Still preparing the reader…Long documents can take a little longer to load and typeset.
A grid can miss the entire energy of a sum of sawtooth waves when they cancel at every sample. For the first N harmonics on Q equally spaced points, the worst relative error in mean square is at most K N^{3/2}/Q. Exponentially fine grids attain this order. On grids 2\le Q\le N^A coprime to every frequency the bound improves to O_{A,\varepsilon}(N^{1+\varepsilon}/Q) for each fixed A\ge1. With p-1 harmonics on p^2 points, p prime, the worst error has exponent -1. Along a sequence of primes its ratio to 1/p grows faster than every fixed power of \log p. At equal weights the difference between the continuous and sampled mean squares is known to approach one. The positions of the jumps give exact formulas for what remains after that unit is subtracted. A half-power bound on this remainder would determine the finite carry energy to within O_\varepsilon(p^{3/2+\varepsilon}).
Download the PDF · Download the TeX source
How much of a wave’s energy can a grid miss even when every reading is exact? The sum of the first six sawtooth harmonics has mean square 43/10 over one period. At the seven points 0,1/7,\ldots,6/7, every reading of that sum is zero.
Each sawtooth rises from -1 to 1 and takes the value zero at its reset. In one turn of the circle the first wave completes one cycle, the second two, and so on through the sixth. At 1/7 the first wave is a seventh of a cycle past its reset and the sixth is a seventh short of its next. Their heights are -5/7 and 5/7. The second cancels the fifth in the same way, and the third cancels the fourth. At every seventh the phases of each pair add to a whole number of cycles. On the open interval 0<x<1/7 the six waves rise together. Their sum there is 42x-6. Halfway across, all six are below zero and their sum is -3.
Divide each interval into seven. The first new reading, at 1/49, is -36/7. Changing the weights can also break the cancellation. Keep the jumps at the multiples of 1/6 with their original sizes and remove the others. The resulting weights (2,2,1,0,0,1) preserve the jump at zero and give a reading of -12/7 at 1/7. A bound for all weights lets us choose the grid spacing before we know which wave we will measure.
Each rising sawtooth takes the midpoint of its two limits at a reset. For weights c=(c_1,\ldots,c_N)\in\mathbb{R}^N the field is \psi(x)=\begin{cases}\{x\}-\tfrac12,&x\notin\mathbb{Z},\\0,&x\in\mathbb{Z},\end{cases} \qquad F_c(x)=2\sum_{h=1}^Nc_h\psi(hx),\qquad B_N=F_{\mathbf1}. Energy means the mean square of the sum, including the interactions between harmonics. On the circle and on the grid of Q points the corresponding bilinear forms are \begin{aligned} C_N(c,w)&=\int_0^1F_c(x)F_w(x)\,dx, & (C_N)_{ij}&=\frac{\gcd(i,j)^2}{3ij},\\ H_{N,Q}(c,w)&=\frac1Q\sum_{a\bmod Q}F_c(a/Q)F_w(a/Q). \end{aligned}\tag{1} The continuous entries are Franel’s integrals for dilated sawtooths [4], also called the Franel–Landau integrals [2, p. 1534]. For the six unit weights the diagonal entries of C_6 sum to 2 and the off-diagonal entries to 23/10. This gives the 43/10 missed by the grid of sevenths. The weights may be chosen anew at every cutoff. Such fields are finite Davenport series. Their jumps record sums of coefficients over multiples of each denominator [13]. Comparing a continuous norm with finitely many readings is a sampling discretization problem [14]. Here the arithmetic progression of sample addresses is prescribed.
Write D_{N,Q}=C_N-H_{N,Q} and measure the worst relative error by \eta_{N,Q}=\sup_{c\ne0}\frac{|D_{N,Q}(c,c)|}{C_N(c,c)}. Thus even the choice of weights that gives a particular grid its largest error must obey the following bounds.
Theorem 1.1 (Uniform arithmetic sampling theorem). For N\ge2 and every integer Q\ge1, \eta_{N,Q}\le K N^{3/2}/Q with an absolute constant K. For each fixed A\ge1 and \varepsilon>0, if 2\le Q\le N^A and (Q,h)=1 for 1\le h\le N, then \eta_{N,Q}\ll_{A,\varepsilon} N^{1+\varepsilon}/Q. The field e_1+e_N supplies a lower bound on prime squares. For primes p\ge5 the two bounds are p^{-1}+p^{-2}-p^{-3}\ \le\ \eta_{p-1,p^2} \ \ll_\varepsilon\ p^{-1+\varepsilon}.
The first and last harmonics alone fix the exponent on prime squares. One might then expect a constant times 1/p to cover every choice of weights. Theorem 2.5 constructs a sequence of primes on which the worst error eventually exceeds (\log p)^A/p for each fixed A. The error also has a location. On the 49-point grid two cells suffice to see it.
On the same grid one can account for an entire sampling error by measuring two lengths. Give the first three harmonics weights 1,-2,1 and the other three weight zero. Their slopes cancel since 1-4+3=0. The field steps from zero to -2 at 1/3, to 2 at 1/2, and back to zero at 2/3. Its square is four on two open intervals of length 1/6. Each contains eight samples. Thus C_6(c,c)=4/3,\qquad H_{6,49}(c,c)=64/49,\qquad C_6(c,c)-H_{6,49}(c,c)=4/147 .
The reset at 1/3 lies one third of the way across its cell. The field square occupies the other two thirds at height four (Figure 1). Move the reset halfway across, keeping both endpoint values fixed, and the integral falls by (4/49)(2/3-1/2)=2/147. The reflected cell at 2/3 supplies the other 2/147. The reset at 1/2 is already centered. With the resets centered, the integral equals the sampled energy. Restoring the two original lengths accounts for the whole difference.
In base p, the equal-digit words 00,11,\ldots,(p-1)(p-1) mark the addresses (p+1)d, 0\le d<p, on a circle of p^2 places. For a multiplier a, multiply the marked addresses by a and count how many cross zero on advancing another a places, including arrival at zero. Subtract a/p from that count. The sum of the squared readings over 0\le a<p^2 is the carry energy E_p. At the address a\equiv(1-p)A\bmod p^2 the centered count equals B_{p-1}(A/p^2). Indeed the fractional parts left after subtracting the floors pair as \{hA/p^2\}-\{-hA/p^2\}=2\psi(hA/p^2), 1\le h<p [9, 12]. The grid gives the common reset at zero a positive mass.
For a prime p\ge5, put N=p-1, Q=p^2 and write \mathsf C_p=C_p=C_N,\qquad \mathsf G_p=G_p=QH_{N,Q},\qquad \mathcal H_Q=H_Q=H_{N,Q}. We use the same symbols for these matrices and their bilinear forms, with one argument for quadratic evaluation. The defect of a source and its value at the equal-weight, or coherent, source are \mathcal S_p(c)=C_N(c,c)-H_{N,Q}(c,c),\qquad \mathcal S_p=\mathcal S_p(\mathbf1). At zero the coherent field reads zero. Its one-sided squares are both N^2, giving a contact contribution N^2/Q that tends to one. The known rate [12] and the proposed rate are \begin{align*} \mathcal S_p&=1+O_\varepsilon(p^{-1/4+\varepsilon}),\tag{2}\\ \mathcal S_p&=1+O_\varepsilon(p^{-1/2+\varepsilon}).\tag{3}\end{align*} The second estimate remains open. Hilberdink, Luca and Tóth [6, Theorem 5.1] give C_N(\mathbf1,\mathbf1)=p+O(\log^2p) by summing \gcd(i,j)^2/(3ij). For the sampled carry table the Cubic Law [9] gives E_p=G_p(\mathbf1,\mathbf1)=p^3+O(p^2\log^2p). Theorem 1.1 therefore bounds the absolute error at this source by O_\varepsilon(p^\varepsilon). Even the best uniform relative constant, multiplied by this source’s energy, is at least 1-o(1) by the two-wave lower bound. For the carry table, (3) is equivalent to E_p=p^2 C_{p-1}(\mathbf1,\mathbf1)-p^2 +O_\varepsilon(p^{3/2+\varepsilon}). It asks how closely the integral predicts a finite count once the known contribution of the common reset is removed. Since C_{p-1}(\mathbf1,\mathbf1)\sim p, this requires relative precision of order p^{-3/2+\varepsilon} about the corrected value.
One can see where the unit enters the energy matrix. For every 1\le h<p, multiplication by h permutes the Q grid. Thus the single-harmonic defect is the same in every diagonal entry, \frac13-\frac1Q\sum_{a\bmod Q}[2\psi(a/Q)]^2 =\frac1Q-\frac{2}{3Q^2}. The p-1 diagonal entries together contribute only O(1/p). By (2) the sum over ordered pairs of distinct harmonics tends to one. The rate after that unit is removed is a question about their signed interaction.
We can isolate some of those interactions by grouping resets according to their denominators. At each retained jump the entire field still contributes its height. Section (3.1) bounds the signed contribution of such a forest in terms of the number of denominators it contains. The contribution from its complement is the one that still has to be estimated.
Within one sample interval the grid records a single value, however far the square of the field rises or falls before the next sample. That variation bounds the area the grid can miss. There is more information on a coprime grid. The position of a jump at a/d in its cell is set by the residue of Qa modulo d. Jumps with the same denominator can therefore be counted together.
Put C=C_N(c,c), J_2(d)=d^2\prod_{r\mid d,\ r\ {\rm prime}}(1-r^{-2}) and L_N=\sum_{d\le N}J_2(d). Introduce four linear readings of the source, q_d=\sum_{d\mid h}\frac{c_h}{h},\qquad m_d=\sum_{d\mid h}c_h,\qquad s_c=\sum_hc_h,\qquad \ell_c=\sum_hhc_h . All harmonic sums run to N. Summing J_2 or \varphi over the divisors of h gives h^2 or h, respectively. The resulting divisor coordinates of the continuous Gram [4, 8] satisfy 3C=\sum_{d\le N}J_2(d)q_d^2,\qquad \ell_c=\sum_{d\le N}J_2(d)q_d,\qquad s_c=\sum_{d\le N}\varphi(d)q_d .\tag{4} Indeed, expand the square and sum the Jordan totients over d\mid(i,j). The map c\mapsto(q_d) is triangular and invertible. Since \varphi(d)^2\le J_2(d), these identities give s_c^2\le3NC, \ell_c^2\le3L_NC and q_1^2\le3C. Here d^2/2\le J_2(d)\le d^2 by the Euler product for \zeta(2).
Theorem 2.1 (Farey variation inequality). The Jordan variation around the circle, with the assigned midpoint field values included, satisfies V(F_c^2)\le \bigl(4\sqrt3+12\sqrt2\,\zeta(3/2)\bigr)N^{3/2}C . Consequently |D_{N,Q}(c,c)|\le V(F_c^2)/Q for every Q\ge1.
Proof. Between jumps F_c'=2\ell_c. Thus the smooth part of the variation is at most 4|\ell_c|\int|F_c|\le4\sqrt{3L_N}\,C.
For the jumps, set y_h=hq_h and \|y\|_w^2=\sum_{h\le N}h\,y_h^2. Divisor inversion gives c_h=\sum_{k\le N/h}\mu(k)y_{hk}/k. With y extended by zero, \|(y_{hk})_h\|_w\le k^{-1/2}\|y\|_w. The triangle inequality and (4) now give \sum_hh c_h^2\le\zeta(3/2)^2\|y\|_w^2 \le6\zeta(3/2)^2 NC . A jump at denominator d has size -2m_d and occurs \varphi(d) times, counting the origin once with d=1. By Cauchy–Schwarz on the multiples of d, \sum_{d\le N} d\varphi(d)m_d^2 \le N\sum_{h\le N} c_h^2\sum_{d\mid h}\varphi(d) =N\sum_{h\le N}h c_h^2 .\tag{5}
Let u_\xi,v_\xi be the left and right traces at every reduced reset \xi of denominator d_\xi, including zero. A gap between consecutive resets of denominators d,e has length a\ge1/(de), so both 1/d and 1/e are at most Na. If the affine field on that gap starts at v and ends at u, its squared integral is a(u^2+uv+v^2)/3\ge a(u^2+v^2)/6. Summing gives \sum_{\xi}\frac{u_\xi^2+v_\xi^2}{d_\xi}\le6NC . The actual square visits u^2,((u+v)/2)^2,v^2 at a jump. Its variation there is at most 2|m_d|(|u|+|v|), including when u=-v. Apply Cauchy–Schwarz with weights d_\xi,1/d_\xi and use (5). The sum is at most 12\sqrt2\,\zeta(3/2)N^{3/2}C. Since L_N\le N^3, the two parts give the asserted variation.
On [a/Q,(a+1)/Q] the difference between the integral of the square and its left sample times 1/Q is at most the variation there divided by Q. Addition around the circle proves the energy bound, including grids that land on jumps. This last step is Koksma’s one-dimensional quadrature inequality [15, 1]. ◻
Assume now that Q\ge2 and (Q,h)=1 for 1\le h\le N. In particular Q>N. At a reduced fraction \xi=a/d the energy jump is -4m_d F_c(\xi). Put \mathcal U_d=(\mathbb{Z}/d\mathbb{Z})^\times and \begin{aligned} j_d(c)&=\sum_{a\in\mathcal U_d}F_c(a/d)\psi(Qa/d),\\ R_B(h,d)&=\sum_{a\bmod d}\psi(ha/d)\psi(Ba/d),\qquad K_{d,h}=R_Q(h,d). \end{aligned} Set j_1=0 and write W_d(c)=(Kc)_d. A complete d-clock, its d numerators counted once each, splits into its reduced clocks. Hence 2W_d(c)=\sum_{e\mid d}j_e(c) and \sum_d m_d j_d=2c^TKc. The exact reset identity is QD_{N,Q}(c,c)=s_c^2+4\sum_{d=2}^Nm_dj_d(c) -\frac{2\ell_c q_1}{3Q}.\tag{6} It holds even when several different resets occupy one sample cell.
For two harmonics use the Dedekind sum s(h,k)=\sum_{a\bmod k}\psi(a/k)\psi(ha/k),\qquad s(h,1)=0 . Write i=gu, j=gv, g=(i,j). On the grid QH_{N,Q}(e_i,e_j)=4s(ij^{-1},Q). Rademacher’s three-term reciprocity [5] for u,v,Q then gives QD_{N,Q}(e_i,e_j) =1+4s(vQ^{-1},u)+4s(uQ^{-1},v)-\frac{u^2+v^2}{3uvQ}.\tag{7} The inverses belong to the modulus of their respective sums, and a modulus-one sum is zero. To pass from a pair to (6), partition the i-clock by residue modulo u. Summing its g lifts uses \sum_{k=0}^{g-1}\psi((x+k)/g)=\psi(x), with the midpoint convention also at integers. Changing the remaining variable by Q gives K_{i,j}=s(vQ^{-1},u). Interchanging i,j gives the other clock. Summing (7) against c_i c_j proves (6). Related weighted Farey identities are given in [7, 12, 10].
Theorem 2.2 (Coprime clock bound). Suppose Q\ge2 is coprime to every h\le N. Write H_*=\sum_{j\le\lfloor N/2\rfloor}1/j, T_N=\max_{h\le N}\tau(h) and D_*=\max_{1\le n\le N(N+Q)/2}\tau(n), where \tau counts positive divisors. Then Q\eta_{N,Q}\le 3N+48\zeta(2)T_N N H_*^2\sqrt{D_*/8} +2\sqrt{L_N}/Q. In particular, for each fixed A\ge1 and \varepsilon>0, \eta_{N,Q}\ll_{A,\varepsilon}N^{1+\varepsilon}/Q when Q\le N^A.
Proof. The finite Fourier coefficients of a\mapsto\psi(a/d) have zero mean. For 0<|u|\le d/2 their absolute values are |\cot(\pi u/d)|/(2d)\le1/(4|u|), by differentiating the finite geometric sum. The even midpoint frequency has coefficient zero. Inserting the two finite Fourier series into the Dedekind correlation K_{d,h} and summing over a gives the cotangent bound [3] |K_{d,h}|\le\frac d{16} \sum_{\substack{0<|u|,|v|\le d/2\\ d\mid hu+Qv}}\frac1{|uv|}.\tag{8} Over one period of h, put g=(u,d). The congruence has g solutions when g\mid v and none otherwise, since (Q,d)=1. The sum of 1/|v| over these multiples is at most 2H_*/g. The remaining u-sum is at most 2H_*. There are at most \lceil N/d\rceil periods, so every absolute row sum is at most NH_*^2/2. Over a single period the bound is dH_{\lfloor d/2\rfloor}^2/4, where H_j=\sum_{k=1}^j1/k. This is the full-clock row estimate of [10, Proposition 6.1].
For a column, fix h,u,v and sum over d. The integer hu+Qv is nonzero. Indeed, hu+Qv=0 would give Q\mid u because (Q,h)=1, whereas 0<|u|\le N/2<Q. Its absolute value is at most N(N+Q)/2. Only divisors of that integer can contribute. Replacing d by N in the prefactor of (8) bounds the absolute column sum by ND_*H_*^2/4. This count by a nonzero integer also enters [12, Lemma 3.4]. The Schur test, or Cauchy–Schwarz with the absolute rows and columns, therefore gives \|K\|_{2\to2}\le NH_*^2\sqrt{D_*/8}.
Apply Cauchy–Schwarz to the divisor inversion. This gives c_h^2\le\zeta(2)\sum_{k\le N/h}y_{hk}^2. Each y_j^2 occurs \tau(j) times, so \|c\|_2^2\le6\zeta(2)T_N C. It follows from (6) that \frac{Q|D_{N,Q}(c,c)|}{C} \le 3N+48\zeta(2)T_N\|K\|_{2\to2} +\frac{2\sqrt{L_N}}Q \qquad (c\ne0). For Q\le N^A the arguments of both divisor counts are polynomial in N. The elementary bound \tau(n)\ll_\gamma n^\gamma for each \gamma>0 follows by bounding (k+1)/r^{\gamma k} at the finitely many primes r<2^{1/\gamma} and using k+1\le2^k at the others. Choose \gamma small enough for the given A,\varepsilon and absorb the two harmonic sums. This improvement over the all-grid exponent is asymptotic. ◻
There is a short test of the exponent. Give weight one to the slowest wave and to the fastest, and zero to every wave between them. The prime-square grid is one modulo the last frequency.
Theorem 2.3 (Prime-square sampling exponent). For N=p-1, Q=p^2 and primes p\ge5, \eta_{N,Q}=p^{-1+o(1)}. The field with coefficients c=e_1+e_N has relative defect exactly p^{-1}+p^{-2}-p^{-3}. More generally its relative defect is (N+2)/Q-(N+1)/Q^2 on every Q\equiv1\pmod N, irrespective of the other frequencies.
Proof. Write Q=kn+1>1 and look only at frequencies 1 and n. Each nonzero sample has a unique address a=kj+r, 0\le j<n, 1\le r\le k. The residue of na modulo Q is nr-j. Both addresses have mean Q/2 on this rectangle. Their covariance is n\,\operatorname{Var}(r)-k\,\operatorname{Var}(j) =\frac{n(k^2-1)-k(n^2-1)}{12}=\frac{(k-n)Q}{12}. The sample at zero contributes zero. Multiplying this covariance by 4(Q-1)/Q^3 gives the sampled mixed entry. Subtraction from 1/(3n) gives D_{1n}=\frac{n+2/n}{3Q}-\frac{n+1/n}{3Q^2}. Multiplication by either frequency permutes the grid, so each has diagonal defect 1/Q-2/(3Q^2). With n=N the continuous energy of their sum is 2(1+1/N)/3. Adding both diagonals and twice the mixed entry gives \frac{D_{N,Q}(c,c)}{C_N(c,c)} =\frac{N+2}{Q}-\frac{N+1}{Q^2}.\tag{9} At Q=1 both sides are one. At N=p-1, Q=p^2 the right side is 1/p+1/p^2-1/p^3. Since p^2\le N^3, the coprime clock bound gives the matching exponent and completes the proof of Theorem 1.1. ◻
On some much finer grids every harmonic sees the very same residue as on a grid of size one. Two divisor coordinates then determine the departure from the identity in the normalized energy matrix.
Theorem 2.4 (Residue-one extremal formula). Let M_N=\operatorname{lcm}(1,\ldots,N). On grids with Q\equiv1\pmod{M_N} one has \eta_{N,Q}=\frac{1+(1-Q^{-1})(1+\sqrt{L_N})}{Q}. Thus the power N^{3/2} in the unrestricted bound is sharp along arbitrarily fine grids.
Proof. At Q=1 the only sample is zero and the formula gives one. Otherwise Q^{-1}\equiv1 in both clock sums of (7). Two-term Dedekind reciprocity gives 4s(v,u)+4s(u,v)=-1+(u/v+v/u+1/(uv))/3, including a modulus one. In source coordinates this yields QD_{N,Q}(c,c)=C_N(c,c)+\tfrac23(1-Q^{-1})\ell_c q_1 .\tag{10} Set x_d=\sqrt{J_2(d)}q_d, a=e_1 and v_d=\sqrt{J_2(d)}. Then 3C=\|x\|^2, \ell_c=v^Tx, q_1=a^Tx. The relative quadratic form for QD is represented by \mathrm I+(1-Q^{-1})(av^T+va^T). The last matrix in parentheses has eigenvectors a\pm v/\sqrt{L_N} with eigenvalues 1\pm\sqrt{L_N}, and vanishes on their orthogonal complement. The upper eigenvalue of the whole expression dominates in absolute value and is attained because the divisor coordinates are invertible. Finally L_N\asymp N^3 and arbitrarily large multiples of \operatorname{lcm}(1,\ldots,N) can be increased by one. ◻
Every grid in this residue class beyond Q=1 exceeds M_N. Indeed \log M_N=\sum_{r^j\le N,\ r\ {\rm prime}}\log r\sim N by the prime number theorem. This proves sharpness for the unrestricted bound on exponentially fine grids. It leaves room for smaller errors in intermediate ranges and on grids with shared divisors.
On a prime square, one fast wave gives a mixed error of order 1/p against the first harmonic. Several fast waves can give errors of the same sign. To make them count, their continuous energy must stay small enough that dividing by it does not erase the gain.
Theorem 2.5 (Extremal growth on prime squares). Put B(p)=p\,\eta_{p-1,p^2}. With the limsup over primes, 0<\limsup_{p\to\infty} \frac{\log B(p)\,\log\log p}{\log p}<\infty . In particular B(p) exceeds every fixed power of \log p along one sequence of primes.
Proof. For large x, pair consecutive primes in [x,2x] as a_1<b_1<\cdots<a_m<b_m, dropping the last prime if necessary. The prime number theorem gives m\asymp x/\log x. Since the intervals [a_i,b_i] are disjoint and ordered, \prod_i b_i/a_i\le b_m/a_1\le2.
Choose a prime p satisfying p\equiv-1\bmod a_i and p\equiv1\bmod b_i for all i. These congruences specify a reduced class modulo M=\prod_i a_i b_i. For each subset S take the frequency h_S=(p-1)\prod_{i\in S}a_i/b_i . It is an integer divisor of Q-1=p^2-1, because the complementary divisor is (p+1)\prod_{i\in S}b_i/a_i. Unique factorization distinguishes the 2^m frequencies, and (p-1)/2\le h_S\le p-1. The mixed-entry calculation in Theorem 2.3 gives D_{1h_S}\ge h_S(1-1/Q)/(3Q)\gg 1/p for every S.
Set w=\sum_S e_{h_S}. In the ratio h_S/h_T only the pairs indexed by the symmetric difference S\mathbin{\triangle}T survive reduction. The continuous Gram therefore gives \begin{aligned} 3C(e_{h_S},e_{h_T})&=\prod_{i\in S\mathbin{\triangle}T}(a_i b_i)^{-1},\\ 3\sum_T C(e_{h_S},e_{h_T}) &=\prod_i(1+1/(a_i b_i))\le \exp(m/x^2). \end{aligned} Thus C(w,w)\ll 2^m while D(e_1,w)\gg 2^m/p. The symmetric matrix C^{-1/2}DC^{-1/2} has operator norm \eta_{N,Q}. Hence |D(e_1,w)|\le\eta_{N,Q}\sqrt{C(e_1,e_1)C(w,w)}. It follows that B(p)\gg2^{m/2}.
Dirichlet’s theorem already supplies primes in the indicated class for each x and proves unboundedness. To compare the growth with p, take the least such prime. Linnik’s theorem gives p\ll M^L for an absolute L [16, Theorem 1.1]. Also \prod_i b_i\mid p-1. These two bounds give \log p\asymp x. Consequently B(p)\ge\exp(c\log p/\log\log p) along a sequence, with c>0 absolute.
For the upper bound use Theorem 2.2 with Q=p^2. Both divisor maxima there have arguments at most p^3. For n\le X, split its prime factors at \sqrt{\log X}. The smaller primes contribute O(\sqrt{\log X}\log\log X) to \log\tau(n). At larger primes the sum of the exponents is at most 2\log X/\log\log X, and \log(1+j)\le j\log2. Hence \max_{n\le X}\tau(n)\le\exp(O(\log X/\log\log X)) for large X. This bound also absorbs H_*^2 at X=p^3. Multiplying the resulting bound for \eta by p proves the upper half of the assertion. ◻
The construction picks both the primes and the weights. The typical size of B(p), and whether it tends to infinity on all primes, remain open.
The cancellation at 1/7 recurs on every prime grid. At a/p the waves of frequencies h and p-h cancel when their weights agree. With unit weights their sum has a single downward jump per coarse cell, which can fall at either endpoint. The fine grid records how far each pair has risen before it jumps. Two pairs may jump at the same fraction, and their coincidence enters the square of the sum.
The second set of weights in Section 1 comes from the jumps the six waves share. At a fraction of reduced denominator d, the harmonics that reset are exactly the multiples of d. For six unit weights their counts, in order of denominator, are (6,3,2,1,1,1). Keeping the jumps only at denominators dividing six gives (6,3,2,0,0,1). Read this row backwards. The sixth harmonic keeps weight one, the fifth and fourth drop out, and the third needs weight one. The count at denominator two then gives the second harmonic weight two. The total at the origin gives the first the remaining two. These are the weights (2,2,1,0,0,1).
We can also leave all six weights in place and count the contribution those same jumps make to the original field. At each retained fraction the entire field enters the energy jump. A reset forest selects a set of denominators for which this signed contribution can be bounded [10]. Let \operatorname{rough}_N(k) be the largest divisor of k coprime to N. For a positive integer D, adjoin the origin coordinate and put \begin{gathered} \mathcal F=\{1\}\cup\{2\le d\le N:\operatorname{rough}_N(d)\mid D\}, \\ \Omega=N^{\lceil\log_2N\rceil}\operatorname{rough}_N(D). \end{gathered} Then d\in\mathcal F exactly when d\mid\Omega, for d\le N. Its source selector is P_{\mathcal F}e_h=e_{(h,\Omega)}. For any denominator mask \sigma, let A_\sigma(c)=\sum_{d=2}^N\sigma(d)m_d(c)j_d(c). If (\mathsf Zc)_d=\sum_{d\mid h}c_h, the corresponding selector is P_\sigma=\mathsf Z^{-1}\operatorname{diag}(\sigma)\mathsf Z. The inverse of the divisor matrix is (\mathsf Z^{-1}m)_h=\sum_{j\le N/h}\mu(j)m_{hj}. Write A_{\mathcal F} and A_{\rm other} for the indicator masks of the forest and its complement, and A_{\rm all} for the full mask. Partitioning each full clock into reduced clocks gives A_\sigma(c)=2(P_\sigma c)^TKc. The first factor carries the selection. The second reads the original field, including harmonics whose own reset has been omitted.
There are \varphi(d) reduced fractions of denominator d, each with the same jump -2m_d. Counting the mass once at each of those fractions controls one of the two factors in the forest sum.
Proposition 3.1 (Reset mass inequality). For every real source and N\ge2, with T_N=\max_{h\le N}\tau(h), \sum_{d\le N}\varphi(d)m_d^2\le12NT_N C_N(c,c).
Proof. Inverting q_d gives m_d=d\sum_{j\le N/d}\varphi(j)q_{dj}. The matrix A carrying \sqrt{J_2(k)}q_k to \sqrt{\varphi(d)}m_d has nonnegative entries at most \sqrt{2d} for d\mid k, and zero otherwise. In the weighted Schur test with w_k=k^{-1/2} the row sums \sum_k A_{dk}w_k/w_d are at most 2\sqrt{2N} and the column sums \sum_d A_{dk}w_d/w_k are at most \sqrt{2N}T_N. Their product bounds the squared operator norm. Apply (4). ◻
For a mask \sigma, Cauchy–Schwarz now gives |A_\sigma(c)|^2\le 12NT_N C_N(c,c)\sum_{d=2}^N\frac{\sigma(d)^2j_d(c)^2}{\varphi(d)} . The right-hand sum still measures the original field at the selected fractions. Section (4.3) shows how its midpoint readings can be recovered from the samples. At equal weights the signed sum over the complementary clocks can be written directly.
Define the fixed linear readout and its coherent value by f_p(c)=\sum_{s\notin\mathcal F}W_s(c),\qquad U_{\rm other}=f_p(\mathbf1).\tag{11} If b=P_{\mathcal F}\mathbf1, the restoration is E_{\rm restore}=2\sum_{d\in\mathcal F}(1-b_d)W_d(\mathbf1). Thus A_{\rm other}(\mathbf1)=2U_{\rm other}+E_{\rm restore}. The fixed linear readout and the quadratic A_\sigma have different source variations.
The number of selected denominators is enough to bound both signed terms. Write f=|\mathcal F\cap[2,N]|. In W_d(\mathbf1) every complete block of d harmonics sums to zero. Indeed, for each numerator a the residues ha\bmod d run over a subgroup with equal multiplicities, and opposite sawtooth values cancel there. The one-period row estimate of Theorem 2.2 bounds the one partial block by |W_d(\mathbf1)|\le dH_{\lfloor d/2\rfloor}^2/4\qquad(d\ge2). For d\in\mathcal F, b_d counts the h\le N with (h,\Omega)=d. It lies between 1 and N/d, and W_1=0. Apply the last inequality to A_{\mathcal F}=2\sum b_d W_d(\mathbf1) and to the restoration to obtain |A_{\mathcal F}|,\ |E_{\rm restore}|\le NfH_N^2/2. In (6) the coherent readings are s_c=N, \ell_c=Np/2 and q_1=H_N. Splitting the clock sum gives Q(\mathcal S_p-1)=8U_{\rm other} +4(A_{\mathcal F}+E_{\rm restore})+(N^2-Q)-NH_N/(3p). In particular, \left|\mathcal S_p-1-8U_{\rm other}/Q\right| \le (2+H_N+4fH_N^2)/p. Whenever f\ll_\varepsilon p^\varepsilon for every \varepsilon>0, the error here is O_\varepsilon(p^{-1+\varepsilon}). Every clock left in U_{\rm other} still reads the full coherent field.
Equal weights are unchanged by the exchange h\leftrightarrow p-h. The six weights selected in Section 1 change under that exchange. Splitting them into their even and odd parts shows where a mixed energy can enter.
Put n=(p-1)/2 and define (J_pc)_h=c_{p-h},\qquad A_+e_h=e_h+e_{p-h},\quad A_-e_h=e_h-e_{p-h} \quad(1\leq h\leq n). Write V_\pm=\operatorname{ran} A_\pm. These are the algebraic eigenspaces of the cutoff reflection. This reflection acts on the cutoff harmonic labels. It exchanges the labels while leaving the sample address fixed.
Proposition 3.2 (Cutoff reflection). The coherent cutoff source is A_+\boldsymbol1=\boldsymbol1. Its two work matrices in pair coordinates are C_+=A_+^T\mathsf C_pA_+,\qquad G_+=A_+^T\mathsf G_pA_+.\tag{12} Both are positive definite. Every field F_c with c\in V_+ vanishes at every coarse address a/p. In particular all nonunit addresses of its p^2 sample contribute zero. For a general source the decomposition c=A_+u+A_-v,\qquad u_h=\frac{c_h+c_{p-h}}2,\quad v_h=\frac{c_h-c_{p-h}}2 retains all four work blocks A_\sigma^T\mathsf C_pA_\tau and A_\sigma^T\mathsf G_pA_\tau, \sigma,\tau\in\{+,-\}.
Proof. Each source label belongs to exactly one pair. This proves the decomposition and the identity at the coherent source. Pulling back the continuous and sampled forms gives (12). For the sampled form the recovery identity (34) reconstructs the coefficients from their samples. Its proof uses the separation of the jumps and divisor inversion. Thus the form is positive definite on the full space. Continuous positivity follows from (4), and A_+ is injective. At a coarse address, \psi(ha/p)+\psi((p-h)a/p)=0. Oddness includes the integer midpoint value, so the assertion holds at every coarse address. Expanding either quadratic form after the displayed decomposition gives its four blocks. ◻
The algebraic parities need not be orthogonal in either cutoff work. For example, at p=7 the continuous mixed entry between e_1+e_6 and e_2-e_5 is 1/5. Even and odd coefficients can therefore contribute a mixed term to the energy on the circle.
The source and detector have different centers on a reset clock. The difference is half a residue step. Pairing their reflections therefore produces an adjacent detector difference, including on an even clock where division by two in the residue ring is unavailable.
Let K_{s,h}=R_Q(h,s) as in Section (2.2) and put W_s(c)=\sum_{h=1}^{p-1}K_{s,h}c_h,\qquad n_s^c(a)=\sum_{\substack{1\leq h<p\\h\equiv pa\ (s)}}c_h.
Proposition 3.3 (Clock reflection formula). For c_\pm=(c\pm J_pc)/2 one has \begin{aligned} W_s(c)=\frac12\sum_{a\bmod s}\bigl\{ &n_s^{c_+}(a)[R_p(a,s)-R_p(a-1,s)]\notag\\ {}+{}&n_s^{c_-}(a)[R_p(a,s)+R_p(a-1,s)]\bigr\}. \end{aligned}\tag{13} The formula holds at every reset clock 1\leq s<p, with its zero, nonunit and midpoint states included. The source reflection a\mapsto1-a and detector reflection a\mapsto-a have product a one-residue translation.
Proof. Multiplication by p permutes the full s-clock. Changing that variable gives R_Q(h,s)=R_p(hp^{-1},s), hence W_s(c)=\sum_{a\bmod s}n_s^c(a)R_p(a,s). The change h\mapsto p-h gives n_s^{J_pc}(a)=n_s^c(1-a). On the other hand R_p(-a,s)=-R_p(a,s), including self-opposite addresses. Average the last display with its reflection separately on the two parities. Since R_p(1-a,s)=-R_p(a-1,s), the result is (13). The two reflections compose to a\mapsto a+1 in one order and a\mapsto a-1 in the other. On the doubled residue coordinate their centers are 1 and 0, a separation of 1/(2s) in the normalized unit circle. No inverse of 2 was used. ◻
For the coherent source the second line of (13) vanishes. The first is a first difference around a full clock. Its unweighted sum would telescope. The actual sum weights each difference by n_s^{c_+}(a), the number of harmonics at that residue for unit coefficients. These populations survive the reflection.
For a change w in the source, put A=\mathrm I-P_{\mathcal F}. The contribution 8(Ac)^TKc has derivative 8[(Aw)^TKc+(Ac)^TKw]. The reflection formula applies to the clock in each term. When c=A_+u+A_-v, the corresponding symmetric form with matrix 4(A^TK+K^TA) retains the mixed product between A_+u and A_-v. By comparison the linear reading (11) holds the clock populations fixed.
Follow one pair across a coarse cell. Its two waves add to a line of fixed slope with a single downward step, which may fall at either endpoint. The position of the step is determined by a residue modulo p. Steps belonging to different pairs can coincide.
For 0\leq a<p and 1\leq h\leq n, let r_h(a) be the representative of ha\pmod p in \{0,\ldots,p-1\} and set \alpha_h(a)= \begin{cases} r_h(a)/(p-h),&r_h(a)\leq p-h,\\ (p-r_h(a))/h,&r_h(a)>p-h. \end{cases}\tag{14} Let \chi(t) equal 0, 1/2 or 1 according as t<0, t=0 or t>0.
Theorem 3.4 (Single-jump formula). For 0<u<1 the paired cutoff field satisfies 2\psi\left(\frac{h(a+u)}p\right) +2\psi\left(\frac{(p-h)(a+u)}p\right) =2u-2\chi(u-\alpha_h(a)).\tag{15} At every coarse address the pair has value zero. Its jump positions obey \alpha_h(p-1-a)=1-\alpha_h(a).\tag{16} For the coherent source, the complete cell profile is B_N((a+u)/p)=Nu-2\sum_{h=1}^n \chi(u-\alpha_h(a)).\tag{17} Thus the p^2 readings use this same profile at u=b/p, 1\leq b<p, together with the coarse zeros. An interior jump at a reduced position x=k/d, 2\leq d\leq N, has multiplicity \lfloor N/d\rfloor in its cell.
Proof. Put r=\{h(a+u)/p\}. Away from its two possible trace boundaries, the sum of the two sawtooths is 2u-2 for 0<r<u and 2u for u<r<1. At r=0 or r=u the midpoint convention gives 2u-1. As u increases from 0 to 1, the inequality r<u starts at u=r_h(a)/(p-h) if r_h(a)\leq p-h. In the other case r first wraps at u=(p-r_h(a))/h, after which r<u. This proves (15), including the endpoint positions 0 and 1 on the open cell. At a coarse address oddness gives zero separately. For a'=p-1-a, the residue is p-h-r_h(a) reduced modulo p. Substitution in the two cases of (14) gives (16). Summing the pair profiles proves (17). The sampling grid has a/p+b/p^2=(a+b/p)/p, proving the sampling assertion. At a reduced interior position k/d, precisely the harmonics divisible by d reset. A pair cannot contain two such harmonics, since that would give d\mid p. Thus exactly \lfloor N/d\rfloor distinct pairs jump there. Since d<p, this is inside a unique coarse cell. ◻
In the single-jump formula, the two step functions for pairs i and j overlap from the later jump to the end of the cell. Their common tail has length 1-\max(\alpha_i,\alpha_j). On the grid we can count the sample addresses in that same tail.
Corollary 3.5 (Jump overlap formula). Put k_h(a)=\min(p-1,\lfloor p\alpha_h(a)\rfloor). Then \begin{aligned} (C_+)_{ij}=\frac4p\sum_{a=0}^{p-1} \left[\frac13+\frac{\alpha_i(a)^2+\alpha_j(a)^2}{2} -\max(\alpha_i(a),\alpha_j(a))\right]. \end{aligned}\tag{18} The raw sampled Gram is \begin{aligned} (G_+)_{ij}=4\biggl[&\frac{(p-1)(2p-1)}6\notag\\ &+\frac1{2p}\sum_{a=0}^{p-1} \{k_i(a)(k_i(a)+1)+k_j(a)(k_j(a)+1)\}\notag\\ &-\sum_{a=0}^{p-1}\max(k_i(a),k_j(a))\biggr]. \end{aligned}\tag{19} For each h the marginal positions satisfy \sum_a\alpha_h(a)=\frac p2,\qquad \sum_a\alpha_h(a)^2=\frac p3+\frac1{6h}+\frac1{6(p-h)}.\tag{20} Both formulas retain the joint positions of different source pairs.
Proof. In one cell the integral of the product in (15) is 4\left[\frac13+\frac{\alpha_i^2+\alpha_j^2}2 -\max(\alpha_i,\alpha_j)\right]. Each cell has width 1/p, giving (18). At u=b/p, no interior jump can be hit. Its reduced denominator divides h or p-h, both coprime to p. The indicator is therefore 1_{b>k_h(a)}. The two sums needed for its product are \sum_{b=k+1}^{p-1}\frac bp =\frac{p(p-1)-k(k+1)}{2p},\qquad \sum_{b=1}^{p-1}\frac{b^2}{p^2} =\frac{(p-1)(2p-1)}{6p}. Substitution and summation over the p cells give (19). The clipping at p-1 retains a jump at \alpha=1 correctly. The coarse sample contributes zero by Proposition 3.2. As a varies, r_h(a) takes every residue once. Its jump positions are the grid j/(p-h), 0\leq j\leq p-h, together with the interior grid j/h, 1\leq j<h. Their first and second sums give (20). ◻
The jump overlap formula uses the positions \alpha_h(a) in the integral and their bins k_h(a) in the sampled sum. Both count coincident jumps with their multiplicities. The sampling defect is \mathcal S_p=\boldsymbol1^T(C_+-G_+/p^2)\boldsymbol1. The sampled numerator is \boldsymbol1^TG_+\boldsymbol1=E_p. To compare the two energies, it remains to measure how rounding moves each jump relative to the others in its cell.
Theorem 3.6 (Permutation distance formula). For every h, the vector (k_h(a))_{a=0}^{p-1} is a permutation of 0,\ldots,p-1. Put \rho_h=\frac1h+\frac1{p-h},\qquad D^\alpha_{ij}=\sum_a|\alpha_i(a)-\alpha_j(a)|, D^k_{ij}=\sum_a|k_i(a)-k_j(a)|. The two cutoff work matrices are \begin{aligned} (C_+)_{ij}&=\frac23+\frac{\rho_i+\rho_j}{3p} -\frac2pD^\alpha_{ij},\notag\\ (G_+)_{ij}&=\frac{2(p^2-1)}3-2D^k_{ij}. \end{aligned}\tag{21} In particular every cutoff pair has raw sampled energy 2(p^2-1)/3. With \mathcal T_p=\sum_{i,j=1}^n(D^k_{ij}-pD^\alpha_{ij}), \qquad H_N=\sum_{h=1}^N\frac1h, the full sampling defect is \mathcal S_p=\frac{NH_N}{3p}+\frac{N^2}{6p^2} +\frac{2\mathcal T_p}{p^2}.\tag{22} Define its centered distance discrepancy by \mathcal Z_p=\mathcal T_p+\frac{pNH_N}6+\frac{N^2}{12}-\frac{p^2}2.\tag{23} Then \mathcal S_p-1=2\mathcal Z_p/p^2. The full half-power target is equivalent to \mathcal Z_p=O_\varepsilon(p^{3/2+\varepsilon}).
Proof. For 1\leq b<p, the marginal grids in the proof of Corollary 3.5 contain \#\{a:\alpha_h(a)<b/p\} =1+\left\lfloor\frac{(p-h)b}p\right\rfloor +\left\lfloor\frac{hb}p\right\rfloor=b positions. Both displayed fractions are nonintegers with sum b. No interior jump is on the sample grid. Consecutive differences therefore give one position in every quantization bin, with the last endpoint retained by clipping. This proves the permutation assertion. Consequently \sum_a k_h(a)=\frac{p(p-1)}2,\qquad \sum_a k_h(a)(k_h(a)+1)=\frac{p(p-1)(p+1)}3. Apply 2\max(x,y)=x+y+|x-y| in (18) and (19). The marginal moments (20) and the last display give (21). Summing its continuous minus normalized sampled entries, and using \sum_{h=1}^n\rho_h=H_N and 2n=N, proves (22). Subtraction of the permanent unit gives (23) and the stated equivalence. ◻
Corollary 3.7 (Paired diagonal defect). In the prescribed pair coordinates the raw sampled diagonal is \operatorname{tr} G_+=\frac{N(p^2-1)}3=\frac{p^3}3+O(p^2).\tag{24} The Cubic Law therefore gives the same limiting diagonal fraction 1/3 after the cutoff pairs are assembled. Their complete diagonal contribution to the sampling defect is \sum_{h=1}^n\left[(C_+)_{hh}-\frac{(G_+)_{hh}}{p^2}\right] =\frac{2H_N}{3p}+\frac{N}{3p^2} =O\left(\frac{\log p}{p}\right).\tag{25} Thus the remaining permanent unit and its full half-power question lie in the interactions between distinct prescribed pairs, after this explicit diagonal term is retained.
Proof. The diagonal distances in (21) vanish. Each sampled diagonal entry is 2(p^2-1)/3. Summing over n=N/2 gives (24). Its continuous counterpart is 2/3+2\rho_h/(3p). Subtract the normalized sampled value and use \sum_h\rho_h=H_N to obtain (25). The Cubic Law has complete leading energy p^3, so the displayed raw trace has limiting fraction 1/3. ◻
For a source v with \boldsymbol1^Tv=0, v^TG_+v=-2v^TD^kv. Thus the permutation distances carry a positive work norm on the balanced source plane. For a general source the constant term in (21) retains its coherent mean. Both terms are required by the cutoff work.
The remaining signed quantity also has an explicit rounding form. Within each cell order the \alpha_h(a) increasingly, retaining repeated positions with their multiplicities. Write (r) for that order and \epsilon_{(r)}=p\alpha_{(r)}-k_{(r)}. Since quantization preserves order, \mathcal T_p=-2\sum_{a=0}^{p-1}\sum_{r=1}^n (2r-n-1)(\epsilon_{(r)}-\tfrac12).\tag{26} Indeed, the ordered-distance identity is \sum_{i,j}|x_i-x_j|=2\sum_r(2r-n-1)x_{(r)}. Subtract its two versions and use \sum_r(2r-n-1)=0. The result is independent of how tied positions are ordered. Each marginal distribution is fixed. The remaining signed sum pairs each jump’s rank with its rounding error. Since \mathcal S_p-1=2\mathcal Z_p/p^2, the estimate in Section (3.1) gives \mathcal Z_p=4U_{\rm other}+O_\varepsilon(p^{1+\varepsilon}) whenever f\ll_\varepsilon p^\varepsilon for every \varepsilon>0. The ordered distances and the complementary clocks thus measure the same signed discrepancy to this precision.
At 1/3 on the 49-point grid, the square occupies the last two thirds of a sample cell. Move the jump to the middle and the same square occupies half the cell. Both samples keep their values. We can make the same change at every reset that lies alone between two samples. The new coordinate records the jumps in a common position. A density restores the length each side originally occupied. We use N=p-1, Q=p^2 unless other grid sizes are stated.
Mark both endpoints of a cell and the reset between them. The increasing map that fixes the endpoints, centers the reset and is linear on each side is already determined by those three marks.
Let \mathcal R_N=\{a/d:2\leq d\leq N,\ 1\leq a<d,\ (a,d)=1\}. For \xi=a/d\in\mathcal R_N, write Q\xi=b+\theta_\xi, 0<\theta_\xi<1. In its cell use coordinates x=(b+u)/Q, y=(b+v)/Q and set v=\begin{cases} u/(2\theta_\xi),&0\leq u\leq\theta_\xi,\\ \tfrac12+(u-\theta_\xi)/(2(1-\theta_\xi)), &\theta_\xi\leq u\leq1. \end{cases}\tag{27} Use the identity in a cell without a boundary.
Theorem 4.1 (Reset centering map). The law (27) defines an increasing piecewise affine circle coordinate \Theta_p. It fixes every Q-grid point and sends every potential interior reset to its cell’s midpoint. It is the unique map affine on the indicated pieces with those marked values. Moreover \|\Theta_p-\mathrm I\|_\infty\leq\frac1{2Q},\qquad \Theta_p(1-x)=1-\Theta_p(x).\tag{28} Writing X_p=\Theta_p^{-1}, the restoring density is \varpi_p(y)=X_p'(y)= \begin{cases} 2\theta_\xi,&0<v<1/2,\\ 2(1-\theta_\xi),&1/2<v<1, \end{cases}\tag{29} in an occupied cell, and 1 in an empty cell. It has mean one in each cell and satisfies 2/N\leq\varpi_p\leq2-2/N almost everywhere.
Proof. For distinct denominators, the distance of two reduced resets is at least 1/(dd')\geq1/[N(N-1)]. For equal denominators it is at least 1/N. The first bound is attained by 1/N and 1/(N-1), so the minimum gap is exactly 1/[N(N-1)] for N\geq3. It exceeds 1/Q in the sampling problem. Distinct boundaries therefore occupy different fine cells. None lies on the grid because d<p and (d,Q)=1. The local maps are increasing and agree at the grid endpoints, so they assemble to a circle coordinate. Two marked endpoints and the marked interior point determine both affine pieces. Its displacement attains its maximum at that interior mark and is at most 1/(2Q). The reflected boundary has cell Q-1-b and fractional position 1-\theta_\xi, proving the reflection identity. Differentiation of the two inverse affine pieces proves (29). Since \theta_\xi=r/d with 1\leq r<d\leq N, the stated bounds follow. The two half-cell densities have average one. ◻
The same construction defines \Theta_{N,Q} for any N\geq2 and Q>N(N-1) with (Q,d)=1 for 1\leq d\leq N. The separation proof, fixed sample marks and density law use only these conditions. The cutoff coordinate is \Theta_p=\Theta_{p-1,p^2}.
At \xi=1/4 in Figure 3, half of the centered cell accounts for only a quarter of its original length. Integrating the aligned field with uniform measure gives a different weight to that part of the field. Subtracting this unweighted integral from the original one measures the change for each pair of sources.
For every real cutoff source put \widehat F_c=F_c\circ X_p and s_c=\sum_hc_h,\qquad t_c=\frac2Q\sum_hh c_h,\qquad C_p^0(c,w)=\int_0^1\widehat F_c(y)\widehat F_w(y)\,dy. The cell-length correction is \mathfrak K_p(c,w)=\int_0^1(\varpi_p(y)-1) \widehat F_c(y)\widehat F_w(y)\,dy. Write I_L,I_R for the left and right halves of a centered cell and g=\widehat F_c\widehat F_w. By (29) the cell containing a reset at fractional position \theta contributes (2\theta-1)\left(\int_{I_L}g(y)\,dy-\int_{I_R}g(y)\,dy\right). A cell with a large difference in its two original lengths can contribute little if the field gives its two halves nearly equal work. The bilinear formula below evaluates that difference at each reset.
Theorem 4.2 (Cell-length correction formula). The original energies satisfy \begin{aligned} c^T\mathsf C_pw&=C_p^0(c,w)+\mathfrak K_p(c,w),\notag\\ c^T\mathsf G_pw&=\sum_{b\bmod Q} \widehat F_c(b/Q)\widehat F_w(b/Q). \end{aligned}\tag{30} For \xi=a/d\in\mathcal R_N, write f_c=F_c(\xi), m_c=m_d(c) and \delta_\xi=2\theta_\xi-1. The complete bilinear correction is \begin{aligned} \mathfrak K_p(c,w)=\frac1Q\sum_{\xi\in\mathcal R_N} \biggl[&\delta_\xi \left(f_cm_w+f_wm_c-\frac{f_ct_w+f_wt_c}{4}\right)\notag\\ {}+{}&\delta_\xi^2 \left(\frac{t_ct_w}{6}-\frac{t_cm_w+t_wm_c}{4}\right)\biggr]. \end{aligned}\tag{31} In particular, writing \mathfrak K_p(c)=\mathfrak K_p(c,c), \mathfrak K_p(c)=\frac1Q\sum_{\xi\in\mathcal R_N} \left[\delta_\xi f_c(2m_c-t_c/2) +\delta_\xi^2(t_c^2/6-t_cm_c/2)\right].\tag{32} Its directional derivative at c along w is 2\mathfrak K_p(c,w). In the cell containing \xi, write L_c=\widehat F_c(b/Q) and R_c=\widehat F_c((b+1)/Q) for the endpoint samples. Then \begin{aligned} 2m_c&=t_c+L_c-R_c,\qquad f_c=\frac{L_c+R_c}{2}+t_c\psi(Q\xi),\notag\\ 4m_cf_c\psi(Q\xi) &=[t_c(L_c+R_c)-(R_c^2-L_c^2)]\psi(Q\xi) +t_cm_c\delta_\xi^2. \end{aligned}\tag{33} For every N\ge2 and Q>N(N-1) coprime to each frequency the same identities hold. In this case use \Theta_{N,Q} and its inverse in the definitions of C^0,\varpi and \mathfrak K, and take the original energy matrices to be C_N and QH_{N,Q}.
Proof. Change variables by x=X_p(y) in the continuous bilinear integral. The pushforward of the continuous measure is \varpi_p(y)\,dy. Every grid point is fixed, proving the second identity without changing an atomic sample weight. In an occupied cell let f_c be the midpoint trace. The two traces are f_c+m_c and f_c-m_c. The left endpoint is f_c+m_c-t_c\theta_\xi and the right endpoint is f_c-m_c+t_c(1-\theta_\xi). The unweighted half-cell integrals of \widehat F_c^2, in unit-cell coordinates, have difference f_c(2m_c-t_c/2)+\delta_\xi(t_c^2/6-t_cm_c/2). Multiply by \delta_\xi/Q from (29) and sum the occupied cells. Empty cells have zero density correction. This proves (32). Polarization gives (31) and the full source variation. Subtracting the two endpoint expressions gives the reset formula in (33). Their average gives its midpoint formula. Substitution and \delta_\xi=2\psi(Q\xi) give the last identity. The common coordinate was defined using every potential boundary, so the formula also includes cells where a particular source has m_c=0. ◻
Even a cell in which a particular source has no jump can be moved since another harmonic may reset there. The same X_p therefore acts on both fields in (31), including when their sets of actual jumps differ.
The first two nonzero sample addresses lie before the first possible interior reset. They give the common rise t_c=2\ell_c/Q and the total mass s_c. The samples on either side of 1/d give the mass at denominator d. In full, with y_b=F_c(b/Q) and b_d=\lfloor Q/d\rfloor, one has \begin{aligned} t_c&=y_2-y_1, &m_1&=t_c-y_1,\\ m_d&=(t_c+y_{b_d}-y_{b_d+1})/2\quad(2\le d\le N),& c_h&=\sum_{j\le N/h}\mu(j)m_{hj}. \end{aligned}\tag{34} The first line follows from 2/Q<1/N. The endpoint identity (33) proves the second line, followed by divisor inversion. At most 2N samples suffice for recovery on the prime square.
The same endpoints bound the midpoint readings in the forest sum. On the circle we can approach each reset from both sides along its own short interval. On a grid the neighboring samples serve that purpose when the reset has a cell to itself.
Proposition 4.3 (Farey midpoint bounds). For N\ge2 and every real source, \sum_{\xi\in\mathcal R_N}F_c(\xi)^2\le(2N^2+N)C_N(c,c). If Q>N(N-1) is coprime to every h\le N, then, with t_c=2\ell_c/Q, \sum_{d=2}^N\frac{j_d(c)^2}{\varphi(d)} \le\frac{2QH_{N,Q}(c,c)+Qt_c^2/2}{9}. At zero slope the right side improves to QH_{N,Q}(c,c)/9.
Proof. Take a=1/(2N^2). Intervals of radius a about the interior reduced fractions are disjoint and avoid the origin. On either side of \xi, the trace there is F_c(x)+2\ell_c(\xi-x). Squaring and averaging on the half-interval bounds its square by twice the mean of F_c^2 there plus 8\ell_c^2a^2/3. The assigned value is the mean of the two traces. Average the two bounds and sum over at most N^2/2 resets. The integrals contribute at most C_N(c,c)/a and \ell_c^2\le3N^3C_N(c,c) bounds the remaining term by NC_N(c,c).
Let L,R be the two readings on either side of \xi. Equation (33) gives F_c(\xi)=(L+R)/2+t_c\psi(Q\xi), and hence F_c(\xi)^2\le L^2+R^2+t_c^2/2. Each endpoint is counted at most twice. At zero slope use (L+R)^2/4\le(L^2+R^2)/2 instead. To pass to the sum over denominators put V_d=\sum_{a\in\mathcal U_d}\psi(Qa/d)^2. Multiplication by Q permutes the units. Inverting the complete sum \sum_{a=1}^{d-1}\psi(a/d)^2=(d-1)(d-2)/(12d) over divisors gives \frac{V_d}{\varphi(d)} =\frac1{12}+\frac{\kappa(d)}{6d\varphi(d)}\le\frac19,\qquad \kappa(d)=\prod_{\substack{r\mid d\\r\ {\rm prime}}}(1-r). A positive correction requires at least two distinct prime factors, so d\ge6 and |\kappa(d)|\le\varphi(d) suffice. Cauchy–Schwarz on the dth denominator gives j_d(c)^2\le V_d\sum_{a\in\mathcal U_d}F_c(a/d)^2. ◻
Together with the reset mass inequality, the sampled midpoint bound now measures the energy jump on any selected set of denominators. For |\sigma(d)|\le1 and the separated grids of the proposition it gives |A_\sigma(c)|^2\le \tfrac43 NT_N C_N(c,c)\bigl(2QH_{N,Q}(c,c)+Qt_c^2/2\bigr). If several resets share a cell, its endpoints record only their total jump after the rise is subtracted. In a cell touching zero, the inward trace must also replace the assigned midpoint there. The first wave alone gives F_{e_1}(1/11)-F_{e_1}(0)=2/11-1 across the first cell.
For example, take N=7, Q=11 and c=e_5+e_6-e_4-e_7. Each pair of frequencies adds to eleven. Oddness of the sawtooth makes every sample zero, and the two frequency sums cancel in the slope. Yet (1) gives C_N(c,c)=6/5. At the thirds the field takes values 1 and -1, giving j_3(c)=1/3. The sampled energy on the right of the zero-slope midpoint bound is zero. The term from denominator three on the left is already positive.
A reset can move at most half a grid step, even when its original position divides the cell very unequally. Squaring each half-cell density’s deviation from one and summing over reduced fractions measures how much of this imbalance persists as the grid gets finer.
Write H_n=\sum_{j=1}^n1/j and put L_0(p)=\sum_{\xi\in\mathcal R_N}\delta_\xi^2,\qquad L_1(p)=\sum_{\substack{\xi=a/d\in\mathcal R_N}} \lfloor N/d\rfloor\delta_\xi^2 .
Proposition 4.4 (Variance of the restoring density). The density has the exact variance \int_0^1(\varpi_p-1)^2\,dy =\frac{L_0(p)}Q =\frac1{\pi^2}+O\left(\frac{\log p}{p}\right).\tag{35} Its two phase moments are \begin{aligned} L_0(p)&=\frac13\sum_{d=2}^N\varphi(d) +\frac23\sum_{d=2}^N\frac{\kappa(d)}d,\notag\\ L_1(p)&=\frac{Np}{6}-N+\frac23H_N. \end{aligned}\tag{36} For a fixed Lipschitz function g, \left|\int_0^1g(y)(\varpi_p(y)-1)\,dy\right| \leq\frac{\operatorname{Lip}(g)}{2Q}.
Proof. Each occupied cell has density deviations \delta_\xi,-\delta_\xi on equal halves, giving its variance \delta_\xi^2/Q. At denominator d, summing these squares gives 4V_d. The evaluation of V_d in the proof of Proposition 4.3 gives 4V_d=\varphi(d)/3+2\kappa(d)/(3d), proving the first moment formula. Partitioning each full denominator clock by reduced denominators gives L_1=\sum_{h=1}^N\frac{(h-1)(h-2)}{3h}, which is the second formula. Moreover |\kappa(d)|\leq d. Counting ordered coprime pairs by divisor inversion gives 2\sum_{d=1}^N\varphi(d)-1 =\sum_{r=1}^N\mu(r)\lfloor N/r\rfloor^2. Replace the squares by N^2/r^2, with total error O(N\log N). The omitted tail is O(N). The absolutely convergent inverse product \sum_r\mu(r)/r^2=1/\zeta(2)=6/\pi^2 therefore gives \sum_{d\leq N}\varphi(d)=3N^2/\pi^2+O(N\log N). Substitute in (36) and divide by Q. Finally the transported integral of g is \int g(\Theta_p(x))\,dx. The displacement bound (28) gives the Lipschitz estimate. ◻
More generally, in the same (N,Q) class, \int_0^1(\varpi_{N,Q}-1)^2\,dy =\frac{N^2}{\pi^2Q}+O\left(\frac{N\log N}{Q}\right). This follows from the same primitive moment sum, which is independent of Q when multiplication by Q permutes every denominator’s units.
For a fixed Lipschitz test the two half-cell contributions cancel up to O(1/Q). Squaring the density deviations adds them instead. Counting coprime pairs gives 1/(6\zeta(2))=1/\pi^2 here as the limiting variance. The same coprimality factor enters the coefficient -1/\pi^2 of p^2(\log p)^2 in the cubic law [11]. The energy of a particular field still depends on the signed product in (31).
The reset forest occupies a small collection of the fine cells. Both its density variation and its signed contribution can be bounded there. The signed estimate retains the complete clock sums before their magnitudes are taken.
Let \mathfrak K_{\mathcal F} restrict the cell-length correction to cells whose reset denominator belongs to \mathcal F. Write f=|\mathcal F\cap[2,N]| and T_N=\max_{d\leq N}\tau(d).
Theorem 4.5 (Forest density bound). The forest density variation satisfies \|1_{\mathcal F\text{ cells}}(\varpi_p-1)\|_2^2 \leq\frac{Nf}{Q}.\tag{37} For the coherent cutoff source and t=N/p, |\mathfrak K_{\mathcal F}(\boldsymbol1)| \leq\frac{Nf}{Q} \left[(2+T_N/2)H_N^2+\frac23\right].\tag{38} If f\ll_\varepsilon p^\varepsilon for every \varepsilon>0, the two estimates give an O_\varepsilon(p^{-1/2+\varepsilon}) density norm and an O_\varepsilon(p^{-1+\varepsilon}) signed work correction.
Proof. For each denominator d, its phase-square sum is at most \varphi(d). Thus L_{0,\mathcal F}\leq Nf, proving (37). The weighted moment also satisfies L_{1,\mathcal F}\leq Nf, since \lfloor N/d\rfloor\varphi(d)\le N. Put j_d=\sum_{a\in\mathcal U_d}B_N(a/d)\psi(Qa/d),\qquad J_{\mathcal F}=\sum_{\substack{d\in\mathcal F\\d>1}}j_d. Partitioning the full clocks and using divisor inversion gives j_d=2\sum_{e\mid d}\mu(d/e)W_e(\boldsymbol1),\qquad |j_d|\leq\frac12\sigma_1(d)H_{\lfloor d/2\rfloor}^2. Here we used the complete-clock estimate |W_e(\mathbf1)|\le eH_{\lfloor e/2\rfloor}^2/4 from Section (3.1) for e\ge2 and W_1=0. Consequently |J_{\mathcal F}|\leq NfT_NH_N^2/2, while the earlier signed forest estimate gives |A_{\mathcal F}|\leq NfH_N^2/2. The density formula now reads exactly \mathfrak K_{\mathcal F}(\boldsymbol1)=\frac1Q \left(4A_{\mathcal F}-tJ_{\mathcal F} +\frac{t^2L_{0,\mathcal F}}6 -\frac{tL_{1,\mathcal F}}2\right). Use the preceding bounds and 0<t<1 to prove (38). The assumed bound on f and the divisor bound for T_N give both exponents. ◻
The density norm and the signed field work are different observables. The latter benefits from cancellation within complete physical clocks. The complementary cells carry the remaining density variation and its full interaction with the coherent field.
After centering, a field of zero slope spends half of each occupied cell at one trace and half at the other. Its mean square in that cell is the mean of the two squared endpoint values. The common reset at zero and the rise across a cell account for the terms that can spoil this agreement. All the interior midpoints enter through one signed sum.
The mass s_c fixes the two traces at the origin. The rise t_c is the same in every cell. For the third reading, weight the field at each interior reset by its signed displacement from the cell midpoint, in units of the cell width, and add.
Write \mathcal B_p(c,w)=C_p^0(c,w)-G_p(c,w)/Q and \mathcal J_p(c)=\sum_{d=2}^Nj_d(c). The notation L_0 retains the primitive squared phase moment of (36).
Theorem 5.1 (Three-reading formula). For every cutoff source, \mathcal B_p(c,c)=\frac{s_c^2}{Q}+\frac{t_c^2}{12} -\frac{s_ct_c}{2Q}+\frac{t_c\mathcal J_p(c)}{Q} -\frac{t_c^2L_0}{6Q}.\tag{39} Its rank is at most three, with source readings s_c,t_c,\mathcal J_p(c). On the paired space \operatorname{ran} A_+ its rank is at most two. In particular, C_p^0(c,w)=G_p(c,w)/Q \quad\text{if }s_c=t_c=s_w=t_w=0.\tag{40}
Proof. Put \theta=\{Q\xi\} and \psi=\theta-1/2 in an occupied cell. Integration of its two affine halves shows that its unweighted energy, multiplied by Q, differs from (\ell^2+r^2)/2 by -\frac{t_c^2}{6}+\frac{t_cm_d(c)}2 +\frac{t_c}2(\ell+r)\psi+\frac{t_c^2}3\psi^2. In an empty cell that difference is -t_c^2/6. The trapezoidal sum, using origin traces -s_c,+s_c, is G_p(c,c)+s_c^2. Write L_{d,a},R_{d,a} for the endpoint samples in the cell containing a/d. Summing the masses and the phase-weighted endpoint values gives \sum_{d=2}^N\varphi(d)m_d(c)=\sum_h(h-1)c_h =Qt_c/2-s_c, \sum_{d,a}(L_{d,a}+R_{d,a})\psi(Qa/d) =2\mathcal J_p(c)-t_cL_0/2. Sum the cell identities, use \sum_{d,a}\psi(Qa/d)^2=L_0/4, and simplify to prove (39). Its polarization factors through the three stated linear readings, proving the rank bound. For paired sources \sum_hh c_h=ps_c/2, hence t_c=s_c/p. Only s_c and \mathcal J_p(c) remain. Restriction to s=t=0 and polarization prove (40). ◻
In Figure 1 the weights have zero mass and slope. The theorem explains why centering the jumps there removes the whole flat discrepancy. Any source with those two vanishing sums has the same exact agreement with its samples in the centered coordinate.
Many fields give the same three readings. Choose the one of least sampled energy and subtract it from the original field. Their difference has all three readings zero. The two pieces are orthogonal on the grid, yet restoring the cell lengths can give them a mixed energy.
Theorem 5.2 (Least-energy projection). On the full source take Uc=(s_c,t_c,\mathcal J_p(c)). On the paired source take U_+u=(s_{A_+u},\mathcal J_p(A_+u)). Choose independent rows spanning the row space of U or U_+, and write G for the corresponding raw sample Gram. Put T=G^{-1}U^T(UG^{-1}U^T)^{-1},\qquad \Pi=TU.\tag{41} For b=\Pi c, z=(\mathrm I-\Pi)c, the source b uniquely minimizes sampled work subject to Ub=Uc. Moreover Uz=0,\qquad G(b,z)=0,\qquad \mathcal B_p(z,w)=0\quad\hbox{for every source }w. Consequently C_p^0(b,z)=0 and C_p^0(z,z)=G(z,z)/Q. For every source c=b+z, its sampling defect is \mathcal S_p(c)=\mathcal B_p(b,b)+\mathfrak K_p(b,b) +2\mathfrak K_p(b,z)+\mathfrak K_p(z,z).\tag{42}
Proof. The independence of the rows of U and the positivity of G make UG^{-1}U^T positive definite. Direct multiplication gives UT=\mathrm I, \Pi^2=\Pi and \Pi^TG=G\Pi. Thus Uz=0 and G(b,z)=0. Any other source with the same readings differs from b by a vector in \ker U, whose positive sampled energy adds orthogonally. This proves uniqueness. Theorem 5.1 factors \mathcal B_p through U, so z lies in its radical, giving both assertions about C_p^0. Expand C_p=C_p^0+\mathfrak K_p and subtract the complete sampled energy to obtain (42). ◻
The inverse in (41) has a direct meaning. For a single reading r^Tc, its greatest square per unit of continuous energy is r^TC^{-1}r, where C=C_N. In sampled energy the corresponding value is r^TH^{-1}r, with H=G/Q on the full source. For the three readings together these values and their pairings form UH^{-1}U^T=QUG^{-1}U^T. On the opening grid the projection lets us measure the mixed energy after restoring the cell lengths.
Return to the six weights in Figure 1. The mass and slope vanish. Denominator three contributes 1/3 to the third reading, and denominator five contributes -2/5. The other denominators contribute zero, giving \mathcal J_7(c)=-1/15. The fifths enter this reading even though the field has no jump there. The least-energy projection (41) gives b=\frac1{1054}(-81,77,79,-86,32,-21),\qquad z=c-b . Here G_{ij} is the raw sum of the sample products in (1). The third row of U is the sum of 2\psi(ha/d)\psi(49a/d) over a\in\mathcal U_d, 2\le d\le6. Then Ub=Uc, Uz=0 and b^TGz=0. The four terms of (42), in their order there, are 0,\qquad \frac{22082}{204130815},\qquad -\frac{2896}{1388645},\qquad \frac{397214}{13608721}. For example the third entry is 2b^TC_6z, since both the sampled and the flat mixed energies vanish. The four entries sum to 4/147, the area already found in two cells. The original lengths give the two orthogonal sampled fields a negative interaction of their own.
Give every harmonic weight one. At the origin all p-1 waves reset together. The field reads zero there, between traces of opposite sign whose squares are both (p-1)^2. The prime-square grid assigns that one address weight 1/p^2. Its contribution to the defect approaches one. The same unit has a second description. The individual harmonic defects add to only O(1/p), so in the square of the sum the unit is carried by the interactions between distinct harmonics.
Two harmonics suffice to reach relative error 1/p to leading order. Along the primes of Theorem 2.5, whole families of harmonics reinforce the same signed error and exceed that scale by more than any fixed power of \log p. At equal weights the continuous energy grows like p. The carry table asks for the absolute error after its unit has been subtracted. That finer precision depends on how the other interactions add in this one field.
The jump positions already give one unconditional reading of the remainder, \mathcal S_p-1=2\mathcal Z_p/p^2. The signed distance \mathcal Z_p measures how far rounding moves the pairs of jumps relative to one another. In the original coordinate, Section (3.1) gives \mathcal S_p-1=8U_{\rm other}/Q+O_\varepsilon(p^{-1+\varepsilon}) whenever its f selected interior denominators satisfy f\ll_\varepsilon p^\varepsilon for every \varepsilon>0. In the centered coordinate put \mathcal S_p^0=\mathcal B_p(\mathbf1,\mathbf1) and let \mathfrak K_{\rm other} sum the restoring correction over the complementary cells. By (30) and (38), the proposed absolute half-power bound is equivalent to (\mathcal S_p^0-1)+\mathfrak K_{\rm other}(\mathbf1) =O_\varepsilon(p^{-1/2+\varepsilon}). This is one signed sum. The three-reading formula evaluates its flat part exactly. The forest estimate controls the part of the restored lengths already removed. A bound on the displayed combination remains open.
On the 49-point grid, the longer piece of the cell at 1/3 lies on the right. At 2/3 it lies on the left. In both cells the square sits on the longer piece. Reflection adds the two corrections. As p grows, no reset moves more than half a grid step under centering, yet the variance of the density that restores those lengths tends to 1/\pi^2. On the p^2-point grid the signed readings determine every harmonic weight. To bound the equal-weight remainder, we still have to explain which excesses of area cancel over a period and which add as these two do.
[1]C. Aistleitner and J. Dick, Functions of bounded variation, signed measures, and a general Koksma–Hlawka inequality, Acta Arith. 167 (2015), 143–171. https://arxiv.org/abs/1406.0230.
[2]C. Aistleitner, I. Berkes and K. Seip, GCD sums from Poisson integrals and systems of dilated functions, J. Eur. Math. Soc. 17 (2015), no. 6, 1517–1546.\newline https://doi.org/10.4171/JEMS/537.
[3]M. Beck, Dedekind cotangent sums, Acta Arith. 109 (2003), no. 2, 109–130.\newline https://doi.org/10.4064/aa109-2-1.
[4]J. Franel, Les suites de Farey et le problème des nombres premiers, Nachr. Ges. Wiss. Göttingen, Math.-Phys. Kl. (1924), 198–201.
[5]H. Rademacher, Generalization of the reciprocity formula for Dedekind sums, Duke Math. J. 21 (1954), 391–397.
[6]T. Hilberdink, F. Luca and L. Tóth, On certain sums concerning the gcd's and lcm's of k positive integers, Int. J. Number Theory 16 (2020), 77–90. https://doi.org/10.1142/S1793042120500049.
[7]M. N. Huxley, Identities involving Farey fractions, Proc. Steklov Inst. Math. 276 (2012), no. 1, 125–139.\newline https://doi.org/10.1134/S0081543812010105.
[8]J. S. Ovall, An analysis of GCD and LCM matrices via the LDL^T-factorization, Electron. J. Linear Algebra 11 (2004), 51–58.\newline https://doi.org/10.13001/1081-3810.1121.
[9]A. S. Petty, The Cubic Law for Digit-Collision Energy, 2026.\newline https://doi.org/10.5281/zenodo.20547910.
[10]A. S. Petty, The Sampling Defect and Reset Forests, 2026.\newline https://doi.org/10.5281/zenodo.24282329.
[11]A. S. Petty, The Secondary Term in Digit-Collision Energy, 2026.\newline https://doi.org/10.5281/zenodo.21875096.
[12]A. S. Petty, The Unit Sampling Defect in Digit-Collision Energy, 2026.\newline https://doi.org/10.5281/zenodo.23285306.
[13]A. Durand and S. Jaffard, Multivariate Davenport series, arXiv:1205.2343. https://arxiv.org/abs/1205.2343.
[14]B. Kashin, E. Kosov, I. Limonova and V. Temlyakov, Sampling discretization and related problems, J. Complexity 71 (2022), 101653. https://doi.org/10.1016/j.jco.2022.101653.
[15]J. F. Koksma, Een algemeene stelling uit de theorie der gelijkmatige verdeeling modulo 1, Mathematica (Zutphen) 11B (1942/43), 7–11.
[16]T. Xylouris, On the least prime in an arithmetic progression and estimates for the zeros of Dirichlet L-functions, Acta Arith. 150 (2011), 65–91. https://doi.org/10.4064/aa150-1-4.
Discussion
Sign in to join the discussion.