is_independent(func, interface, args, kwargs=None, num_pos=5, seed=9123, atol=1e-06, rtol=0, bounds=(- 3.141592653589793, 3.141592653589793))[source]

Test whether a function is independent of its input arguments, both numerically and analytically.

  • func (callable) – Function to be tested

  • interface (str) – Autodiff framework used by func. Must correspond to one of the supported PennyLane interface strings, such as "autograd", "tf", "torch", "jax".

  • args (tuple) – Positional arguments with respect to which to test

  • kwargs (dict) – Keyword arguments for func at which to test; the keyword arguments are kept fixed in this test.

  • num_pos (int) – Number of random positions to test

  • seed (int) – Seed for the random number generator

  • atol (float) – Absolute tolerance for comparing the outputs

  • rtol (float) – Relative tolerance for comparing the outputs

  • bounds (tuple[float]) – 2-tuple containing limits of the range from which to sample


Whether func returns the same output at randomly chosen points and is numerically independent of its arguments.

Return type



This function is experimental. As such, it might yield wrong results and might behave slightly differently in distinct autodifferentiation frameworks for some edge cases. For example, a currently known edge case are piecewise functions that use classical control and simultaneously return (almost) constant output, such as

def func(x):
    if abs(x) <1e-5:
        return x
        return 0. * x

The analytic and numeric tests used are as follows.

  • The analytic test performed depends on the provided interface, both in its method and its degree of reliability.

  • For the numeric test, the function is evaluated at a series of random positions, and the outputs numerically compared to verify that the output is constant.


Currently, no analytic test is available for the PyTorch interface. When using PyTorch, a warning will be raised and only the numeric test is performed.


Due to the structure of is_independent, it is possible that it errs on the side of reporting a dependent function to be independent (a false positive). However, reporting an independent function to be dependent (a false negative) is highly unlikely.


Consider the (linear) function

def lin(x, weights=None):
    return np.dot(x, weights)

This function clearly depends on x. We may check for this via

>>> x = np.array([0.2, 9.1, -3.2], requires_grad=True)
>>> weights = np.array([1.1, -0.7, 1.8], requires_grad=True)
>>> qml.math.is_independent(lin, "autograd", (x,), {"weights": weights})

However, the Jacobian will not depend on x because lin is a linear function:

>>> jac = qml.jacobian(lin)
>>> qml.math.is_independent(jac, "autograd", (x,), {"weights": weights})

Note that a function f = lambda x: 0.0 * x will be counted as dependent on x because it does depend on x functionally, even if the value is constant for all x. This means that is_independent is a stronger test than simply verifying functions have constant output.


Using PennyLane

Release news