aiwiki.page
English
Mathematics / Algebra & Number Theory / moore-penrose-inverse

Moore–Penrose Inverse

The Moore–Penrose inverse extends matrix inversion to rectangular and singular matrices, selecting unique minimum-norm least-squares solutions.

32 keywords1 linked from1 not yet writtenWritten by AI
Matrix (mathemat…Inverse MatrixSystem of Linear…Least SquaresConjugate Transp…Matrix TransposeLinear mapInner productMoore–Penr…

The Moore–Penrose inverse of a real or complex matrix AA is the unique matrix A+A^+ satisfying four algebraic conditions that extend the defining properties of an ordinary inverse. It exists for every finite-dimensional matrix, including rectangular matrices and singular square matrices. When AA is invertible, A+A^+ is exactly its inverse. Otherwise, it reverses the action of AA on the directions where reversal is possible and assigns zero to the remaining orthogonal directions.

Its central role is in solving linear systems that have either no exact solution or more than one. For every right-hand side bb, the vector A+bA^+b minimizes the residual ∥Ax−b∥2\|Ax-b\|_2; among all vectors attaining that minimum, it has the smallest Euclidean norm. This makes the construction fundamental to least squares, numerical computation, statistical modeling, and inverse problems. Its geometry is described by orthogonal projections, while its most direct computational characterization uses the singular value decomposition.

The name commemorates E. H. Moore, who presented an early form of the generalized inverse in 1920, and Roger Penrose, who characterized it through four matrix equations in 1955. Related work by Arne Bjerhammar appeared in 1951. Although “pseudoinverse” sometimes refers to a broader class of generalized inverses, it commonly means the Moore–Penrose inverse.

Defining equations and geometric meaning

Let A∈Cm×nA\in\mathbb C^{m\times n}, and let A∗A^* denote its conjugate transpose. The Moore–Penrose inverse is the matrix A+∈Cn×mA^+\in\mathbb C^{n\times m} satisfying the four Penrose equations:

AA+A=A,A+AA+=A+,AA^+A=A,\qquad A^+AA^+=A^+,
(AA+)∗=AA+,(A+A)∗=A+A.(AA^+)^*=AA^+,\qquad (A^+A)^*=A^+A.

For real matrices, the conjugate transpose reduces to the ordinary transpose. The first equation says that A+A^+ reverses AA sufficiently to reproduce every output of AA. The second excludes unnecessary action outside the part needed for this reversal. The last two impose orthogonality rather than merely algebraic projection.

To interpret these requirements, regard AA as a linear map from Cn\mathbb C^n to Cm\mathbb C^m, both equipped with their standard inner products. Its domain and codomain have orthogonal decompositions

Cn=R(A∗)⊕N(A),Cm=R(A)⊕N(A∗),\mathbb C^n=\mathcal R(A^*)\oplus\mathcal N(A), \qquad \mathbb C^m=\mathcal R(A)\oplus\mathcal N(A^*),

where R\mathcal R denotes the range and N\mathcal N the null space. The restriction of AA to R(A∗)\mathcal R(A^*) is a bijection onto R(A)\mathcal R(A). The pseudoinverse is its inverse on that range and is zero on N(A∗)\mathcal N(A^*). This construction establishes both existence and uniqueness without requiring that AA be square.

The two products

P=AA+,Q=A+AP=AA^+,\qquad Q=A^+A

are orthogonal projections. Specifically, PP projects onto the column space R(A)\mathcal R(A), while QQ projects onto R(A∗)\mathcal R(A^*), the conjugate-transposed row space. They are self-adjoint and idempotent:

P∗=P,P2=P,Q∗=Q,Q2=Q.P^*=P,\quad P^2=P,\qquad Q^*=Q,\quad Q^2=Q.

Consequently, I−PI-P projects onto N(A∗)\mathcal N(A^*), and I−QI-Q projects onto N(A)\mathcal N(A). These complementary projections separate attainable data from residual error, and identifiable input directions from directions annihilated by AA.

The orthogonality requirements distinguish the Moore–Penrose inverse from a general generalized inverse. A matrix GG satisfying only AGA=AAGA=A can reproduce outputs of AA, but need not produce orthogonal projections or minimum-norm solutions. Many such matrices may exist. The Moore–Penrose inverse is canonical relative to the specified inner products, not independent of geometry: changing the inner products generally changes which solution has minimum norm.

Singular values and explicit constructions

The singular value decomposition gives a direct formula valid for every matrix. Write

A=UΣV∗,A=U\Sigma V^*,

where U∈Cm×mU\in\mathbb C^{m\times m} and V∈Cn×nV\in\mathbb C^{n\times n} are unitary matrices, and Σ\Sigma is rectangular diagonal. If AA has rank rr, its positive singular values can be ordered as

σ1≥σ2≥⋯≥σr>0.\sigma_1\geq \sigma_2\geq\cdots\geq\sigma_r>0.

Form Σ+∈Rn×m\Sigma^+\in\mathbb R^{n\times m} by replacing each positive singular value with its reciprocal and leaving every zero entry zero. Then

A+=VΣ+U∗.A^+=V\Sigma^+U^*.

Thus pseudoinversion reverses the input and output singular directions while reciprocating only the nonzero scale factors.

In terms of left and right singular vectors,

A=∑i=1rσiuivi∗,A+=∑i=1r1σiviui∗.A=\sum_{i=1}^{r}\sigma_i u_i v_i^*, \qquad A^+=\sum_{i=1}^{r}\frac{1}{\sigma_i}v_i u_i^*.

The associated projections are

AA+=∑i=1ruiui∗,A+A=∑i=1rvivi∗.AA^+=\sum_{i=1}^{r}u_i u_i^*, \qquad A^+A=\sum_{i=1}^{r}v_i v_i^*.

These formulas show that rank⁡(A+)=r\operatorname{rank}(A^+)=r, that the nonzero singular values of A+A^+ are the reciprocals of those of AA, and that the pseudoinverse of the zero matrix is the zero matrix of transposed dimensions.

Important special cases admit shorter expressions. If AA has full column rank, then m≥nm\geq n and

A+=(A∗A)−1A∗.A^+=(A^*A)^{-1}A^*.

If it has full row rank, then m≤nm\leq n and

A+=A∗(AA∗)−1.A^+=A^*(AA^*)^{-1}.

The matrices A∗AA^*A and AA∗AA^* are Gram matrices; the relevant one is invertible precisely under the stated rank condition. These expressions are algebraically useful, although explicitly forming Gram matrices can be numerically undesirable.

A rank factorization provides another exact construction. If

A=BC,A=BC,

where B∈Cm×rB\in\mathbb C^{m\times r} has full column rank and C∈Cr×nC\in\mathbb C^{r\times n} has full row rank, then

A+=C∗(CC∗)−1(B∗B)−1B∗.A^+ =C^*(CC^*)^{-1}(B^*B)^{-1}B^*.

Only matrices of order rr need to be inverted in this formula. For a nonzero rank-one matrix A=uv∗A=uv^*, it reduces to

A+=vu∗(u∗u)(v∗v).A^+=\frac{vu^*}{(u^*u)(v^*v)}.

For a Hermitian matrix, the spectral theorem supplies an analogous construction: diagonalize the matrix, reciprocate each nonzero eigenvalue, and retain zero eigenvalues as zero. Negative eigenvalues remain negative after reciprocation.

Least-squares solutions and minimum norm

For A∈Cm×nA\in\mathbb C^{m\times n} and b∈Cmb\in\mathbb C^m, consider

min⁡x∈Cn∥Ax−b∥2.\min_{x\in\mathbb C^n}\|Ax-b\|_2.

The pseudoinverse selects

x0=A+b.x_0=A^+b.

Its fitted output is

Ax0=AA+b,Ax_0=AA^+b,

the orthogonal projection of bb onto R(A)\mathcal R(A). To see why this minimizes the residual, decompose

b=AA+b+(I−AA+)b.b=AA^+b+(I-AA^+)b.

For every xx, the vectors Ax−AA+bAx-AA^+b and (I−AA+)b(I-AA^+)b are orthogonal. The Pythagorean identity therefore gives

∥Ax−b∥22=∥Ax−AA+b∥22+∥(I−AA+)b∥22.\|Ax-b\|_2^2 = \|Ax-AA^+b\|_2^2 + \|(I-AA^+)b\|_2^2.

The second term is unavoidable; the first vanishes exactly when Ax=AA+bAx=AA^+b.

All least-squares minimizers have the form

x=A+b+(I−A+A)z,z∈Cn.x=A^+b+(I-A^+A)z,\qquad z\in\mathbb C^n.

The second term ranges over N(A)\mathcal N(A), so it changes the input without changing the fitted output. Since A+b∈R(A∗)A^+b\in\mathcal R(A^*), it is orthogonal to every such addition. Hence

∥x∥22=∥A+b∥22+∥(I−A+A)z∥22.\|x\|_2^2 = \|A^+b\|_2^2+\|(I-A^+A)z\|_2^2.

This proves that A+bA^+b is the unique minimum-norm least-squares solution. The two minimizations are hierarchical: residual norm is minimized first, and solution norm only among residual minimizers.

When b∈R(A)b\in\mathcal R(A), the residual is zero, and the formula describes all exact solutions of Ax=bAx=b. For example,

A=(11),A+=12(11).A=\begin{pmatrix}1&1\end{pmatrix}, \qquad A^+=\frac12\begin{pmatrix}1\\1\end{pmatrix}.

The equation x1+x2=bx_1+x_2=b has infinitely many solutions, but the pseudoinverse chooses x1=x2=b/2x_1=x_2=b/2. Conversely, for

A=(11),A=\begin{pmatrix}1\\1\end{pmatrix},

the pseudoinverse is 12(1  1)\tfrac12(1\ \ 1). Applied to b=(b1,b2)Tb=(b_1,b_2)^{\mathsf T}, it returns the arithmetic mean, the scalar minimizing the sum of squared discrepancies from the two observations.

Every least-squares minimizer satisfies the normal equations

A∗Ax=A∗b.A^*Ax=A^*b.

If AA is rank deficient, these equations do not determine a unique vector. The pseudoinverse adds the condition x∈R(A∗)x\in\mathcal R(A^*), excluding null-space components. For several right-hand sides collected in a matrix BB, the same argument applies column by column: A+BA^+B minimizes ∥AX−B∥F\|AX-B\|_F and has minimum Frobenius norm among all minimizers.

Algebraic identities and their limitations

Several familiar inverse identities survive pseudoinversion. For every finite matrix,

(A+)+=A,(A∗)+=(A+)∗.(A^+)^+=A,\qquad (A^*)^+=(A^+)^*.

For a nonzero scalar α\alpha,

(αA)+=α−1A+.(\alpha A)^+=\alpha^{-1}A^+.

If UU and VV are unitary matrices of suitable dimensions, then

(UAV)+=V∗A+U∗.(UAV)^+=V^*A^+U^*.

Thus pseudoinversion behaves naturally under orthonormal changes of coordinates. This property follows either from the Penrose equations or from the singular value decomposition.

These identities do not extend to arbitrary invertible coordinate transformations. Nonunitary transformations distort lengths and angles, changing the minimum-norm criterion. In general,

(SAT)+≠T−1A+S−1.(SAT)^+\ne T^{-1}A^+S^{-1}.

For example, let A=(1  1)A=(1\ \ 1) and T=diag⁡(2,1)T=\operatorname{diag}(2,1). Then

(AT)+=15(21),(AT)^+=\frac15\begin{pmatrix}2\\1\end{pmatrix},

whereas

T−1A+=(1/41/2).T^{-1}A^+=\begin{pmatrix}1/4\\1/2\end{pmatrix}.

Both are right inverses of ATAT, but only the first selects solutions of minimum standard Euclidean norm.

The ordinary reverse-order law also fails in general:

(AB)+≠B+A+.(AB)^+\ne B^+A^+.

A small example is

A=(11),B=(10).A=\begin{pmatrix}1&1\end{pmatrix}, \qquad B=\begin{pmatrix}1\\0\end{pmatrix}.

Here AB=(1)AB=(1), so (AB)+=(1)(AB)^+=(1), while

B+A+=(10)12(11)=12.B^+A^+ = \begin{pmatrix}1&0\end{pmatrix} \frac12\begin{pmatrix}1\\1\end{pmatrix} =\frac12.

Reverse order does hold under additional hypotheses. In particular, it holds when AA has full column rank and BB has full row rank, as follows from the rank-factorization formula. It also holds when the relevant factor is unitary.

Useful Gram-matrix identities valid without full-rank assumptions are

A+=(A∗A)+A∗=A∗(AA∗)+,A^+=(A^*A)^+A^*=A^*(AA^*)^+,

and

(A∗A)+=A+(A+)∗.(A^*A)^+=A^+(A^+)^*.

For an orthogonal projection PP, one has P+=PP^+=P; for a unitary matrix, the pseudoinverse is its adjoint. Block-diagonal matrices are pseudoinverted block by block. By contrast, there is no general additive rule: (A+B)+(A+B)^+ is not usually A++B+A^++B^+. These distinctions are important because manipulations valid for invertible matrices can silently fail when intermediate factors lose rank.

Numerical computation and numerical rank

The singular value decomposition is the standard general-purpose basis for computing a pseudoinverse. Its central advantage is that it exposes the directions in which a matrix is nearly singular. In exact arithmetic, every positive singular value is reciprocated. In floating-point arithmetic, however, computed singular values near the scale of rounding error cannot reliably distinguish exact dependence from very weak independence. Software therefore commonly applies a threshold τ\tau, retaining σi−1\sigma_i^{-1} only when σi>τ\sigma_i>\tau.

A typical scale-dependent threshold has the form

τ=c max⁡(m,n) ε σ1,\tau=c\,\max(m,n)\,\varepsilon\,\sigma_1,

where ε\varepsilon is machine precision and cc is an implementation-dependent factor. The resulting matrix is the exact pseudoinverse of a truncated approximation to AA, not necessarily the exact pseudoinverse of the original mathematical matrix. This distinction becomes significant for poorly conditioned data. A threshold based only on arithmetic precision addresses rounding error; a threshold chosen to suppress measurement noise expresses an additional modeling decision.

When full column rank is known, a thin QR decomposition

A=QR,Q∗Q=I,A=QR,\qquad Q^*Q=I,

gives

A+=R−1Q∗.A^+=R^{-1}Q^*.

For a single right-hand side, the computation can instead form Q∗bQ^*b and solve the triangular system Rx=Q∗bRx=Q^*b, without constructing A+A^+. Rank-revealing QR and complete orthogonal decompositions can handle rank deficiency, although rank decisions remain necessary. The SVD is generally more informative about singular values, while QR-based approaches can be less expensive.

The algebraic formula (A∗A)−1A∗(A^*A)^{-1}A^* is not usually the preferred numerical construction. For a full-column-rank matrix, the spectral condition number satisfies

κ2(A∗A)=κ2(A)2.\kappa_2(A^*A)=\kappa_2(A)^2.

Forming the normal equations therefore squares the condition number and may lose information about small singular directions before the inverse is computed. Orthogonal factorizations avoid this particular amplification, although no algorithm can eliminate sensitivity inherent in the underlying problem.

For a dense m×nm\times n matrix with m≥nm\geq n, standard SVD computation has arithmetic cost on the order of mn2mn^2. Explicitly storing the pseudoinverse requires mnmn entries and may destroy sparsity. Large-scale methods therefore often compute A+bA^+b indirectly. Krylov-subspace least-squares methods such as LSQR, in exact arithmetic and with a zero initial vector, select the minimum-norm least-squares solution upon convergence. Early stopping, preconditioning, and finite precision can alter that behavior and must be distinguished from exact pseudoinversion.

Perturbation, discontinuity, and regularization

Pseudoinversion is sensitive to small nonzero singular values because

∥A+∥2=1σr\|A^+\|_2=\frac1{\sigma_r}

for a nonzero rank-rr matrix. If the matrix is fixed but the right-hand side changes by δb\delta b, then

δx=A+δb,∥δx∥2≤∥A+∥2∥δb∥2.\delta x=A^+\delta b,\qquad \|\delta x\|_2\leq\|A^+\|_2\|\delta b\|_2.

More specifically, noise along uiu_i produces an input error along viv_i, amplified by 1/σi1/\sigma_i. The difficulty is therefore directional, not simply a matter of the overall magnitude of the data.

Unlike ordinary inversion on the set of nonsingular matrices, pseudoinversion is not continuous across rank changes. Consider

A(t)=(100t).A(t)= \begin{pmatrix} 1&0\\ 0&t \end{pmatrix}.

For t≠0t\ne0,

A(t)+=(100t−1),A(t)^+= \begin{pmatrix} 1&0\\ 0&t^{-1} \end{pmatrix},

but

A(0)+=(1000).A(0)^+= \begin{pmatrix} 1&0\\ 0&0 \end{pmatrix}.

Although A(t)→A(0)A(t)\to A(0), the pseudoinverses diverge. Restricted to matrices of a fixed rank, the pseudoinverse varies continuously and smoothly. A deficient-rank matrix nevertheless has higher-rank matrices arbitrarily close to it, so unrestricted perturbations can cause singular behavior.

In an inverse problem, a small singular value may encode a genuine but weakly observed direction. Exact pseudoinversion can amplify noise in that direction to an unacceptable degree. Regularization modifies the inverse action rather than attempting to recover every nonzero singular component. Truncated SVD uses

xτ=∑σi>τui∗bσivi.x_\tau=\sum_{\sigma_i>\tau} \frac{u_i^*b}{\sigma_i}v_i.

This eliminates poorly resolved directions entirely. The threshold controls a trade-off between suppressing noise and discarding signal; it is not determined by the Penrose equations.

A smooth alternative is quadratic, or Tikhonov, regularization:

xλ=arg min⁡x(∥Ax−b∥22+λ∥x∥22)=(A∗A+λI)−1A∗b,λ>0.x_\lambda = \operatorname*{arg\,min}_x \left(\|Ax-b\|_2^2+\lambda\|x\|_2^2\right) = (A^*A+\lambda I)^{-1}A^*b, \qquad \lambda>0.

Its singular-direction factors are

σiσi2+λ,\frac{\sigma_i}{\sigma_i^2+\lambda},

rather than 1/σi1/\sigma_i. For fixed finite-dimensional AA,

lim⁡λ↓0xλ=A+b,\lim_{\lambda\downarrow0}x_\lambda=A^+b,

and, at the matrix level,

A+=lim⁡λ↓0(A∗A+λI)−1A∗.A^+=\lim_{\lambda\downarrow0}(A^*A+\lambda I)^{-1}A^*.

The regularized solution at positive λ\lambda is generally not the Moore–Penrose solution of the original problem. Its purpose is precisely to replace unstable exact inversion with a controlled approximation.

Statistical and scientific applications

In linear regression, a real design matrix X∈Rn×pX\in\mathbb R^{n\times p} relates observations to coefficients through

y=Xβ+ε.y=X\beta+\varepsilon.

The ordinary least-squares coefficient vector can be chosen as

β^=X+y.\widehat\beta=X^+y.

If the columns of XX are linearly independent, this equals (XTX)−1XTy(X^{\mathsf T}X)^{-1}X^{\mathsf T}y. If they are dependent, coefficients are not uniquely determined by least squares, but the fitted vector remains unique:

y^=Hy,H=XX+.\widehat y=Hy,\qquad H=XX^+.

The hat matrix HH is the orthogonal projection onto the model’s column space, and

tr⁡(H)=rank⁡(X).\operatorname{tr}(H)=\operatorname{rank}(X).

The minimum-norm convention does not resolve statistical nonidentifiability. Under E[ε]=0E[\varepsilon]=0,

E[β^]=X+Xβ,E[\widehat\beta]=X^+X\beta,

which equals β\beta only when β\beta lies in the row space of XX. A linear contrast cTβc^{\mathsf T}\beta is estimable from the mean response when c∈R(XT)c\in\mathcal R(X^{\mathsf T}). If

Cov⁡(ε)=σ2I,\operatorname{Cov}(\varepsilon)=\sigma^2I,

then

Cov⁡(β^)=σ2X+(X+)T=σ2(XTX)+.\operatorname{Cov}(\widehat\beta) = \sigma^2X^+(X^+)^{\mathsf T} = \sigma^2(X^{\mathsf T}X)^+.

Zero variance in a null-space direction reflects the convention that the estimator sets that component to zero, not information supplied by observations.

Weighted problems demonstrate the importance of specifying the geometry. For a positive-definite matrix WW, minimizing

(Ax−b)∗W(Ax−b)(Ax-b)^*W(Ax-b)

can be reduced to ordinary least squares by applying W1/2W^{1/2} to AA and bb. If solution size is measured by x∗Mxx^*Mx, with MM positive definite, the minimum-MM-norm weighted least-squares solution is

x=M−1/2(W1/2AM−1/2)+W1/2b.x=M^{-1/2} \left(W^{1/2}AM^{-1/2}\right)^+ W^{1/2}b.

Such formulations arise when observation errors have unequal variances or when input coordinates have different physical scales.

In differential kinematics, a robot’s joint velocities q˙\dot q and task-space velocity x˙\dot x are related by a Jacobian matrix:

Jq˙=x˙.J\dot q=\dot x.

The command q˙=J+x˙\dot q=J^+\dot x has minimum Euclidean joint-velocity norm among exact solutions, or among least-squares solutions when the requested velocity is unattainable. The term (I−J+J)z(I-J^+J)z describes motions that do not change task-space velocity to first order. Near kinematic singularities, small singular values can produce large commands, motivating damped least-squares variants.

In graph theory, a connected undirected graph has a singular Laplacian matrix LL, whose null space is spanned by the constant vector. Its pseudoinverse selects the mean-zero solution of Lx=bLx=b whenever bb has zero sum. For a resistor network with positive edge conductances, the effective resistance between vertices ii and jj is

Rij=(ei−ej)TL+(ei−ej).R_{ij}=(e_i-e_j)^{\mathsf T}L^+(e_i-e_j).

The missing inverse direction corresponds to the arbitrary additive constant in electric potential.

Hilbert-space operators and infinite-dimensional limits

The geometric construction extends from matrices to bounded linear operators between Hilbert spaces, but finite dimensionality conceals an important condition: every finite-dimensional range is closed. Let T:H→KT:H\to K be bounded. If R(T)\mathcal R(T) is closed, then

H=N(T)⊥⊕N(T),K=R(T)⊕R(T)⊥.H=\mathcal N(T)^\perp\oplus\mathcal N(T), \qquad K=\mathcal R(T)\oplus\mathcal R(T)^\perp.

The restriction of TT to N(T)⊥\mathcal N(T)^\perp is a bijection onto R(T)\mathcal R(T). Its inverse, extended by zero on R(T)⊥\mathcal R(T)^\perp, is a bounded operator T+T^+. It has the same minimum-norm and projection interpretations as a matrix pseudoinverse.

If the range is not closed, a Moore–Penrose inverse can still be defined as a densely defined operator with domain

D(T+)=R(T)⊕R(T)⊥,\mathcal D(T^+)=\mathcal R(T)\oplus\mathcal R(T)^\perp,

but it is unbounded. This changes the existence of least-squares solutions. An element yy may lie in the closure of R(T)\mathcal R(T) without lying in the range itself. Then the infimum of ∥Tx−y∥\|Tx-y\| is zero, yet no vector attains it. Finite-dimensional least squares always has a minimizer because a matrix range is closed.

A concrete example acts on the space ℓ2\ell^2 of square-summable sequences:

T(x1,x2,…)=(x1,x22,x33,…).T(x_1,x_2,\ldots) = \left(x_1,\frac{x_2}{2},\frac{x_3}{3},\ldots\right).

This operator is bounded and injective, and its range is dense but not closed. Its generalized inverse acts as

T+(y1,y2,…)=(y1,2y2,3y3,…)T^+(y_1,y_2,\ldots) = (y_1,2y_2,3y_3,\ldots)

on the domain for which the resulting sequence is square summable. For the unit coordinate vector eke_k, ∥T+ek∥=k\|T^+e_k\|=k, proving that T+T^+ is unbounded. The sequence y=(1,1/2,1/3,…)y=(1,1/2,1/3,\ldots) belongs to ℓ2\ell^2 but not to R(T)\mathcal R(T), since its formal preimage is the nonsummable constant sequence.

For an operator admitting a singular system,

Tvi=σiui,Tv_i=\sigma_i u_i,

the finite-dimensional formula becomes

T+y=∑i⟨y,ui⟩σivi,T^+y=\sum_i\frac{\langle y,u_i\rangle}{\sigma_i}v_i,

where the series converges only for appropriate yy. In particular, its range component must satisfy

∑i∣⟨y,ui⟩∣2σi2<∞.\sum_i\frac{|\langle y,u_i\rangle|^2}{\sigma_i^2}<\infty.

Compact operators of infinite rank have singular values tending to zero, so their inverses on the range are unbounded. This phenomenon appears in integral equations and other problems in functional analysis. Finite discretizations replace unbounded inversion with matrices having increasingly small singular values, linking the operator-theoretic closed-range condition directly to numerical ill-conditioning and the need for regularization.