Take two consecutive intervals of width \(e^{-L}\) in the source variable. Allow a smooth zero-mean perturbation supported only inside those two cells. Then, on any fixed collection of compact windows, that tiny perturbation can independently prescribe:
- the canonical Poisson defect in the critical strip;
- the full Mellin carrier after division by \(\Xi\);
- the completed-zeta multiplier in \(\Re s>1\);
- and any finite list of values and derivatives on the arithmetic side.
The source interval can sit near \(e^L/2\), so its total width is \(O(e^{-L})\). It eventually disappears from every fixed source compact. Yet its transforms remain locally universal.
This is already a wonderfully hostile fact. A proposed RH criterion can watch a long zero-free Stokes corridor, the real discriminant, the normalized strip quotient, and finitely many prime-side jets, and the same two microscopic cells can still counterfeit all of them at once.
But the counterfeit is not free. The exact completion has a compulsory global shape:
$$ \mathcal A_q(s)=\zeta(s)M_q(s). $$
Every nontrivial zeta zero divides it. Its logarithmic tail cannot terminate. Fixed-support approximation pays gamma-scale source mass. And after dividing the full carrier by \(\Xi\), every nonzero annular source contributes a density-one divisor pinned superfactorially close to the reflected odd gamma lattice.
Locally, almost anything. Globally, an infinite arithmetic mess. I like this seam a lot.
The two-cell subtraction trick#
Let
$$ I_{N,L}=((N-1)e^{-L},Ne^{-L}) $$
and let \(J\) contain \(I_{N,L}\) and \(I_{N+1,L}\). For an even extension of a smooth zero-mean source \(q\) supported in \(J\), the canonical defect contains a Hurwitz-zeta kernel. On one cell the relevant index is \(N\); on the next it is \(N+1\).
The entire mechanism turns on the tiny identity
$$ \zeta(s,N)-\zeta(s,N+1)=N^{-s}. $$
Suppose a Hahn--Banach functional annihilates every source image. Write its two cell identities and subtract them. The full Mellin terms cancel, while the Hurwitz difference leaves a pure exponential. That forces all even cosine moments of the defect-side measure to vanish. Polynomial approximation then kills that measure completely.
What remains is the full-carrier measure. Inside \(|\Im z|<1/2\), its exponential growth is strictly slower than the \(e^{x/2}\) term that a nonzero constant annihilator would require. That growth gap kills the constant, and a second approximation argument kills the carrier measure.
The resulting density theorem is
$$ \overline{ \left\{ \left( \mathsf D_q^{\rm P}|_{K_D}, \mathsf X_q^{\rm P}|_{K_X} \right) \right\} } \supset A_{\rm ev}(K_D)\times \Xi A_{\rm ev}(K_X). $$
The closure holds uniformly and in every fixed \(C^m\) norm. So the defect can be prescribed locally while the full carrier is independently prescribed subject only to its exact local \(\Xi\)-divisibility.
That is much stronger than saying a flexible family of entire functions can approximate a target. These are exact compact sources, constrained by the Poisson architecture, living in two adjacent arithmetic cells.
The arithmetic window is not safer#
Write
$$ \mathsf X_q^{\rm P}(z)=\Xi(z)E_q(z) $$
and move to the arithmetic coordinate
$$ \mathcal E_q(s)=E_q\!\left(-i(s-\tfrac12)\right). $$
The de-archimedean completion is
$$ Z_{\mathsf X_q^{\rm P}}(s)=\zeta(s)\mathcal E_q(s). $$
The two-cell argument can be run jointly on a strip compact and a compact subset of \(\Re s>1\). A one-sided exponential-support lemma separates the right-half-plane contribution from the slower strip contribution. The joint map
$$ q\longmapsto \left( \mathsf D_q^{\rm P}|_{K_D}, E_q|_{K_E}, \mathcal E_q|_{K_+} \right) $$
is dense in the full product of the three target algebras.
Finite arithmetic jets are therefore not a rescue. Hermite interpolation lets one prescribe any finite list
$$ \mathcal E_q^{(j)}(s_\nu)=c_{\nu,j}, \qquad \Re s_\nu>1, $$
exactly, while retaining arbitrary compact-open strip data.
So a finite packet of prime-side derivatives does not make a local geometric certificate arithmetic. The missing condition must be genuinely noncompact: an entire half-plane law, a Hardy or Dirichlet norm, coefficient control, divisor tightness, or a coercive source topology.
Exact counterfeits, not just density#
The density theorem can be turned into concrete hostile objects.
One construction realizes a moving zero-free Stokes chamber of length \(2S_L\). Its projective phase variation stays bounded, its Stokes set is one scaled-\(C^2\) graph, the literal transform has no zero in the chamber, and a real interval is protected from double zeros. The source correction still lives in two microscopic cells and vanishes on every fixed source compact.
The price is global nontameness:
$$ \log M_L(Y)\ge cLS_L-O(1), $$
with corresponding exponential growth in high Sobolev norms.
Another construction is even ruder. For a grid-flat source, the central Euler operator
$$ \mathcal Z=-\left(x\frac d{dx}+\frac12\right)^2 $$
intertwines exactly with multiplication by \(z^2\). Thus a polynomial \(Q(\mathcal Z)\) multiplies the bulk, defect, and literal transform by the same even polynomial \(Q(z^2)\).
A second source, placed below the literal sampling threshold, has identically zero literal transform but nonzero canonical bulk. Adding it removes the common factor in the canonical pair. The inserted off-axis zeros become simple, reduced, projectively regular literal zeros, while the complete real zero and double-zero sets remain unchanged.
This is not an external multiplication trick. The false divisor is realized inside the exact source architecture.
Then the global completion retaliates#
Define the arithmetic sample sum
$$ A_q(v)=\sum_{n\ge1}q(nv). $$
For an exact-\(\mathcal S_0\) source it is smooth, compactly supported at large \(v\), and flat at zero. Its Mellin transform is
$$ \boxed{ \mathcal A_q(s) = \int_0^\infty A_q(v)v^{s-1}\,dv = \zeta(s)M_q(s). } $$
This is the compulsory completion. Every nontrivial zero of \(\zeta\) divides \(\mathcal A_q\), with at least its zeta multiplicity.
If the positive logarithmic tail terminated, then \(A_q(e^x)\) would have compact support in \(x\). Its Mellin transform would be an entire function of finite exponential type, with only \(O(r)\) zeros in a disk of radius \(r\). But the zeta divisor contributes order \(r\log r\) zeros. Contradiction.
Therefore every nonzero canonical tail is unbounded in both logarithmic directions.
This is the exact answer to the local universality: the far tail is not an optional remainder. It is glued to the finite bulk at every zeta zero.
The whole ideal is still too large#
Fix the source support inside \([-2,2]\), and normalize
$$ f_q(s)=\frac{M_q(s)}{A(s)}, \qquad A(s)=\frac12s(s-1)\pi^{-s/2}\Gamma(s/2). $$
The family \(f_q\) is compact-open dense, with every fixed derivative, in the real-symmetric entire functions. Consequently the source-generated completion ideal is exactly
$$ (\xi), $$
and the even full-carrier ideal is exactly
$$ (\Xi). $$
Two exact sources suffice for an entire Bezout identity.
This sounds powerful until one notices the topology. Compact-open approximation can imitate a quartic false-RH multiplier on larger and larger disks, but fixed support forces
$$ \log\|q_n\|_1 \ge \frac n2\log n-O(n). $$
The unrestricted holomorphic ideal is saturated precisely because its coefficient functions and source norms are allowed to become violent.
Ideal membership alone sees the compulsory zeta divisor. It does not control the extra divisor.
The extra divisor sits on the gamma lattice#
For a nonzero real-even smooth source \(q\), supported away from the origin and with zero integral, write
$$ \mathcal E_q(s) = \frac12\left( \frac{M_q(s)}{A(s)} + \frac{M_q(1-s)}{A(1-s)} \right) = E_q\!\left(-i(s-\tfrac12)\right), $$
where
$$ A(s)=\frac12s(s-1)\pi^{-s/2}\Gamma(s/2). $$
Gamma reflection gives the exact normal form
$$ \boxed{ s(s-1)\mathcal E_q(s) = U(s)+V(s)\cos\frac{\pi s}{2} } $$
and duplication gives
$$ \frac{U(s)}{V(s)} = \frac{(2\pi)^s}{2\Gamma(s)} \frac{M_q(s)}{M_q(1-s)}. $$
This formula is wonderfully lopsided near large odd integers. The cosine vanishes there. The reciprocal gamma factor is superfactorially small. The only possible spoiler is the reflected Mellin factor \(H(s)=M_q(1-s)\).
After translating the left endpoint of the source support, \(H\) is an exponential times a bounded analytic function in the right half-plane. Its zeros obey a half-plane Blaschke condition, so only \(o(T)\) of the odd cells between \(T\) and \(2T\) contain a zero of \(H\).
Cartan--Boutroux minimum modulus does the second job. Around all but another \(o(T)\) odd centers, one can choose a shrinking circle on which \(H\) is not too small. Uniform Stirling then makes
$$ \left|\frac{U(s)}{V(s)}\right| \le \exp\left(-\frac12T\log T\right), $$
while the cosine on that circle is only polynomially small. Rouché now forces exactly one multiplier zero in each good odd cell.
Conjugation symmetry makes that zero real and the one-zero count makes it simple. If \(m\) is a good odd integer, the zero \(\sigma_m\) satisfies
$$ \boxed{ |\sigma_m-m| \le \exp(-c_qm\log m). } $$
Returning to the \(z\)-coordinate gives simple pure-imaginary zeros
$$ \pm i\left(\sigma_m-\frac12\right) $$
of \(E_q\), with count at least \(T-o(T)\). Since \(E_q\) has order at most one, its zero divisor has convergence exponent exactly one.
These are zeros of the source multiplier, not zeros of \(\Xi\). They may escape every compact window, and the complete Poisson tail may cancel their effect in the literal transform. That is exactly why all the local tests can look perfect.
The pinning mechanism itself needs much less regularity than the Poisson architecture. If \(q\) is merely a nonzero real \(L^1\) density on \([a,R]\), with \(0<a<R\) and zero integral, the same density-one theorem still holds. Compact support makes every Mellin derivative legitimate, zero mean removes the apparent completed-factor singularity, and the reflected Mellin transform is still a bounded half-plane function. No integration by parts, endpoint derivative, or smooth cutoff enters the pinning proof.
There is a pleasantly blunt witness:
$$ q=1_{[1,2]}-1_{[2,3]}. $$
Its Mellin transform is
$$ M_q(s)=\frac{2\,2^s-3^s-1}{s}. $$
This source does better than the generic density-one theorem. For every sufficiently large odd integer \(m\), its completed multiplier has exactly one zero \(\sigma_m\) in
$$ |s-m|<\frac14. $$
The zero is real and simple, with alternating signed displacement and
$$ |\sigma_m-m| \sim \frac{(6\pi)^m}{\pi\Gamma(m)}. $$
The proof is delightfully direct. Uniformly in every large odd cell, \(M(1-s)\) is zero-free and \(M(s)/M(1-s)\sim-3^s\). Stirling makes the reflected correction smaller than the cosine on the fixed quarter-circle, so Rouché gives one root in every cell, not merely after deleting a density-zero exceptional set. The asymptotic crosses into rapid decay near \(6\pi e\approx51.2\).
This explicit source is not smooth. The theorem is only about the analytic completed Mellin multiplier; it does not give an arbitrary \(L^1\) density the exact literal Poisson or Stokes machinery used above.
A compact literal cannot swallow the lattice#
Now return to one fixed smooth exact source and one fixed admissible literal scale. Write \(X=\Xi E_q\), \(F=F_q^{\rm lit}\), and \(D=F-X\). Remove common factors from \((X,D)\) before forming
$$ \Pi_{\rm red}=-\frac{X_{\rm red}}{D_{\rm red}}. $$
In Mellin coordinates, normalize the compact literal transform by its right endpoint:
$$ B_F(s) = e^{-L(s-1/2)} F\!\left(-i(s-\tfrac12)\right). $$
Because the literal kernel is supported in \([-L,L]\),
$$ |B_F(s)| \le \|\kappa_q\|_{L^1[-L,L]} \qquad (\Re s>1/2). $$
So unless \(F\equiv0\), this is a nonzero bounded analytic half-plane function. Its positive real zeros must obey the Blaschke budget
$$ \sum_{B_F(x)=0}\frac1x<\infty. $$
The pinned gamma cells have \(\sigma_n\sim2n\), so their harmonic mass diverges. A compact literal can absorb only a density-zero subset of them. At every unabsorbed pinned cell,
$$ X(z_n)=0, \qquad D(z_n)=F(z_n)\ne0. $$
Common-factor reduction therefore cannot erase the zero, and \(\Pi_{\rm red}\) has a simple zero there. The dichotomy is exact:
$$ \boxed{ F\equiv0\ \Longrightarrow\ \Pi_{\rm red}\equiv1, } $$
or else
$$ \boxed{ N_0(\Pi_{\rm red};|z|\le T)\ge T-o(T). } $$
I keep staring at the mismatch between those two zero budgets. The compact literal has finite Blaschke mass. The compulsory lattice has infinite harmonic mass. Almost all of the bill has nowhere to go except the reduced projective quotient.
This also rules out a hidden analytic shortcut: in Mellin coordinates the reduced quotient is not meromorphic of bounded type in \(\Re s>1\). Its logarithmic derivative has a linear-density family of simple poles there, so it cannot secretly be a holomorphic absolutely convergent ordinary Dirichlet series or a finite positive-gap Laplace law.
The vertical lattice overdraws its growth budget#
There is a second retaliation. Form the paired product over the pinned pure-imaginary zeros:
$$ P_q^\Gamma(z) = \prod_{n\in\mathcal G_q} \left(1+\frac{z^2}{y_n^2}\right). $$
Here
$$ y_n=2n+\frac12+o(1). $$
Along the real axis it has exact type
$$ \log P_q^\Gamma(x) = \left(\frac{\pi}{2}+o_q(1)\right)x. $$
But the full carrier \(\Xi E_q\) is bounded on the real axis, and the completed-gamma decay leaves the multiplier only a \(\pi/4\) average allowance. Integrated over \([T,2T]\), the pinned product demands
$$ \left(\frac{3\pi}{4}+o_q(1)\right)T^2, $$
while the archimedean budget permits only
$$ \frac{3\pi}{8}T^2+O_q(T\log^2T). $$
That factor of two is rude and decisive. Since \(E_q\) is even, a zero-free exponential factor cannot pay the deficit. Additional paired zeros must shrink the product by a quadratic logarithmic amount.
Set
$$ H(\rho)=(\Re\rho)^2-(\Im\rho)^2. $$
The geometry of one paired factor is exact:
$$ \left|1-\frac{x^2}{\rho^2}\right|<1 \iff x^2<2H(\rho). $$
So every compensating zero lies beyond the horizontal hyperbola. A Jensen tail estimate then turns the quadratic deficit into a linear divisor count: for each fixed source, there are constants \(K_q,T_q\) and an absolute \(c_0>0\). Let \(\Omega_q(T)\) be the zeros satisfying
$$ |\rho|\le K_qT\log(eT),\ H(\rho)>\frac{T^2}{2}. $$
Then, for \(T\ge T_q\),
$$ \boxed{ N_{E_q}^{\rm res}(\Omega_q(T))\ge c_0T. } $$
Multiplicity is counted. These zeros are additional to the pure-imaginary gamma-pinned divisor.
At any fixed literal scale, local valuation bookkeeping conserves that multiplicity:
$$ N_F(\Omega_q(T)) + N_0(\Pi_{\rm red};\Omega_q(T)) \ge c_0T. $$
The transverse zeros may appear literally or projectively, but they cannot disappear. The constants and onset are source-dependent, moving source families can postpone them, and every zero here still belongs to the source multiplier rather than to \(\Xi\). None of this proves RH. It does make the global price of the local counterfeit much more specific: one vertical density-one lattice, plus another linear divisor forced into a transverse sector.
Two sources can cancel the lattice and still fake arithmetic#
The obvious escape is to divide two completed source multipliers. Their common gamma-pinned zeros may cancel. Write
$$ R_{p,q}(s)=\frac{E_p(s)}{E_q(s)} $$
and form the cross-Mellin commutator
$$ \mathcal C_{p,q}(s) \begin{aligned} &=M_p(s)M_q(1-s)\\ &\quad-M_q(s)M_p(1-s). \end{aligned} $$
There is now an exhaustive dichotomy. If this commutator is nonzero, the quotient still pays asymptotically half a lattice in each direction:
$$ N_0(R_{p,q};(1,T])\ge \frac T2-o(T). $$
$$ N_\infty(R_{p,q};(1,T])\ge \frac T2-o(T). $$
The zeros and poles are simple, real, and superfactorially close to the positive odd integers. If the residual divisor is finite instead, that can happen only through an exact Euler-operator relation. With
$$ Z=-\left(x\frac d{dx}+\frac12\right)^2, $$
there must be nonzero real polynomials \(A,B\) such that
$$ B(Z)p=A(Z)q. $$
Equivalently, the quotient is rational in \((s-\tfrac12)^2\). Entire finite-divisor quotients reduce further to \(p=Q(Z)q\), and a zero-free one is just a constant multiple.
But the infinite-residual branch can counterfeit arithmetic remarkably well. Let \(T_c\) be centered dilation, choose \(c=\log 2\), and set
$$ q=\frac12(T_ch+T_{-c}h), $$
$$ p=q+\lambda h, \qquad 0<\lambda<1. $$
Then every common gamma-lattice factor cancels and
$$ R_{p,q}(s) =1+\frac{\lambda}{\cosh(c(s-\tfrac12))}. $$
Its remaining simple zeros and poles interlace on the critical line. Yet \(-R'_{p,q}/R_{p,q}\) is an absolutely convergent ordinary Dirichlet series in \(\Re s>1/2\), supported only on powers of \(2\). At \(\lambda=1/(2\sqrt2)\), its first two coefficients are
$$ b_2=\log 2, \qquad b_4=-\log 2. $$
It gets the first prime right and the first prime power wrong.
This counterfeit uses fixed smooth compact sources, not a moving family or a gamma-scale norm explosion. Exact gamma-lattice cancellation, right-half-plane holomorphy and zero-freeness, a strict frequency gap, an ordinary Dirichlet logarithm, and the correct first coefficient still do not force arithmetic. The missing hypothesis is global divisor elimination: entireness, not merely good behavior to the right of the critical strip.
And the wrong coefficient at \(4\) is not the real limit of the trick. Fix any \(m\). A reciprocal paraorthogonal completion produces exact source multipliers whose quotient has
$$ b_{2^n}=\log 2 \qquad(1\le n\le m), $$
while every reduced zero and pole still lies on the critical line. The completion is made from two reciprocal polynomials. One begins with the degree-\(m\) geometric jet
$$ 1+\frac{x}{\sqrt2}+\cdots+\frac{x^m}{2^{m/2}}, $$
then adds only reciprocal terms above degree \(m\). A finite Blaschke-product argument puts every completed root exactly on \(|x|=1\), and logarithmic differentiation preserves the requested jet. So no fixed number of genuine power-of-two coefficients detects the counterfeit either.
The part that made me sit up is that the construction multiplies across primes. For every cutoff \(N\), choose the local depth
$$ m_p=\max\{k:p^k\le N\} $$
for each prime \(p\le N\), make the corresponding paraorthogonal quotient, and take their finite product. Its ordinary Dirichlet logarithm then satisfies
$$ \boxed{b_n=\Lambda(n)\qquad(2\le n\le N).} $$
Not just the prime coefficients. Not just one prime-power ray. The entire finite von Mangoldt prefix is exact.
This finite product is still one exact compact-source quotient. Each local factor is an even finite exponential polynomial in \(w=s-\tfrac12\), and multiplying them gives another finite exponential polynomial. Its exponentials pull back to finitely many centered dilations of one base source.
The price remains visible and wonderfully stubborn: every local factor has infinite zero and pole progressions on the critical line. Progressions belonging to distinct primes can meet at most once, because two meetings would make \(\log p/\log q\) rational. Cross-prime cancellation can therefore erase only finitely many points from each progression. Infinitely many zeros and poles survive.
So an RH bridge may inspect any prescribed finite initial segment of the complete von Mangoldt law, exact source realizability, reflection invariance, right-half-plane zero-freeness, absolute Dirichlet convergence, and critical-line localization of every residual divisor, and still be looking at a counterfeit. The finite source cost grows rapidly with \(N\), the quotient is meromorphic rather than entire, and it does not match the infinite coefficient law. Those are the seams left to grab.
What arithmetic would have to do#
An ordinary Dirichlet logarithmic law is rigid enough to determine \(\Xi\) itself. But the smooth canonical defect cannot carry such a law: on a vertical line it is the Fourier transform of an absolutely continuous measure, whereas an absolutely convergent ordinary Dirichlet series is the Fourier transform of an atomic measure with a mandatory atom at zero.
There are even explicit ghost packets whose strip transforms and every fixed weighted source \(L^1\) norm go to zero, while a prescribed right-half-plane arithmetic derivative remains \(O(1)\). Their high derivatives blow up exponentially. Again the missing datum is global regularity.
So the surviving RH bridge has become sharply nonlocal: control the pinned projective lattice and its transverse compensator while controlling both complete logarithmic tails strongly enough to force an ordinary Dirichlet, prime-gap, or comparable arithmetic law.
Two source cells have made the local geometry too flexible. The proof, if it comes through this architecture, has to live in the topology that those cells cannot counterfeit.
Notebook references: C-0249, C-0250, C-0251, C-0252, C-0253, C-0254, C-0255, C-0256, C-0257, C-0258, C-0259, C-0260, K-0181, K-0182, K-0183, K-0184, K-0185, K-0186, O-0328, C-0248, K-0179