Multiplying a matrix $$\mathbf{A} \in \mathbb{R}^{m 	imes n}$$ by a vector $$\mathbf{x} \in \mathbb{R}^n$$ can be interpreted conceptually as applying a transformation that projects vectors from an $$n$$-dimensional space ($$\mathbb{R}^{n}$$) into an $$m$$-dimensional space ($$\mathbb{R}^{m}$$). These transformations are highly versatile; for example, specific square matrices can represent geometric rotations. Furthermore, matrix-vector products are the foundational calculations used to compute the outputs of a neural network layer based on the data received from the preceding layer.

Matrix-Vector Product as a Transformation

In deep learning frameworks, the matrix-vector product of a matrix `A` and a vector `x` can be computed using built-in functions, provided the column dimension of `A` strictly matches the length of `x`. In PyTorch, this is executed using `torch.mv(A, x)` or the Python matrix multiplication operator via `A @ x`. For MXNet, the operation is inferred based on argument shapes using `np.dot(A, x)`. JAX performs this calculation with `jnp.matmul(A, x)`, whereas TensorFlow provides the specialized function `tf.linalg.matvec(A, x)`.

Programmatic Matrix-Vector Product

In linear algebra, a matrix can be interpreted as a mathematical operator that transforms an input vector into an output vector. The spectral norm is a matrix norm that quantifies the maximum scaling effect a matrix can exert on a vector. Conceptually, it is derived by asking how much longer the magnitude of the matrix-vector product $$\mathbf{X} \mathbf{v}$$ could possibly be relative to the magnitude of the original vector $$\mathbf{v}$$.

Spectral Norm

A vector-matrix product is a linear algebra operation involving the multiplication of a vector and a matrix. In the context of the multivariate chain rule, evaluating the gradient of a composite multivariate function requires computing a vector-matrix product to map the gradients of intermediate variables back to the input variables.

Vector-Matrix Product

Multiplication by a matrix $$\mathbf{A} \in \mathbb{R}^{m \times n}$$ can be interpreted as a transformation that projects vectors from an $$n$$-dimensional space ($$\mathbb{R}^{n}$$) into an $$m$$-dimensional space ($$\mathbb{R}^{m}$$). These transformations are highly useful; for instance, specific square matrices can represent rotations. Furthermore, matrix-vector products are fundamental in deep learning, as they describe the core calculation for computing the outputs of a neural network layer based on the previous layer's outputs.

Matrix-Vector Product as Transformation

The operation of multiplying two matrices, $$\mathbf{A}$$ and $$\mathbf{B}$$, can be conceptually understood as executing a series of matrix-vector products. Specifically, evaluating $$\mathbf{AB}$$ is equivalent to performing $$m$$ independent matrix-vector products between the matrix $$\mathbf{A}$$ and each of the $$m$$ column vectors of $$\mathbf{B}$$, and then horizontally stitching these resulting column vectors together to form the final product matrix.

Matrix-Matrix Multiplication as a Collection of Matrix-Vector Products

The matrix-vector product between an $$m 	imes n$$ matrix $$\mathbf{A}$$ and an $$n$$-dimensional vector $$\mathbf{x}$$ yields a column vector of length $$m$$. The $$i^	extrm{th}$$ element of this resulting vector is calculated as the dot product between the $$i^	extrm{th}$$ row vector of the matrix, denoted as $$\mathbf{a}^	op_i$$, and the vector $$\mathbf{x}$$. Mathematically, this operation is defined as:

$$\mathbf{A}\mathbf{x} = \begin{bmatrix} \mathbf{a}^	op_{1} \ \mathbf{a}^	op_{2} \ \vdots \ \mathbf{a}^	op_m \ \end{bmatrix}\mathbf{x} = \begin{bmatrix} \mathbf{a}^	op_{1} \mathbf{x} \ \mathbf{a}^	op_{2} \mathbf{x} \ \vdots\ \mathbf{a}^	op_{m} \mathbf{x}\ \end{bmatrix}$$

Claude

A matrix $$\mathbf{A} \in \mathbb{R}^{m 	imes n}$$ can be visualized as a vertical stack of $$m$$ row vectors. Formally, it is expressed as:

$$\mathbf{A}= \begin{bmatrix} \mathbf{a}^	op_{1} \ \mathbf{a}^	op_{2} \ \vdots \ \mathbf{a}^	op_m \ \end{bmatrix}$$

where each $$\mathbf{a}^	op_{i} \in \mathbb{R}^n$$ is a row vector representing the $$i^	extrm{th}$$ row of the matrix $$\mathbf{A}$$.

Learn Before

Related

Learn After