`local_add_neg_to_sub` rewrite gives wrong results with negative constants

[ricardoV94]

Reported in https://discourse.pymc.io/t/shape-issue-with-custom-logp-in-densitydist/13608/5?u=ricardov94

import numpy as np

import pytensor
import pytensor.tensor as pt

x = pt.scalar("x")
neg_const = pt.as_tensor(np.full((10,), -0.1))
out = neg_const + x 

pytensor.function([x], out)

<<!! BUG IN FGRAPH.REPLACE OR A LISTENER !!>> <class 'TypeError'> The type of the replacement (Vector(float64, shape=(1,))) must be compatible with the type of the original Variable (Vector(float64, shape=(10,))). local_add_neg_to_sub
ERROR (pytensor.graph.rewriting.basic): Rewrite failure due to: local_add_neg_to_sub
ERROR (pytensor.graph.rewriting.basic): node: Add([-0.1 -0.1 ... -0.1 -0.1], ExpandDims{axis=0}.0)
ERROR (pytensor.graph.rewriting.basic): TRACEBACK:
ERROR (pytensor.graph.rewriting.basic): Traceback (most recent call last):
  File "/home/ricardo/Documents/Projects/pytensor/pytensor/graph/rewriting/basic.py", line 1968, in process_node
    fgraph.replace_all_validate_remove(  # type: ignore
  File "/home/ricardo/Documents/Projects/pytensor/pytensor/graph/features.py", line 626, in replace_all_validate_remove
    chk = fgraph.replace_all_validate(replacements, reason=reason, **kwargs)
  File "/home/ricardo/Documents/Projects/pytensor/pytensor/graph/features.py", line 571, in replace_all_validate
    fgraph.replace(r, new_r, reason=reason, verbose=False, **kwargs)
  File "/home/ricardo/Documents/Projects/pytensor/pytensor/graph/fg.py", line 508, in replace
    self.change_node_input(
  File "/home/ricardo/Documents/Projects/pytensor/pytensor/graph/fg.py", line 419, in change_node_input
    raise TypeError(
TypeError: The type of the replacement (Vector(float64, shape=(1,))) must be compatible with the type of the original Variable (Vector(float64, shape=(10,))).

pytensor/pytensor/tensor/rewriting/math.py

Lines 1897 to 1922 in c52154d

	@register_specialize
	@node_rewriter([add])
	def local_add_neg_to_sub(fgraph, node):
	"""
	-x + y -> y - x
	x + (-y) -> x - y
	"""
	# This rewrite is only registered during specialization, because the
	# `local_neg_to_mul` rewrite modifies the relevant pattern during canonicalization
	# Rewrite is only applicable when there are two inputs to add
	if node.op == add and len(node.inputs) == 2:
	# Look for pattern with either input order
	for first, second in (node.inputs, reversed(node.inputs)):
	if second.owner:
	if second.owner.op == neg:
	pre_neg = second.owner.inputs[0]
	new_out = sub(first, pre_neg)
	return [new_out]
	# Check if it is a negative constant
	const = get_constant(second)
	if const is not None and const < 0:
	new_out = sub(first, np.abs(const))
	return [new_out]

The helper get_constant actually calls get_unique_constant_value under the hood, which in this case returns a single scalar, not the filled homogenous vector. The rewrite itself is also expecting only a scalar constant, otherwise it should be using np.all

[Armavica]

pytensor/pytensor/tensor/rewriting/math.py

Lines 1918 to 1922 in c52154d

	# Check if it is a negative constant
	const = get_constant(second)
	if const is not None and const < 0:
	new_out = sub(first, np.abs(const))
	return [new_out]

Do we actually want to keep this part? If it replaces x + neg_const with x - pos_const, is this actually useful?
Plus, doesn't this prevent things such as -x + -6 from being rewritten as -6 - x (actually saving an op here)?

[ricardoV94]

Good point, it doesn't save us anything with a constant. Only reason would be for canonicalization but this is a specialization rewrite.

Will remove that branch and see if anything was relying on it