A fixed border for every quotient of order four

What Lean checks

The full signed-automorphism group excludes order 479, with both norm proofs checked. The fixed-border common multiplier group is cyclic of order dividing four and every preserving subgroup leaves at least 151 sign orbits. On any finite abelian group with a subgroup of index four, the original four sign rows and raw correlation equations yield an actual bordered Hadamard matrix; normalization, the literal finite border, and all fourteen size-149 baseline patterns are proved. The length-598 order-eleven obstruction is complete. The actual two-row Hadamard construction and quotient transport are checked. For length 1197, the actual common multiplier order is proved to divide 27; exclusion of order 27 and the finite quotient/no-lift certificates remain candidates.

This entry is a partial result on an open problem. It does not claim a solution of the general conjecture.

Conditional construction. Let GG be any finite abelian group with a subgroup SS of size ss and quotient Q=G/SQ=G/S of order four. Suppose four sign-valued functions on GG have combined periodic autocorrelation −12-12 on the nonzero elements of SS and +4+4 off SS. They yield a Hadamard matrix of order 16s+1216s+12. For order 2396 the outstanding task in this construction is to find the four functions at s=149s=149.

The bordered array and its three matrix conditions are Theorem 7 of Egan, Kotsireas, London and Ó Catháin. Kumar gives a border construction for compressions having one exceptional coset per nonconstant row and discusses how symmetry can enforce that form. We show directly that for either group of order four the autocorrelation equations alone force this form up to independent row translations and sign changes. Our border data adapt Kumar's explicit order-2636 certificate. Theorem 7 · Kumar, search notes §102 · Source certificate

Normalization of the coset sums

Write ci(q)c_i(q) for the sum of row ii over coset qq. Summing the original autocorrelations within cosets shows that the combined periodic autocorrelation of these four compressed rows is 4s+124s+12 at zero and 4s4s at each other shift. Let ee be the parity of ss and put xi(q)=(ci(q)−e)/2x_i(q)=(c_i(q)-e)/2. All these entries are integers. At each nonzero shift tt,

∑i,q(xi(q+t)−xi(q))2=14(2(4s+12)−2(4s))=6.\sum_{i,q}(x_i(q+t)-x_i(q))^2=\tfrac14\big(2(4s+12)-2(4s)\big)=6.

Summing over the three shifts counts each unordered pair of distinct cosets twice. Hence the four nonnegative integers

Di=∑q<r(xi(q)−xi(r))2=4∑qxi(q)2−(∑qxi(q))2D_i=\sum_{q\lt r}(x_i(q)-x_i(r))^2 =4\sum_q x_i(q)^2-\big(\sum_q x_i(q)\big)^2

sum to nine. Each is zero or three modulo four. Exactly three are three modulo four, so those three are each three and the remaining one is zero. Four integers with pairwise squared differences summing to three have three equal entries and a fourth differing by one. Indeed any difference of magnitude two already contributes four, and a two-against-two split contributes four.

The zero row variance gives a constant compression. Independently translate each exceptional coset to zero and negate the corresponding row if necessary. Such translations lift to GG and the changes preserve every autocorrelation. After a row permutation the compressions consequently have the form

d,a+2δ0,b+2δ0,c+2δ0.d,\qquad a+2\delta_0,\qquad b+2\delta_0,\qquad c+2\delta_0.

At the trivial character the combined squared row sums equal 16s+1216s+12, giving

(4d)2+(4a+2)2+(4b+2)2+(4c+2)2=16s+12.(4d)^2+(4a+2)^2+(4b+2)^2+(4c+2)^2=16s+12.

At s=149s=149 the four baselines are odd. Negate the constant row to make d>0d>0 and sort the other three baselines. Exactly 14 possibilities remain. The verifier enumerates every admissible baseline and prints the list. Each possibility still requires binary rows on all 596 group elements.

Border certificate

Use the Goethals–Seidel signs in Theorem 3 of the cited survey. On vectors constant on cosets, its core acts by the quotient array of the cic_i. Within each of the four quotient blocks restrict to vectors whose four entries sum to zero. A constant compression acts by zero there and a compression a+2δ0a+2\delta_0 acts by 2I2I. The constant row is the diagonal generator and vanishes on this subspace. Every off-diagonal generator in the array includes the same reflection RQ:q↦−qR_Q:q\mapsto -q. Hence the induced core there is Γ⊗RQ\Gamma\otimes R_Q, where

Γ=2(0111−101−1−1−101−11−10).\Gamma=2\begin{pmatrix} 0&1&1&1\\-1&0&1&-1\\-1&-1&0&1\\-1&1&-1&0 \end{pmatrix}.

Let K,E,FK,E,F be the sign matrices in the companion certificate and let J4J_4 be the all-ones matrix. The finite certificate satisfies

KKT=12I,EET=16I,FFT=16I−4(I4⊗J4),KK^T=12I,\quad EE^T=16I,\quad FF^T=16I-4(I_4\otimes J_4),

(I4⊗J4)ET=0,(Γ⊗I4)ET+FKT=0.(I_4\otimes J_4)E^T=0,\qquad (\Gamma\otimes I_4)E^T+FK^T=0.

Take the left border factor to be FQ=(I4⊗RQ)FF_Q=(I_4\otimes R_Q)F. Reflection permutes sign entries and preserves the Gram identity for FF. Multiplying the last certificate identity on the left by I4⊗RQI_4\otimes R_Q gives the required cross term. The balanced columns of ETE^T therefore satisfy the cross-term equation of Theorem 7 after replication on cosets. All its hypotheses hold for the normalized core. This proves the conditional construction, including for nonsplit extensions.

The certificate uses Kumar's XX as KK and his RhatR_{\mathrm{hat}} as EE with the second and third four-column blocks negated. We computed F=−(Γ⊗I4)ETK/12F=-(\Gamma\otimes I_4) E^T K/12. The verifier checks every identity in integer arithmetic and assembles two whole matrices at order 60, one for each quotient group, as controls. Border data · Executable check

Symmetries available at order 2396

Here G=Q×C149G=Q\times C_{149} since the two factor orders are coprime. Suppose all four rows are fixed by a common group UU of group automorphisms in the chosen origins. Then |U|≤4|U|\le4.

First, a prime-order symmetry acting only on QQ must act as an involution or a three-cycle. In C4C_4 the only possibility is inversion, and in the Klein four group the action permutes the three nonzero points. In every case there is an element hh of order two such that one pair x,x+hx,x+h belongs to the same nontrivial orbit, while a fixed point must lie in the other pair. For an involution take its swapped pair. For a three-cycle pick any two nonzero points, whose difference is the third. Thus in every row the two length-149 strings on the first pair are identical. For the other pair let gig_i be the number of positions where its two signs agree. The combined correlation at (h,0)(h,0) is ∑i4gi=4\sum_i 4g_i=4. The absolute sum of the two coset sums in the other pair is at most 2gi2g_i. A row with a unique exceptional coset must have that coset fixed, so the other pair contains the exceptional position. Row negations preserve both the agreement counts and the absolute sums. Using the baselines above gives

|d|+|a+1|+|b+1|+|c+1|≤1.|d|+|a+1|+|b+1|+|c+1|\le1.

None of the 14 candidate patterns satisfies this. Consequently UU injects into Aut(C149)\mathop{\mathrm{Aut}}(C_{149}), which is cyclic of order 148.

If 37 divides |U||U|, an element of order 37 acts trivially on QQ and has four length-37 orbits and one singleton on each C149C_{149} fiber. A fiber sum must then be congruent to +1+1 or −1-1 modulo 37. Every candidate fiber sum has magnitude less than 36, so every one must be +1+1 or −1-1. That would require d=1d=1 and a=b=c=−1a=b=c=-1, contrary to the squared-row-sum equation. Thus |U||U| divides four.

Even the largest permitted common multiplier group leaves at least 151 sign orbits per generator. The 592 positions with nonzero C149C_{149} coordinate contribute at least 148. On the four remaining positions a cyclic group of order dividing four has image of order at most two on QQ, with at least three orbits. This count concerns fixed common multipliers. Other forms of symmetry may impose different constraints.

Lean verification

The symmetry theorem in the preceding section is fully formalized, including arbitrary finite abelian domains and every common multiplier subgroup. The 12-module closure passed strict source checks, a 216-theorem standard-axiom audit, and independent replay by the unchanged pinned Lean kernel checker. Receipt · Exact checked scope. The earlier normalization and matrix border construction remain separate formalization obligations.

The formal proof uses the raw correlations and does not require the fourteen normalized baseline patterns. Write Ci(q)=∑zFi(q,z)C_i(q)=\sum_z F_i(q,z) for the 16 fiber sums. Summing over fiber shifts gives

∑i,qCi(q)2=608,∑i,qCi(q+h)Ci(q)=596(h≠0).\sum_{i,q} C_i(q)^2=608,\qquad \sum_{i,q} C_i(q+h)C_i(q)=596\quad(h\ne0).

Thus |Ci(q)|≤24|C_i(q)|\le24. A nontrivial quotient automorphism preserving all rows has a moved pair a,a+ha,a+h, with hh of order two and Q={0,h,a,a+h}Q=\{0,h,a,a+h\}. The actual correlation at (h,0)(h,0) and the compressed correlation at aa imply

∑i|Ci(0)+Ci(h)|≤2,∑iCi(a)(Ci(0)+Ci(h))=298.\sum_i |C_i(0)+C_i(h)|\le2,\qquad \sum_i C_i(a)\bigl(C_i(0)+C_i(h)\bigr)=298.

The absolute value of the latter sum is at most 24⋅2=4824\cdot2=48, a contradiction. This proves injectivity of the actual multiplier projection. An element of order 37 would force each fiber sum to be congruent to 11 or −1-1 modulo 37. The same magnitude bound then makes all 16 fiber sums signs, contradicting energy 608. The unit group is cyclic of order 148, so the common multiplier group is cyclic of order dividing four.

For the orbit bound, the origin has a singleton orbit. A nonzero point in the characteristic four-torsion subgroup has an orbit of size at most two. Every orbit has size at most four. If there are mm actual orbits, these two deficits give 596+5≤4m596+5\le4m, hence m≥151m\ge151.

For an original group GG and subgroup SS, the formalization constructs the product coordinates from the actual quotient map and the retraction g↦g−149gg\mapsto g-149g onto SS. Conjugation identifies the full row-preserving subgroup and its orbit quotient with those in product coordinates. No splitting or prescribed coordinate action is an extra hypothesis.

The certificate's transport to an arbitrary quotient is formalized in HadamardFixedBorderTransport.lean and HadamardFixedBorderFin16.lean. Given the explicit finite certificate fields and a quotient of cardinality four, the theorem constructs the actual relabeled E and reflected F, preserves sign-valuedness, and proves all five transported identities. The two-module closure passed strict source checks, a 66-theorem axiom audit, and unchanged independent kernel replay. Receipt.

The step from compression shapes to actual normalized rows is also formalized. Given one constant compressed row and three compressed rows with exceptional jumps of 2 or -2, the theorem constructs the row permutation, sign changes, and lifted translations on the original group. It proves the exact four normalized compressions, sign-valuedness, and preservation of every combined periodic autocorrelation. No split extension or positive constant baseline is assumed. The two-module closure passed strict source checks, a 14-theorem axiom audit, and independent pinned-kernel replay. Receipt.

The variance argument producing the compression shapes, the literal finite certificate, and final bordered matrix assembly remain separate obligations under issue 17.