Identify the quantities and the coordinate system
A vector can represent a geometric displacement or an ordered collection of measurements. In two dimensions, write its coordinates as a column and define addition and scaling component by component. Coordinates depend on a chosen basis: the same abstract vector can have different numeric descriptions in different bases. Before multiplying matrices, state what each component measures and which coordinate systems the input and output use.
Read the images of basis vectors from columns
A linear map preserves vector addition and scalar multiplication. If the images of basis vectors e1 and e2 are known, then the image of x times e1 plus y times e2 is x times the first image plus y times the second. These two images form the columns of the matrix. The familiar multiply-and-add procedure is therefore a way to combine transformed basis vectors, rather than an arbitrary arithmetic rule.
Work through one transformation
Let the columns of A be (2, 0) and (1, 3). Applying A to (4, 2) gives four times (2, 0) plus two times (1, 3), which is (10, 6). Notice that the second input component contributes to both output coordinates. A map from two input coordinates to three output coordinates instead needs three rows and two columns. Checking dimensions first makes it easier to recognize invalid products and to interpret the resulting values.
Recognize where linear reasoning stops
A transformation written as A times x plus b is generally affine rather than linear when b is nonzero. Every linear map sends the zero vector to zero, whereas a translation does not. Neural networks combine matrix operations with nonlinear functions because repeated linear maps alone remain linear. Finally, distinguish symbolic algebra from floating-point computation. An invertible matrix can still be numerically troublesome when small input or rounding errors cause large changes in the computed solution.
Test a necessary condition without overclaiming
Sending zero to zero is necessary for linearity, but it is not sufficient. Consider the one-dimensional map T(x)=x squared. It sends zero to zero, yet T(1+1) equals four while T(1)+T(1) equals two. One counterexample disproves additivity. This is a useful check before assigning a matrix to an unfamiliar transformation: return to both defining identities rather than treating a single convenient property as a complete proof of linearity.
Composition fixes the multiplication order
With column vectors, applying A first and B second gives BAv. The rightmost operation acts first. Let A double only the horizontal coordinate, and let B rotate a vector counterclockwise by ninety degrees. Starting at(1,0), A then B gives(0,2). Reversing the order first gives(0,1), which horizontal stretching leaves unchanged. Thus AB and BA generally represent different transformations. This is not a numerical rounding issue: the intended mathematical operations themselves differ. In graphics and robotics, document the coordinate frame and the order of every transformation. A dimensionally valid product can still represent the wrong sequence, so test a simple basis direction rather than relying only on matrix shapes.
Recognize lost information
If two independent input directions collapse onto one output line, the output cannot distinguish every input. For A=[[1,2],[2,4]], the second column is twice the first. Inputs(2,0) and(0,1) both produce(2,4), so a unique inverse is impossible. The determinant is zero, agreeing with the column-dependence argument. A nonsingular square matrix has a mathematical inverse, but the magnitude of its determinant alone is not a reliable measure of numerical conditioning. Scaling, input perturbations and the chosen solver matter. Separate the question of whether an inverse exists from the question of whether a computed reconstruction is accurate enough. For a numerical system, inspect residuals and sensitivity rather than only the existence of an inverse symbol.
Change coordinates without moving the object
Choose basis vectors b1=(1,1) and b2=(1,−1). The vector whose standard coordinates are(4,2) equals 3b1+1b2, giving coordinates(3,1) in the new basis. The object has not moved; only its description changed. If P contains the new basis vectors as columns, standard coordinates equal P times new coordinates. The matrix representing A in that basis is P-inverse times A times P. Follow the three stages: convert the input into standard coordinates, apply A, then convert the output back. Mixing coordinates from different frames can yield a dimensionally valid yet meaningless calculation. Label coordinate systems alongside arrays when implementing transformations, especially when several devices or models use different conventions.
Verify the interpretation with a small program
For each output component, identify which input components contribute to its sum. The program below constructs(10,6) as a linear combination of matrix columns and compares it with the usual matrix-vector multiplication. This is an exact small-integer check, not a performance comparison between numerical libraries. General floating-point inputs require an appropriate tolerance because evaluation order can change rounding. Agreement on several examples also does not prove that an unknown transformation is linear on every input. Distinguish checking an implementation of a known linear map from establishing linearity as a mathematical property. Use basis vectors and a few nontrivial combinations to catch transposition, dimension and multiplication-order mistakes before trusting a more complicated pipeline.
A = ((2, 1), (0, 3))
v = (4, 2)
product = tuple(sum(row[j] * v[j] for j in range(2)) for row in A)
columns = (4 * 2 + 2 * 1, 4 * 0 + 2 * 3)
assert product == columns == (10, 6)
print(product)