In Set Theory

In set theory, the projection function serves as a fundamental surjective mapping that extracts a specific component from elements of a Cartesian product. For nonempty sets AAA and BBB, the Cartesian product A×BA \times BA×B is the set of all ordered pairs (a,b)(a, b)(a,b) where a∈Aa \in Aa∈A and b∈Bb \in Bb∈B. The first projection π1:A×B→A\pi_1: A \times B \to Aπ1​:A×B→A is defined by π1(a,b)=a\pi_1(a, b) = aπ1​(a,b)=a, while the second projection π2:A×B→B\pi_2: A \times B \to Bπ2​:A×B→B is defined by π2(a,b)=b\pi_2(a, b) = bπ2​(a,b)=b.[7]

This concept generalizes to finite products of sets. Given sets A1,A2,…,AnA_1, A_2, \dots, A_nA1​,A2​,…,An​, the iii-th projection πi:A1×⋯×An→Ai\pi_i: A_1 \times \cdots \times A_n \to A_iπi​:A1​×⋯×An​→Ai​ maps an nnn-tuple (a1,…,an)(a_1, \dots, a_n)(a1​,…,an​) to its iii-th component aia_iai​. For instance, in the Cartesian product R×R=R2\mathbb{R} \times \mathbb{R} = \mathbb{R}^2R×R=R2, the projection πx:R2→R\pi_x: \mathbb{R}^2 \to \mathbb{R}πx​:R2→R is given by πx(x,y)=x\pi_x(x, y) = xπx​(x,y)=x, which isolates the first coordinate along the x-axis.[7]

Projections are inherently surjective. To verify this for π1:A×B→A\pi_1: A \times B \to Aπ1​:A×B→A, consider any a∈Aa \in Aa∈A; assuming BBB is nonempty, select any b∈Bb \in Bb∈B, so (a,b)∈A×B(a, b) \in A \times B(a,b)∈A×B and π1(a,b)=a\pi_1(a, b) = aπ1​(a,b)=a, ensuring every element of AAA is attained.[7]

These projections are essential for conceptualizing and working with Cartesian products and ordered tuples, as they enable the precise identification and isolation of individual components within composite structures, forming the basis for coordinate-based definitions in set theory. This pure set-theoretic projection extends as a foundational idea to idempotent endofunctions in algebraic contexts.[7]

As Idempotent Mappings

In algebraic structures, a projection is abstractly defined as an idempotent endomorphism p:X→Xp: X \to Xp:X→X, meaning p∘p=pp \circ p = pp∘p=p. This condition ensures that applying the map twice yields the same result as applying it once, capturing the intuitive notion of "projecting" elements onto a subspace or substructure without further alteration.[8] Such projections arise naturally in categories like sets, groups, rings, and modules; the terminology draws from the component-selection projections in set theory, where the focus is on idempotent mappings onto substructures.[9]

Examples abound in ring and module theory. For a ring RRR, an idempotent element e∈Re \in Re∈R satisfies e2=ee^2 = ee2=e and induces a projection onto the principal left ideal ReReRe, splitting R=Re⊕R(1−e)R = Re \oplus R(1 - e)R=Re⊕R(1−e) as left RRR-modules.[10] Similarly, in the category of modules over RRR, an idempotent endomorphism p:M→Mp: M \to Mp:M→M projects onto a direct summand submodule im⁡(p)\operatorname{im}(p)im(p), with M=im⁡(p)⊕ker⁡(p)M = \operatorname{im}(p) \oplus \ker(p)M=im(p)⊕ker(p), where the decomposition is internal and respects the module structure. These constructions are fundamental for decomposing complex algebraic objects into simpler components.

The image of a projection ppp coincides exactly with its fixed-point set {x∈X∣p(x)=x}{ x \in X \mid p(x) = x }{x∈X∣p(x)=x}. Indeed, for any y∈im⁡(p)y \in \operatorname{im}(p)y∈im(p), there exists x∈Xx \in Xx∈X such that y=p(x)y = p(x)y=p(x), so p(y)=p(p(x))=p(x)=yp(y) = p(p(x)) = p(x) = yp(y)=p(p(x))=p(x)=y, placing yyy among the fixed points; conversely, if p(x)=xp(x) = xp(x)=x, then x=p(x)∈im⁡(p)x = p(x) \in \operatorname{im}(p)x=p(x)∈im(p).[8] In additive settings like modules, the kernel ker⁡(p)={x∈M∣p(x)=0}\ker(p) = { x \in M \mid p(x) = 0 }ker(p)={x∈M∣p(x)=0} complements the image as a direct summand, ensuring every element decomposes uniquely as a fixed point plus a kernel element.[11]

From an algebraic perspective, projections relate to retractions in categorical terms: a retraction r:X→Ar: X \to Ar:X→A for an inclusion i:A→Xi: A \to Xi:A→X satisfies r∘i=id⁡Ar \circ i = \operatorname{id}_Ar∘i=idA​, and the composite i∘ri \circ ri∘r is an idempotent endomorphism on XXX projecting onto AAA. This view bridges algebra to topology, where continuous retractions analogously yield idempotent maps.[12][13]

Orthogonal Projections

In finite-dimensional inner product spaces, an orthogonal projection of a vector vvv onto a subspace WWW is defined as the unique vector w∈Ww \in Ww∈W that minimizes the Euclidean distance ∥v−w∥|v - w|∥v−w∥, with the additional property that the error vector v−wv - wv−w is perpendicular to every vector in WWW.[14] This perpendicularity condition, expressed as ⟨v−w,u⟩=0\langle v - w, u \rangle = 0⟨v−w,u⟩=0 for all u∈Wu \in Wu∈W, ensures that the projection aligns with the geometry induced by the inner product.[15]

The uniqueness of this projection follows from the Pythagorean theorem in inner product spaces: if w1w_1w1​ and w2w_2w2​ are two such projections, then ∥v−w1∥2=∥v−w2∥2|v - w_1|^2 = |v - w_2|^2∥v−w1​∥2=∥v−w2​∥2 implies ∥w1−w2∥2=0|w_1 - w_2|^2 = 0∥w1​−w2​∥2=0, since the difference lies in WWW and the error terms are orthogonal to WWW.[16] Thus, w1=w2w_1 = w_2w1​=w2​, confirming a single closest point in WWW.[17]

A concrete example occurs when projecting onto a line in R2\mathbb{R}^2R2. For a nonzero vector uuu spanning the line and any v∈R2v \in \mathbb{R}^2v∈R2, the orthogonal projection is given by proj⁡uv=v⋅u∥u∥2u\operatorname{proj}_u v = \frac{v \cdot u}{|u|^2} uproju​v=∥u∥2v⋅u​u.[18] This scalar multiple of uuu lies on the line and satisfies the perpendicularity condition, as the dot product of the error v−proj⁡uvv - \operatorname{proj}_u vv−proju​v with uuu vanishes.

The orthogonal projection induces a linear operator PPP on the space such that P(v)=wP(v) = wP(v)=w, with key properties including idempotence P2=PP^2 = PP2=P—making PPP an idempotent mapping—and orthogonality ⟨Pv,z⟩=0\langle P v, z \rangle = 0⟨Pv,z⟩=0 for all z∈W⊥z \in W^\perpz∈W⊥, the orthogonal complement of WWW.[2] These ensure PPP preserves WWW and annihilates W⊥W^\perpW⊥.[1]

Oblique Projections

In linear algebra, an oblique projection is a linear operator PPP on a vector space VVV that is idempotent, satisfying P2=PP^2 = PP2=P, and maps VVV onto a subspace WWW along a complementary subspace UUU, such that V=W⊕UV = W \oplus UV=W⊕U with W∩U={0}W \cap U = {0}W∩U={0}.[19] Unlike orthogonal projections, the subspaces WWW and UUU are not required to be orthogonal complements, meaning the direction of projection is generally not perpendicular to WWW.[20] For any vector v∈Vv \in Vv∈V, there exists a unique decomposition v=w+uv = w + uv=w+u with w∈Ww \in Ww∈W and u∈Uu \in Uu∈U, and Pv=wP v = wPv=w.[19] The operator PPP is determined by the choice of UUU, and different choices of complementary subspace yield different oblique projections onto the same WWW.[21]

To compute an oblique projection matrix in a finite-dimensional space with respect to a basis, let the columns of matrix AAA form a basis for WWW and the columns of BBB form a basis for UUU, where the combined matrix M=[A B]M = [A , B]M=[AB] is invertible since WWW and UUU are complementary. The projection matrix PPP is then given by P=A[I 0]M−1P = A [I , 0] M^{-1}P=A[I0]M−1, where III is the identity matrix of appropriate dimension and the zero block matches the dimensions of UUU.[22] (Note that this construction relies on the direct sum decomposition and does not assume an inner product.) In contrast, the familiar formula P=A(ATA)−1ATP = A (A^T A)^{-1} A^TP=A(ATA)−1AT for the projection onto the column space of AAA holds only in the special case where UUU is the orthogonal complement of WWW with respect to the standard inner product, reducing to an orthogonal projection.[20]

A concrete example in R2\mathbb{R}^2R2 is the shearing projection onto the x-axis (W=span⁡{(1,0)}W = \operatorname{span}{(1,0)}W=span{(1,0)}) along the slanted line U=span⁡{(1,1)}U = \operatorname{span}{(1,1)}U=span{(1,1)}. Here, A=(10)A = \begin{pmatrix} 1 \ 0 \end{pmatrix}A=(10​) and B=(11)B = \begin{pmatrix} 1 \ 1 \end{pmatrix}B=(11​), so M=(1101)M = \begin{pmatrix} 1 & 1 \ 0 & 1 \end{pmatrix}M=(10​11​) with M−1=(1−101)M^{-1} = \begin{pmatrix} 1 & -1 \ 0 & 1 \end{pmatrix}M−1=(10​−11​). The projection matrix is P=A[1 0]M−1=(1000)(1−101)=(1−100)P = A [1 , 0] M^{-1} = \begin{pmatrix} 1 & 0 \ 0 & 0 \end{pmatrix} \begin{pmatrix} 1 & -1 \ 0 & 1 \end{pmatrix} = \begin{pmatrix} 1 & -1 \ 0 & 0 \end{pmatrix}P=A[10]M−1=(10​00​)(10​−11​)=(10​−10​). For a vector v=(xy)v = \begin{pmatrix} x \ y \end{pmatrix}v=(xy​), Pv=(x−y0)P v = \begin{pmatrix} x - y \ 0 \end{pmatrix}Pv=(x−y0​), which lies on the x-axis, and the difference v−Pv=(yy)v - P v = \begin{pmatrix} y \ y \end{pmatrix}v−Pv=(yy​) lies along UUU.[19] More generally, such projections can be represented in block form as P=(IC00)P = \begin{pmatrix} I & C \ 0 & 0 \end{pmatrix}P=(I0​C0​) relative to bases adapted to W⊕UW \oplus UW⊕U, where CCC encodes the "slant" determined by UUU.[21]

Oblique projections do not minimize the Euclidean distance from vvv to WWW unless U=W⊥U = W^\perpU=W⊥, distinguishing them from orthogonal projections, which achieve the shortest distance in normed spaces.[20] This property arises because the decomposition prioritizes the direct sum structure over perpendicularity, leading to applications in contexts requiring specific directional constraints rather than metric optimization.[19]

Idempotence and Decomposition

In linear algebra, a projection P:V→VP: V \to VP:V→V on a vector space VVV is characterized by its idempotence, meaning P2=PP^2 = PP2=P. This property ensures that applying PPP twice yields the same result as applying it once, distinguishing projections from other linear maps. Idempotence implies that PPP acts as the identity on its image and as the zero map on its kernel, facilitating a canonical decomposition of the space.[23][24]

To see this decomposition explicitly, consider any vector v∈Vv \in Vv∈V. Write v=(v−P(v))+P(v)v = (v - P(v)) + P(v)v=(v−P(v))+P(v), where P(v)∈im⁡(P)P(v) \in \operatorname{im}(P)P(v)∈im(P) and v−P(v)∈ker⁡(P)v - P(v) \in \ker(P)v−P(v)∈ker(P) since P(v−P(v))=P(v)−P2(v)=P(v)−P(v)=0P(v - P(v)) = P(v) - P^2(v) = P(v) - P(v) = 0P(v−P(v))=P(v)−P2(v)=P(v)−P(v)=0. Thus, V=im⁡(P)+ker⁡(P)V = \operatorname{im}(P) + \ker(P)V=im(P)+ker(P). Moreover, im⁡(P)∩ker⁡(P)={0}\operatorname{im}(P) \cap \ker(P) = {0}im(P)∩ker(P)={0}, because if w∈im⁡(P)∩ker⁡(P)w \in \operatorname{im}(P) \cap \ker(P)w∈im(P)∩ker(P), then w=P(u)w = P(u)w=P(u) for some u∈Vu \in Vu∈V and P(w)=0P(w) = 0P(w)=0, so P2(u)=0P^2(u) = 0P2(u)=0 implies P(u)=0P(u) = 0P(u)=0, hence w=0w = 0w=0. Therefore, V=im⁡(P)⊕ker⁡(P)V = \operatorname{im}(P) \oplus \ker(P)V=im(P)⊕ker(P). On im⁡(P)\operatorname{im}(P)im(P), PPP acts as the identity: for w=P(u)w = P(u)w=P(u), P(w)=P2(u)=P(u)=wP(w) = P^2(u) = P(u) = wP(w)=P2(u)=P(u)=w. On ker⁡(P)\ker(P)ker(P), PPP acts as zero by definition.[23][25]

Conversely, any idempotent linear map PPP is a projection onto im⁡(P)\operatorname{im}(P)im(P) along ker⁡(P)\ker(P)ker(P), meaning it splits VVV into this direct sum where vectors in the image are fixed and those in the kernel are annihilated. This equivalence holds in finite-dimensional spaces and extends to the infinite-dimensional case under suitable conditions. Oblique projections provide concrete examples of such decompositions where the summands are not orthogonal.[24][25]

In matrix terms, if PPP is represented by a matrix in a basis adapted to the decomposition V=im⁡(P)⊕ker⁡(P)V = \operatorname{im}(P) \oplus \ker(P)V=im(P)⊕ker(P), then PPP takes the block diagonal form (Ir000n−r)\begin{pmatrix} I_r & 0 \ 0 & 0_{n-r} \end{pmatrix}(Ir​0​00n−r​​), where r=dim⁡(im⁡(P))r = \dim(\operatorname{im}(P))r=dim(im(P)) and n=dim⁡(V)n = \dim(V)n=dim(V). The eigenvalues of PPP are thus 1 with multiplicity rrr and 0 with multiplicity n−rn - rn−r. Since the minimal polynomial of PPP divides x(x−1)x(x-1)x(x−1), which has distinct roots, PPP is diagonalizable, and its Jordan canonical form is diagonal with blocks of 1's and 0's.[26] For instance, the matrix (1100)\begin{pmatrix} 1 & 1 \ 0 & 0 \end{pmatrix}(10​10​) is idempotent, projecting R2\mathbb{R}^2R2 onto the x-axis along the line y=−xy = -xy=−x, with eigenvalues 1 and 0.[27]

Orthogonal Projection Theorem

In a Hilbert space HHH, equipped with an inner product ⟨⋅,⋅⟩\langle \cdot, \cdot \rangle⟨⋅,⋅⟩, consider a closed subspace W⊆HW \subseteq HW⊆H. For every vector v∈Hv \in Hv∈H, there exists a unique vector p=\projWv∈Wp = \proj_W v \in Wp=\projW​v∈W such that v−pv - pv−p is orthogonal to WWW, meaning ⟨v−p,w⟩=0\langle v - p, w \rangle = 0⟨v−p,w⟩=0 for all w∈Ww \in Ww∈W. This ppp is called the orthogonal projection of vvv onto WWW, and it satisfies ∥v−p∥=inf⁡{∥v−w∥:w∈W}| v - p | = \inf { | v - w | : w \in W }∥v−p∥=inf{∥v−w∥:w∈W}.[28]

The proof relies on the Riesz representation theorem and the completeness of HHH. Since WWW is a closed subspace, it is itself a Hilbert space. The map ϕ:W→C\phi: W \to \mathbb{C}ϕ:W→C (or R\mathbb{R}R) defined by ϕ(w)=⟨v,w⟩\phi(w) = \langle v, w \rangleϕ(w)=⟨v,w⟩ is a bounded linear functional on WWW, with ∥ϕ∥≤∥v∥| \phi | \leq | v |∥ϕ∥≤∥v∥. By the Riesz representation theorem, there exists a unique p∈Wp \in Wp∈W such that ϕ(w)=⟨p,w⟩\phi(w) = \langle p, w \rangleϕ(w)=⟨p,w⟩ for all w∈Ww \in Ww∈W. This implies ⟨v−p,w⟩=0\langle v - p, w \rangle = 0⟨v−p,w⟩=0 for all w∈Ww \in Ww∈W, establishing orthogonality. The uniqueness of ppp follows directly from the uniqueness of the representing element in the Riesz theorem. The minimizing property holds because, for any w∈Ww \in Ww∈W,

with equality at t=0t = 0t=0.[28]

A key corollary is that HHH decomposes as the orthogonal direct sum H=W⊕W⊥H = W \oplus W^\perpH=W⊕W⊥, where W⊥={u∈H:⟨u,w⟩=0 ∀w∈W}W^\perp = { u \in H : \langle u, w \rangle = 0 \ \forall w \in W }W⊥={u∈H:⟨u,w⟩=0 ∀w∈W} is the orthogonal complement of WWW. Specifically, every v∈Hv \in Hv∈H can be uniquely written as v=p+(v−p)v = p + (v - p)v=p+(v−p) with p∈Wp \in Wp∈W and v−p∈W⊥v - p \in W^\perpv−p∈W⊥, and W∩W⊥={0}W \cap W^\perp = { 0 }W∩W⊥={0} follows from the positive definiteness of the inner product. This decomposition extends the finite-dimensional case to infinite dimensions.[28]

The Riesz representation theorem, underpinning this result, was independently established by Maurice Fréchet and Frigyes Riesz in 1907, with Erhard Schmidt providing an early explicit proof of the projection theorem in the context of ℓ2\ell^2ℓ2 spaces in 1908.[29][30]

In Geometry

In geometry, orthogonal projection represents the process of dropping a perpendicular from a point to a line, plane, or subspace, yielding the foot of that perpendicular as the projected point. This concept is fundamental in Euclidean spaces, where it minimizes the distance between the original point and its image on the target subspace. For instance, in the Euclidean plane, orthogonal projection onto a line computes heights or distances by constructing right angles, essential for measuring altitudes in triangles or resolving vectors into components along axes.[31]

Projections in geometry extend beyond orthogonality to include parallel and perspective types, each offering distinct spatial interpretations. Parallel projections, which are affine transformations, map points using rays perpendicular or at a fixed angle to the projection plane, preserving parallelism of lines without convergence. In contrast, perspective projections simulate human vision by directing rays from a fixed viewpoint (center of projection) through the object to the image plane, causing parallel lines to converge at vanishing points on the horizon line. Parallel projections emerge as the limiting case of perspective projections when the center of projection recedes to infinity, transitioning from projective to affine geometry.[32][3]

A key application of parallel projection is orthographic projection in technical drawing, where three-dimensional objects are represented in multiple two-dimensional views (e.g., front, top, side) by projecting orthogonally onto perpendicular planes, ensuring accurate scale without distortion from depth. The formula for the orthogonal projection of a vector v\mathbf{v}v onto a plane with unit normal vector n\mathbf{n}n is given by

which subtracts the component along the normal to yield the in-plane vector. This method facilitates precise engineering visualizations by eliminating perspective foreshortening.[33]

In crystallography, projections visualize periodic lattice structures by mapping three-dimensional crystal arrangements onto two-dimensional planes, aiding in the identification of symmetry and unit cells. For cubic crystals, standard projections along principal axes, such as the direction, reveal square or rectangular lattices, while clinographic views highlight atomic packing without overlap. These techniques, often orthographic, are crucial for analyzing diffraction patterns and material properties.[34][35]

Oblique projections, involving non-perpendicular rays, model slanted shadows in geometric scenes but lack the distance-minimizing property of orthogonal ones.[3]

In Statistics and Optimization

In statistics, projections are fundamental to methods that approximate data by minimizing errors in a least squares sense, particularly in linear regression where the goal is to find the best linear fit to observed data points. The ordinary least squares (OLS) estimator projects the response vector yyy onto the column space of the design matrix XXX, yielding the fitted values y^=Py\hat{y} = P yy^​=Py, where P=X(XTX)−1XTP = X (X^T X)^{-1} X^TP=X(XTX)−1XT is the projection matrix. This projection minimizes the Euclidean distance ∥y−y^∥22| y - \hat{y} |_2^2∥y−y^​∥22​, ensuring the residuals are orthogonal to the column space of XXX. The coefficient vector is given by β^=(XTX)−1XTy,\hat{\beta} = (X^T X)^{-1} X^T y,β^​=(XTX)−1XTy, assuming XTXX^T XXTX is invertible.[36]

This approach traces back to the method of least squares, originally developed by Adrien-Marie Legendre in 1805 for determining comet orbits by minimizing squared deviations from predicted positions, and independently by Carl Friedrich Gauss around 1795, though published later. In modern linear algebra terms, the geometric interpretation as an orthogonal projection onto a subspace became prominent in the 20th century, aligning with Hilbert space theory.[37][38]

In optimization, projections play a key role in handling constraints, especially in convex programming where one projects points onto feasible sets to enforce constraints iteratively. A classic example is the alternating projections method, which finds a point in the intersection of two closed convex sets by successively projecting onto each set; for subspaces, it converges to the projection onto their intersection. This method, analyzed by John von Neumann in 1933 for operator theory in Hilbert spaces, underpins algorithms like the convex feasibility problem solvers.

Principal component analysis (PCA) exemplifies projections in data reduction, involving successive orthogonal projections onto the principal component subspaces to capture maximum variance. Introduced by Karl Pearson in 1901 as the "principal axes" of a point cloud minimizing squared distances, PCA projects data onto a lower-dimensional space spanned by eigenvectors of the covariance matrix, ordered by eigenvalues. This technique, further developed by Harold Hotelling in 1933 for multivariate analysis, is widely used for dimensionality reduction while preserving essential structure.[39]

In Functional Analysis

In functional analysis, projections are generalized to bounded linear operators on normed linear spaces that satisfy the idempotence condition P2=PP^2 = PP2=P. Such operators are called bounded projections or continuous idempotents, and they map the space onto their range while fixing points in that range. For a non-zero bounded projection PPP, the operator norm satisfies ∥P∥≥1|P| \geq 1∥P∥≥1, since for any xxx in the range of PPP, Px=xPx = xPx=x implies ∥x∥≤∥P∥∥x∥|x| \leq |P| |x|∥x∥≤∥P∥∥x∥. In Hilbert spaces, the orthogonal projection theorem ensures that every closed subspace admits an orthogonal projection, which is self-adjoint and has norm exactly 1.

The spectral theorem for self-adjoint operators on a Hilbert space provides a key application of projections, associating to each bounded self-adjoint operator AAA a resolution of the identity, which is a projection-valued measure EEE. The spectral projections E(Δ)E(\Delta)E(Δ) for Borel sets Δ⊆R\Delta \subseteq \mathbb{R}Δ⊆R are orthogonal projections onto the subspaces corresponding to the spectral support in Δ\DeltaΔ, satisfying A=∫σ(A)λ dE(λ)A = \int_{\sigma(A)} \lambda , dE(\lambda)A=∫σ(A)​λdE(λ).[40] These projections commute with AAA and decompose the space according to the spectrum.

A concrete example arises in the space L2([0,2π])L^2([0, 2\pi])L2([0,2π]), where the partial sum operators of the Fourier series define orthogonal projections onto the finite-dimensional subspaces of trigonometric polynomials of degree at most nnn. Specifically, the operator Snf=∑k=−nnf^(k)eikxS_n f = \sum_{k=-n}^n \hat{f}(k) e^{ikx}Sn​f=∑k=−nn​f^​(k)eikx projects fff onto span⁡{eikx:∣k∣≤n}\operatorname{span}{e^{ikx} : |k| \leq n}span{eikx:∣k∣≤n}, preserving the L2L^2L2 norm on that subspace.[41]

Unlike Hilbert spaces, where every closed subspace is complemented by a bounded (orthogonal) projection, Banach spaces in general do not possess this property. A closed subspace YYY of a Banach space XXX is complemented if there exists a bounded projection from XXX onto YYY; however, non-complemented subspaces exist in many Banach spaces. For instance, Lindenstrauss and Tzafriri proved that a Banach space in which every closed subspace is complemented must be isomorphic to a Hilbert space.[42] Dvoretzky's theorem implies the existence of finite-dimensional subspaces nearly isometric to Euclidean spaces in every infinite-dimensional Banach space, highlighting that such "Hilbert-like" structures do not guarantee complementation in non-Hilbert settings, as non-Hilbert spaces necessarily contain uncomplemented closed subspaces.

In Category Theory

In category theory, projections arise prominently in the context of products, which are a type of limit. Given a family of objects {Xi}i∈I{X_i}{i \in I}{Xi​}i∈I​ in a category C\mathcal{C}C, their product ∏i∈IXi\prod{i \in I} X_i∏i∈I​Xi​ is an object equipped with projection morphisms πi:∏i∈IXi→Xi\pi_i: \prod_{i \in I} X_i \to X_iπi​:∏i∈I​Xi​→Xi​ for each i∈Ii \in Ii∈I. These projections satisfy the universal property: for any object QQQ and family of morphisms fi:Q→Xif_i: Q \to X_ifi​:Q→Xi​, there exists a unique morphism f:Q→∏i∈IXif: Q \to \prod_{i \in I} X_if:Q→∏i∈I​Xi​ such that the diagrams commute, i.e., πi∘f=fi\pi_i \circ f = f_iπi​∘f=fi​ for all iii. This ensures that the product cone (∏i∈IXi,{πi}i∈I)(\prod_{i \in I} X_i, {\pi_i}{i \in I})(∏i∈I​Xi​,{πi​}i∈I​) is universal among all cones over the diagram {Xi}i∈I{X_i}{i \in I}{Xi​}i∈I​.

Projections also connect to the notion of idempotents in categories. An idempotent morphism e:X→Xe: X \to Xe:X→X satisfies e∘e=ee \circ e = ee∘e=e, generalizing projectors from linear algebra. A split idempotent is one that factors as a retraction-section pair: there exist morphisms r:X→Yr: X \to Yr:X→Y and s:Y→Xs: Y \to Xs:Y→X such that e=r∘se = r \circ se=r∘s, s∘r=idYs \circ r = \mathrm{id}_Ys∘r=idY​, and r∘s=er \circ s = er∘s=e, where YYY is the image of eee. In this case, sss is a section and rrr a retraction, allowing a direct sum decomposition X≅Y⊕ZX \cong Y \oplus ZX≅Y⊕Z in categories supporting such splits, like abelian categories.[8] Categories where all idempotents split are called idempotent complete.

In abelian categories, projections relate closely to split short exact sequences. Consider a short exact sequence 0→K→iE→pQ→00 \to K \xrightarrow{i} E \xrightarrow{p} Q \to 00→Ki​Ep​Q→0; it splits if there exists a morphism s:Q→Es: Q \to Es:Q→E such that p∘s=idQp \circ s = \mathrm{id}_Qp∘s=idQ​, making ppp a retraction and sss a section. This yields an isomorphism E≅K⊕QE \cong K \oplus QE≅K⊕Q, with ppp serving as the projection onto QQQ along KKK. Such splits correspond to idempotent endomorphisms on EEE, where the projection ppp embodies the decomposition.

More broadly, projections form the components of product cones in the limit-colimit duality of categories. While products are limits (universal cones into the diagram), coproducts are colimits (universal cones out of the diagram), and in categories with both, like the category of sets, the projections πi\pi_iπi​ contrast with the coproduct inclusions. This relational structure underscores projections' role in universal constructions across categorical limits.

In Set Theory

In set theory, the projection function serves as a fundamental surjective mapping that extracts a specific component from elements of a Cartesian product. For nonempty sets AAA and BBB, the Cartesian product A×BA \times BA×B is the set of all ordered pairs (a,b)(a, b)(a,b) where a∈Aa \in Aa∈A and b∈Bb \in Bb∈B. The first projection π1:A×B→A\pi_1: A \times B \to Aπ1​:A×B→A is defined by π1(a,b)=a\pi_1(a, b) = aπ1​(a,b)=a, while the second projection π2:A×B→B\pi_2: A \times B \to Bπ2​:A×B→B is defined by π2(a,b)=b\pi_2(a, b) = bπ2​(a,b)=b.[7]

This concept generalizes to finite products of sets. Given sets A1,A2,…,AnA_1, A_2, \dots, A_nA1​,A2​,…,An​, the iii-th projection πi:A1×⋯×An→Ai\pi_i: A_1 \times \cdots \times A_n \to A_iπi​:A1​×⋯×An​→Ai​ maps an nnn-tuple (a1,…,an)(a_1, \dots, a_n)(a1​,…,an​) to its iii-th component aia_iai​. For instance, in the Cartesian product R×R=R2\mathbb{R} \times \mathbb{R} = \mathbb{R}^2R×R=R2, the projection πx:R2→R\pi_x: \mathbb{R}^2 \to \mathbb{R}πx​:R2→R is given by πx(x,y)=x\pi_x(x, y) = xπx​(x,y)=x, which isolates the first coordinate along the x-axis.[7]

Projections are inherently surjective. To verify this for π1:A×B→A\pi_1: A \times B \to Aπ1​:A×B→A, consider any a∈Aa \in Aa∈A; assuming BBB is nonempty, select any b∈Bb \in Bb∈B, so (a,b)∈A×B(a, b) \in A \times B(a,b)∈A×B and π1(a,b)=a\pi_1(a, b) = aπ1​(a,b)=a, ensuring every element of AAA is attained.[7]

These projections are essential for conceptualizing and working with Cartesian products and ordered tuples, as they enable the precise identification and isolation of individual components within composite structures, forming the basis for coordinate-based definitions in set theory. This pure set-theoretic projection extends as a foundational idea to idempotent endofunctions in algebraic contexts.[7]

As Idempotent Mappings

In algebraic structures, a projection is abstractly defined as an idempotent endomorphism p:X→Xp: X \to Xp:X→X, meaning p∘p=pp \circ p = pp∘p=p. This condition ensures that applying the map twice yields the same result as applying it once, capturing the intuitive notion of "projecting" elements onto a subspace or substructure without further alteration.[8] Such projections arise naturally in categories like sets, groups, rings, and modules; the terminology draws from the component-selection projections in set theory, where the focus is on idempotent mappings onto substructures.[9]

Examples abound in ring and module theory. For a ring RRR, an idempotent element e∈Re \in Re∈R satisfies e2=ee^2 = ee2=e and induces a projection onto the principal left ideal ReReRe, splitting R=Re⊕R(1−e)R = Re \oplus R(1 - e)R=Re⊕R(1−e) as left RRR-modules.[10] Similarly, in the category of modules over RRR, an idempotent endomorphism p:M→Mp: M \to Mp:M→M projects onto a direct summand submodule im⁡(p)\operatorname{im}(p)im(p), with M=im⁡(p)⊕ker⁡(p)M = \operatorname{im}(p) \oplus \ker(p)M=im(p)⊕ker(p), where the decomposition is internal and respects the module structure. These constructions are fundamental for decomposing complex algebraic objects into simpler components.

The image of a projection ppp coincides exactly with its fixed-point set {x∈X∣p(x)=x}{ x \in X \mid p(x) = x }{x∈X∣p(x)=x}. Indeed, for any y∈im⁡(p)y \in \operatorname{im}(p)y∈im(p), there exists x∈Xx \in Xx∈X such that y=p(x)y = p(x)y=p(x), so p(y)=p(p(x))=p(x)=yp(y) = p(p(x)) = p(x) = yp(y)=p(p(x))=p(x)=y, placing yyy among the fixed points; conversely, if p(x)=xp(x) = xp(x)=x, then x=p(x)∈im⁡(p)x = p(x) \in \operatorname{im}(p)x=p(x)∈im(p).[8] In additive settings like modules, the kernel ker⁡(p)={x∈M∣p(x)=0}\ker(p) = { x \in M \mid p(x) = 0 }ker(p)={x∈M∣p(x)=0} complements the image as a direct summand, ensuring every element decomposes uniquely as a fixed point plus a kernel element.[11]

From an algebraic perspective, projections relate to retractions in categorical terms: a retraction r:X→Ar: X \to Ar:X→A for an inclusion i:A→Xi: A \to Xi:A→X satisfies r∘i=id⁡Ar \circ i = \operatorname{id}_Ar∘i=idA​, and the composite i∘ri \circ ri∘r is an idempotent endomorphism on XXX projecting onto AAA. This view bridges algebra to topology, where continuous retractions analogously yield idempotent maps.[12][13]

Orthogonal Projections

In finite-dimensional inner product spaces, an orthogonal projection of a vector vvv onto a subspace WWW is defined as the unique vector w∈Ww \in Ww∈W that minimizes the Euclidean distance ∥v−w∥|v - w|∥v−w∥, with the additional property that the error vector v−wv - wv−w is perpendicular to every vector in WWW.[14] This perpendicularity condition, expressed as ⟨v−w,u⟩=0\langle v - w, u \rangle = 0⟨v−w,u⟩=0 for all u∈Wu \in Wu∈W, ensures that the projection aligns with the geometry induced by the inner product.[15]

The uniqueness of this projection follows from the Pythagorean theorem in inner product spaces: if w1w_1w1​ and w2w_2w2​ are two such projections, then ∥v−w1∥2=∥v−w2∥2|v - w_1|^2 = |v - w_2|^2∥v−w1​∥2=∥v−w2​∥2 implies ∥w1−w2∥2=0|w_1 - w_2|^2 = 0∥w1​−w2​∥2=0, since the difference lies in WWW and the error terms are orthogonal to WWW.[16] Thus, w1=w2w_1 = w_2w1​=w2​, confirming a single closest point in WWW.[17]

A concrete example occurs when projecting onto a line in R2\mathbb{R}^2R2. For a nonzero vector uuu spanning the line and any v∈R2v \in \mathbb{R}^2v∈R2, the orthogonal projection is given by proj⁡uv=v⋅u∥u∥2u\operatorname{proj}_u v = \frac{v \cdot u}{|u|^2} uproju​v=∥u∥2v⋅u​u.[18] This scalar multiple of uuu lies on the line and satisfies the perpendicularity condition, as the dot product of the error v−proj⁡uvv - \operatorname{proj}_u vv−proju​v with uuu vanishes.

The orthogonal projection induces a linear operator PPP on the space such that P(v)=wP(v) = wP(v)=w, with key properties including idempotence P2=PP^2 = PP2=P—making PPP an idempotent mapping—and orthogonality ⟨Pv,z⟩=0\langle P v, z \rangle = 0⟨Pv,z⟩=0 for all z∈W⊥z \in W^\perpz∈W⊥, the orthogonal complement of WWW.[2] These ensure PPP preserves WWW and annihilates W⊥W^\perpW⊥.[1]

Oblique Projections

In linear algebra, an oblique projection is a linear operator PPP on a vector space VVV that is idempotent, satisfying P2=PP^2 = PP2=P, and maps VVV onto a subspace WWW along a complementary subspace UUU, such that V=W⊕UV = W \oplus UV=W⊕U with W∩U={0}W \cap U = {0}W∩U={0}.[19] Unlike orthogonal projections, the subspaces WWW and UUU are not required to be orthogonal complements, meaning the direction of projection is generally not perpendicular to WWW.[20] For any vector v∈Vv \in Vv∈V, there exists a unique decomposition v=w+uv = w + uv=w+u with w∈Ww \in Ww∈W and u∈Uu \in Uu∈U, and Pv=wP v = wPv=w.[19] The operator PPP is determined by the choice of UUU, and different choices of complementary subspace yield different oblique projections onto the same WWW.[21]

To compute an oblique projection matrix in a finite-dimensional space with respect to a basis, let the columns of matrix AAA form a basis for WWW and the columns of BBB form a basis for UUU, where the combined matrix M=[A B]M = [A , B]M=[AB] is invertible since WWW and UUU are complementary. The projection matrix PPP is then given by P=A[I 0]M−1P = A [I , 0] M^{-1}P=A[I0]M−1, where III is the identity matrix of appropriate dimension and the zero block matches the dimensions of UUU.[22] (Note that this construction relies on the direct sum decomposition and does not assume an inner product.) In contrast, the familiar formula P=A(ATA)−1ATP = A (A^T A)^{-1} A^TP=A(ATA)−1AT for the projection onto the column space of AAA holds only in the special case where UUU is the orthogonal complement of WWW with respect to the standard inner product, reducing to an orthogonal projection.[20]

A concrete example in R2\mathbb{R}^2R2 is the shearing projection onto the x-axis (W=span⁡{(1,0)}W = \operatorname{span}{(1,0)}W=span{(1,0)}) along the slanted line U=span⁡{(1,1)}U = \operatorname{span}{(1,1)}U=span{(1,1)}. Here, A=(10)A = \begin{pmatrix} 1 \ 0 \end{pmatrix}A=(10​) and B=(11)B = \begin{pmatrix} 1 \ 1 \end{pmatrix}B=(11​), so M=(1101)M = \begin{pmatrix} 1 & 1 \ 0 & 1 \end{pmatrix}M=(10​11​) with M−1=(1−101)M^{-1} = \begin{pmatrix} 1 & -1 \ 0 & 1 \end{pmatrix}M−1=(10​−11​). The projection matrix is P=A[1 0]M−1=(1000)(1−101)=(1−100)P = A [1 , 0] M^{-1} = \begin{pmatrix} 1 & 0 \ 0 & 0 \end{pmatrix} \begin{pmatrix} 1 & -1 \ 0 & 1 \end{pmatrix} = \begin{pmatrix} 1 & -1 \ 0 & 0 \end{pmatrix}P=A[10]M−1=(10​00​)(10​−11​)=(10​−10​). For a vector v=(xy)v = \begin{pmatrix} x \ y \end{pmatrix}v=(xy​), Pv=(x−y0)P v = \begin{pmatrix} x - y \ 0 \end{pmatrix}Pv=(x−y0​), which lies on the x-axis, and the difference v−Pv=(yy)v - P v = \begin{pmatrix} y \ y \end{pmatrix}v−Pv=(yy​) lies along UUU.[19] More generally, such projections can be represented in block form as P=(IC00)P = \begin{pmatrix} I & C \ 0 & 0 \end{pmatrix}P=(I0​C0​) relative to bases adapted to W⊕UW \oplus UW⊕U, where CCC encodes the "slant" determined by UUU.[21]

Oblique projections do not minimize the Euclidean distance from vvv to WWW unless U=W⊥U = W^\perpU=W⊥, distinguishing them from orthogonal projections, which achieve the shortest distance in normed spaces.[20] This property arises because the decomposition prioritizes the direct sum structure over perpendicularity, leading to applications in contexts requiring specific directional constraints rather than metric optimization.[19]

Idempotence and Decomposition

In linear algebra, a projection P:V→VP: V \to VP:V→V on a vector space VVV is characterized by its idempotence, meaning P2=PP^2 = PP2=P. This property ensures that applying PPP twice yields the same result as applying it once, distinguishing projections from other linear maps. Idempotence implies that PPP acts as the identity on its image and as the zero map on its kernel, facilitating a canonical decomposition of the space.[23][24]

To see this decomposition explicitly, consider any vector v∈Vv \in Vv∈V. Write v=(v−P(v))+P(v)v = (v - P(v)) + P(v)v=(v−P(v))+P(v), where P(v)∈im⁡(P)P(v) \in \operatorname{im}(P)P(v)∈im(P) and v−P(v)∈ker⁡(P)v - P(v) \in \ker(P)v−P(v)∈ker(P) since P(v−P(v))=P(v)−P2(v)=P(v)−P(v)=0P(v - P(v)) = P(v) - P^2(v) = P(v) - P(v) = 0P(v−P(v))=P(v)−P2(v)=P(v)−P(v)=0. Thus, V=im⁡(P)+ker⁡(P)V = \operatorname{im}(P) + \ker(P)V=im(P)+ker(P). Moreover, im⁡(P)∩ker⁡(P)={0}\operatorname{im}(P) \cap \ker(P) = {0}im(P)∩ker(P)={0}, because if w∈im⁡(P)∩ker⁡(P)w \in \operatorname{im}(P) \cap \ker(P)w∈im(P)∩ker(P), then w=P(u)w = P(u)w=P(u) for some u∈Vu \in Vu∈V and P(w)=0P(w) = 0P(w)=0, so P2(u)=0P^2(u) = 0P2(u)=0 implies P(u)=0P(u) = 0P(u)=0, hence w=0w = 0w=0. Therefore, V=im⁡(P)⊕ker⁡(P)V = \operatorname{im}(P) \oplus \ker(P)V=im(P)⊕ker(P). On im⁡(P)\operatorname{im}(P)im(P), PPP acts as the identity: for w=P(u)w = P(u)w=P(u), P(w)=P2(u)=P(u)=wP(w) = P^2(u) = P(u) = wP(w)=P2(u)=P(u)=w. On ker⁡(P)\ker(P)ker(P), PPP acts as zero by definition.[23][25]

Conversely, any idempotent linear map PPP is a projection onto im⁡(P)\operatorname{im}(P)im(P) along ker⁡(P)\ker(P)ker(P), meaning it splits VVV into this direct sum where vectors in the image are fixed and those in the kernel are annihilated. This equivalence holds in finite-dimensional spaces and extends to the infinite-dimensional case under suitable conditions. Oblique projections provide concrete examples of such decompositions where the summands are not orthogonal.[24][25]

In matrix terms, if PPP is represented by a matrix in a basis adapted to the decomposition V=im⁡(P)⊕ker⁡(P)V = \operatorname{im}(P) \oplus \ker(P)V=im(P)⊕ker(P), then PPP takes the block diagonal form (Ir000n−r)\begin{pmatrix} I_r & 0 \ 0 & 0_{n-r} \end{pmatrix}(Ir​0​00n−r​​), where r=dim⁡(im⁡(P))r = \dim(\operatorname{im}(P))r=dim(im(P)) and n=dim⁡(V)n = \dim(V)n=dim(V). The eigenvalues of PPP are thus 1 with multiplicity rrr and 0 with multiplicity n−rn - rn−r. Since the minimal polynomial of PPP divides x(x−1)x(x-1)x(x−1), which has distinct roots, PPP is diagonalizable, and its Jordan canonical form is diagonal with blocks of 1's and 0's.[26] For instance, the matrix (1100)\begin{pmatrix} 1 & 1 \ 0 & 0 \end{pmatrix}(10​10​) is idempotent, projecting R2\mathbb{R}^2R2 onto the x-axis along the line y=−xy = -xy=−x, with eigenvalues 1 and 0.[27]

Orthogonal Projection Theorem

In a Hilbert space HHH, equipped with an inner product ⟨⋅,⋅⟩\langle \cdot, \cdot \rangle⟨⋅,⋅⟩, consider a closed subspace W⊆HW \subseteq HW⊆H. For every vector v∈Hv \in Hv∈H, there exists a unique vector p=\projWv∈Wp = \proj_W v \in Wp=\projW​v∈W such that v−pv - pv−p is orthogonal to WWW, meaning ⟨v−p,w⟩=0\langle v - p, w \rangle = 0⟨v−p,w⟩=0 for all w∈Ww \in Ww∈W. This ppp is called the orthogonal projection of vvv onto WWW, and it satisfies ∥v−p∥=inf⁡{∥v−w∥:w∈W}| v - p | = \inf { | v - w | : w \in W }∥v−p∥=inf{∥v−w∥:w∈W}.[28]

The proof relies on the Riesz representation theorem and the completeness of HHH. Since WWW is a closed subspace, it is itself a Hilbert space. The map ϕ:W→C\phi: W \to \mathbb{C}ϕ:W→C (or R\mathbb{R}R) defined by ϕ(w)=⟨v,w⟩\phi(w) = \langle v, w \rangleϕ(w)=⟨v,w⟩ is a bounded linear functional on WWW, with ∥ϕ∥≤∥v∥| \phi | \leq | v |∥ϕ∥≤∥v∥. By the Riesz representation theorem, there exists a unique p∈Wp \in Wp∈W such that ϕ(w)=⟨p,w⟩\phi(w) = \langle p, w \rangleϕ(w)=⟨p,w⟩ for all w∈Ww \in Ww∈W. This implies ⟨v−p,w⟩=0\langle v - p, w \rangle = 0⟨v−p,w⟩=0 for all w∈Ww \in Ww∈W, establishing orthogonality. The uniqueness of ppp follows directly from the uniqueness of the representing element in the Riesz theorem. The minimizing property holds because, for any w∈Ww \in Ww∈W,

with equality at t=0t = 0t=0.[28]

A key corollary is that HHH decomposes as the orthogonal direct sum H=W⊕W⊥H = W \oplus W^\perpH=W⊕W⊥, where W⊥={u∈H:⟨u,w⟩=0 ∀w∈W}W^\perp = { u \in H : \langle u, w \rangle = 0 \ \forall w \in W }W⊥={u∈H:⟨u,w⟩=0 ∀w∈W} is the orthogonal complement of WWW. Specifically, every v∈Hv \in Hv∈H can be uniquely written as v=p+(v−p)v = p + (v - p)v=p+(v−p) with p∈Wp \in Wp∈W and v−p∈W⊥v - p \in W^\perpv−p∈W⊥, and W∩W⊥={0}W \cap W^\perp = { 0 }W∩W⊥={0} follows from the positive definiteness of the inner product. This decomposition extends the finite-dimensional case to infinite dimensions.[28]

The Riesz representation theorem, underpinning this result, was independently established by Maurice Fréchet and Frigyes Riesz in 1907, with Erhard Schmidt providing an early explicit proof of the projection theorem in the context of ℓ2\ell^2ℓ2 spaces in 1908.[29][30]

In Geometry

In geometry, orthogonal projection represents the process of dropping a perpendicular from a point to a line, plane, or subspace, yielding the foot of that perpendicular as the projected point. This concept is fundamental in Euclidean spaces, where it minimizes the distance between the original point and its image on the target subspace. For instance, in the Euclidean plane, orthogonal projection onto a line computes heights or distances by constructing right angles, essential for measuring altitudes in triangles or resolving vectors into components along axes.[31]

Projections in geometry extend beyond orthogonality to include parallel and perspective types, each offering distinct spatial interpretations. Parallel projections, which are affine transformations, map points using rays perpendicular or at a fixed angle to the projection plane, preserving parallelism of lines without convergence. In contrast, perspective projections simulate human vision by directing rays from a fixed viewpoint (center of projection) through the object to the image plane, causing parallel lines to converge at vanishing points on the horizon line. Parallel projections emerge as the limiting case of perspective projections when the center of projection recedes to infinity, transitioning from projective to affine geometry.[32][3]

A key application of parallel projection is orthographic projection in technical drawing, where three-dimensional objects are represented in multiple two-dimensional views (e.g., front, top, side) by projecting orthogonally onto perpendicular planes, ensuring accurate scale without distortion from depth. The formula for the orthogonal projection of a vector v\mathbf{v}v onto a plane with unit normal vector n\mathbf{n}n is given by

which subtracts the component along the normal to yield the in-plane vector. This method facilitates precise engineering visualizations by eliminating perspective foreshortening.[33]

In crystallography, projections visualize periodic lattice structures by mapping three-dimensional crystal arrangements onto two-dimensional planes, aiding in the identification of symmetry and unit cells. For cubic crystals, standard projections along principal axes, such as the direction, reveal square or rectangular lattices, while clinographic views highlight atomic packing without overlap. These techniques, often orthographic, are crucial for analyzing diffraction patterns and material properties.[34][35]

Oblique projections, involving non-perpendicular rays, model slanted shadows in geometric scenes but lack the distance-minimizing property of orthogonal ones.[3]

In Statistics and Optimization

In statistics, projections are fundamental to methods that approximate data by minimizing errors in a least squares sense, particularly in linear regression where the goal is to find the best linear fit to observed data points. The ordinary least squares (OLS) estimator projects the response vector yyy onto the column space of the design matrix XXX, yielding the fitted values y^=Py\hat{y} = P yy^​=Py, where P=X(XTX)−1XTP = X (X^T X)^{-1} X^TP=X(XTX)−1XT is the projection matrix. This projection minimizes the Euclidean distance ∥y−y^∥22| y - \hat{y} |_2^2∥y−y^​∥22​, ensuring the residuals are orthogonal to the column space of XXX. The coefficient vector is given by β^=(XTX)−1XTy,\hat{\beta} = (X^T X)^{-1} X^T y,β^​=(XTX)−1XTy, assuming XTXX^T XXTX is invertible.[36]

This approach traces back to the method of least squares, originally developed by Adrien-Marie Legendre in 1805 for determining comet orbits by minimizing squared deviations from predicted positions, and independently by Carl Friedrich Gauss around 1795, though published later. In modern linear algebra terms, the geometric interpretation as an orthogonal projection onto a subspace became prominent in the 20th century, aligning with Hilbert space theory.[37][38]

In optimization, projections play a key role in handling constraints, especially in convex programming where one projects points onto feasible sets to enforce constraints iteratively. A classic example is the alternating projections method, which finds a point in the intersection of two closed convex sets by successively projecting onto each set; for subspaces, it converges to the projection onto their intersection. This method, analyzed by John von Neumann in 1933 for operator theory in Hilbert spaces, underpins algorithms like the convex feasibility problem solvers.

Principal component analysis (PCA) exemplifies projections in data reduction, involving successive orthogonal projections onto the principal component subspaces to capture maximum variance. Introduced by Karl Pearson in 1901 as the "principal axes" of a point cloud minimizing squared distances, PCA projects data onto a lower-dimensional space spanned by eigenvectors of the covariance matrix, ordered by eigenvalues. This technique, further developed by Harold Hotelling in 1933 for multivariate analysis, is widely used for dimensionality reduction while preserving essential structure.[39]

In Functional Analysis

In functional analysis, projections are generalized to bounded linear operators on normed linear spaces that satisfy the idempotence condition P2=PP^2 = PP2=P. Such operators are called bounded projections or continuous idempotents, and they map the space onto their range while fixing points in that range. For a non-zero bounded projection PPP, the operator norm satisfies ∥P∥≥1|P| \geq 1∥P∥≥1, since for any xxx in the range of PPP, Px=xPx = xPx=x implies ∥x∥≤∥P∥∥x∥|x| \leq |P| |x|∥x∥≤∥P∥∥x∥. In Hilbert spaces, the orthogonal projection theorem ensures that every closed subspace admits an orthogonal projection, which is self-adjoint and has norm exactly 1.

The spectral theorem for self-adjoint operators on a Hilbert space provides a key application of projections, associating to each bounded self-adjoint operator AAA a resolution of the identity, which is a projection-valued measure EEE. The spectral projections E(Δ)E(\Delta)E(Δ) for Borel sets Δ⊆R\Delta \subseteq \mathbb{R}Δ⊆R are orthogonal projections onto the subspaces corresponding to the spectral support in Δ\DeltaΔ, satisfying A=∫σ(A)λ dE(λ)A = \int_{\sigma(A)} \lambda , dE(\lambda)A=∫σ(A)​λdE(λ).[40] These projections commute with AAA and decompose the space according to the spectrum.

A concrete example arises in the space L2([0,2π])L^2([0, 2\pi])L2([0,2π]), where the partial sum operators of the Fourier series define orthogonal projections onto the finite-dimensional subspaces of trigonometric polynomials of degree at most nnn. Specifically, the operator Snf=∑k=−nnf^(k)eikxS_n f = \sum_{k=-n}^n \hat{f}(k) e^{ikx}Sn​f=∑k=−nn​f^​(k)eikx projects fff onto span⁡{eikx:∣k∣≤n}\operatorname{span}{e^{ikx} : |k| \leq n}span{eikx:∣k∣≤n}, preserving the L2L^2L2 norm on that subspace.[41]

Unlike Hilbert spaces, where every closed subspace is complemented by a bounded (orthogonal) projection, Banach spaces in general do not possess this property. A closed subspace YYY of a Banach space XXX is complemented if there exists a bounded projection from XXX onto YYY; however, non-complemented subspaces exist in many Banach spaces. For instance, Lindenstrauss and Tzafriri proved that a Banach space in which every closed subspace is complemented must be isomorphic to a Hilbert space.[42] Dvoretzky's theorem implies the existence of finite-dimensional subspaces nearly isometric to Euclidean spaces in every infinite-dimensional Banach space, highlighting that such "Hilbert-like" structures do not guarantee complementation in non-Hilbert settings, as non-Hilbert spaces necessarily contain uncomplemented closed subspaces.

In Category Theory

In category theory, projections arise prominently in the context of products, which are a type of limit. Given a family of objects {Xi}i∈I{X_i}{i \in I}{Xi​}i∈I​ in a category C\mathcal{C}C, their product ∏i∈IXi\prod{i \in I} X_i∏i∈I​Xi​ is an object equipped with projection morphisms πi:∏i∈IXi→Xi\pi_i: \prod_{i \in I} X_i \to X_iπi​:∏i∈I​Xi​→Xi​ for each i∈Ii \in Ii∈I. These projections satisfy the universal property: for any object QQQ and family of morphisms fi:Q→Xif_i: Q \to X_ifi​:Q→Xi​, there exists a unique morphism f:Q→∏i∈IXif: Q \to \prod_{i \in I} X_if:Q→∏i∈I​Xi​ such that the diagrams commute, i.e., πi∘f=fi\pi_i \circ f = f_iπi​∘f=fi​ for all iii. This ensures that the product cone (∏i∈IXi,{πi}i∈I)(\prod_{i \in I} X_i, {\pi_i}{i \in I})(∏i∈I​Xi​,{πi​}i∈I​) is universal among all cones over the diagram {Xi}i∈I{X_i}{i \in I}{Xi​}i∈I​.

Projections also connect to the notion of idempotents in categories. An idempotent morphism e:X→Xe: X \to Xe:X→X satisfies e∘e=ee \circ e = ee∘e=e, generalizing projectors from linear algebra. A split idempotent is one that factors as a retraction-section pair: there exist morphisms r:X→Yr: X \to Yr:X→Y and s:Y→Xs: Y \to Xs:Y→X such that e=r∘se = r \circ se=r∘s, s∘r=idYs \circ r = \mathrm{id}_Ys∘r=idY​, and r∘s=er \circ s = er∘s=e, where YYY is the image of eee. In this case, sss is a section and rrr a retraction, allowing a direct sum decomposition X≅Y⊕ZX \cong Y \oplus ZX≅Y⊕Z in categories supporting such splits, like abelian categories.[8] Categories where all idempotents split are called idempotent complete.

In abelian categories, projections relate closely to split short exact sequences. Consider a short exact sequence 0→K→iE→pQ→00 \to K \xrightarrow{i} E \xrightarrow{p} Q \to 00→Ki​Ep​Q→0; it splits if there exists a morphism s:Q→Es: Q \to Es:Q→E such that p∘s=idQp \circ s = \mathrm{id}_Qp∘s=idQ​, making ppp a retraction and sss a section. This yields an isomorphism E≅K⊕QE \cong K \oplus QE≅K⊕Q, with ppp serving as the projection onto QQQ along KKK. Such splits correspond to idempotent endomorphisms on EEE, where the projection ppp embodies the decomposition.

More broadly, projections form the components of product cones in the limit-colimit duality of categories. While products are limits (universal cones into the diagram), coproducts are colimits (universal cones out of the diagram), and in categories with both, like the category of sets, the projections πi\pi_iπi​ contrast with the coproduct inclusions. This relational structure underscores projections' role in universal constructions across categorical limits.