Open problem. Does a Hadamard matrix of order 2396 exist? The results below restrict two sequence constructions. Both unrestricted existence at this order and the full Hadamard conjecture remain open. These arguments are unrefereed.
Legendre pairs of length 1197
Let G be any abelian group of order 1197 = 9·7·19. Write R for the subgroup of order nine, which is either cyclic or elementary abelian. A candidate consists of two sign-valued functions A,B on G whose periodic autocorrelations sum to −2 at every nonzero shift. The two row sums can be normalized to +1, since their squares sum to 2. A pair would give a Hadamard matrix of order 2396 by the two-circulant-core construction (group-developed matrices work over an abelian group).
Restriction. The group of common automorphisms of the two sequences, allowing a separate translation of each sequence, has order at most nine. Here each automorphism preserves each row up to translation. Row interchange is outside the statement.
To remove translations, center each row f separately by its first moment, the element Σ f(g)g of G. Translating by t adds t to the moment because the row sum is one. A group automorphism carries the moment to its image, so centered rows are fixed by every common multiplier.
Even orders. An involution on the dual of an abelian group of odd order splits it into the subgroups fixed and negated by that involution (apply (1±h)/2). A nonidentity involution therefore negates a character of prime order p in {3,7,19}. For either invariant row, its Fourier value at that character is real. Dividing the value by two gives an algebraic integer, since at a nontrivial character the transform of the constant function vanishes and f equals 1 minus twice an indicator. Denote the resulting two values by x,y. Fourier transforming the correlation equation gives x²+y²=599. Both elements belong to the real cyclotomic field K of conductor p and odd degree (p−1)/2. Take the norm of x+iy from K(i) to Q(i). The result is a Gaussian integer whose squared absolute value is 599 to that odd degree, since conjugation commutes with the norm. This is impossible because 599 is a prime congruent to 3 modulo 4. Cauchy's theorem now excludes all even-order multiplier groups. For related prior work on compression and the p=3 sum-of-squares test see Multiplier obstructions for Legendre pairs of length 333.
Projection away from R. An odd-order automorphism subgroup on R has order at most three (the ambient orders are 6 and 48). Any nontrivial such subgroup has three singleton orbits in R and permutes each of the other two cosets of a subgroup of size three transitively. In the cyclic case this is multiplication by 4 or 7. In the vector-space case it is a nonidentity unipotent 2-by-2 matrix over the field of three elements. The image and kernel of its difference from the identity are the same line.
If a multiplier of order three acts only on R, each of the 133 copies of R has two constant three-point fibers in each row. The sum of squares of the entries after summing each three-point fiber is at least 133(9+9+1)=2527 for even one row. The corresponding energy for both rows is exactly 2394−2·2=2390. Indeed compression along a subgroup of size d sums the correlations on every coset of that subgroup. The combined energy on the quotient is 2396−2d, and the other combined correlations are −2d. Thus projection of the common multiplier group to the 7- and 19-components is injective. Its image lies in the groups of squares of orders 3 and 9, respectively. Only image order 27 still needs to be excluded to obtain the stated bound.
Full image of order 27. An image of that order acts as the independent square multipliers on the factors of orders 7 and 19. Its action on R fixes a quotient C3 pointwise, by the preceding description, including when the action is trivial. Two restrictions of the same original sequences can now be tested in succession.
- Compress along all of R to C7 × C19. Each row is specified by nine odd integers between −9 and 9, one per class in the product of the zero, square and nonsquare classes. Each row has weighted sum one. For the pair the combined energy must be 2378 and the combined correlation at each other shift must be −18. Exactly 32 ordered pairs of nine-entry rows pass.
- In C3 × C7 × C19, obtained by compressing only the kernel of R → C3, each of the nine entries must split into three entries from {−3,−1,1,3} with the prescribed sum. On each of the three layers the same nine square classes apply. After further compression to C3 × C7, the three columns of a layer (zero, squares, nonsquares modulo 19) carry weights 1,9,9. The resulting combined energy must be 2282 and the other twenty combined correlations must be −114. None of the 32 pairs admits even these projected margins.
The first enumeration covers every possible nine-entry normalized row. One form eliminates the zero-class entry using the row sum. The other joins triples of digits by their weighted sum. The second enumeration allows every splitting described in step 2, so failure excludes all lifts. The file quotient133.pairs contains the 32 survivors of the first test. The C++ and Python enumerations use separate correlation computations on the 133-point quotient. The audit works on the expanded cyclic rows of lengths 133 and 21 and checks all 32 without symmetry reduction.
Bordered quadruples of length 598
Here the group is C26 × C23. The four binary rows would have combined periodic autocorrelation −4 at each nonzero shift. This is Theorem 5 in the construction survey. Their row sums have absolute values 2,0,0,0.
A common multiplier of order 11 cannot fix all four rows. Every order-11 multiplier acts on the 23-component as the quadratic residues and fixes the 26-component. The following argument assumes the rows are fixed in the chosen origins. It also applies if each row can be individually translated to such an origin.
In a 23-point fiber the compressed entry is in {±1,±21,±23}. The squared entries of the four length-26 compressions have sum 2392−4·22=2304. Relative to a baseline of 104, the excess is 2200. Hence 440a+528b=2200 for counts a,b of entries of absolute sizes 21,23. The sole nonnegative solution is (5,0).
Write each compression as c+22e, with c the signs at the zero positions of the 23-point fibers and e supported on those five exceptional positions across all rows. Its nonzero entries are ±1, opposite to the corresponding c. At every nonzero shift of C26 the compressed combined correlation is −92. The c-c correlation has absolute value at most 104. The two cross-correlations contribute at most 22·10. Since 104+220+92<484, the combined e-e correlation, whose coefficient is 484, must vanish. Summing over all shifts shows that the four row sums of e have squares summing to five. One therefore has absolute value at least two. For this row the original row sum has absolute value at least 44−26=18, contradicting the upper bound two.