inner() function in the stokes package
inner## function (M)
## {
## ktensor(spray(expand.grid(seq_len(nrow(M)), seq_len(ncol(M))),
## c(M)))
## }
## <bytecode: 0x7fc3a6e66f10>
## <environment: namespace:stokes>
Spivak, in a memorable passage, states:
The reader is already familiar with certain tensors, aside from members of \(V^*\). The first example is the inner product \(\left\langle{,}\right\rangle\in{\mathcal J}^2(\mathbb{R}^n)\). On the grounds that any good mathematical commodity is worth generalizing, we define an inner product on \(V\) to be a 2-tensor \(T\) such that \(T\) is symmetric, that is \(T(v,w)=T(w,v)\) for \(v,w\in V\) and such that \(T\) is positive-definite, that is, \(T(v,v) > 0\) if \(v\neq 0\). We distinguish \(\left\langle{,}\right\rangle\) as the usual inner product on \(\mathbb{R}^n\).
- Michael Spivak, 1969 (Calculus on Manifolds, Perseus books). Page 77
Function inner() returns the inner product corresponding to a matrix \(M\). Spivak’s definition requires \(M\) to be positive-definite, but that is not necessary in the package. The inner product of two vectors \(\mathbf{x}\) and \(\mathbf{y}\) is usually written \(\left\langle\mathbf{x},\mathbf{y}\right\rangle\) or \(\mathbf{x}\cdot\mathbf{y}\), but the most general form would be \(\mathbf{x}^TM\mathbf{y}\). Noting that inner products are multilinear, that is \(\left\langle\mathbf{x},a\mathbf{y}+b\mathbf{z}\right\rangle=a\left\langle\mathbf{x},\mathbf{y}\right\rangle + b\left\langle\mathbf{x},\mathbf{z}\right\rangle\) and \(\left\langle a\mathbf{x} + b\mathbf{y},\mathbf{z}\right\rangle=a\left\langle\mathbf{x},\mathbf{z}\right\rangle + b\left\langle\mathbf{y},\mathbf{z}\right\rangle\) we see that the inner product is indeed a multilinear map, that is, a tensor.
We can start with the simplest inner product, the identity matrix:
inner(diag(7))## A linear map from V^2 to R with V=R^7:
## val
## 1 1 = 1
## 4 4 = 1
## 5 5 = 1
## 2 2 = 1
## 3 3 = 1
## 6 6 = 1
## 7 7 = 1
Note how the rows of the tensor appear in arbitrary order. Verify:
x <- rnorm(7)
y <- rnorm(7)
V <- cbind(x,y)
LHS <- sum(x*y)
RHS <- as.function(inner(diag(7)))(V)
c(LHS=LHS,RHS=RHS,diff=LHS-RHS)## LHS RHS diff
## 5.503805 5.503805 0.000000
Above, we see agreement between \(\mathbf{x}\cdot\mathbf{y}\) calculated directly [LHS] and using inner() [RHS]. A more stringent test would be to use a general matrix:
M <- matrix(rnorm(49),7,7)
f <- as.function(inner(M))
LHS <- quad.3form(M,x,y)
RHS <- f(V)
c(LHS=LHS,RHS=RHS,diff=LHS-RHS)## LHS RHS diff
## -3.410660e+00 -3.410660e+00 -1.332268e-15
(function emulator::quad.3form() evaluates matrix products efficiently; quad.3form(M,x,y) returns \(x^TMy\)). Above we see agreement, to within numerical precision, of the dot product calculated two different ways: LHS uses quad.3form() and RHS uses inner(). Of course, we would expect inner() to be a homomorphism:
M1 <- matrix(rnorm(49),7,7)
M2 <- matrix(rnorm(49),7,7)
g <- as.function(inner(M1+M2))
LHS <- quad.3form(M1+M2,x,y)
RHS <- g(V)
c(LHS=LHS,RHS=RHS,diff=LHS-RHS)## LHS RHS diff
## -5.418253 -5.418253 0.000000
Above we see numerical verification of the fact that \(I(M_1+M_2)=I(M_1)+I(M_2)\), by evaluation at \(\mathbf{x},\mathbf{y}\), again with LHS using direct matrix algebra and RHS using inner(). Now, if the matrix is symmetric the corresponding inner product should also be symmetric:
h <- as.function(inner(M1 + t(M1))) # send inner() a symmetric matrix
LHS <- h(V)
RHS <- h(V[,2:1])
c(LHS=LHS,RHS=RHS,diff=LHS-RHS)## LHS RHS diff
## -22.52436 -22.52436 0.00000
Above we see that \(\mathbf{x}^TM\mathbf{y} = \mathbf{y}^TM\mathbf{x}\). Further, a positive-definite matrix should return a positive quadratic form:
M3 <- crossprod(matrix(rnorm(56),8,7)) # 7x7 pos-def matrix
as.function(inner(M3))(kronecker(rnorm(7),t(c(1,1))))>0 # should be TRUE## [1] TRUE
Above we see the second line evaluating \(\mathbf{x}^TM\mathbf{x}\) with \(M\) positive-definite, and correctly returning a non-negative value.
The inner product on an antisymmetric matrix should be alternating:
jj <- matrix(rpois(49,lambda=3.2),7,7)
M <- jj-t(jj) # M is antisymmetric
f <- as.function(inner(M))
LHS <- f(V)
RHS <- -f(V[,2:1]) # NB negative as we are checking for an alternating form
c(LHS=LHS,RHS=RHS,diff=LHS-RHS) ## LHS RHS diff
## 19.50013 19.50013 0.00000
Above we see that \(\mathbf{x}^TM\mathbf{y} = -\mathbf{y}^TM\mathbf{x}\) where \(M\) is antisymmetric.