Projection (linear algebra)

Imagine you cast a shadow on a wall.

This is like a math trick. It makes a shape look flat.

You can do this trick many ways. One way is a straight drop. This is called an orthogonal projection.

Another way is a slanted drop. This is an oblique projection. It makes the shadow look different.

If you do the trick once, you find a spot. If you do it again, you stay in that same spot. This is a special rule for these tricks.

Math uses these tricks to study shapes. It helps us see things in a new way.

Imagine you cast a shadow on a wall.

In math, this is called a projection. It is a way to move things from one space to another. A projection follows a special rule. If you apply it once, you find a spot. If you apply it again, you stay in that same spot. This rule is called idempotence.

There are two main ways to do this. The first is an orthogonal projection. This is like a straight drop. It is the shortest path to a surface. It is used to find the smallest distance between a point and a plane.

The second way is an oblique projection. This is a slanted drop. It does not use the shortest path.

Mathematicians use matrices to show these tricks. A square matrix is a projection matrix if it stays the same when multiplied by itself. These matrices help us study shapes and spaces. They are very useful in fields like machine learning.

Imagine you are holding a flashlight in a dark room. When you shine it on an object, it casts a shadow on the wall. In mathematics, this idea is called a projection. A projection is a special kind of movement called a linear transformation. It moves things from one space to another.

This movement follows one very important rule. If you apply the projection once, you reach a new spot. If you apply it a second time, you stay exactly where you are. Mathematicians call this property idempotence. This means the projection leaves its own image unchanged.

There are two main ways to perform a projection. The first type is called an orthogonal projection. This is like a shadow made by a light shining straight down. It follows the shortest path to a surface. This type of projection is very useful for finding the smallest distance between a point and a plane.

The second type is called an oblique projection. This is like a shadow made by a light shining at an angle. It is a slanted drop rather than a straight one. It does not use the shortest path to reach the surface.

Mathematicians use tools called matrices to describe these movements. A square matrix is called a projection matrix if it is equal to its own square. This is a way of writing the idempotence rule using math symbols. For an orthogonal projection, the matrix has even more special rules. In a real matrix, it must be equal to its transpose. This ensures the movement stays perfectly straight and perpendicular.

These matrices help us understand how different spaces relate to each other.

Every projection can be described by two parts. One part is the range, which is where the points land. The other part is the kernel, which is the set of points that get moved to zero. These two parts are complementary. This means they work together to split a larger space into smaller pieces. In a finite-dimensional space, every vector can be split into one piece from the range and one from the kernel. This helps us see how a complex space is built from simpler parts.

Projections are not just for classroom math problems. They are used in many modern technologies. For example, they are very helpful in the field of machine learning. They are also used in statistics to find the best fit for data. Scientists use orthogonal projections to solve problems involving least squares. This helps them find the most accurate answers when data is messy. By using these mathematical shadows, we can make sense of a complicated world.

In linear algebra, a projection is a specific type of linear transformation called an endomorphism. This means the transformation maps a vector space back onto itself. The defining characteristic of a projection is a property called idempotence. A transformation is idempotent if applying it twice results in the same outcome as applying it once. Mathematically, if we call the transformation $P$, then $P^2 = P$. This means that once a vector has been projected into its new position, any further attempts to project it will leave it unchanged.

To understand how a projection works, we must look at its two fundamental components: the range and the kernel. The range, or image, is the subspace where all the projected vectors land. The kernel, or null space, consists of all the vectors that the transformation moves to zero. These two subspaces are complementary. In a finite-dimensional vector space, this means every single vector can be decomposed uniquely into two parts. One part belongs to the range, and the other belongs to the kernel. This relationship allows mathematicians to split a large, complex space into two distinct, manageable pieces.

Projections are categorized into two main types: orthogonal and oblique. An orthogonal projection is a special case where the range and the kernel are orthogonal to each other. This means every vector in the range is perpendicular to every vector in the kernel. In a Hilbert space, which is a complete vector space with an inner product, an orthogonal projection is also self-adjoint. This means the transformation behaves symmetrically with respect to the inner product.

An oblique projection is any projection that does not satisfy this perpendicular requirement. In an oblique projection, the range and kernel are still complementary, but they meet at an angle rather than at a right angle.

We can represent these transformations using square matrices. A matrix is a projection matrix if it equals its own square. For real matrices, an orthogonal projection matrix must also be equal to its transpose. If the matrix is complex, it must be equal to its adjoint, also known as the Hermitian transpose. The eigenvalues of any projection matrix are restricted to only two possible values: 0 or 1. The eigenspace corresponding to the eigenvalue 1 is the range of the projection. Meanwhile, the eigenspace corresponding to the eigenvalue 0 is the kernel.

There are many ways to calculate an orthogonal projection, depending on the information available. If you are projecting onto a line defined by a unit vector $u$, the operation is a simple outer product of $u$ with itself. For more complex subspaces, you can use an orthonormal basis. If you have a matrix $Q$ whose columns form an orthonormal basis for the subspace, the projection matrix is calculated as $QQ^T$. If the basis is not orthonormal, a normalizing factor is required to recover the correct norm. This is often expressed using the formula $M(M^TM)^{-1}M^T$.

In practical applications, orthogonal projections are vital for finding the shortest distance between a point and a subspace. This is a core concept in the Hilbert projection theorem. This principle is used heavily in the method of least squares. In statistics, ordinary least squares regression relies on orthogonal projections to find fitted values. However, some specialized tasks, like instrumental variables regression, actually require the use of oblique projections. These mathematical tools allow us to approximate complex data by finding the closest possible representation within a simpler subspace.

Projections also interact with other mathematical operations in specific ways. For example, the product of two projections is not always a projection itself. However, if two projections commute, their product is guaranteed to be a projection. If they are both orthogonal projections and they commute, their product is also an orthogonal projection. These rules help mathematicians understand how different subspaces overlap and interact. From machine learning to spherical trigonometry, the ability to project data into different dimensions is a fundamental part of modern science.