. Someone can see every mark I have made and then choose the weights to give me the worst possible reading. The bound still holds. For the grids that arise in a prime base , there are harmonics and readings. As the base grows, even that worst relative error goes to zero. Its exponent is exactly minus one.
There is some arithmetic tucked inside that exponent. Give weight one to just the slowest and fastest waves. Their relative error is a little greater than . One might expect that a sufficiently large constant times would cover all the other choices. On certain primes many fast waves put their errors against the slowest wave on the same side. Those waves have little overlap with one another over a full turn. Their errors can accumulate faster than their shared energy would suggest. Along one sequence of primes even a fixed power of multiplying is eventually too small.
Let be the mean square over the circle and the mean of the squared samples. Write for the largest value of over nonzero real weight vectors on the first harmonics of a -point grid. For every and ,
The constant is absolute. Fix and . When and is coprime to every frequency through , the bound improves to . At , , with prime, the slowest and fastest harmonics together give the exact relative error . The matching exponent for the worst choice is .
For one choice of weights you can measure the entire error on the 49-point grid with a ruler. Give the first three waves weights and turn the other three off. Their slopes cancel. The sum sits at zero until one third of a turn, steps down to minus two, then up to plus two at the halfway mark. At two thirds it returns to zero.
Its square stands at height four on two intervals, each one sixth of a turn long. Each contains eight samples. The circle gives and the grid gives . Where, exactly, is the missing ?
Look at the gap from to . The step at one third comes early enough that the high part fills two thirds of the gap. If the step had fallen in the middle, the two endpoint readings would have been exactly the same. The extra area is four times one sixth of a gap, or . Across the circle the step down at two thirds comes late. It leaves the high part the very same extra length. Reflection puts the steps on opposite sides of their midpoints. Both areas count in the same direction.
Return all the weights to one and let the prime grow. At the origin every wave resets together. The grid reads zero there. From either side the squared height approaches . That single reading has weight , enough to leave one whole unit in the limit even as the relative error goes to zero.
The individual waves account for very little of that unit. Add up the sampling errors of their separate squares and the total tends to zero. The unit is carried by the products of different waves when their sum is squared. After the unit is removed, the quarter-power bound is known. A half-power bound for the remaining signed error is still open.
The little on the 49-point grid can be drawn on paper. It is the area of those two gold strips. The unit at the common reset has an equally definite address. To see where it comes from, we have to add the waves before we square. Square them one at a time and the unit slips out of the calculation. What the grid reads at that one mark belongs to the waves together.
Revised November 20, 2026.
Discussion
Sign in to join the discussion.