Cheat to compute ground truth.

All-pairwise inner product version.

Duplicate individual elements in A.

Duplicate individual elements in A.

Pre-rotate A, all shifts, mult with diagonals from B.

Pre-rotate A, baby step giant step version, mult with diagonals from B.

Generalized to levels >= 2.

Multiply X and Y to add to Acc without modular reduction.

index specifies which diagonal to extract applies right-rotation by nrot positions before encoding.

Return if a diagonal vector exists without extracting/encoding the vectors.

index 0 is the main diagonal max size of Block is dim by dim and index ranges from 0 to dim-1 (mod dim) If given diagonal does not overlap with X (matrix might be smaller), returns false.

Return if a diagonal vector exists without extracting elements.

Since required precision scales with 1/sqrt(n) maintain sqrt(n)*v for unit vectors v.

NetDQRplain returns Q all zeros (or nil) (for pid=0), else returns share of Q for each party.

Second output is a flag indicating whether write is attempted when the file already exists.

Multiply Q (kp by nsnp) with X (nsnp by nind) with lazy normalization of X Compute Q * S * (X - m * 1^T) as (Q * S) * X - ((Q * S) * m) * 1^T S: diagonal matrix containing 1/stdev of each SNP, m: column vector containing mean of each SNP.

Multiply Q (kp by nind) with X^T (nind by nsnp) with lazy normalization of X Compute Q * (X^T - 1 * m^T) * S as ((Q * X^T) - ((Q * 1) * m^T)) * S S: diagonal matrix containing 1/stdev of each SNP, m: column vector containing mean of each SNP TODO: multiply with S AFTER aggregation across parties, that way bootstrap once for all.

Cache structure for MatMult4 First index corresponds to row index of blocks of size slots-by-slots Second index of indexMap corresponds to index of diagonals (0..slots-1), The table maps a diag index to the encoded PlainVector If a given index has empty data, stored element is nil.