Introduction

Generating synthetic data from a linear function is a common step in plotting, testing regression code, and creating toy datasets for machine learning experiments. In NumPy, the basic pattern is simple: generate x values, apply the linear equation y = mx + b, and optionally add noise to simulate measurement error.

Generate Clean Linear Data

A linear function has the form y = mx + b, where m is the slope and b is the intercept.

1import numpy as np
2
3m = 2.5
4b = -1.0
5x = np.linspace(0, 10, 6)
6y = m * x + b
7
8print(x)
9print(y)

np.linspace is a good default because it gives evenly spaced input values across a chosen interval. If you already know the step size you want, np.arange is also fine.

Use Random x Values When Uniform Spacing Is Not Required

If the goal is to simulate scattered observations instead of plotting a neat line, random x values are often better.

1import numpy as np
2
3rng = np.random.default_rng(42)
4x = rng.uniform(0, 10, size=8)
5y = 3.0 * x + 5.0
6
7print(x)
8print(y)

This produces points that still lie on the same linear relationship but do not appear evenly spaced.

Add Noise for More Realistic Data

Real data almost never sits perfectly on a line. Add Gaussian noise to simulate observation error.

1import numpy as np
2
3rng = np.random.default_rng(42)
4x = np.linspace(0, 10, 20)
5noise = rng.normal(loc=0.0, scale=1.5, size=x.shape)
6y = 4.0 * x + 2.0 + noise
7
8print(y[:5])

This keeps the underlying linear trend while making the data look more like a real regression dataset.

Stack the Result into a Dataset Shape

If you want to pass the generated data into another tool, it is often convenient to combine the columns.

1import numpy as np
2
3x = np.linspace(0, 5, 6)
4y = 1.2 * x + 0.7
5
6data = np.column_stack((x, y))
7print(data)

This produces a two-column array where each row is one observation.

Generate Multifeature Linear Targets

For machine learning experiments, you may want several input features and one target generated by a linear combination.

1import numpy as np
2
3rng = np.random.default_rng(7)
4X = rng.normal(size=(100, 3))
5weights = np.array([2.0, -1.0, 0.5])
6bias = 3.0
7noise = rng.normal(scale=0.2, size=100)
8
9y = X @ weights + bias + noise
10
11print(X.shape)
12print(y.shape)

This is a useful pattern for quickly creating regression training data with a known ground-truth relationship.

Plotting and Testing Often Need Different Data Shapes

For plotting, evenly spaced x values from linspace usually make the line easy to inspect visually. For model testing, randomly sampled inputs plus noise are often better because they stress the code in a less artificial way. The same linear function can support both goals, but the generated dataset shape should match the purpose.

Common Pitfalls

• Using Python loops when NumPy vectorization is simpler and faster.
• Choosing np.arange with floating-point steps and then being surprised by endpoint behavior.
• Forgetting that perfectly clean linear data is often too artificial for testing real pipelines.
• Mixing up slope and intercept signs when reading back the generated data.
• Adding noise with the wrong shape and accidentally broadcasting it in an unintended way.

Summary

• Generate x values first, then compute y = mx + b with vectorized NumPy operations.
• Use np.linspace for evenly spaced points and random sampling for scattered observations.
• Add Gaussian noise when you need realistic synthetic data.
• Use np.column_stack or matrix formulas when the data will feed a downstream model.
• Keep the generation logic explicit so the synthetic relationship is easy to reason about later.