torch.cov

torch.cov(input, *, correction=1, fweights=None, aweights=None) → Tensor

Estimates the covariance matrix of the variables given by the input matrix, where rows are the variables and columns are the observations.

A covariance matrix is a square matrix giving the covariance of each pair of variables. The diagonal contains the variance of each variable (covariance of a variable with itself). By definition, if input represents a single variable (Scalar or 1D) then its variance is returned.

The unbiased sample covariance of the variables $x$ and $y$ is given by:

$$\text{cov}_w(x,y) = \frac{\sum^{N}_{i = 1}(x_{i} - \bar{x})(y_{i} - \bar{y})}{N~-~1}$$

where $\bar{x}$ and $\bar{y}$ are the simple means of the $x$ and $y$ respectively.

If fweights and/or aweights are provided, the unbiased weighted covariance is calculated, which is given by:

$$\text{cov}_w(x,y) = \frac{\sum^{N}_{i = 1}w_i(x_{i} - \mu_x^*)(y_{i} - \mu_y^*)}{\sum^{N}_{i = 1}w_i~-~1}$$

where $w$ denotes fweights or aweights based on whichever is provided, or $w = fweights \times aweights$ if both are provided, and $\mu_x^* = \frac{\sum^{N}_{i = 1}w_ix_{i} }{\sum^{N}_{i = 1}w_i}$ is the weighted mean of the variable.

Parameters:

input (Tensor) – A 2D matrix containing multiple variables and observations, or a Scalar or 1D vector representing a single variable.

Keyword Arguments:

• correction (int, optional) – difference between the sample size and sample degrees of freedom. Defaults to Bessel’s correction, correction = 1 which returns the unbiased estimate, even if both fweights and aweights are specified. correction = 0 will return the simple average. Defaults to 1.

• fweights (tensor, optional) – A Scalar or 1D tensor of observation vector frequencies representing the number of times each observation should be repeated. Its numel must equal the number of columns of input. Must have integral dtype. Ignored if None. Defaults to ``None`.

• aweights (tensor, optional) – A Scalar or 1D array of observation vector weights. These relative weights are typically large for observations considered “important” and smaller for observations considered less “important”. Its numel must equal the number of columns of input. Must have floating point dtype. Ignored if None. Defaults to ``None`.

Returns:

(Tensor) The covariance matrix of the variables.

See also

torch.corrcoef() normalized covariance matrix.

Example::

>>> x = torch.tensor([[0, 2], [1, 1], [2, 0]]).T
>>> x
tensor([[0, 1, 2],
        [2, 1, 0]])
>>> torch.cov(x)
tensor([[ 1., -1.],
        [-1.,  1.]])
>>> torch.cov(x, correction=0)
tensor([[ 0.6667, -0.6667],
        [-0.6667,  0.6667]])
>>> fw = torch.randint(1, 10, (3,))
>>> fw
tensor([1, 6, 9])
>>> aw = torch.rand(3)
>>> aw
tensor([0.4282, 0.0255, 0.4144])
>>> torch.cov(x, fweights=fw, aweights=aw)
tensor([[ 0.4169, -0.4169],
        [-0.4169,  0.4169]])