# The small differences: a second exploration

10 November 2026 PT. Follow-up to Grant's objection that one natural extension is not the whole landscape.

[known] = established mathematics with sources. [project] = derived or numerically checked here, not a novelty claim. [speculative] = a direction not established. No claim of RH progress. “The intrinsic lift is equivalent to complex data” is a statement about a construction, not a verdict that hypercomplex exploration is a dead end.

## 1. Horns, infinity, and what the earlier local picture missed

**[known/project]** The specific intrinsic xi zero set in A=H or O is a locally finite union of standard round S^{d-2}, d=4 or 8. This explicit description, not smoothness alone, rules out a horned-sphere embedding at an ordinary finite zero. Xi has no real zeros. Let P be its distinct upper-half-plane complex zeros, ignoring multiplicity only for the topology, and let Z be their circularization.

**[project; exact fundamental-group calculation]** Remove the real axis temporarily. There is a homeomorphism

    A \ (Z union R) = [(R x (0,infinity)) \ P] x S^{d-2}.

The sphere factor is simply connected for d=4,8. Restoring the real axis does not change pi_1: by general position, a loop and a disk homotopy can avoid a smooth real line of codimension d-1>=3. Consequently

    pi_1(A\Z) = pi_1(upper half-plane \ P).

For n distinct zero spheres this is the free group F_n, not Z^n. The meridian loops need not commute. For the locally finite infinite zeta configuration it is the free group on countably many generators. Every loop and every disk homotopy has compact image; the punctures have no finite accumulation, so this is a direct-limit finite-word calculation, NOT the Hawaiian-earring group. This is a proof for this rotational arrangement, not a claim about arbitrary codimension-two links. Standard tool: van Kampen; compare Hatcher, Algebraic Topology, §1.2. Homology alone would not prove the free-group assertion.

**[project]** Thus there is a genuinely richer topology than one sphere's Z, already in H and O. It is still the same free-group information as the punctured orbit half-plane. Distinguish noncommuting TOPOLOGICAL loops from noncommuting quaternion multiplication; the two have different definitions.

**[project; infinity statement with proof]** In the one-point compactification S^d, form K=Z union {infinity}. Its complement fails the following local simple-connectivity condition at infinity: for a neighborhood U of infinity there is no smaller neighborhood V such that every loop in V\K contracts in U\K. Every V contains a small meridian of a sufficiently distant zero sphere. That loop already has a nonzero winding about the corresponding orbit puncture and cannot contract even in the full complement. The same argument works for all sufficiently large exterior neighborhoods: new essential generators keep appearing arbitrarily far out. One may describe a cofinal end system by free groups on tail punctures and inclusion-induced maps, rather than pretend there is just a single stable loop at infinity.

**[project; terminology matters]** This IS an infinity phenomenon missed by a purely local discussion. It resembles one diagnostic used in wild-embedding theory, but K is NOT a sphere: its finite spherical components are disconnected from one another and from the accumulation point. It is not locally a manifold at infinity and is not locally connected there. Calling it “an Alexander horned sphere” would be false. The obstruction here arises from infinitely many ordinary holes, not from proof of infinite interlocking horns. It also occurs for suitably arranged tame spheres with no arithmetic content, so it is not RH evidence.

**[project; checked compactification geometry]** Inversion q -> q/|q|² turns a sphere a+bS^{d-2} into

    [a/(a²+b²)] + [b/(a²+b²)] S^{d-2}.

At computed zero indices 1,10,100,1000, the inverted radii are approximately .07065933, .02008885, .004227878, .0007045118. The entire sphere's distance from the inversion origin is 1/sqrt(a²+b²). They shrink toward the new infinity point while each retains a meridian. This calculation uses known on-line sample zeros, not RH globally; bounded real parts and unbounded heights suffice for the infinity argument.

### An axis link that is easy to miss

**[project; exact]** Each zero sphere links the COMPACTIFIED real axis once. Its standard spanning (d-1)-disk {a+v:|v|<=b} intersects the real axis transversely at a, in one point. After compactifying the line to S1, the dimensions (d-2)+1=d-1 are precisely those for ordinary integer linking in S^d. Equivalently, A\R retracts to S^{d-2}, and each zero sphere is a generator of pi_{d-2}(A\R). This is a real geometric role for the axis, not a horn analogy.

**[project; caveats]** The two zero spheres themselves have no ordinary pairwise integer linking number of that type: 2(d-2) is not d-1 for d=4,8. A finite collection in this rotational configuration is an unlink (move their distinct orbit points apart in the half-plane and lift the isotopy). The compactified axis is disjoint from each individual sphere, but meets K at infinity, so do not assign a single linking number to the axis and the WHOLE compactified infinite set. Individual trivial-zero points and the pole at 1 are not closed linking loops; removing isolated points in dimensions >=4 adds no pi_1 generators.

### Level sets rather than zeros

**[known/project]** For intrinsic h_A, |h_A(x+yI)|=|h(x+iy)|. Near a simple complex zero, a sufficiently small regular norm-level curve |h|=c is a circle in the normal two-coordinate plane. Its circularization is S1 x S^{d-2}: a 3-dimensional tube in H or a 7-dimensional tube in O. Thus the thin-tube count has actual geometric level surfaces behind it. Near a multiple zero one must use multiplicity/critical-level analysis; do not assume every level is regular.

**[known/project]** On compact sets, these particular real-analytic level sets have tame local analytic geometry; they cannot conceal a horned embedding at every smaller scale near an ordinary point. The relevant general reference is Łojasiewicz's local triangulation of semianalytic sets, not the false assertion “smooth implies tame.” Global level sets can change topology at critical values and have infinitely complicated ends. The norm formula reduces their radial profile to a two-dimensional level-curve problem; whether the actual global xi level graph has informative end structure is OPEN here.

**[project]** A complex argument is slice-specific. A quaternionic polar angle also needs an imaginary direction, becomes ambiguous on negative real values, and cannot simply be treated as one global scalar phase. An “argument-level set” requires a stated slice or branch convention before topology can be compared.

## 2. i, j and k give genuinely different outputs for non-real coefficients

**[known/project; exact angular law]** Fix the coefficient plane C_J. For a right-coefficient L lift, put U=L(x+yJ,chi), V=L(x-yJ,chi), A=(U+V)/2, B=J(V-U)/2. Then f(x+yI)=A+IB. With c=<I,J>,

    |f(x+yI)|² = [(1+c)/2]|U|² + [(1-c)/2]|V|².

This is an affine function of latitude c, not merely a numerical fit. When U=0 and V!=0, |f|=|V|sin(theta/2), so only I=J is zero. A constant-norm set on the input sphere is a LATITUDE sphere; it is a great equator only at c=0. In H these are circles; in O they are S5 latitudes on S6. The endpoints are single points.

**[project]** For arbitrary A,B, the map I -> A+IB takes the imaginary-unit sphere to a round affine sphere centered at A with radius |B| (when B!=0), because right multiplication by B scales norms. This remains true in O; two-factor norm multiplicativity is enough. Thus an isolated zero is a translated image sphere just touching/passing through the origin at one input direction, rather than the entire image collapsing to zero.

**[project; checked]** angles.py tests two zeros each for the order-4 mod-5 character chi(2)=i and an order-3 mod-7 character chi(3)=exp(2pi i/3). Values at 60 digits:

| modulus | positive zero height | norm at perpendicular j or k | norm at opposite -i |
|---|---:|---:|---:|
| 5 | 6.18357819545085 | 1.99211636624293 | 2.81727798296616 |
| 5 | 8.45722917442323 | 1.30156580632751 | 1.84069201562944 |
| 7 | 4.35640162473628 | 1.78542603876920 | 2.52497371864147 |
| 7 | 8.78555471449908 | 1.08442461011056 | 1.53360799098951 |

All sampled roots have real part 1/2 to working precision. Random directions in both H and O satisfy the exact latitude formula within finite-precision residuals recorded in angles-results.json. These are not interval-certified roots and not claims about all character zeros.

**[project; an actual j/k difference]** At the first mod-5 zero with J=i, the outputs are approximately

    f(1/2+t j) = 1.134986 - .834308 i + .834308 j -1.134986 k,
    f(1/2+t k) = 1.134986 - .834308 i +1.134986 j + .834308 k.

They have the SAME norm but DIFFERENT components. “Nothing changes with direction” was true for intrinsic zeta's norms, not for every output of every hypercomplex L-function.

**[known/project; exception to “always one point”]** For coefficients in one C_J, a given conjugacy sphere contains: no zero; one zero at one of its two C_J intersections; or the whole sphere when U=V=0. A coincident zero of the character and its conjugate can give the latter case. We have not proved that such coincidences never occur. For general quaternionic coefficients not confined to one coefficient plane, an isolated zero can point in another direction: solve I=-A B^{-1}, when that is a pure unit. Slice zero theory gives at most one unless the whole sphere vanishes. Source: Ghiloni–Perotti (2009), Corollaries 2/5 and Theorem 6.

## 3. Order and branch choices: familiar formulas really do break

**[project; exact counterexample]** At the REAL input s=2, give prime 2 coefficient i and prime 3 coefficient j. The two Euler factors (1-i/4)^{-1} and (1-j/9)^{-1} do not commute. Their products have the same scalar/i/j components but k components +.02582496413 and -.02582496413. Swapping changes the result by norm .05164992826. More fundamentally, a completely multiplicative arithmetic coefficient a_6 would have to be a_2 a_3=k AND a_3 a_2=-k. This is inconsistent. One can define an ordered product, but that is new data, not the usual order-independent Euler product for ordinary integers.

**[known/project]** For intrinsic zeta at ONE q, every p^{-q} lies in the same commutative C_I, so its Euler product survives unchanged in Re q>1. For the complex-character lift outside its coefficient plane, even a single expression n^{-q}chi(n) need not commute with another factor. A coefficientwise Dirichlet/slice product can preserve algebraic convolution identities when defined correctly; a pointwise Euler product is not automatically that product. Always state which multiplication is intended.

**[project; checked]** exp(i+j) differs from exp(i)exp(j), with norm difference .799214173966059. Exponential addition is valid for commuting arguments, not arbitrary different directions. By contrast, p^{-q}p^{-q}=p^{-2q} is safe because both exponents are in C_I and commute.

**[known/project; a sharp branch example]** The logarithms of -1 in H include pi I for EVERY imaginary unit I (and all odd multiples). If a^q is defined as exp(q log a), the expression (-1)^j is not a single number without a branch choice:

    log(-1)=pi i  -> exp(j pi i)=exp(-pi k)=-1,
    log(-1)=pi j  -> exp(j pi j)=exp(-pi)=.04321391826377....

Both are valid under their stated conventions; neither is “the” unique quaternionic power. The ordinary principal quaternion logarithm is not uniquely defined on negative real inputs. At a zeta zero, log zeta is undefined regardless of direction. This distinguishes the logarithm of the input -2n from the logarithm of the zero VALUE zeta(-2n).

**[known/project; functional-equation audit]** The intrinsic lift of zeta's classical functional equation survives away from its singular factors: Gamma(q), powers of positive real bases, sin(pi q/2), and zeta(1-q) all belong to the same C_I at a fixed q. Apparent singularities at trivial zeros must be interpreted by continuation/limits, not by multiplying 0 by infinity on a computer. If one replaces these factors by unrelated i/j/k-valued functions, commutations and cancellations cannot be silently reused; one has changed the hypotheses.

**[known/project; integral audit]** In Re q>1 the Mertens inverse formula remains q int_1^infinity M(x)x^{-q-1}dx because its factors lie in one C_I. Move an arbitrary non-real coefficient inside the integral and left/right order can matter. The original Mertens function is real, so it does not introduce that change. The Riesz RH criterion remains on the positive real ray; arbitrary directional growth estimates are different statements and fail already on the negative real ray. These domains and orders are substantive assumptions, not cosmetic notation.

## 4. Quaternionic arithmetic can force a zero exactly at 1/2

**[known; importantly different construction]** The quaternion GROUP Q8={±1,±i,±j,±k} can occur as the Galois symmetry group of a number field N/Q. An Artin L-function packages how primes act through a representation of that group. This is NOT “evaluate Riemann zeta at a quaternion.” It is another arithmetic L-function, of an established kind, whose symmetry really is quaternionic.

**[known/project; representation verified]** Q8 has one irreducible two-dimensional complex character psi of quaternionic/symplectic type. Explicit matrices are

    rho(i)=diag(i,-i), rho(j)=[[0,1],[-1,0]].

They anticommute, square to -identity, and preserve the skew form [[0,1],[-1,0]]. The Frobenius–Schur indicator (1/8)sum_g trace rho(g²) is -1, checked exactly by matrix arithmetic here. “Quaternionic type” here is a precise representation-theoretic condition, not a metaphor for choosing another slice.

**[known]** Artin local factors use det(1-rho(Frob_p)p^{-s})^{-1}, not an unordered product of noncommuting quaternion scalars. Determinants turn the conjugacy-invariant matrix data into complex scalar Euler factors. For elements i,j,k the factor polynomial is 1+t²; for -1 it is (1+t)². This is one reason genuine nonabelian arithmetic avoids the contradictory a_6 assignment above.

**[known; Armitage/Fröhlich]** Quaternionic Artin L-functions can have root number W=-1. For a holomorphic self-dual completed L-function with functional equation

    Lambda(s)=W Lambda(1-s),

W=-1 forces Lambda(1/2)=-Lambda(1/2), hence Lambda(1/2)=0 exactly. In the Q8 case the degree-two representation is induced from a nontrivial degree-one character of a cyclic subgroup, giving the needed holomorphic Artin L-function via its Hecke realization. Quaternionic type alone does NOT force W=-1: the sign depends on the extension, and +1 examples also exist. Armitage (1972) established number-field zeta functions with a central zero; Fröhlich studied the associated Artin root numbers and integral bases. The special arithmetic sign, not bare dimension or “j instead of i,” is what changes the answer at 1/2.

**[known/project]** Artin factorization for a Q8 Galois field reads

    zeta_N(s)=zeta_Q(s) L(s,chi_1)L(s,chi_2)L(s,chi_3)L(s,psi)^2.

A forced central zero of L(s,psi) therefore supplies at least a double central zero of zeta_N. This is consistent with zeta_N's sign +1: the negative-sign factor is squared. There is no assertion that zeta_Q itself has acquired that zero. Numerically zeta_Q(1/2)=-1.4603545088095868..., which is nonzero.

### The tiny change in the mirror changes the allowed outputs

**[project; exact deduction]** Suppose the completed self-dual Artin L-function is of real type, as for the real-valued Q8 character. For W=+1 its intrinsic lift obeys F(1-conj q)=conj F(q); the critical hyperplane maps into R. For W=-1 it instead obeys

    F(1-conj q)=-conj F(q).

At a fixed input Re q=1/2 this forces F(q)=-conj F(q), so the output is PURE IMAGINARY, not real. At the central REAL input q=1/2, real type also forces the output real; it must therefore be zero. On a positive-radius critical sphere the intrinsic output has form I beta(r): it sweeps an imaginary sphere, rather than collapsing every direction to the same real scalar. This is precisely a change in which outputs satisfy the symmetry constraints.

**[project; derivative consequence]** The earlier Jacobian collapse for xi used beta=0 throughout the critical hyperplane. With negative mirror sign, beta need not vanish there. Angular derivatives can survive, and the full determinant need not vanish off the central zero. A forced central zero of odd order m is an isolated real zero with local degree m, not a positive-radius spherical zero contributing 2m. “Real on the critical plane” is a plus-sign theorem, not a universal L-function fact.

**[project; checked exact models]** The illustrative even/odd models cosh(q-1/2), sinh(q-1/2) make this visible without claiming a numerical Artin computation. On q=1/2+rI they give cos r and I sin r. At r=.7 the odd output has norm .6442176872... and rotates with I, while the even output is the same real .7648421873... in every direction. The odd model's centre has degree 1 and its critical-sphere Jacobian determinant is positive in d=4 and d=8; see angles-results.json. These are symmetry models, not the actual arithmetic L-function.

**[project; verification limits and a rejected example]** We verified the matrices, symplectic form, Frobenius–Schur indicator and local factor polynomials, NOT an explicit field's Artin L-value. Installing PARI/GP was attempted through the ordinary package manager; package indexes were unavailable and Debian repository refresh failed with HTTP 500, so no PARI computation is claimed. Historical original-paper bibliographic records were read, but their full scans were inaccessible through the supplied archive/publisher links. A full 2016 thesis reproducing Fröhlich's theorem was read as secondary exposition. Its Example 2.9 prints alpha=(5+sqrt5)(41+sqrt41)/4 and calls Q(sqrt alpha) a tame Q8 field; we did NOT adopt that formula. Its conjugation ratios introduce sqrt410, not contained in Q(sqrt5,sqrt41), giving a normality warning. A certified explicit field/root-number example remains a follow-up, not a fabricated numerical result.

**[speculative; constructive next experiment]** Obtain an independently certified Q8 field polynomial and ramification data from the historical literature or a reputable arithmetic database, verify its Galois group and root number, then compute the odd completed function near 1/2. This is a bounded, meaningful quaternionic-arithmetic experiment. It addresses the central point for another L-function; it is not an RH proof strategy merely because that point is shared.

## 5. Octonions: parentheses are part of the question

**[known/project; exact checked example]** With Cayley–Dickson units e1,e2,e4,

    (e1 e2)e4= e7,     e1(e2 e4)=-e7.

The associator has norm 2. Thus a three-factor octonionic Euler product requires a bracketing convention even if the order is fixed. No familiar reordering, regrouping, inverse cancellation or determinant construction is licensed without checking its hypotheses. Artin's theorem covers TWO generators, not three arbitrary directions.

**[known]** Alternativity and the Moufang identities provide important controlled reassociations, e.g. (xy)(zx)=x(yz)x with the usual consistent bracketing. They do not make every product associative. The previously reproduced octonionic slice-product/camshaft example is an established instance where the correct slice product moves an isolated zero to another point of its conjugacy sphere.

**[project]** Intrinsic zeta never uses three independent imaginary coefficients at a single input, so this exact associator does not modify its functional equation. It demonstrates where a newly defined octonionic arithmetic object would need extra structure before classical formulas were even well-defined.

## 6. What this changes about the project

**[project]** We now have distinct live questions, not a blanket dead end: (1) characterize the topology of infinite zero-sphere ends without misnaming them horned spheres; (2) examine non-real-coefficient angular geometry, including equal norms with different outputs; (3) track branch and multiplication assumptions explicitly; (4) investigate genuine quaternionic Artin arithmetic and central root-number forcing; (5) test genuinely nonassociative constructions with stated order and brackets. These directions differ in scope and should not be collapsed into the original intrinsic-lift reduction theorem.

**[speculative]** Best next steps: a certified Q8 negative-root-number example; a rigorously stated pro-fundamental-group/end description of the actual xi sphere complement and its norm-level tubes; or a motivated multi-direction coefficient system with a well-defined Euler theory. None currently supplies an obstruction to off-critical zeros of the original Riemann zeta function.

## Sources and exact scope of access

**[known; primary citations]**
- J. W. Alexander, “An example of a simply connected surface bounding a region which is not simply connected,” PNAS 10 (1924), 8–10. https://doi.org/10.1073/pnas.10.1.8 . Historical horned-sphere reference; not freshly read in this follow-up. The zero-sphere/infinity statements above have their own explicit proofs.
- S. Łojasiewicz, “Triangulation of semi-analytic sets,” Annali della Scuola Normale Superiore di Pisa (3) 18 (1964), 449–474. Cited for local analytic tameness, not for smooth sets generally. https://www.numdam.org/item/ASNSP_1964_3_18_4_449_0/
- R. Ghiloni and A. Perotti, “Zeros of regular functions of quaternionic and octonionic variable: a division lemma and the camshaft effect,” https://arxiv.org/html/0904.2667 . Full author preprint read in the first pass; supplies the zero-structure and product distinctions used here.
- J. V. Armitage, “Zeta functions with a zero at s=1/2,” Invent. Math. 15 (1972), 199–205. DOI https://doi.org/10.1007/BF01404125 . Primary bibliographic record read at https://eudml.org/doc/142128 ; full scan link failed. Its theorem is cited as established literature, not represented as independently re-proved or numerically reproduced here.
- A. Fröhlich, “Artin-root numbers and normal integral bases for quaternion fields,” Invent. Math. 17 (1972), 143–166. Primary bibliographic record: https://geodesic.mathdoc.fr/item/IM_1972__17_142165/ ; full scan link failed. Also “Artin root numbers, conductors, and representations for generalized quaternion groups,” Proc. LMS (3) 28 (1974), 402–438, https://doi.org/10.1112/plms/s3-28.3.402 .
- A. Fröhlich and J. Queyrut, “On the functional equation of the Artin L-function for characters of real representations,” Invent. Math. 20 (1973), 125–138. Contextual reference; not freshly read. We do not attribute a numerical Q8 field example to this paper without checking it.
- J.-P. Serre, Linear Representations of Finite Groups, Springer (English edition 1977), chapters on real/quaternionic type and the Frobenius–Schur criterion. Standard primary-author exposition, not freshly read; the Q8 matrix calculation is reproduced explicitly.

**[project; supporting exposition actually read]** R. Fregoli, “On Frohlich's conjecture for tamely ramified quaternion extensions,” University of Pisa thesis (2016), introduction, Theorem 0.1 and pp. 35–37. https://etd.adm.unipi.it/theses/available/etd-11182016-190630/ ; downloaded public PDF under sources/fregoli2016.pdf and pdftotext output. This is secondary exposition and its suspect Example 2.9 was explicitly not used as a verified field example. Allen Hatcher's Algebraic Topology is a standard reference for the van Kampen calculation; the problem-specific quotient argument is given above.

**[project; reproducibility]** angles.py -> angles-results.json / angles-run.log. Character numerics and matrix/algebra checks use 60-digit mpmath, deterministic random directions, no interval arithmetic. Compactification sample uses zero indices 1,10,100,1000. No script output is a numerical Artin L-function evaluation.
