Still preparing the reader…Long documents can take a little longer to load and typeset.
A continuous energy model can miss a finite contribution from an arithmetic sampling grid, even as that grid becomes finer. In digit-collision energy, the source grows with the grid, and a simultaneous reset of all its waves leaves a discrepancy tending to one. After this contact is accounted for, the rate of agreement depends on the signed interaction among the other resets. The established global rate is quarter-power; the target is half-power. Complete arithmetically selected families of resets attain that target. For forests whose surviving unit clocks divide an integer of polynomial height, the normalized absolute boundary sum is O_\varepsilon(p^{-1/2+\varepsilon}), while its signed contribution and restoration are O_\varepsilon(p^{-1+\varepsilon}). A specified growing return window obeys the same bounds. Divisor inversion retains the source responsible for each selection, and primitive tori realize its continuous gcd energy. A separate packet decomposition controls the complete cutoff fringe while keeping the Cubic Law’s full energy budget, including mixed terms. These estimates identify where the finite collision table already agrees with its continuous model at the desired rate. The full half-power problem remains an explicit signed reading over the complementary resets of the same field and grid.
Replacing a finite arithmetic table by a continuous energy integral requires more than a leading asymptotic. A single sampled reset can leave a finite discrepancy, and the interactions at the other resets determine how accurately the corrected integral represents the table. The sampling defect measures that difference. Bounding it separates the energy explained by the continuous model from the signed arithmetic information still carried by the finite observation.
Digit-collision energy supplies a concrete instance. Its carry table can be rearranged exactly into equally spaced readings of a sum of repeating ramps. Each ramp rises from minus one toward plus one and then resets. The ramps are added before the field is squared, so their interactions contribute to the energy. The continuous model integrates that square; the finite table averages it at its prescribed grid points. The straightening identity in Theorem 2.4 identifies this comparison with the original collision energy.
Use the midpoint sawtooth \psi(x)= \begin{cases}\{x\}-\tfrac12,&x\notin\mathbb{Z},\\0,&x\in\mathbb{Z},\end{cases} \qquad B_N(x)=2\sum_{j=1}^N\psi(jx). The jth ramp completes j circuits on 0\le x\le1. For an odd prime p, put N=p-1, Q=p^2, and \mathcal S_p=\int_0^1B_N(x)^2\,dx -\frac1Q\sum_{a\bmod Q}B_N(a/Q)^2.
At p=5 there are four ramps and twenty-five samples. Their sum is linear between the reset addresses 0,\tfrac14,\tfrac13,\tfrac12,\tfrac23,\tfrac34,1. Integrating its square on those intervals and summing the twenty-five sampled squares gives \int_0^1 B_4(x)^2\,dx=\frac{23}{9},\qquad \frac1{25}\sum_{a=0}^{24}B_4(a/25)^2=\frac{48}{25},\qquad \mathcal S_5=\frac{143}{225}. At the origin B_4(0)=0, while the right and left traces on the circle are -4 and 4. The square approaches 16 on either side, although the sample reads zero. Replacing that squared sample by 16 adds 16/25 to the finite average. The corrected difference is \frac{23}{9}-\left(\frac{48}{25}+\frac{16}{25}\right) =-\frac1{225}. The correction slightly overshoots at this cutoff. Figure 1 shows both the missed origin contact and the samples away from it.
For general p, the same common reset contributes N^2/Q=(p-1)^2/p^2, which tends to one. Section 4 separates this contact exactly from the remaining reset interactions and the curvature term. Changing one point does not change an integral, but it changes a finite average. Here the height of that one squared reading grows like the number of samples, so its normalized contribution survives refinement. Both the field and its grid grow with p; convergence for a fixed function does not settle their difference.
The established estimate is \mathcal S_p=1+O_\varepsilon(p^{-1/4+\varepsilon}) for every \varepsilon>0 [7]. The full target is \mathcal S_p=1+O_\varepsilon(p^{-1/2+\varepsilon}). Since N^2/Q=1+O(p^{-1}), either rate also describes the error after the finite contact correction. A half-power bound would show that the continuous model, with this correction, tracks the growing collision table more closely than the quarter-power estimate guarantees. The remaining question is how much cancellation the arithmetic forces among the other boundaries.
The boundary identities in Collision Calculus [4] retain the point readings, reset jumps, curvature and interactions needed to ask that question. Here whole families of resets meet the half-power bound even when charged at absolute size. Keeping their signs gives the stronger first-power rate. The selection must retain the complete field read at every chosen address, so that its estimate applies to the original discrepancy.
A clock is one component 2\psi(jx) of the field. At a reduced address a/k, every clock with k\mid j resets together. For a set \mathcal A of nonorigin reset denominators put \mathcal A_{\mathcal A} =\sum_{k\in\mathcal A}\left\lfloor\frac Nk\right\rfloor \sum_{a\in(\mathbb{Z}/k\mathbb{Z})^\times}B_N(a/k)\psi(Qa/k). Write \operatorname{rough}_N(k) for the largest divisor of k coprime to N. The reset forest selected by a positive integer D is \mathcal F(N,D)=\{2\le k\le N:\operatorname{rough}_N(k)\mid D\}. The terminology records what remains of a denominator after repeated multiplication by N: its primes dividing N disappear, leaving \operatorname{rough}_N(k). No restriction is imposed on the source harmonics read at those addresses.
At the prime thirty-one, the field has thirty clocks. One return-selected forest contains every denominator from two through thirty except seventeen, twenty-three and twenty-nine. The selection rule and its growing-cutoff estimate are given below. Both parts still read all thirty waves. The issue is what permits an entire family of their reset interactions to satisfy the stronger bound.
Theorem 1.1 (Controlled forests and the complementary reading). Let D=D(p) satisfy 1\le D(p)\le Cp^K for fixed C,K, and let \mathcal F=\mathcal F(N,D). For every \varepsilon>0, \begin{aligned} \sum_{k\in\mathcal F}\left\lfloor\frac Nk\right\rfloor \sum_{a\in(\mathbb{Z}/k\mathbb{Z})^\times} |B_N(a/k)\psi(Qa/k)|&=O_\varepsilon(p^{3/2+\varepsilon}), \\ |\mathcal A_{\mathcal F}|&=O_\varepsilon(p^{1+\varepsilon}). \end{aligned} The implied constants may also depend on C,K. If \mathcal A_{\rm other} is the same signed reading over 2\le k\le N outside \mathcal F, then \mathcal S_p-1=\frac4{p^2}\mathcal A_{\rm other} +O_\varepsilon(p^{-1+\varepsilon}). Consequently (5) is equivalent to \mathcal A_{\rm other}=O_\varepsilon(p^{3/2+\varepsilon}).
The proof is given at the end of Section 9.
After sampling normalization, the selected forest meets the half-power target even in absolute work. Keeping its signed interactions improves its normalized contribution to first-power scale. A specified return window may also grow while retaining these bounds; its hypotheses are given in Section 8. Under those hypotheses the forest has subpower population. The large complementary population remains, with the same source and detector, and carries the remaining half-power question.
The Cubic Law [5] supplies the energy available to these interactions, E_p=p^3+O\bigl(p^2(\log p)^2\bigr). Here E_p=\sum_{a\bmod Q}B_N(a/Q)^2 by the collision straightening identity in Theorem 2.4. The continuous energy has its own secondary terms. The sampling defect is the difference between the two complete readings, not the entire remainder in the Cubic Law. The leading harmonic diagonal and off-diagonal masses of E_p are p^3/3 and 2p^3/3. Primitive torus fields realize the continuous gcd pairing in counting coordinates. The complete-clock bound then prices the selected resets. For the complement, lower and upper arithmetic branches organize the remaining clocks into packets of two or three, with their entire fringe within the half-power allowance. The final source-dependent readout criterion states a sufficient bound for their signed interaction.
These arguments use classical inputs in a specified cutoff problem. The divisor factorization is a gcd-matrix factorization [3], and the finite sawtooth transforms belong to the Dedekind-sum setting [1]. The complete reset identity is the coprime prime-square specialization of Huxley’s weighted identity [2], in the midpoint convention explained in [7]. The projection calculations at the end are finite-dimensional orthogonal decompositions. The forest and packet arguments retain the original source and both energy readings through these changes of coordinates.
Energy is quadratic in the complete field. Waves add before their sum is squared. Their mixed terms therefore travel with every change of source, coordinates or sampler. Below, native energy means the energy of the specified field under its specified reading and weights.
Selecting resets can change the source that produces them. To distinguish a genuine part of the sampling error from the error of a different field, both energy readings must be carried through that selection. Source coefficients keep this information. The associated fibre path makes the interactions between reset directions visible.
For c\in\mathbb{R}^N, set F_c(x)=2\sum_{j=1}^N c_j\psi(jx),\qquad \mathcal I(c,e)=\int_0^1F_cF_e, \mathcal H_d(c,e)=\frac1d\sum_{a\bmod d}F_c(a/d)F_e(a/d). Write \mathbf1 for the complete coefficient source and \mathsf C_p,\mathsf G_p for the matrices of \mathcal I and Q\mathcal H_Q. Thus (\mathsf C_p)_{ij}=\frac{(i,j)^2}{3ij},\qquad (\mathsf G_p)_{ij}=\sum_{a\bmod Q}F_{e_i}(a/Q)F_{e_j}(a/Q). The continuous formula follows by expanding the midpoint waves in Fourier series and retaining their common frequencies. Equivalently, divide i,j by their gcd and integrate over their common period. The bilinear form measures cross-work as well as squared energy. Let H_n=\sum_{j=1}^n1/j, with H_0=0.
Lemma 2.1 (Positivity of the energy forms). Both \mathsf C_p and \mathsf G_p are positive definite.
Proof. For a continuous zero-energy source, F_c vanishes on each interval between its resets. Its jump at a reduced denominator k is -2\sum_{k\mid j}c_j. Starting with the largest nonzero harmonic shows that all coefficients vanish.
For sampled work, any two distinct reset fractions with denominators at most N are separated by at least N^{-2}>Q^{-1}. Every open reset interval therefore contains a Q-grid point. The first interval (0,1/N) contains at least two such points. If all sampled values vanish, those two points make the common slope of F_c zero. Each remaining interval is constant and has a zero sample, hence is zero. The same jump argument gives c=0. ◻
Proposition 2.2 (The complete fibre coordinates). For 0\le v<p, put z_c(u)_v=F_c((v+u)/p)-F_c(u)/p,\qquad \mathcal V_p=\{z\in\mathbb{R}^p:\ \sum_vz_v=0\}. Then z_c(u)\in\mathcal V_p. The map from c to its entire path is injective and reconstructs both energy readings: \begin{aligned} \int_0^1z_c\cdot z_e&=(p-p^{-1})\mathcal I(c,e),\\ \mathcal H_{p^2}(c,e)&=\frac{\mathcal H_p(c,e)}{p^2} +\frac1{p^2}\sum_{r\bmod p}z_c(r/p)\cdot z_e(r/p). \end{aligned} In particular every mixed source term and every nonunit address remains.
Proof. For 1\le j<p, multiplication by j permutes the p residues. The midpoint distribution identity therefore gives \sum_vF_c((v+u)/p)=F_c(u). Subtract the orthogonal mean direction. Integrating the sum of products with dx=du/p proves (9). Grouping the sample labels as a=pv+r proves (10). If the path is zero, its continuous work is zero. The largest nonzero harmonic coefficient has an uncancelled denominator jump; a source supported at 1 alone is also nonzero. Hence the finite sawtooth fields are independent and c=0. ◻
Proposition 2.3 (The reset directions). Put w_v=e_v-\mathbf1/p and m_k(c)=\sum_{k\mid j}c_j. Between resets the path z_c is constant. At a reduced reset t=a/k, 2\le k\le N, \Delta z_c(t)=-2m_k(c)w_v,\qquad v=-ak^{-1}\pmod p,\qquad w_v\cdot w_b=\delta_{vb}-p^{-1}. For an integral source, pz_c is an integral zero-sum vector. For the complete source, z_{\mathbf1}(0+)=-Nw_0,\qquad z_{\mathbf1}(1-)=Nw_{p-1},\qquad \|z_{\mathbf1}(0+)\|^2=N^3/p. At that reset, with b=ap^{-1}\pmod k, z_c(a/k)_v+F_c(a/k)/p=F_c(b/k).
Proof. The derivative of each fibre component equals the derivative of its subtracted mean. At a/k, precisely the harmonics divisible by k reset in the single fibre satisfying p\mid vk+a. This proves the jump and its counting Gram. Direct evaluation gives pz_{e_j}(0+)_v=2(jv\bmod p)-N, proving integrality. For the whole source its nonzero p-grid readings vanish; odd reflection supplies the final trace. Finally vk+a=pb proves (11), including midpoint values. ◻
The path has p-1 dimensions and retains the negative interactions \delta_{vb}-1/p between distinct reset directions. At p=31 the source therefore has thirty dimensions, regardless of how many coordinates a particular boundary record uses.
Define the threshold count of the collision table by f_p(a)=\sum_{d=0}^{p-1} \left[ \left\lfloor\frac{((p+1)d+1)a}{Q}\right\rfloor -\left\lfloor\frac{(p+1)da}{Q}\right\rfloor \right]. Its centered value gives a pointwise link between the collision table and the complete sampled field. This is the straightening identity of [6]. The formulation below also records the full source Gram, including its nonunit rows.
Theorem 2.4 (Collision-to-Bernoulli straightening). Let f_p(a) be the threshold count of the Cubic Law [5]. For a\in(\mathbb{Z}/Q\mathbb{Z})^\times, its class-centered value is f_p(a)-a/p=B_N((p+1)a/Q). The collision energy is therefore E_p=Q\mathcal H_Q(\mathbf1,\mathbf1)= \mathbf1^T\mathsf G_p\mathbf1,\qquad (\mathsf G_p)_{ij}=\sum_{a\bmod Q}F_{e_i}(a/Q)F_{e_j}(a/Q). Dilation F_c(x)\mapsto F_c((p+1)x) preserves both cutoff bilinear work forms for every source. The exact harmonic diagonal is \operatorname{tr}\mathsf G_p=\frac{N(Q-1)(Q-2)}{3Q}. The nonunit Gram \mathsf G_p^\nu remains part of this full form. It satisfies \mathsf G_p^\nu\mathbf1=0,\qquad \operatorname{tr}\mathsf G_p^\nu=\frac{N^2(p-2)}{3p}. Its off-diagonal sum is the negative of its diagonal sum.
Proof. Put t=(p+1)a/Q and \alpha=a/p, so pt=a+\alpha. The threshold formula gives f_p(a)=\sum_{d=0}^N \left[\lfloor(d+1)t-\alpha\rfloor-\lfloor dt\rfloor\right]. Since (d+1)t-\alpha=a-(N-d)t, and jt is nonintegral for 1\le j<p, this equals pa-N-2\sum_{j=1}^N\lfloor jt\rfloor. Expanding B_N(t) proves (12). The class mean in the collision table is -(p-s)/p at a\equiv s\pmod p, so the left side is precisely its class-centered value. Multiplication by p+1 permutes every Q-grid label and preserves continuous integration by integer dilation. The complete field is zero at each nonunit a=pb, by reflection of its full p-period. This proves (13) and both bilinear work identities.
Each j<p permutes the Q-grid. Summing (2a-Q)^2/Q^2 for 1\le a<Q proves (14). On the nonunit grid the same calculation has modulus p, giving N(p-1)(p-2)/(3p). Its full Gram annihilates \mathbf1, since the complete source vanishes there. Hence its off-diagonal sum cancels its trace. ◻
The full diagonal is the Cubic Law’s one-third leading mass. The complementary harmonic interactions carry the other two thirds. Nonunit rows vanish on \mathbf1 but remain in source variations and in the two energy forms.
The carry table and the coherent field now share a complete source Gram. A forest can be separated within this work, with its interactions accounted for. The arithmetic structure of that Gram supplies the next coordinates.
The continuous energy has arithmetic coordinates in which its primitive parts are orthogonal. These coordinates also have a literal finite geometry. At denominator two, the primitive pairs are (0,1),(1,0),(1,1) modulo two. Their first-coordinate projection has one point at zero and two at one half. Applying the same sawtooth to those three points gives 2\psi(x)+4\psi(x-\tfrac12)=4\psi(2x)-2\psi(x). The construction at every denominator supplies a basis for the cutoff work. Its counting weights give the continuous energy of each primitive component before that field is restricted to a reset grid.
For this construction N is any positive integer. Write C_N for the continuous Gram of the first N sawtooth sources and put J_2(n)=\sum_{d\mid n}\mu(n/d)d^2,\qquad g_n(x)=2\sum_{d\mid n}d\mu(n/d)\psi(dx). The finite arithmetic torus is (\mathbb{Z}/n\mathbb{Z})^2, with primitive points defined by \gcd(a,b,n)=1.
Theorem 3.1 (Primitive torus fields and their work). The fields g_1,\ldots,g_N are an orthogonal basis of the cutoff source space, with \langle g_n,g_m\rangle=\frac{J_2(n)}3\,\mathbf1_{n=m}. They have the literal projection formula g_n(x)=2\sum_{\substack{a,b\bmod n\\(a,b,n)=1}}\psi(x-a/n). There are J_2(n) primitive points. At a first coordinate whose reduced denominator is r\mid n, the marginal count is r\varphi(n/r).
Proof. Let E_{h,d}=\mathbf1_{d\mid h} and D=\operatorname{diag}(1,\ldots,N). The divisor identity \sum_{d\mid k}J_2(d)=k^2 gives the standard incidence factorization of a gcd matrix [3], C_N=\tfrac13D^{-1}E\operatorname{diag}(J_2(d))E^TD^{-1}. Consequently, with q_d(c)=\sum_{d\mid h}c_h/h, C_N(c,c)=\tfrac13\sum_{d\le N}J_2(d)q_d(c)^2. The source of g_n has c_h=h\mu(n/h) for h\mid n and zero otherwise. Its q vector is e_n, by divisor inversion. This proves the Gram identity and the basis assertion.
In (16), exclude common divisors of a,b,n. For k\mid n the unrestricted b sum gives n/k copies. The midpoint distribution identity \sum_{a=0}^{d-1}\psi(x-a/d)=\psi(dx) holds also at resets. The resulting expression is 2\sum_{d\mid n}d\mu(n/d)\psi(dx)=g_n(x). The same exclusion counts the points as \sum_{k\mid n}\mu(k)(n/k)^2=J_2(n). For a fixed a, put g=(a,n) and r=n/g. The admissible second coordinates are those coprime to g, numbering n\varphi(g)/g=r\varphi(n/r). ◻
The gcd pairing is therefore the projection of a finite counting geometry. Its primitive masses determine the work of each component. Restricting those components to a reset grid changes their orthogonality. The defect asks how that restriction reads the coherent sum, including the interactions that the full-circle coordinates separate.
A reset changes the field through all harmonics sharing its denominator. Recording the resulting energy jump at the sampler’s phase gives an exact formula for the defect. This fixes the surviving unit, the curvature correction and the source derivative in the same units. The identity specializes Huxley’s weighted formula [2]; the midpoint conversion and a direct proof appear in [7].
At a reduced reset t=a/k, put r_k(c)=\sum_{k\mid j}c_j,\qquad t_k(c)=\sum_{a\in\mathcal U_k}F_c(a/k)\psi(Qa/k), \qquad \mathcal U_k=(\mathbb{Z}/k\mathbb{Z})^\times . For a fixed clock collection \mathcal A, define D_{\mathcal A}(c)=4\sum_{k\in\mathcal A}r_k(c)t_k(c). Its coherent clock contribution is q_k=4m_k t_k(\mathbf1),\qquad m_k=\lfloor N/k\rfloor. The midpoint field at a reset is the mean of its two traces.
Theorem 4.1 (The complete reset identity). For s_c=\sum c_j, D_c=2\sum jc_j, h_c=\sum c_j/j, \mathcal I(c,c)-\mathcal H_Q(c,c) =\frac{s_c^2+D_{\rm all}(c)-D_ch_c/(3Q)}Q. In particular, \mathcal S_p=\frac{N^2+\sum_{k=2}^Nq_k}{p^2} -\frac{NH_N}{3p^3}. For u_j=j, v_j=1/j, put \mathsf R_p=(uv^T+vu^T)/(3Q). If \mathsf B_{\rm all} is the matrix of D_{\rm all}, the full sampling half-gradient at \mathbf1 is \ell_{\mathcal S} =\frac{N\mathbf1+\mathsf B_{\rm all}\mathbf1-\mathsf R_p\mathbf1}{Q}.
Proof. The field jump at a/k is -2r_k(c); its squared-energy jump is j_t=-4r_k(c)F_c(t). Let L_Q(\rho)=\sum_{\{Qt\}\le\rho}j_t over nonorigin resets, and P_c(\rho)=\int_0^\rho F_c. For the shifted grid energy \mathcal H_Q^{[\rho]}=Q^{-1}\sum_aF_c((a+\rho)/Q)^2, for 0<\rho<1 with \rho\ne\{Qt\} at every nonorigin reset, the midpoint distribution identity gives L_Q(\rho)=Q\mathcal H_Q^{[\rho]}-Q\mathcal H_Q^{[0]} -s_c^2-\frac{2D_c}{Q}P_c(\rho). Its one-sided jumps and its derivative between resets agree on both sides. At a phase that hits a reset, the square of the midpoint field need not equal the average of its one-sided squares. The displayed equality is used away from those finitely many phases; their values do not affect the integrals below. The origin changes from its actual midpoint zero to -s_c, fixing the constant. Reflection makes the total nonorigin jump mass zero, so \int_0^1L_Q=-\sum_tj_t\psi(Qt)=D_{\rm all}(c). Also \int\mathcal H_Q^{[\rho]}=\mathcal I(c,c) and \int P_c=-h_c/6. Integrating proves (19). Substitute c=\mathbf1 for (20). The curvature quadratic form is c^T\mathsf R_pc=D_ch_c/(3Q); differentiation proves (21). ◻
The factor N^2/p^2 tends to the common-reset unit. The rest of the defect is now a sum of complete clock readings and an explicit curvature correction. A bound for a selected collection applies directly to this identity. The next step is to expose the arithmetic structure inside each of those readings.
Each clock reads the unfinished part of the cutoff after complete periods have cancelled. Reflection folds that word again. Centering the prime residue then gives a common phase coordinate for the two branches, while their denominators and reset loads remain distinct.
Theorem 5.1 (The reflected source word). For a clock 2\le k\le N, set r=N\bmod k,\qquad s=\min(r,k-1-r),\qquad m=\lfloor N/k\rfloor. For every integer a, including nonunit addresses, B_N(a/k)=B_s(a/k),\qquad s\le\lfloor(N-1)/3\rfloor. With a tie assigned to the lower branch, the complete clock has \begin{array}{c|c|c|c} &\text{carrier}&\text{load}&\text{multiplier modulo }k\\ \text{lower}&k\mid N-s,\ k\ge2s+1&(N-s)/k&(s+1)^2\\ \text{upper}&k\mid p+s,\ k\ge2s+2&(p+s)/k-1&s^2 . \end{array} Both multipliers are represented by \kappa_{p,s}=p+s(s+1) =(N-s)+(s+1)^2=(p+s)+s^2. At k=2s+1, source, phase and load agree.
Proof. A full k-period of the midpoint word sums to zero. Odd reflection also gives B_r(a/k)=B_{k-1-r}(a/k). This proves the source identity. Since r\ge s, k\ge2s+1, and N\ge k+r, one has N\ge3s+1. Substituting r=s or r=k-1-s in N=mk+r proves both carriers and loads. Squaring p modulo k gives the two multipliers, and (22) gives their common representation. At the tie their difference is 2s+1=k, and (p+s)/k-1=(N-s)/k. Thus it is counted once. ◻
Proposition 5.2 (The centered prime phase). For 2\le k\le N, set h=\min(p\bmod k,k-(p\bmod k)). Then 1\le h\le H=\lfloor p/3\rfloor. With a tie assigned below, the two branches are \begin{array}{c|c|c|c} &\text{clock condition}&\text{source word}&\text{load}\\ -&k\mid p-h,\ k\ge2h&B_{h-1}&(p-h)/k\\ +&k\mid p+h,\ k>2h&B_h&(p+h)/k-1 . \end{array} The actual sampling multiplier is h^2\bmod k on both branches. For A=p-h, B=p+h, g=(A,B) and L=\operatorname{lcm}(A,B), g\in\{1,2\},\qquad gL+h^2=p^2.
Proof. Write p=mk+h below and p=(m+1)k-h above. Centeredness gives p\ge3h, with strict inequality above. The word and load formulas follow from the reflected source word. Squaring p\equiv\pm h\pmod k gives the phase. Finally g\mid2p and g\mid2h; since h<p and p is prime, g is one or two. The product identity (p-h)(p+h)=p^2-h^2=gL proves (23). ◻
Proposition 5.3 (Integer phase lift and carry work). For every real x and integer s\ge0, s(s+1)x-B_s(x) =s+2\sum_{j=1}^s\lfloor jx\rfloor -\#\{j\le s:jx\in\mathbb{Z}\}\in\mathbb{Z}. Thus the phase lift retains the carries. At an actual unit root x=a/k, put u=\{px\},\quad v=B_N(x),\quad w=\lfloor u+v\rfloor,\quad C(u,w)=\frac{w(w+1-2u)}2. Then C(u,w)\ge0 and the clock work is q_k=2m_k\sum_{a\in\mathcal U_k}B_N(a/k)^2 -4m_k\sum_{a\in\mathcal U_k}C(u,w).
Proof. Sum the defining sawtooths to obtain the integer lift. Since p+N(N+1)=Q, it gives \psi(Qx)=\psi(px+B_N(x)). For \Phi(y)=(\{y\}^2-\{y\})/2, direct expansion gives v\psi(u+v)=\Phi(u+v)-\Phi(u)+\frac{v^2}2-C(u,w). The actual unit-root phase Qx is nonintegral. Multiplication by p and by Q permutes the clock units, so the potential terms sum to zero. Multiply by 4m_k for (24). For w\ge1 or w\le-1 the displayed product defining C is nonnegative; it is zero at w=0. ◻
The source energy and carry moment in (24) are measured on the actual root grid. Their difference gives each clock’s signed work. Complete periods and reflection have reduced the coherent reading to two short-word branches with a common phase. That agreement allows whole clocks to be estimated together. For varying coefficients, the full source still has to accompany the shorter root word.
The signed work of a complete clock is much smaller than the sum of the absolute contributions of its individual addresses. This cancellation supplies a uniform budget for small populations of clocks and phase cells. It also identifies the whole population that remains after those allowances are taken.
Put R_B(j,d)=\sum_{a\bmod d}\psi(ja/d)\psi(Ba/d). These finite correlations are sawtooth forms of Dedekind sums; the cotangent transform used below is classical [1].
Write \tau(n) for the number of positive divisors and \sigma_1(n) for their sum.
Proposition 6.1 (The complete clock row). For 2\le d<p, \sum_{a\bmod d}|R_p(a,d)|\le\frac d4H_{\lfloor d/2\rfloor}^2. Put W_d=\sum_{j=1}^N R_Q(j,d), n_d(v)=\#\{1\le j<p:j=pv\pmod d\}, and \Gamma_d(a)=n_d(a)-n_d(a+1). Then |\Gamma_d|\le1 and W_d=\sum_{a=1}^{\lfloor(d-1)/2\rfloor}\Gamma_d(a)R_p(a,d).
Proof. The midpoint clock Fourier coefficient has magnitude at most 1/(4|u|) on signed frequencies 0<|u|\le d/2; its mean and even Nyquist coefficient vanish. The geometric sum gives i\cot(\pi u/d)/(2d); use \sin(\pi|u|/d)\ge2|u|/d. Thus |R_p(a,d)|\le\frac d{16} \sum_{u,v}\frac{\mathbf1_{d\mid au+pv}}{|uv|}. For fixed u, let g=(u,d). There are g solutions in a if g\mid v, and none otherwise. Their weighted v-sum is at most 2H_{\lfloor d/2\rfloor}; summing the signed u-weights proves the bound, including all nonunits. Change variable by the unit p in the definition of W_d. The counts n_d(v) differ by at most one and satisfy n_d(v)=n_d(1-v), by j\mapsto p-j. Pair the odd readings at v,-v to obtain (25). ◻
Theorem 6.2 (The complete-clock bound). For 2\le k\le N, |q_k|\le2NH_N^2\frac{\sigma_1(k)}k. The sum of absolute complete-clock works in any one centered phase cell is O_\varepsilon(p^{1+\varepsilon}), uniformly in the cell. Set R=\min(\lfloor\sqrt p\rfloor,H),\quad a_0=R+2,\quad b_0=H-1, \quad \mathcal J_p=\{k:a_0\le h_k\le b_0\}. Then \mathcal S_p-1=\frac1{p^2}\sum_{k\in\mathcal J_p}q_k +O_\varepsilon(p^{-1/2+\varepsilon}).
Proof. Let C_k=\sum_{a\in\mathcal U_k}B_N(a/k)\psi(Qa/k). Partitioning a complete d-grid by its reduced denominators gives W_d=\frac12\sum_{k\mid d}C_k. Finite divisor inversion and the preceding clock-row bound give C_k=2\sum_{d\mid k}\mu(k/d)W_d,\qquad |C_k|\le\tfrac12H_N^2\sigma_1(k). Multiply by 4m_k\le4N/k to obtain (26).
Clocks in a phase cell divide p-h or p+h. With T=\max_{n\le2p}\tau(n), their total absolute work is at most 4NH_N^2T^2, since \sigma_1(k)/k\le\tau(k). The divisor bound makes this O_\varepsilon(p^{1+\varepsilon}). The cells outside \mathcal J_p number at most R+2=O(\sqrt p). Their work is therefore at half scale. Insert this estimate in (20); the common-reset correction and smooth curvature are smaller than that scale. ◻
Corollary 6.3 (The reset-load tail). For any collection of reset clocks with m_k\ge M, \sum|q_k|\le 2NH_N^2K H_K^{(2)}\le\frac{4N^2H_N^2}{M}, \quad K=\lfloor N/M\rfloor,\quad H_K^{(2)}=\sum_{j=1}^Kj^{-2}.
Proof. The load condition implies k\le K, and \sum_{k\le K}\frac{\sigma_1(k)}k =\sum_{d\le K}\frac{\lfloor K/d\rfloor}{d} \le K H_K^{(2)}\le2K. Apply (26). ◻
This tail contributes O(\log^2p/M) after the specified normalization. The dense low-load population still needs its signed estimate. Equation (27) retains all of it.
Whole clocks and boundary cells now have a budget at the desired scale. An arithmetic rule selecting few enough surviving clocks can therefore control every reset feeding into them. Reset forests provide such a rule.
A reset clock reads the whole cutoff field. A bound for that coherent reading can control its incoming folds without pricing its waves separately. The finite Fourier calculation below keeps nonunit frequencies and sample addresses as part of the same grid.
Write \tau(k) for the divisor count and H_k=\sum_{j=1}^k j^{-1}.
Lemma 7.1 (Reset-grid work). For k\ge2 and c\in\mathbb{C}^{k-1}, \frac1k\sum_{a\bmod k} \left|2\sum_{h=1}^{k-1}c_h\psi(ha/k)\right|^2 \le \tau(k)^2H_k^2\sum_{h=1}^{k-1}|c_h|^2. For the actual cutoff, put r=N\bmod k. Then \sum_{a\bmod k}|B_N(a/k)|^2\le kr\tau(k)^2H_k^2, and, for every integer Q, \left\lfloor\frac Nk\right\rfloor \sum_{a\in(\mathbb{Z}/k\mathbb{Z})^\times}|B_N(a/k)\psi(Qa/k)| \le\frac N2\sqrt{k}\,\tau(k)H_k.
Proof. Use the normalized k-point Fourier transform. The midpoint double sawtooth has coefficient i\cot(\pi j/k)/k for 1\le j<k and zero at j=0, by a finite geometric sum. For g=(h,k), d=k/g and h'=h/g, its h column is supported on g\mid j, with coefficient \frac{i}{d}\cot\left(\frac{\pi(j/g)(h')^{-1}}d\right). Reflection and \sin(\pi v/d)\ge2v/d for v\le d/2 give \frac1d\sum_{v=1}^{d-1}|\cot(\pi v/d)|\le H_d. Every absolute column sum is therefore at most H_k.
For a fixed nonzero j, group columns by g\mid(j,k). Set d=k/g, w=(j/g,d) and e=d/w. Reduction of units from d to e is onto, with fibers \varphi(d)/\varphi(e). This group’s absolute sum is at most \frac{\varphi(d)}d\frac{e}{\varphi(e)}H_e \le\tau(e)H_k. There are at most \tau(k) groups and \tau(e)\le\tau(k). The absolute row sum is at most \tau(k)^2H_k. Weighted Cauchy–Schwarz bounds the squared matrix norm by the product of its row and column bounds. Parseval proves (28), including all nonunits.
A complete block of k source harmonics sums to zero at every k-grid address, including its midpoint resets. Thus B_N(a/k)=B_r(a/k). Apply (28) to its r unit coefficients. A final Cauchy–Schwarz inequality, |\psi|\le1/2 and \lfloor N/k\rfloor\le N/k give (29). ◻
This is a bound on the whole cutoff field at each denominator, ready to be summed over a forest.
The arithmetic folds separate a denominator into its incoming part and its surviving unit clock. When the possible surviving clocks are divisors of an integer of polynomial size, their entire reset forest has half-power work. This includes every mode at each retained denominator, rather than only its invariant cycle means.
Retain \operatorname{rough}_N and \mathcal F(N,D) from the opening. For D\ge1 write \mathcal F(N,D)=\{2\le k\le N:\operatorname{rough}_N(k)\mid D\}, \qquad \tau_3(D)=\sum_{\ell\mid D}\tau(\ell).
Lemma 8.1 (The incoming denominator budget). For every \eta>0, \sum_{\substack{m\le N\\ \text{every prime factor of }m\text{ divides }N}}\tau(m) \le C_\eta N^{2\eta}. Also \tau_3(D)\le C_\eta D^\eta.
Proof. The first sum is at most N^\eta\prod_{\ell\mid N,\ \ell\ {\rm prime}} (1-\ell^{-\eta})^{-2}. For \ell^\eta\ge4 each factor is at most \ell^\eta. Their product is bounded by \operatorname{rad}(N)^\eta\le N^\eta. The remaining finitely many primes give a constant depending on \eta.
For the second assertion, \tau_3(\ell^a)=(a+1)(a+2)/2\le3^a. Sufficiently large primes absorb this in \ell^{\eta a}. At each of the finitely many smaller primes the same bound holds with a finite factor, since exponential growth dominates the quadratic polynomial in a. Multiplication proves the claim. ◻
Theorem 8.2 (Whole-forest half-power work). For N=p-1, Q=p^2 and every D\ge1, \begin{aligned} &\sum_{k\in\mathcal F(N,D)} \left\lfloor\frac Nk\right\rfloor \sum_{a\in(\mathbb{Z}/k\mathbb{Z})^\times}|B_N(a/k)\psi(Qa/k)|\notag\\ &\hspace{25mm}\le \frac{C_\eta}{2}N^{3/2+2\eta}H_N\tau_3(D). \end{aligned} If D=D(p)\le Cp^K with fixed C,K, this is O_\varepsilon(p^{3/2+\varepsilon}) for every \varepsilon>0. The forest’s contribution to the sampling difference is therefore O_\varepsilon(p^{-1/2+\varepsilon}).
Proof. Write k=m\ell, where all prime factors of m divide N, (\ell,N)=1 and \ell\mid D. These factors are coprime, so \tau(k)=\tau(m)\tau(\ell). Bound \sqrt{k}\le\sqrt N and H_k\le H_N in (29). Sum over m using Lemma 8.1, and then over \ell\mid D. This gives (30). The polynomial-height assertion follows by choosing \eta sufficiently small and using H_N\le1+\log N. The cutoff boundary reading has factor 4/Q. ◻
A whole class of resets has reached the target scale. The estimate is uniform as the source and grid grow, and it includes all the incoming denominators above each surviving clock. A return rule gives a concrete instance of this class.
A concrete family is selected by relative-return congruences. Multiplying a reduced reset address repeatedly by NQ^{-1} leaves its denominator coprime to N after the incoming factors have disappeared. On that surviving denominator, a return after j steps means (NQ^{-1})^j\equiv1. Here Q is invertible modulo every denominator at most N. For an integer L\ge3 put D_{p,L}=15(NQ+1)(N^2-N-1) \prod_{\substack{1\le j\le L\\j\ {\rm odd}}}(Q^j-N^j). For a unit denominator k, the congruence NQ\equiv-1\pmod k puts k among the divisors of NQ+1. If NQ^{-1} has odd return length j\le L, then Q^j-N^j\equiv0\pmod k. The recurrence condition N^2\equiv N+1\pmod k gives the remaining factor. The fixed factor 15 includes those small divisors as well. These congruences are sufficient selection rules for this example; the general forest estimate applies to any positive divisor parameter. Any additional divisor roots stay in the bounded population.
At p=31,L=3, the unit roots below 30 are precisely 7,11,13,19. The whole reset forest consists of every denominator from 2 through 30 except 17,23,29. Both the retained and complementary clocks read the full thirty-harmonic field.
The three surviving denominators still read every source harmonic. The forest estimate has settled the other denominators together; the complementary population now carries the question.
Corollary 8.3 (A growing return window). The forest \mathcal F(N,D_{p,L}) has O_\varepsilon(p^{3/2+\varepsilon}) absolute work for each fixed L. The same conclusion holds for any specified integer window L=L(p)\ge3 with L(p)=o\bigl(\sqrt{\log\log p}\bigr). Its sampling contribution has the half-power bound.
Proof. The fixed-L polynomial in (31) has fixed degree. For a growing window, \log D_{p,L}=O(L^2\log p). An elementary divisor estimate gives \log\tau_3(n)=O(\log n/\log\log n). To verify it, put X=\log n. The primes at most \sqrt X number at most \sqrt X and each local factor is at most (1+X/\log2)^2. Their logarithmic contribution is O(\sqrt X\log X). At larger primes use \tau_3(\ell^a)\le3^a and total exponent at most 2X/\log X. This proves the estimate.
Since \log\log D_{p,L}\ge\log\log p-O(1), the stated window gives \tau_3(D_{p,L})=p^{o(1)}. Lemma 8.1, valid for every fixed \eta>0, also makes the incoming budget subpower. Apply (30). ◻
For example, \max(3,\lfloor(\log\log p)^{1/3}\rfloor) gives an unbounded admissible window, extended by 3 over its finite initial range. The bound keeps the full incoming trees above those roots.
Define the remaining complete boundary reading by \mathcal A_{\rm other}= \sum_{\substack{2\le k\le N\\ k\notin\mathcal F(N,D_{p,L})}} \left\lfloor\frac Nk\right\rfloor \sum_{a\in(\mathbb{Z}/k\mathbb{Z})^\times}B_N(a/k)\psi(Qa/k).
Corollary 8.4 (The complementary forest target). For a fixed or growing window of Corollary 8.3, \mathcal S_p-1=\frac4{p^2}\mathcal A_{\rm other} +O_\varepsilon(p^{-1/2+\varepsilon}). The full uniform half-power sampling estimate is equivalent to \mathcal A_{\rm other}=O_\varepsilon(p^{3/2+\varepsilon}).
Proof. The midpoint boundary identity for the unweighted cutoff is \mathcal S_p=\frac{N^2}{Q} +\frac4Q\sum_{k=2}^N\left\lfloor\frac Nk\right\rfloor \sum_{a\in(\mathbb{Z}/k\mathbb{Z})^\times}B_N(a/k)\psi(Qa/k) -\frac{N(N+1)H_N}{3Q^2}. This is the complete reset identity [7, 4]. Partition its denominator sum once by the forest. Theorem 8.2 and Corollary 8.3 bound the retained part. The common-origin term satisfies N^2/Q=1+O(1/p), and the last term is O(\log p/p^2). This proves (32) and both implications. ◻
The forest contribution already has the desired half-power estimate. Its source coefficients remain to be identified. Doing so will retain the signs through the selection and improve the bound further. The later quotient-parent partition has its own source ownership; combining its allowance with this one requires a map between the populations. Both partitions retain the complete Cubic observation, its one-third diagonal and two-thirds off-diagonal energy, and the full source.
A set of reset addresses selects a part of the boundary reading. Its source coefficients need not be the indicator of those addresses. Divisor inversion determines the source and its interaction with the rest of the field before any energy is assigned to it. For source coordinates, adjoin the denominator 1 and write \mathcal F=\{1\}\cup\mathcal F(N,D) throughout this section. Its nonorigin boundary reading is unchanged because W_1=0. A saturated divisor forest means a set \{d\le N:d\mid\Omega\} for some positive integer \Omega. For the present forest one may take \Omega=N^{\lceil\log_2N\rceil}\operatorname{rough}_N(D).
Proposition 9.1 (Reset selection and source coordinates). Let (\mathsf Zc)_k=\sum_{k\mid j}c_j. The unique coefficient map selecting a reset mask \sigma is P_\sigma=\mathsf Z^{-1}\operatorname{diag}(\sigma)\mathsf Z,\qquad (\mathsf Z^{-1}m)_d=\sum_{j\le N/d}\mu(j)m_{dj}. For a saturated divisor forest \mathcal F, choose \Omega with k\in\mathcal F iff k\mid\Omega, k\le N. Then P_{\mathcal F}e_j=e_{\pi(j)}, \pi(j)=(j,\Omega). The vectors e_d, d\in\mathcal F, and e_j-e_{\pi(j)}, j\notin\mathcal F, form an integral basis of the full source. Both work forms and their covectors are carried by their full Gram in that basis.
Proof. Finite divisor inversion follows from \sum_{d\mid n}\mu(d)=\mathbf1_{n=1}. The proposed gcd column has reset mass \mathbf1_{k\mid\Omega}\mathbf1_{k\mid j}, proving its formula by uniqueness. In forest-first order the basis matrix is \left[\begin{smallmatrix}I&-U\\0&I\end{smallmatrix}\right], where U_{dj}=\mathbf1_{d=\pi(j)}, and has determinant one. For either cutoff matrix W, its chart is J^TWJ. This retains its full cross block and covector 2J^TWJ\xi. For the actual coherent source its covariance is \xi\xi^T, where root coordinates count their collapse fibres and outside coordinates equal one. ◻
The selected resets thus determine an integral source decomposition. Each original harmonic sends its forest component to the frequency given by its gcd with the forest carrier. This identifies the field belonging to the selection and retains its mixed work with the complement.
Proposition 9.2 (Ordered boundary work). Define K_{dj}=R_Q(j,d), where R_B(j,d)=\sum_{a\bmod d}\psi(ja/d)\psi(Ba/d). For any coefficient source, A_\sigma(c)=\sum_{k=2}^N\sigma(k)m_k(c) \sum_{a\in\mathcal U_k}F_c(a/k)\psi(Qa/k) =2(P_\sigma c)^TKc. Its skew part is retained in an individual sector. For the whole source let b=P_{\mathcal F}\mathbf1. Then b_d counts its actual collapse fibre, b_d\le N/d, and \sum b_d=N.
Proof. Partition the complete d-grid into reduced denominators and exchange its finite sums. The selected reset mass is exactly \mathsf ZP_\sigma c=\operatorname{diag}(\sigma)\mathsf Zc. This proves the ordered identity. Split K into symmetric and skew parts; only the sum of complementary sectors cancels their skew readings. The population assertions follow from the gcd map. ◻
Theorem 9.3 (Signed forest work and restoration). For the divisor forest of Section 8, with 1 adjoined as above, |A_{\mathcal F}|,\ |E_{\mathrm{restore}}| \le\frac N2\sum_{\substack{d\in\mathcal F\\d\ge2}} H_{\lfloor d/2\rfloor}^2, where A_{\mathrm{other}}=2\sum_{d\notin\mathcal F}W_d+E_{\mathrm{restore}}, \qquad E_{\mathrm{restore}}=2\sum_{d\in\mathcal F}(1-b_d)W_d. Under the polynomial-height or stated growing-window hypotheses, both are O_\varepsilon(p^{1+\varepsilon}). Consequently \mathcal S_p-1=\frac8{p^2} \sum_{\substack{d\notin\mathcal F\\1\le a<d/2}} \Gamma_d(a)R_p(a,d)+O_\varepsilon(p^{-1+\varepsilon}). For any positive divisor parameter, a forest population M_p also gives |4A_{\mathcal F}/p^2|\le 2M_p(1+\log N)^2/p.
Proof. The ordered formula gives A_{\mathcal F}=2\sum b_dW_d. Use Theorem 6.1, b_d\le N/d, and 1\le b_d on the forest. This proves both finite inequalities. The incoming-divisor count already proved in Section 8 is subpower under the stated height and return-window hypotheses. It proves the signed estimates. The complete source split and (25), inserted in the reset formula, give (33); the common reset and curvature are O(p^{-1}) and O(\log p/p^2). Counting the finite bound proves the last assertion. ◻
Keeping that source decomposition has improved the normalized forest estimate from half-power absolute work to first-power signed work. Restoration to the original complete clocks has the same stronger allowance. The forest selection has therefore been resolved within the original sampling identity, including the cost of restoring its complete-clock reading.
The change of clock in a finite fold has its own source cost. To make that transport explicit, let V_N have basis e_{d,a} for 2\le d\le N, 1\le a\le d/2, including noncoprime pairs. Choose the nearest integer q to d/a, taking the smaller integer at a tie, and write r=d-qa, t=|r|. Define T e_{d,a}= \begin{cases} \operatorname{sgn}(r)e_{a,t},&a\ge2,\ t>0,\\ 0,&a=1\text{ or }t=0. \end{cases} A nonterminal edge strictly decreases the denominator and has t\le a/2. Thus T^N=0 and (I+T)^{-1}=\sum_{j=0}^{N-1}(-T)^j. For M_\beta e_{d,a}=\beta_d e_{d,a}, a nonterminal edge Te_{d,a}=\epsilon e_{a,t} gives [T,M_\beta]e_{d,a}=\epsilon(\beta_d-\beta_a)e_{a,t}, (I+T)^{-1}M_\beta-M_\beta(I+T)^{-1} =-(I+T)^{-1}[T,M_\beta](I+T)^{-1}. The first identity follows on each basis vector, with zero commutator at a terminal state. Multiplication by I+T proves the second. The inverse includes every internal depth. A scalar reading at a terminal exit must be recorded separately. Accumulated coefficients need not have magnitude one, so this transport does not replace the preceding work bound.
Proof of Theorem 1.1. Theorem 8.2 proves (6). The signed estimate is Theorem 9.3; its proof identifies the selected source by divisor inversion and bounds the restoration to complete clocks. Insert the signed estimate into the complete reset identity (20). Its origin term differs from one by O(p^{-1}), and its curvature term is O(\log p/p^2). This gives (8) and both implications. ◻
The opening forest result is now complete. The complementary sum still contains a large clock population. The lower and upper words give a second way to organize the complete cutoff interior, with its own source populations.
The forest estimate leaves the large complementary population. To study the remaining interaction, return to the complete cutoff interior and its two short-word branches. They move at different speeds. An upper word supplies a cell containing all lower words up to the next upper word. This arithmetic order gives complete small packets and controls their unfinished cutoff edges.
Every clock of \mathcal J_p has k\ge2a_0, hence m_k\le M_*=\lfloor N/(2a_0)\rfloor. For an occupied load m, put c_m=\min\left(b_0,\left\lfloor\frac{p-1}{2m+1}\right\rfloor\right). The lower words are s\equiv p-1\pmod m in [a_0-1,c_m-1]; the upper words are s\equiv-p\pmod{m+1} in [a_0,c_m]. Write T_N=\max_{k\le N}\tau(k).
Theorem 10.1 (Small packets and their complete fringe). Keep every upper-word cell [s,s+m] wholly contained in [a_0,c_m-1]. Its packet \mathcal P_{m,s} consists of the upper clock (p+s)/(m+1) and all lower clocks (p-v-1)/m whose words v lie in that cell. For \delta_m(s)=(p-1-s)\bmod m,\qquad 0\le\delta_m(s)<m, the lower words are s+\delta_m(s), together with s+m when \delta_m(s)=0. Every packet therefore contains two or three clocks. Its next offset obeys \delta_m(s+m+1)=\delta_m(s)-1\pmod m. The packets are disjoint. Their complete fringe \mathcal F_p in the cutoff interior satisfies \#\mathcal F_p\le8M_*, \qquad \sum_{k\in\mathcal F_p}|q_k| \le16NH_N^2T_NM_*=O_\varepsilon(p^{3/2+\varepsilon}). Consequently \mathcal S_p-1 =\frac1{p^2}\sum_{m,s}\sum_{k\in\mathcal P_{m,s}}q_k +O_\varepsilon(p^{-1/2+\varepsilon}). The full half-power estimate is equivalent to a uniform O_\varepsilon(p^{3/2+\varepsilon}) bound for this complete signed packet sum. The offset sequence has m steps. Its word period is m(m+1), with m+1 lower entries and m upper entries.
Proof. On either sheet centeredness gives (2m+1)h<p. Equality is impossible for prime p and h\ge a_0\ge3. Solving for h, and imposing the cutoff b_0, gives c_m. Solving for the integer denominator gives the two word congruences and their stated ranges. Conversely each such word recovers its actual clock, proper branch and load m.
Successive upper words differ by m+1, so their cells partition the integer word line. Within a cell of m+1 consecutive integers, the lower congruence has its representative s+\delta_m(s); only \delta_m(s)=0 supplies the second representative s+m. The sole upper word is the left endpoint. There are at most two partial cells at the ends of [a_0,c_m-1], with at most three clocks each. Outside this interval only the lower word a_0-1 and the upper word c_m can occur. There are therefore at most eight fringe clocks per load, also when the common interval is empty. There are at most M_* loads. Apply (26) and M_*=O(\sqrt p) to obtain (35). The complete interior identity (27) then gives (36) and both directions of the equivalence.
Subtracting m+1 modulo m proves (34). It visits every offset once in m steps. One cell has two lower entries and the others have one, giving the stated counts. The unique common-word anchor modulo m(m+1) is r_m\equiv(2m+1)p-(m+1)\pmod{m(m+1)}. It satisfies both word congruences; their coprime moduli give uniqueness. ◻
At load one, the complete packets are \{p-s-1,\ p-s-2,\ (p+s)/2\},\qquad a_0\le s\le b_0-2,\quad s\ {\rm odd}. They leave at most four clocks of that load. At p=31, the one complete packet is \{23,22,19\}, with work -3248/209; its fringe is \{24,20\}. The joining conditions prescribe the readings, including their signs.
This is a second organization of the original defect. Its five-clock interior at this cutoff differs from the three-clock complement of the return-selected forest. Here every omitted endpoint has a half-power allowance, so the estimate is reduced to the interaction of complete packets. Their small size comes from the branch congruences.
The source and observer both change along a loaded branch. The next identity records that change before an estimate is applied.
Proposition 10.2 (The measured branch connection). On adjacent denominators in a fixed load strip, let v_k=B_{s_k},\quad K_k=p+s_k(s_k+1),\quad \mu_k=m\sum_{a\in\mathcal U_k}\delta_{a/k}. With U=v_{k+1}-v_k and \Delta=\psi(K_{k+1}x)-\psi(K_kx), one has \begin{aligned} q_{k+1}-q_k={}&4\int[U\psi(K_kx)+v_k\Delta+U\Delta]\,d\mu_k\\ &+4\int v_{k+1}\psi(K_{k+1}x)\,d(\mu_{k+1}-\mu_k).\notag \end{aligned} The source changes by m harmonics below and m+1 above. Its block slope is the corresponding increment of K_k. For an address A admissible on three consecutive clocks, the actual lifted angle has \theta_k=QA/k,\qquad \theta_{k+2}-2\theta_{k+1}+\theta_k =\frac{2QA}{k(k+1)(k+2)}. The periodic-potential transport of the complete source and measure cancels exactly.
Proof. Expand the next source and phase in the old measure and then change the measure. This gives (37). Below, h=p-mk and s=h-1; above, h=(m+1)k-p and s=h. Subtract their harmonic sums and p+s(s+1) to get the stated block lengths and slopes. The reciprocal second difference gives the angle formula.
Put P_k(x)=\Phi(px+v_k(x))-\Phi(px), with \Phi as in the carry calculation. Its integral in its own measure is zero because the actual root phases are p and Q. Consequently \int(P_{k+1}-P_k)\,d\mu_k+ \int P_{k+1}\,d(\mu_{k+1}-\mu_k)=0. When a foreign phase is integral in the old measure, the midpoint identity has the additional term (v/2)\mathbf1_{u+v\in\mathbb{Z}}. Its source and measure transport has the same add-and-subtract cancellation, since each endpoint’s own unit-root phase is nonintegral. Thus the complete connection retains the energy and carry terms with no lost midpoint correction. ◻
Integrating these changes along a strip retains its multiplicities: \sum_{k=L}^Uq_k=(U-L+1)q_L+ \sum_{k=L}^{U-1}(U-k)(q_{k+1}-q_k). The endpoint allowance applies to the endpoint reading. An accumulated term carries its displayed strip-length factor.
The complete packet population is now fixed, and its entire fringe is controlled. Each packet carries a changing source, phase and root measure. Their joined work is the information a packet record must preserve.
A packet reads two linear quantities at each clock, the resetting mass and the field against its detector. Their product gives its quadratic work. A later packet can see source directions absent from the current record, so those directions and their mixed work must travel with the reading. The following orthogonal decomposition is finite-dimensional linear algebra in the sampled energy metric. Its role is to identify the actual source directions entering the next arithmetic packet, while the continuous mixed work remains visible.
For a t-clock packet let \mathsf A have alternating rows r_k,t_k, and put \mathsf J_t=\operatorname{diag}_{i=1}^t\begin{pmatrix}0&2\\2&0\end{pmatrix}.
Proposition 11.1 (The complete quadratic record). For every cutoff source, D_{\mathcal P}(c)=c^T\mathsf A^T\mathsf J_t\mathsf Ac,\qquad \ell_{\mathcal P}=\mathsf A^T\mathsf J_t\mathsf A\mathbf1 . The source derivative at \mathbf1 is 2\ell_{\mathcal P}. Every harmonic diagonal entry of the query matrix is zero. If \mathsf A has full row rank, the query has rank 2t and signature (t,t). Any linear record followed by a quadratic readout agreeing on all sources has at least 2t coordinates.
Proof. The matrix product gives 4\sum r_kt_k; differentiation gives the complete source derivative. If k\mid j, the jth wave vanishes at every k-root, so its reset and detector coefficients have zero product. This proves the zero diagonal. A surjective record completes to source coordinates in which the query is \mathsf J_t\oplus0. A quadratic readout on d linear records has matrix T^TBT of rank at most d. ◻
Theorem 11.2 (The entering source and its work). Let \mathsf A,\mathsf B be two actual records. Set H=\mathsf A\mathsf G_p^{-1}\mathsf A^T,\quad W=\mathsf G_p^{-1}\mathsf A^TH^+,\quad T=\mathsf BW, E=\mathsf B-T\mathsf A,\qquad S=E\mathsf G_p^{-1}E^T. Here the superscript + means inverse on the matrix image and zero on its orthogonal complement in the fixed record coordinates. For y=\mathsf Ac, \eta=Ec, put u=Wy,\qquad v=\mathsf G_p^{-1}E^TS^+\eta,\qquad z=c-u-v. These three source parts are \mathsf G_p-orthogonal, \mathsf Az=\mathsf Bz=0, and c^T\mathsf G_pc=y^TH^+y+\eta^TS^+\eta+z^T\mathsf G_pz. The later record is Ty+\eta. Its source dimension increase is \operatorname{rank} E=\operatorname{rank}\binom{\mathsf A}{\mathsf B}-\operatorname{rank}\mathsf A. The old and entering lifts are the unique minimum-work sources with their respective readings. The continuous work is \mathcal I(c,c)=\mathcal I(u,u)+\mathcal I(v,v)+\mathcal I(z,z) +2\mathcal I(u,v)+2\mathcal I(u,z)+2\mathcal I(v,z). If the two query matrices in record coordinates are J_0,J_1, their summed query is y^T(J_0+T^TJ_1T)y+2y^TT^TJ_1\eta+\eta^TJ_1\eta. For an onto joint record, all three coefficient blocks are uniquely determined by the cutoff quadratic query.
Proof. Positivity of \mathsf G_p gives \operatorname{ran} H=\operatorname{ran}\mathsf A and \mathsf AWy=y. Also \mathsf A\mathsf G_p^{-1}E^T =(I-HH^+)\mathsf A\mathsf G_p^{-1}\mathsf B^T=0. It follows that EW=0, \mathsf Av=0 and Ev=\eta. The remaining source z is in both record kernels. These identities prove all three work orthogonalities and (38). A different lift adds a kernel vector with positive orthogonal work, proving uniqueness. The residual rows of E give exactly the increase of the joint dual span. Expansion in the continuous form retains its mixed terms. Substitution of Ty+\eta in the next query gives (39). If every joint reading occurs, polarization fixes its whole symmetric matrix. ◻
Proposition 11.3 (The seed interior response). At p=31, the complete remaining interior \{19,20,22,23,24\} has minimum linear record dimension ten for its quadratic work, with signature (5,5). The packet \{23,22,19\} has rank six, and its actual fringe \{24,20\} adds four independent conditional readings. The coherent value is 76889/3135.
Proof. Order the clocks as 23,22,19,24,20. The five detector rows at source columns 1,\ldots,5 are \begin{pmatrix} -7/23&-7/23&17/23&1/23&-77/23\\ 4/11&-2/11&8/11&-10/11&4/11\\ 3/19&-11/19&21/19&-21/19&-9/19\\ 49/36&13/18&1/4&4/9&41/36\\ 7/5&4/5&1&2/5&0 \end{pmatrix}. Its determinant is -88076/216315. The resetting rows vanish at those columns and form the identity at the five distinct clock columns. Thus the full ten-row record is onto. The upper-left three-detector minor is -548/4807, giving rank six for the packet. The two preceding results give minimality, signature and the four entering directions. Direct unit-root sums give packet work -3248/209 and fringe work 601/15, whose sum is the stated value. ◻
The complete source still has thirty coordinates. The ten-reading lift retains a twenty-dimensional source complement and its work. The packet, fringe and mixed contributions are finite parts of the cutoff reading, even when the fringe is within an asymptotic budget.
The packet fringe is controlled, and the full record identifies the source directions entering each successive reading. The quantitative target uses the complete source c=\mathbf1. Its energy is already fixed by the Cubic Law. Measuring the full detector on those directions isolates the remaining cost needed for a half-power bound.
Fix either a dyadic load band M\le m<2M, or the entire complete packet population. Order its packets by load and upper word. Choose a nested prefix partition before evaluating any reading. One useful choice is packet counts 1,2,4,\ldots,L. Let P_r be the \mathsf G_p-orthogonal projection onto the dual span of the physical rows in prefix r, with P_0=0.
Theorem 12.1 (Source-coherent work and readout cost). Put v_r=(P_r-P_{r-1})\mathbf1, e_r=v_r^T\mathsf G_pv_r. Their work adds exactly: \sum_r e_r+ ((I-P_*)\mathbf1)^T\mathsf G_p((I-P_*)\mathbf1)=E_p, where P_* is the final prefix projection. For the rate calculation, fix the reset populations and set f(c)=4\sum_{\text{chosen clocks }k}m_kt_k(c),\qquad D=f(\mathbf1),\qquad \mathscr R=\sum_{e_r>0}\frac{f(v_r)^2}{e_r}. Then |D|^2\le E_{\rm obs}\mathscr R\le E_p\mathscr R, \qquad E_{\rm obs}=\sum_re_r. Writing g=\mathsf G_p^{-1}f^T and Z=\operatorname{span}\{v_r:e_r>0\}, \mathscr R=\|P_Zg\|_{\mathsf G_p}^2 \le f\mathsf G_p^{-1}f^T. Coarsening the fixed nested partition cannot increase \mathscr R.
Proof. Nested orthogonal projections have mutually orthogonal increments. Their sum is P_*, proving the source-work identity. The covector f is a linear combination of the actual final detector rows, so g lies in that observed dual span. Consequently f((I-P_*)\mathbf1)=0 and D=\sum_rf(v_r). Weighted Cauchy gives (40); a zero e_r means v_r=0. Projection onto the mutually orthogonal v_r gives the formula for \mathscr R. If adjacent increments are grouped, their energies and readouts add. The inequality \frac{(\sum d_i)^2}{\sum e_i}\le\sum_{e_i>0}\frac{d_i^2}{e_i} proves monotonicity under coarsening. ◻
These directions are fixed by the source, metric and chosen prefix partition. Dyadic cuts are an analytic decomposition, rather than a physical cadence. The scalar query retains its full quadratic source derivative; f is the linear reading used at the specified coherent source for the rate.
The quantities to estimate can be written directly in the cutoff mixed Gram. Let A_{\rm all} be the final physical record, f=b^TA_{\rm all}, and for prefix J put H_J=A_J\mathsf G_p^{-1}A_J^T,\qquad y_J=A_J\mathbf1. Its source work and predicted detector reading are E_J=y_J^TH_J^+y_J,\qquad F_J=b^T H_{{\rm all},J}H_J^+y_J. All future detector interactions remain in H_{{\rm all},J}. The cost is the sum of \frac{(F_{J_{\rm next}}-F_{J_{\rm prev}})^2} {E_{J_{\rm next}}-E_{J_{\rm prev}}} over the fixed prefix increments.
The positive cost also records how much the bound loses when its source directions are read separately. This can be measured without altering any source or future-detector interaction.
Proposition 12.2 (The signed coarsening loss). For the preceding increments set d_r=f(v_r), e_r=v_r^T\mathsf G_pv_r and E=\sum_r e_r>0. Omit zero-energy increments, whose readings vanish. Then D=\sum_r d_r and E\mathscr R-D^2 =\sum_{r<s}e_re_s \left(\frac{d_r}{e_r}-\frac{d_s}{e_s}\right)^2. For any fixed grouping of these increments, the decrease of cost is \sum_B\frac1{E_B}\sum_{\substack{r<s\\r,s\in B}}e_re_s \left(\frac{d_r}{e_r}-\frac{d_s}{e_s}\right)^2, \qquad E_B=\sum_{r\in B}e_r, with zero-energy groups omitted.
Proof. Expand E\sum_r d_r^2/e_r-(\sum_r d_r)^2. The contribution of an unordered pair is \frac{e_s}{e_r}d_r^2+\frac{e_r}{e_s}d_s^2-2d_rd_s =e_re_s\left(\frac{d_r}{e_r}-\frac{d_s}{e_s}\right)^2. Summing proves (42). Apply the identity within each group and divide by its source energy to obtain the cost decrease. Every d_r still contains its work against the full final detector. ◻
The identity quantifies the difference between the positive cost and the complete signed reading. It supplies no bound on that difference. At the complete load-one chains for p=43,61,73, rational source and detector calculations verify it with all prefix interactions. For example at p=61, the dyadic cost is 3.8176649888\ldots, while D^2/E=0.7201556918\ldots. These finite values measure the loss in the sufficient estimate; neither value has a proved uniform growth law.
Corollary 12.3 (A sufficient half-power criterion). A uniform bound \mathscr R=O_\varepsilon(p^\varepsilon) for the entire complete packet population implies \mathcal S_p=1+O_\varepsilon(p^{-1/2+\varepsilon}). The same bound uniformly for all dyadic load bands also suffices.
Proof. Use E_p\sim p^3 in (40), rescale \varepsilon, and add the proved packet fringe in (36). In the band version there are only O(\log p) bands; absorb that factor by another \varepsilon-rescaling. ◻
The required growth bound remains open. A sufficient local estimate would bound (F_{\rm next}-F_{\rm prev})^2 by a subpower multiple of E_{\rm next}-E_{\rm prev}. The aggregate cost may admit a weaker estimate. A single final source block would only give D^2/E_{\rm obs}, retaining the same uncontrolled signed reading. The stated cost bound is sufficient for the half-power target; no converse is asserted. The fixed multiscale criterion preserves the source and observation while exposing the coupling that needs an arithmetic bound.
The four-clock example begins with a visible discrepancy. A continuous integral sees the square approaching sixteen at the common reset, while the grid reads zero. Restoring that contact leaves the much smaller signed difference -1/225. At a growing cutoff, the corresponding question is whether all the remaining boundary interactions fit within the half-power allowance.
Whole arithmetic families now do. Under the stated height or return-window conditions, the forests have half-power normalized absolute work and first-power signed work. The same stronger bound pays for restoring their reading to complete clocks. Divisor inversion keeps the source of the selection and its interactions with the original field. These are portions of the actual finite-to-continuous discrepancy whose contribution is controlled as the field and grid grow together.
The packet decomposition separately controls the complete cutoff fringe. Its source records and the Cubic Law expose a sufficient bound on the cost of reading the remaining signed work. The forest and packet populations remain distinct. Neither the energy budget alone nor the finite example supplies the missing cancellation estimate.
For the forest decomposition, the outstanding quantity is \mathcal A_{\rm other} in (8). Its O_\varepsilon(p^{3/2+\varepsilon}) bound would complete the half-power comparison between the collision table and the corrected continuous model. Every complementary reset still reads the same growing field. What arithmetic structure distinguishes those surviving clocks, and what does it force in the relation between the source and its observer?
[1]M. Beck, Dedekind cotangent sums, Acta Arith. 109 (2003), no. 2, 109–130.\newline https://doi.org/10.4064/aa109-2-1.
[2]M. N. Huxley, Identities involving Farey fractions, Proc. Steklov Inst. Math. 276 (2012), no. 1, 125–139.\newline https://doi.org/10.1134/S0081543812010105.
[3]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.
[4]A. S. Petty, Collision Calculus: Boundary Readings, Energy and Refinement, 2026.\newline https://doi.org/10.5281/zenodo.23383403.
[5]A. S. Petty, The Cubic Law for Digit-Collision Energy, 2026.\newline https://doi.org/10.5281/zenodo.20547910.
[6]A. S. Petty, The Secondary Term in Digit-Collision Energy, 2026.\newline https://doi.org/10.5281/zenodo.21875096.
[7]A. S. Petty, The Unit Sampling Defect in Digit-Collision Energy, 2026.\newline https://doi.org/10.5281/zenodo.23285306.
Discussion
Sign in to join the discussion.