On Tue, 1 Jun 2021 at 10:53, Neil Girdhar firstname.lastname@example.org wrote:
On Tue, Jun 1, 2021 at 5:39 AM Oscar Benjamin email@example.com wrote:
On Tue, 1 Jun 2021 at 05:16, Neil Girdhar firstname.lastname@example.org wrote:
The problem that the original poster was trying to address with additional syntax is the automatic naming of symbols. He wants to omit this line:
x = symbols("x")
You're right that if you have many one-character symbol names, you can use a shortcut, but this benefit is lost if you want descriptive names like:
momentum = symbols('momentum')
He is proposing new syntax to eliminate the repeated name. The function approach specifies each name exactly once. This is one of the benefits of JAX over TensorFLow.
Second, the function approach allows the function to be a single object that can be used in calcuations. You might ask for:
grad(equation, 2)(2, 3, 4 5) # derivative with respect to parameter 2 of equation evaluated at (2, 3, 4, 5)
With the symbolic approach, you need to keep the equation object as well as the symbols that compose it to interact with it.
This makes more sense in a limited context for symbolic manipulation where symbols only represent function parameters so that all symbols are bound. How would you handle the situation where the same symbols are free in two different expressions that you want to manipulate in tandem though?
In this example we have two different equations containing the same symbols and we want to solve them as a system of equations:
p, m, h = symbols('p, m, h') E = p**2 / 2*m lamda = h / p
E1 = 5 lamda1 = 2 [(p1, m1)] = solve([Eq(E, E1), Eq(lamda, lamda1)], [p, m])
I don't see a good way of doing this without keeping track of the symbols as separate objects. I don't think this kind of thing comes up in Jax because it is only designed for the more limited symbolic task of evaluating and differentiating Python functions.
This is a really cool design question.
One of the things I like about JAX is that they stayed extremely close to NumPy's interface. In NumPy, comparison operators applied to matrices return Boolean matrices.
I would ideally express what you wrote as
def E(p, m): ...
def lamda(h, p): ...
def f(p, m): return jnp.all(E(p, m) == E1) and jnp.all(lamda(h, p) == lamda1)
p1, m1 = solve(f)
So how does solve know to solve for p and m rather than h?
Note that I deliberately included a third symbol and made the parameter lists of E and lamda inconsistent.
Should Jax recognise that the 2nd parameter of lamda has the same name as the 1st parameter of E? Or should symbols at the same parameter index be considered the same regardless of their name?
In Jax everything is a function so I would expect it to ignore the symbol names so that if args = solve([f1, f2]) then f1(*args) == f2(*args) == 0.
This is usually how the API works for numerical rather than symbolic root-finding algorithms. But then how do I solve a system of equations that has a symbolic parameter like `h`?