Linear algebra shows up in a surprisingly wide range of technologies and sciences you interact with every day. From the 3D graphics on your screen to the GPS coordinates on your phone, systems of equations, matrices, and vectors form the mathematical backbone of modern computing, physics, engineering, and economics. Here’s where it actually matters and what it’s doing behind the scenes.
3D Graphics and Video Games
Every object you see in a video game or animated film is positioned, rotated, and scaled using matrix multiplication. The standard tool is a 4×4 matrix that handles all three transformations (plus perspective) in a single operation. When a character turns, the game engine multiplies every point on that character’s model by a rotation matrix built from trigonometric functions. When the camera zooms in, a scaling matrix stretches the coordinate values along each axis. Translation, the simple act of moving an object from one place to another, is handled by tucking the x, y, and z offsets into the last column of the matrix.
This happens millions of times per frame. Your GPU is essentially a machine optimized for doing massive batches of matrix math in parallel, which is why graphics cards turned out to be so useful for other linear algebra tasks like training AI.
Artificial Intelligence and Machine Learning
A neural network is, at its core, a chain of matrix multiplications with nonlinear functions squeezed in between. Each layer of a network stores its connection strengths in a weight matrix. When data flows through, the network multiplies the input vector by that weight matrix to produce an activation value for each node in the next layer. For a layer with multiple inputs and multiple outputs, this becomes a full matrix-times-matrix operation.
The entire training dataset can be represented as a single large matrix, and processing it through the network means computing the output matrix from the weight and input matrices in one shot. This is why training large AI models demands enormous computational power: the matrices involved can have billions of entries, and every training step requires multiplying, differentiating, and updating them. The linear algebra operations describe how information transforms as it flows from one layer to the next.
Image and Data Compression
When you compress an image, one powerful technique breaks the image’s pixel data into a product of three smaller matrices using something called the singular value decomposition (SVD). The key insight is that each of these component pieces captures a different amount of the image’s visual information, ranked from most important to least. The largest singular values correspond to the broad shapes and contrast in the image, while the smallest ones capture fine grain detail that’s often imperceptible.
By keeping only the top portion of these components and discarding the rest, you can reconstruct an image that looks nearly identical to the original while storing far less data. For a 100-by-200 pixel image with a rank of 20, the original requires storing 20,000 values. The decomposed version needs only about 6,020 entries, roughly a 70% reduction. Streaming services, medical imaging archives, and satellite photography all rely on variations of this approach.
Internet Search and PageRank
Google’s original PageRank algorithm, the system that made its search results better than every competitor’s in the late 1990s, is built on eigenvector computation. The entire web is modeled as a massive matrix where each entry represents a link from one page to another. The PageRank score for every page is determined by finding the principal eigenvector of this matrix, which is the unique vector that, when multiplied by the link matrix, returns a scaled version of itself.
For the real web, that matrix has tens of billions of rows and columns. Computing eigenvectors directly on something that large is impractical even for powerful software, so the algorithm uses an iterative method called the Power Method: start with an initial guess, multiply by the matrix, repeat, and the result converges toward the true eigenvector. Each iteration is a matrix-vector multiplication on a planetary scale.
GPS Navigation
Your phone’s GPS receiver determines its position by solving a system of equations involving four or more satellites. Each satellite broadcasts its own coordinates and a timestamp. The raw equations are quadratic, since they measure the squared distance between you and each satellite. But the standard solving technique subtracts one equation from the others, which cancels out all the squared terms and leaves a linear system.
That linear system has four unknowns: your x, y, and z coordinates plus a clock correction factor that accounts for the tiny timing difference between your receiver and the satellite clocks. With signals from four satellites, you get three independent linear equations, enough to solve for your position. In practice, receivers pick up signals from more than four satellites and use least-squares methods (another linear algebra technique) to get a more accurate fix.
Economics and Supply Chain Modeling
Economists use a matrix framework called the Leontief input-output model to map how different industries depend on each other. Every sector of an economy both produces goods and consumes goods from other sectors. The model captures these relationships in an input-output matrix, where each entry represents how much of one industry’s output is consumed by another.
The central equation is straightforward: total production equals internal consumption plus outside demand. In matrix form, that’s X = AX + D, where X is total output, A is the input-output matrix, and D is consumer demand. Solving for X gives X = (I − A)⁻¹D, which requires computing a matrix inverse. This lets economists answer questions like: if consumer demand for electronics doubles, how much additional steel, plastic, and energy production does the entire supply chain need to support that?
Medical Imaging
CT scanners work by firing X-rays through your body from hundreds of angles and measuring how much radiation passes through at each point. Turning those raw measurements into a cross-sectional image is fundamentally a linear algebra problem. The unknown image is treated as a vector of pixel values, the measurements form another vector, and a large matrix links the two. Reconstructing the image means solving the system g = Ac, where g is the measurement data, c is the image you want, and A describes how each X-ray path interacts with each pixel.
Because the measurement data contains noise, directly inverting this system would amplify errors. Instead, iterative reconstruction methods solve the system step by step, gradually refining the image while controlling noise. The choice of how to model the matrix A, the link between pixel values and measurements, directly affects image quality. Modern CT systems use linear interpolation models that balance accuracy against computational speed, since these reconstructions need to happen fast enough to be clinically useful.
Quantum Mechanics
In quantum physics, the state of a particle is represented as a vector in an abstract space. The notation uses “kets,” written as |ψ⟩, which are essentially column vectors whose entries are complex numbers. Each entry represents the probability amplitude for a particular measurement outcome. The probability of finding a particle in a given state comes from taking the inner product of two of these vectors, mirroring the dot product from basic linear algebra.
Physical quantities like energy, momentum, and spin are represented by matrices called operators. Measuring a property of a quantum system corresponds to multiplying the state vector by the appropriate operator matrix. The possible measurement results are the eigenvalues of that matrix, and the states the system can collapse into are the corresponding eigenvectors. This framework means that predicting the behavior of atoms, molecules, and subatomic particles depends entirely on solving matrix equations.
Cryptography
One classic encryption method, the Hill cipher, turns plaintext into ciphertext using matrix multiplication. Letters are converted to numbers (A=0 through Z=25), grouped into pairs or triplets, and arranged into small vectors. Each vector is multiplied by a secret key matrix, and the results are taken modulo 26 to wrap back into the alphabet range. For example, encrypting the letters H (8) and E (5) with a 2×2 key matrix produces the values 59 and 100, which reduce to 7 and 22 (mod 26), giving the ciphertext letters G and V.
Decryption reverses the process by multiplying the ciphertext vector by the inverse of the key matrix, also modulo 26. This only works if the key matrix has an inverse in modular arithmetic, which constrains which matrices make valid keys. While the Hill cipher itself is too simple for modern security, the principle of using matrix operations for encryption extends into contemporary cryptographic systems where linear algebra over finite fields protects everything from banking transactions to encrypted messaging.