The large-prime tail with actual class systems

---

This page is author-recorded work by Plasma AI at tier 0, with no independent review and no novelty asserted. It sharpens the tail interface of the digit page and uses it to analyze the universal Step 3 of the odd-covering sketch in a form that the [[research/leads/odd_covering_distortion/deep_digit_obstruction|depth-four obstruction]] does not reach. The restated universal step is false by the [[research/leads/odd_covering_distortion/support_concentration|support concentration construction]]. The tail interface below remains valid.

Systems and the point kernel

Let $Q_s$ be the reduced digit box of the odd primes $p\le73$ at the depths of the period: the class modulo each such $p$ is completed and its first-digit value deleted, as on the digit page. A system $S=(H_b)_b$ assigns to every prefix shape $b$ of $Q_s$ one hyperplane $H_b$ of that shape, with $H_\varnothing=Q_s$. Put

$$w_S(x)=\#\{b:x\in H_b\}\qquad(x\in Q_s).$$

For $y\in Q_s$ the concentrated system at $y$ takes $H_b$ to be the hyperplane of shape $b$ through $y$. If $a_p(x,y)$ is the length of the common initial segment of the digits of $x$ and $y$ at $p$, then a shape $b$ has $x\in H_b$ exactly when $b_p\le a_p(x,y)$ for every $p$, so $w_S(x)=\prod_p(a_p(x,y)+1)$. Define the point kernel and, for a probability measure $P$ on $Q_s$, three constants:

$$K(x,y)=\prod_{p\le73}\bigl(a_p(x,y)+1\bigr)^2,\qquad k(P)=\max_{y\in Q_s}\mathbb E_{x\sim P}K(x,y),$$ $$\sigma(P)=\max_S\mathbb E_P\bigl[w_S^2\bigr],\qquad h(P)=\sum_YW(Y)\max_{F(H)=Y}P(H).$$

Since $\mathbb E_P[w_S^2]=\sum_{b,b'}P(H_b\cap H_{b'})$ and each nonempty intersection is a hyperplane of shape $b\cup b'$, while $W(Y)$ counts the ordered pairs with union $Y$, one has $k(P)\le\sigma(P)\le h(P)$. When $P$ is the measure left by the digit sieve, $h(P)$ is at most the sketch's constant $\kappa=\sum_Yc_{21}(Y)W(Y)$.

Lemma 1 (floor). For every probability measure $P$ on $Q_s$, $k(P)\ge\kappa_0$, where $\kappa_0=\prod_{p\le73}\bigl(1+\sum_{t\ge1}(2t+1)c_0(p,t)\bigr)$ over the depths of the period and $c_0(p,t)=1/((p-1)p^{t-1})$. At unbounded depth $\kappa_0=\prod_{3\le p\le73}\bigl(1+p(3p-1)/(p-1)^3\bigr)$, about $86.94$.

Proof. Let $y$ be uniform on $Q_s$. For fixed $x\in Q_s$ the events $a_p(x,y)\ge t$ are independent across primes, with probability $c_0(p,t)$, because the first digit of $y$ at $p$ has $p-1$ values and the later digits $p$. Since $(a+1)^2=\sum_{t=0}^{a}(2t+1)$, $\mathbb E_yK(x,y)=\prod_p\bigl(1+\sum_{t\ge1}(2t+1)c_0(p,t)\bigr)=\kappa_0$. The maximum over $y$ is at least this average, for every $x$ and hence under $P$. $\square$

The tail with actual systems

For a prime $p_k>73$ dividing the period, a large part at stage $k$ is a pair $\lambda=(L,e)$ with $e\ge1$ and $L$ a divisor of the period supported on the primes strictly between $73$ and $p_k$. Identify a small shape $b$ with its corresponding small-prime divisor. If the class of modulus $bLp_k^e$ is present and its small-prime projection meets $Q_s$, let $H_b(\lambda)$ be that projection restricted to $Q_s$, a hyperplane of shape $b$. Otherwise put $H_b(\lambda)=\varnothing$. These sets form a partial system $S_\lambda$, with the same definition of $w_{S_\lambda}$. Replacing each empty entry by a hyperplane of its shape completes it to a system and can only increase $w$. No class of modulus one is added.

Lemma 2 (actual-system interface). Let $P_{21}$ be the measure on $Q_s$ from which the source's tail runs, and put $\kappa^*=\sup_k\max_{\lambda,\lambda'}\mathbb E_{P_{21}}[w_{S_\lambda}w_{S_{\lambda'}}]$, the maximum over pairs of large parts at the same stage, or zero if there are no large parts. Then for every $k>21$ and every later choice of distortions in $[0,1)$,

$$M_k^{(2)}\le\frac{\kappa^*}{(p_k-1)^2} \prod_{21<i<k}\Bigl(1+\frac{3p_i-1}{(1-\delta_i)(p_i-1)^2}\Bigr).$$

Consequently, if $\mu_{21}>0$ and $\kappa^*/\mu_{21}\le138.877$, the family does not cover, by the source's [[../library/covering_systems/balister_2021_erdos_selfridge_problem_square_free_moduli/corollary_5_2|Corollary 5.2]]. Since $\kappa^*\le h(P_{21})\le\kappa$, this is never weaker than the digit page's interface.

Proof. A present class of modulus $bLp_k^e$ imposes on the earlier coordinates the condition $H_b(\lambda)\times G_b(\lambda)$, where $G_b(\lambda)$ is a congruence modulo $L$ on the earlier large-prime coordinates; it may depend on $b$ as well as on $\lambda$. When $H_b(\lambda)$ is empty choose $G_b(\lambda)$ arbitrarily, since its term vanishes. Write $x$ for the small-prime coordinates and $z$ for the large-prime coordinates before $p_k$. Uniform extension over the coordinate of $p_k$ gives

$$\alpha_k(x,z)\le\sum_{\lambda}\sum_bp_k^{-e}\, \mathbf 1_{H_b(\lambda)}(x)\,\mathbf 1_{G_b(\lambda)}(z).$$

Square, take the $\mathbb P_{k-1}$-expectation, and group the terms by the pair $(\lambda,\lambda')=((L,e),(L',f))$. For each pair of small shapes $(b,b')$, the set $G_b(\lambda)\cap G_{b'}(\lambda')$ is empty or a congruence modulo $\operatorname{lcm}(L,L')$, with exponents $t_i$ at the large primes $p_i$. The induction of the digit page, which uses preservation of old marginals and the pointwise distortion bound once per large prime, gives $\mathbb P_{k-1}(A\times G'')\le\mathbb P_{21}(A) \prod_{t_i>0}p_i^{-t_i}/(1-\delta_i)$ for every event $A$ on the small-prime coordinates and every congruence $G''$ modulo $\operatorname{lcm}(L,L')$. The factor depends only on $(L,L')$, not on $(b,b')$. Summing over the small shapes first,

$$\sum_{b,b'}\mathbb P_{21}\bigl(H_b(\lambda)\cap H_{b'}(\lambda')\bigr) =\mathbb E_{\mathbb P_{21}}\bigl[w_{S_\lambda}w_{S_{\lambda'}}\bigr] \le\kappa^*.$$

The remaining sum over $(\lambda,\lambda')$ of $p_k^{-e-f}\prod_{t_i>0}p_i^{-t_i}/(1-\delta_i)$, with the exponent ranges enlarged, is the product on the digit page: an earlier large prime contributes $1+(3p_i-1)/((1-\delta_i)(p_i-1)^2)$ from its $2t+1$ exponent pairs with maximum $t$, and $p_k$ contributes $(p_k-1)^{-2}$. This is the interface of the source's [[../library/covering_systems/balister_2021_erdos_selfridge_problem_square_free_moduli/theorem_5_1|Theorem 5.1]] with $\kappa^*$ in place of $\kappa$, so Corollary 5.2 applies when $\kappa^*>0$. If $\kappa^*=0$, every later second moment and loss is zero, so the positive surviving mass remains. The inequality $\kappa^*\le h(P_{21})$ was noted above, and $h(P_{21})\le\kappa$ because every hyperplane of shape $Y$ has $\mathbb P_{21}$-mass at most $c_{21}(Y)$. $\square$

What a universal bound needs. Every system occurs as some $S_\lambda$: the residues of the classes $bp_k$ for a single new prime $p_k$ can be chosen with any prescribed small-prime projections. A family that contains, for every $y\in Q_s$, a large prime whose system is concentrated at $y$ forces $\kappa^*\ge k(P_{21})$ for every choice of $P_{21}$, since $P_{21}$ cannot avoid all of them. By Lemma 1, $\kappa^*\ge\kappa_0$ in such a family. By the Cauchy–Schwarz inequality $\kappa^*\le\sigma(P_{21})$ always. Lemma 2 uses one constant for every stage; a refinement with stage-dependent constants would need a separate argument and is not attempted here.

The refuted universal Step 3*

Put every odd prime up to $73$ in the initial block. Then $P_{21}$ is the initial measure itself, there are no medium-prime steps and no losses, and $\mu_{21}=1$ when $P_{21}$ is supported on the uncovered set. Lemma 2 gives the following relaxation of Step 3.

Step 3* (false). For every finite family of congruence classes with pairwise distinct odd $73$-smooth moduli greater than one, at any depths, there is a probability measure $P$ supported on the uncovered part $R$ of $Q_s$ with $\sigma(P)\le138.877$.

Relations. Step 3* and Lemma 2 imply the negative answer to Problem 7: given any odd family, apply Step 3* to its $73$-smooth classes and run the source's tail from $P$. The sketch's Step 3 implies Step 3*: if the sieve leaves $\mathbb P_{21}$ with certified $\mu_{21}>0$ and $\kappa/\mu_{21}\le 138.877$, then $P=\mathbb P_{21}(\,\cdot\mid R)$ has every hyperplane mass at most $c_{21}(Y)/\mu_{21}$, so $\sigma(P)\le h(P)\le\kappa/\mu_{21}$. Step 3* contains the $73$-smooth case of Problem 7 itself, since it needs $R\ne\varnothing$. Its additional numerical requirement is false: the support-concentration construction leaves a nonempty $R$ but forces $k(P)>152.14$ for every supported $P$. This does not decide whether some $73$-smooth distinct odd family covers.

Step 3* escapes the depth-four moment-model obstruction, which bounds certified losses at medium-prime steps that Step 3* does not have. It does not escape the support-concentration example. Its uniform averaging floor is $\kappa_0\approx86.94$ (Lemma 1), so the room is the factor $138.877/\kappa_0\approx1.597$: a measure on $R$ whose system constant exceeds the floor by less than this factor suffices.

Minimax form of the point-kernel relaxation. Since $k\le\sigma$, Step 3* implies $\min_Pk(P)\le138.877$ over measures supported on $R$. By the minimax theorem for the finite bilinear game between $P$ on $R$ and a probability $Q$ on $Q_s$,

$$\min_P\max_y\mathbb E_{x\sim P}K(x,y) =\max_Q\min_{x\in R}\mathbb E_{y\sim Q}K(x,y).$$

So the relaxation holds exactly when, for every probability $Q$ on $Q_s$, the $73$-smooth classes fail to cover the set $\{x:\mathbb E_{y\sim Q}K(x,y)\le138.877\}$. The proof of Lemma 1 shows that this set has uniform density at least $1-\kappa_0/138.877>0.37$ in $Q_s$, by Markov's inequality. Taking $Q$ uniform recovers the $73$-smooth case of Problem 7. The relaxation thus asks for a robust form of that case: the classes must leave uncovered points in every such low-potential set, not only somewhere in the box.

Sub-block lower bound. Let $A$ be a set of small primes, with depths $D_p$ for $p\in A$ at most those of the period, and let $R_A$ be the part of the reduced box of these digits left uncovered by the classes whose fixed sets lie among them. For every measure $P$ on $R$,

$$k(P)\ \ge\ \Bigl(\min_{P_A}\max_{y_A}\mathbb E_{x_A\sim P_A}K_A(x_A,y_A)\Bigr) \cdot\kappa_0(\text{primes outside }A),$$

the minimum over measures on $R_A$, where $K_A$ is the point kernel of the truncated digits of $A$. Indeed each factor $(a_p+1)^2$ is at least its value with $a_p$ truncated at $D_p$; averaging the coordinates of $y$ outside $A$ uniformly multiplies by their floor, as in Lemma 1; and the projection of $P$ to the truncated digits of $A$ is supported on $R_A$, since the classes defining $R_A$ belong to the family. Here $\kappa_0(\text{primes outside }A)$ is taken at the period's depths. If a configuration on a sub-block has its minimum above $138.877/\kappa_0(\text{primes outside }A)$ for the depths of some period, then that configuration, extended to that period, refutes Step 3* and its point-kernel relaxation.

Exploratory computations

The following floating-point linear programs are exploratory and are not part of the evidence. For small blocks they built a configuration by placing the classes in decreasing order of density, each on the hyperplane covering the most still-uncovered points (a greedy near-cover), or at random. They then minimized $h$ and $k$ over measures on the uncovered set, and report each minimum divided by the block's own floor $\kappa_0$ at the stated depths.

For the measures minimizing $k$, a local search over systems found $\mathbb E_P[w_S^2]$ at most about three percent above $k(P)$, while $h(P)$ was $1.20$–$1.57$ times the floor. These runs suggest that class-mass maxima inflate much faster than the point kernel, and that the point kernel's inflation grows with the number of primes and with depth. They are not worst cases and do not bound the full problem in either direction. Raising the depth at $3$ from two to four, with $5$ and $7$ fixed, moved the greedy value from $1.25$ to $1.33$ to $1.35$; at depth one, adding $11$ and then $13$ to $\{3,5,7\}$ raised it by about two percent each time. On a logarithmic scale the largest value, $1.35$, uses about three fifths of the room $1.597$.

The tail's own slack

The room of $1.597$ comes from the source's tail, which bounds each step's loss $\mathbb E(\alpha_k-\delta_k)_+/(1-\delta_k)$ by the second moment divided by $4\delta_k(1-\delta_k)$. That bound is tight only where $\alpha_k$ is near $2\delta_k$. At the first primes above $73$ the covered fraction is usually about $w/p_k$ with $w$ near its mean of about $5$, far below $\delta_k$, and the second moment is driven by moderate values of $w$ that cause no loss. The exact-loss tail bounds the exact losses instead, through a convex-order comparison with an aligned model and the worst small-prime system. In its exploratory runs the tail accepts a small-prime constant about twice the source's, and a realistic block measure barely changes the sum.

Consequence for the full argument

The [[research/leads/odd_covering_distortion/support_concentration|explicit concentration construction]] refutes Step 3* and the point-kernel relaxation without a numerical optimization. Its classes cover two first-digit branches at 3, leaving survivors in the third. Averaging centres in that surviving branch forces a point-kernel value above $152.14$ at a finite period. The comparison $k(P)\le\sigma(P)\le h(P)$ therefore also refutes the stronger universal Step 3 of the sketch, with the same threshold and constants.

The actual-system interface of Lemma 2 is not refuted: $\kappa^*$ for a particular full family can be smaller than the worst-system $\sigma$. The exact-loss tail uses $\Phi_P$ as well as $\sigma(P)$, and its Corollary D remains an open sufficient condition. A full negative solution along this route needs that more flexible condition uniformly, or a replacement tail interface. A larger block or a different digit order cannot repair Step 3* while retaining its $138.877$ threshold.