add generalized dot product

NEWS.md

@@ -39,6 +39,7 @@ Standard library changes
 
 * `qr` and `qr!` functions support `blocksize` keyword argument ([#33053]).
 
+* `dot` now admits a 3-argument method `dot(x, A, y)` to compute generalized dot products `dot(x, A*y)`, but without computing and storing the intermediate result `A*y` ([#32739]).
 
 #### SparseArrays
 

stdlib/LinearAlgebra/src/bidiag.jl

@@ -647,6 +647,36 @@ function *(A::SymTridiagonal, B::Diagonal)
     A_mul_B_td!(Tridiagonal(zeros(TS, size(A, 1)-1), zeros(TS, size(A, 1)), zeros(TS, size(A, 1)-1)), A, B)
 end
 
+function dot(x::AbstractVector, B::Bidiagonal, y::AbstractVector)
+    require_one_based_indexing(x, y)
+    nx, ny = length(x), length(y)
+    (nx == size(B, 1) == ny) || throw(DimensionMismatch())
+    if iszero(nx)
+        return dot(zero(eltype(x)), zero(eltype(B)), zero(eltype(y)))
+    end
+    ev, dv = B.ev, B.dv
+    if B.uplo == 'U'
+        x₀ = x[1]
+        r = dot(x[1], dv[1], y[1])
+        @inbounds for j in 2:nx-1
+            x₋, x₀ = x₀, x[j]
+            r += dot(adjoint(ev[j-1])*x₋ + adjoint(dv[j])*x₀, y[j])
+        end
+        r += dot(adjoint(ev[nx-1])*x₀ + adjoint(dv[nx])*x[nx], y[nx])
+        return r
+    else # B.uplo == 'L'
+        x₀ = x[1]
+        x₊ = x[2]
+        r = dot(adjoint(dv[1])*x₀ + adjoint(ev[1])*x₊, y[1])
+        @inbounds for j in 2:nx-1
+            x₀, x₊ = x₊, x[j+1]
+            r += dot(adjoint(dv[j])*x₀ + adjoint(ev[j])*x₊, y[j])
+        end
+        r += dot(x₊, dv[nx], y[nx])
+        return r
+    end
+end
+
 #Linear solvers
 ldiv!(A::Union{Bidiagonal, AbstractTriangular}, b::AbstractVector) = naivesub!(A, b)
 ldiv!(A::Transpose{<:Any,<:Bidiagonal}, b::AbstractVector) = ldiv!(copy(A), b)

stdlib/LinearAlgebra/src/diagonal.jl

@@ -637,11 +637,14 @@ end
 
 # disambiguation methods: * of Diagonal and Adj/Trans AbsVec
 *(x::Adjoint{<:Any,<:AbstractVector}, D::Diagonal) = Adjoint(map((t,s) -> t'*s, D.diag, parent(x)))
+*(x::Transpose{<:Any,<:AbstractVector}, D::Diagonal) = Transpose(map((t,s) -> transpose(t)*s, D.diag, parent(x)))
 *(x::Adjoint{<:Any,<:AbstractVector}, D::Diagonal, y::AbstractVector) =
     mapreduce(t -> t[1]*t[2]*t[3], +, zip(x, D.diag, y))
-*(x::Transpose{<:Any,<:AbstractVector}, D::Diagonal) = Transpose(map((t,s) -> transpose(t)*s, D.diag, parent(x)))
 *(x::Transpose{<:Any,<:AbstractVector}, D::Diagonal, y::AbstractVector) =
     mapreduce(t -> t[1]*t[2]*t[3], +, zip(x, D.diag, y))
+function dot(x::AbstractVector, D::Diagonal, y::AbstractVector)
+    mapreduce(t -> dot(t[1], t[2], t[3]), +, zip(x, D.diag, y))
+end
 
 function cholesky!(A::Diagonal, ::Val{false} = Val(false); check::Bool = true)
     info = 0

stdlib/LinearAlgebra/src/generic.jl

@@ -874,6 +874,51 @@ function dot(x::AbstractArray, y::AbstractArray)
     s
 end
 
+"""
+    dot(x, A, y)
+
+Compute the generalized dot product `dot(x, A*y)` between two vectors `x` and `y`,
+without storing the intermediate result of `A*y`. As for the two-argument
+[`dot(_,_)`](@ref), this acts recursively. Moreover, for complex vectors, the
+first vector is conjugated.
+
+!!! compat "Julia 1.4"
+    Three-argument `dot` requires at least Julia 1.4.
+
+# Examples
+```jldoctest
+julia> dot([1; 1], [1 2; 3 4], [2; 3])
+26
+
+julia> dot(1:5, reshape(1:25, 5, 5), 2:6)
+4850
+
+julia> ⋅(1:5, reshape(1:25, 5, 5), 2:6) == dot(1:5, reshape(1:25, 5, 5), 2:6)
+true
+```
+"""
+dot(x, A, y) = dot(x, A*y) # generic fallback for cases that are not covered by specialized methods
+
+function dot(x::AbstractVector, A::AbstractMatrix, y::AbstractVector)
+    (axes(x)..., axes(y)...) == axes(A) || throw(DimensionMismatch())
+    T = typeof(dot(first(x), first(A), first(y)))
+    s = zero(T)
+    i₁ = first(eachindex(x))
+    x₁ = first(x)
+    @inbounds for j in eachindex(y)
+        yj = y[j]
+        if !iszero(yj)
+            temp = zero(adjoint(A[i₁,j]) * x₁)
+            @simd for i in eachindex(x)
+                temp += adjoint(A[i,j]) * x[i]
+            end
+            s += dot(temp, yj)
+        end
+    end
+    return s
+end
+dot(x::AbstractVector, adjA::Adjoint, y::AbstractVector) = adjoint(dot(y, adjA.parent, x))
+dot(x::AbstractVector, transA::Transpose{<:Real}, y::AbstractVector) = adjoint(dot(y, transA.parent, x))
 
 ###########################################################################################
 

stdlib/LinearAlgebra/src/hessenberg.jl

@@ -284,6 +284,37 @@ function logabsdet(F::UpperHessenberg; shift::Number=false)
     return (logdeterminant, P)
 end
 
+function dot(x::AbstractVector, H::UpperHessenberg, y::AbstractVector)
+    require_one_based_indexing(x, y)
+    m = size(H, 1)
+    (length(x) == m == length(y)) || throw(DimensionMismatch())
+    if iszero(m)
+        return dot(zero(eltype(x)), zero(eltype(H)), zero(eltype(y)))
+    end
+    x₁ = x[1]
+    r = dot(x₁, H[1,1], y[1])
+    r += dot(x[2], H[2,1], y[1])
+    @inbounds for j in 2:m-1
+        yj = y[j]
+        if !iszero(yj)
+            temp = adjoint(H[1,j]) * x₁
+            @simd for i in 2:j+1
+                temp += adjoint(H[i,j]) * x[i]
+            end
+            r += dot(temp, yj)
+        end
+    end
+    ym = y[m]
+    if !iszero(ym)
+        temp = adjoint(H[1,m]) * x₁
+        @simd for i in 2:m
+            temp += adjoint(H[i,m]) * x[i]
+        end
+        r += dot(temp, ym)
+    end
+    return r
+end
+
 ######################################################################################
 # Hessenberg factorizations Q(H+μI)Q' of A+μI:
 

stdlib/LinearAlgebra/src/symmetric.jl

@@ -457,6 +457,31 @@ end
 
 *(A::HermOrSym, B::HermOrSym) = A * copyto!(similar(parent(B)), B)
 
+function dot(x::AbstractVector, A::RealHermSymComplexHerm, y::AbstractVector)
+    require_one_based_indexing(x, y)
+    (length(x) == length(y) == size(A, 1)) || throw(DimensionMismatch())
+    data = A.data
+    r = zero(eltype(x)) * zero(eltype(A)) * zero(eltype(y))
+    if A.uplo == 'U'
+        @inbounds for j = 1:length(y)
+            r += dot(x[j], real(data[j,j]), y[j])
+            @simd for i = 1:j-1
+                Aij = data[i,j]
+                r += dot(x[i], Aij, y[j]) + dot(x[j], adjoint(Aij), y[i])
+            end
+        end
+    else # A.uplo == 'L'
+        @inbounds for j = 1:length(y)
+            r += dot(x[j], real(data[j,j]), y[j])
+            @simd for i = j+1:length(y)
+                Aij = data[i,j]
+                r += dot(x[i], Aij, y[j]) + dot(x[j], adjoint(Aij), y[i])
+            end
+        end
+    end
+    return r
+end
+
 # Fallbacks to avoid generic_matvecmul!/generic_matmatmul!
 ## Symmetric{<:Number} and Hermitian{<:Real} are invariant to transpose; peel off the t
 *(transA::Transpose{<:Any,<:RealHermSymComplexSym}, B::AbstractVector) = transA.parent * B

stdlib/LinearAlgebra/src/triangular.jl

@@ -545,6 +545,90 @@ end
 rmul!(A::Union{UpperTriangular,LowerTriangular}, c::Number) = mul!(A, A, c)
 lmul!(c::Number, A::Union{UpperTriangular,LowerTriangular}) = mul!(A, c, A)
 
+function dot(x::AbstractVector, A::UpperTriangular, y::AbstractVector)
+    require_one_based_indexing(x, y)
+    m = size(A, 1)
+    (length(x) == m == length(y)) || throw(DimensionMismatch())
+    if iszero(m)
+        return dot(zero(eltype(x)), zero(eltype(A)), zero(eltype(y)))
+    end
+    x₁ = x[1]
+    r = dot(x₁, A[1,1], y[1])
+    @inbounds for j in 2:m
+        yj = y[j]
+        if !iszero(yj)
+            temp = adjoint(A[1,j]) * x₁
+            @simd for i in 2:j
+                temp += adjoint(A[i,j]) * x[i]
+            end
+            r += dot(temp, yj)
+        end
+    end
+    return r
+end
+function dot(x::AbstractVector, A::UnitUpperTriangular, y::AbstractVector)
+    require_one_based_indexing(x, y)
+    m = size(A, 1)
+    (length(x) == m == length(y)) || throw(DimensionMismatch())
+    if iszero(m)
+        return dot(zero(eltype(x)), zero(eltype(A)), zero(eltype(y)))
+    end
+    x₁ = first(x)
+    r = dot(x₁, y[1])
+    @inbounds for j in 2:m
+        yj = y[j]
+        if !iszero(yj)
+            temp = adjoint(A[1,j]) * x₁
+            @simd for i in 2:j-1
+                temp += adjoint(A[i,j]) * x[i]
+            end
+            r += dot(temp, yj)
+            r += dot(x[j], yj)
+        end
+    end
+    return r
+end
+function dot(x::AbstractVector, A::LowerTriangular, y::AbstractVector)
+    require_one_based_indexing(x, y)
+    m = size(A, 1)
+    (length(x) == m == length(y)) || throw(DimensionMismatch())
+    if iszero(m)
+        return dot(zero(eltype(x)), zero(eltype(A)), zero(eltype(y)))
+    end
+    r = zero(typeof(dot(first(x), first(A), first(y))))
+    @inbounds for j in 1:m
+        yj = y[j]
+        if !iszero(yj)
+            temp = adjoint(A[j,j]) * x[j]
+            @simd for i in j+1:m
+                temp += adjoint(A[i,j]) * x[i]
+            end
+            r += dot(temp, yj)
+        end
+    end
+    return r
+end
+function dot(x::AbstractVector, A::UnitLowerTriangular, y::AbstractVector)
+    require_one_based_indexing(x, y)
+    m = size(A, 1)
+    (length(x) == m == length(y)) || throw(DimensionMismatch())
+    if iszero(m)
+        return dot(zero(eltype(x)), zero(eltype(A)), zero(eltype(y)))
+    end
+    r = zero(typeof(dot(first(x), first(y))))
+    @inbounds for j in 1:m
+        yj = y[j]
+        if !iszero(yj)
+            temp = x[j]
+            @simd for i in j+1:m
+                temp += adjoint(A[i,j]) * x[i]
+            end
+            r += dot(temp, yj)
+        end
+    end
+    return r
+end
+
 fillstored!(A::LowerTriangular, x)     = (fillband!(A.data, x, 1-size(A,1), 0); A)
 fillstored!(A::UnitLowerTriangular, x) = (fillband!(A.data, x, 1-size(A,1), -1); A)
 fillstored!(A::UpperTriangular, x)     = (fillband!(A.data, x, 0, size(A,2)-1); A)

stdlib/LinearAlgebra/src/tridiag.jl

@@ -202,6 +202,27 @@ end
     return C
 end
 
+function dot(x::AbstractVector, S::SymTridiagonal, y::AbstractVector)
+    require_one_based_indexing(x, y)
+    nx, ny = length(x), length(y)
+    (nx == size(S, 1) == ny) || throw(DimensionMismatch())
+    if iszero(nx)
+        return dot(zero(eltype(x)), zero(eltype(S)), zero(eltype(y)))
+    end
+    dv, ev = S.dv, S.ev
+    x₀ = x[1]
+    x₊ = x[2]
+    sub = transpose(ev[1])
+    r = dot(adjoint(dv[1])*x₀ + adjoint(sub)*x₊, y[1])
+    @inbounds for j in 2:nx-1
+        x₋, x₀, x₊ = x₀, x₊, x[j+1]
+        sup, sub = transpose(sub), transpose(ev[j])
+        r += dot(adjoint(sup)*x₋ + adjoint(dv[j])*x₀ + adjoint(sub)*x₊, y[j])
+    end
+    r += dot(adjoint(transpose(sub))*x₀ + adjoint(dv[nx])*x₊, y[nx])
+    return r
+end
+
 (\)(T::SymTridiagonal, B::StridedVecOrMat) = ldlt(T)\B
 
 # division with optional shift for use in shifted-Hessenberg solvers (hessenberg.jl):
@@ -657,3 +678,22 @@ end
 
 Base._sum(A::Tridiagonal, ::Colon) = sum(A.d) + sum(A.dl) + sum(A.du)
 Base._sum(A::SymTridiagonal, ::Colon) = sum(A.dv) + 2sum(A.ev)
+
+function dot(x::AbstractVector, A::Tridiagonal, y::AbstractVector)
+    require_one_based_indexing(x, y)
+    nx, ny = length(x), length(y)
+    (nx == size(A, 1) == ny) || throw(DimensionMismatch())
+    if iszero(nx)
+        return dot(zero(eltype(x)), zero(eltype(A)), zero(eltype(y)))
+    end
+    x₀ = x[1]
+    x₊ = x[2]
+    dl, d, du = A.dl, A.d, A.du
+    r = dot(adjoint(d[1])*x₀ + adjoint(dl[1])*x₊, y[1])
+    @inbounds for j in 2:nx-1
+        x₋, x₀, x₊ = x₀, x₊, x[j+1]
+        r += dot(adjoint(du[j-1])*x₋ + adjoint(d[j])*x₀ + adjoint(dl[j])*x₊, y[j])
+    end
+    r += dot(adjoint(du[nx-1])*x₀ + adjoint(d[nx])*x₊, y[nx])
+    return r
+end

stdlib/LinearAlgebra/src/uniformscaling.jl

@@ -400,3 +400,7 @@ Array(s::UniformScaling, dims::Dims{2}) = Matrix(s, dims)
 ## Diagonal construction from UniformScaling
 Diagonal{T}(s::UniformScaling, m::Integer) where {T} = Diagonal{T}(fill(T(s.λ), m))
 Diagonal(s::UniformScaling, m::Integer) = Diagonal{eltype(s)}(s, m)
+
+dot(x::AbstractVector, J::UniformScaling, y::AbstractVector) = dot(x, J.λ, y)
+dot(x::AbstractVector, a::Number, y::AbstractVector) = sum(t -> dot(t[1], a, t[2]), zip(x, y))
+dot(x::AbstractVector, a::Union{Real,Complex}, y::AbstractVector) = a*dot(x, y)

stdlib/LinearAlgebra/test/bidiag.jl

@@ -455,4 +455,17 @@ end
     @test A * Tridiagonal(ones(1, 1)) == A
 end
 
+@testset "generalized dot" begin
+    for elty in (Float64, ComplexF64)
+        dv = randn(elty, 5)
+        ev = randn(elty, 4)
+        x = randn(elty, 5)
+        y = randn(elty, 5)
+        for uplo in (:U, :L)
+            B = Bidiagonal(dv, ev, uplo)
+            @test dot(x, B, y) ≈ dot(B'x, y) ≈ dot(x, Matrix(B), y)
+        end
+    end
+end
+
 end # module TestBidiagonal

stdlib/LinearAlgebra/test/generic.jl

@@ -409,4 +409,32 @@ end
     @test all(!isnan, lmul!(false, Any[NaN]))
 end
 
+@testset "generalized dot #32739" begin
+    for elty in (Int, Float32, Float64, BigFloat, Complex{Float32}, Complex{Float64}, Complex{BigFloat})
+        n = 10
+        if elty <: Int
+            A = rand(-n:n, n, n)
+            x = rand(-n:n, n)
+            y = rand(-n:n, n)
+        elseif elty <: Real
+            A = convert(Matrix{elty}, randn(n,n))
+            x = rand(elty, n)
+            y = rand(elty, n)
+        else
+            A = convert(Matrix{elty}, complex.(randn(n,n), randn(n,n)))
+            x = rand(elty, n)
+            y = rand(elty, n)
+        end
+        @test dot(x, A, y) ≈ dot(A'x, y) ≈ *(x', A, y) ≈ (x'A)*y
+        @test dot(x, A', y) ≈ dot(A*x, y) ≈ *(x', A', y) ≈ (x'A')*y
+        elty <: Real && @test dot(x, transpose(A), y) ≈ dot(x, transpose(A)*y) ≈ *(x', transpose(A), y) ≈ (x'*transpose(A))*y
+        B = reshape([A], 1, 1)
+        x = [x]
+        y = [y]
+        @test dot(x, B, y) ≈ dot(B'x, y)
+        @test dot(x, B', y) ≈ dot(B*x, y)
+        elty <: Real && @test dot(x, transpose(B), y) ≈ dot(x, transpose(B)*y)
+    end
+end
+
 end # module TestGeneric

stdlib/LinearAlgebra/test/hessenberg.jl

@@ -88,6 +88,11 @@ let n = 10
             @test det(H + shift*I) ≈ det(A + shift*I)
             @test logabsdet(H + shift*I) ≅ logabsdet(A + shift*I)
         end
+
+        HM = Matrix(h)
+        @test dot(b, h, b) ≈ dot(h'b, b) ≈ dot(b, HM, b) ≈ dot(HM'b, b)
+        c = b .+ 1
+        @test dot(b, h, c) ≈ dot(h'b, c) ≈ dot(b, HM, c) ≈ dot(HM'b, c)
     end
 end