Partial Isometries

• Partial isometries are linear maps that preserve norms on a subspace while mapping its orthogonal complement to zero, serving as a cornerstone in both operator theory and geometric analysis.
• They are characterized through properties such as orthogonal projection, singular value decomposition, and polar decomposition, which clarify their structure in finite and infinite dimensions.
• Applications of partial isometries span operator classification, matrix analysis, semigroup theory, and quantum generalizations, providing robust tools for modeling isometric phenomena across disciplines.

A partial isometry is a linear map that preserves norms on a subspace and vanishes on its orthogonal complement. This notion pervades operator theory, matrix analysis, sub-Riemannian geometry, metric geometry, semigroup theory, and noncommutative geometry, serving as a structural backbone for isometric phenomena in both finite and infinite dimensions. The modern theory encompasses not only operator-theoretic, algebraic, and geometric perspectives, but also combinatorial, probabilistic, and quantum generalizations.

1. Operator-Theoretic and Algebraic Foundations

A linear operator $V$ on a Hilbert space $H$ is a partial isometry if any—and therefore all—of the following hold:

• $V^*V$ is an orthogonal projection (the "initial" projection);
• $VV^*$ is an orthogonal projection (the "final" projection);
• $VV^*V=V$ and $V^*VV^*=V^*$;
• $V$ restricts to a unitary map from $\operatorname{Ran}(V^*V)$ onto $\operatorname{Ran}(VV^*)$, acting as zero on the orthogonal complement.

In finite dimensions, a matrix $A\in M_n(\mathbb{C})$ is a partial isometry if $H$0 and $H$1 are orthogonal projections. The singular value decomposition (SVD) immediately yields that a partial isometry has all singular values in $H$2; thus, $H$3 for some $H$4 unitary and rank $H$5 (Garcia et al., 2019).

In $H$6-algebras, an element $H$7 is a partial isometry if $H$8 and $H$9 are projections. The canonical polar decomposition map associates, to each $V^*V$0 with closed range, the unique partial isometry $V^*V$1 such that $V^*V$2, $V^*V$3, $V^*V$4 (Brown, 2016). Partial isometries are exactly the extreme points of the unit ball in $V^*V$5-algebras when for $V^*V$6, one has $V^*V$7.

Partial isometries play a central role in the structure theory of bounded linear operators, model spaces, and functional calculus.

2. Metric, Homotopy, and Continuity Properties in $V^*V$8-Algebras

Norm estimates and geometric properties of the set of partial isometries are sharply characterized. Brown established the inequality: $V^*V$9 for partial isometries $VV^*$0 and $VV^*$1 in a $VV^*$2-algebra; this is sharp even for projections. If $VV^*$3 and $VV^*$4 are extremal partial isometries and $VV^*$5, a stronger inequality holds: $VV^*$6 The constants are optimal (Brown, 2016).

For homotopy, if $VV^*$7 then there exists a norm-continuous path of partial isometries connecting $VV^*$8 to $VV^*$9; if both are extremal, the bound is $VV^*V=V$0 instead of $VV^*V=V$1. These thresholds are again sharp.

The continuity points of the polar decomposition map $VV^*V=V$2 are characterized: $VV^*V=V$3 is a continuity point if and only if a certain “defect corner” $VV^*V=V$4 contains no nonzero closed-range elements (where $VV^*V=V$5, $VV^*V=V$6). In particular, $VV^*V=V$7 is a continuity point (and interior to the domain) if and only if $VV^*V=V$8 is quasi-invertible, equivalently $VV^*V=V$9 is extremal (Brown, 2016).

3. Matrix-theoretic Structure, Numerical Range, and Classification

Matrix models for partial isometries and their iterations ("power partial isometries") lead to precise canonical forms. An $V^*VV^*=V^*$0 matrix $V^*VV^*=V^*$1 is a partial isometry if and only if, up to unitary similarity, $V^*VV^*=V^*$2 with $V^*VV^*=V^*$3 (Gau et al., 2013). For higher powers, a block-shift form in direct sum with a final block encodes the entire sequence of $V^*VV^*=V^*$4, $V^*VV^*=V^*$5, ..., $V^*VV^*=V^*$6 being partial isometries.

A central concept is the ascent $V^*VV^*=V^*$7: the minimal $V^*VV^*=V^*$8 for which $V^*VV^*=V^*$9, i.e., the size of the largest Jordan block at zero. If $V$0, $V$1, ..., $V$2 are all partial isometries up to the ascent, then the numerical range $V$3 is a circular disk centered at $V$4 if and only if $V$5 is unitarily similar to a direct sum of Jordan blocks of sizes up to $V$6.

A correspondence between operator algebraic and block-structural properties connects partial isometries to Kronecker products, essential numerical range phenomena, and specific subclasses such as $V$7-matrices (Gau et al., 2013). For $V$8 nilpotent partial isometries, particularly of rank $V$9 or $\operatorname{Ran}(V^*V)$0, the numerical range is a disk; this holds generically in dimension $\operatorname{Ran}(V^*V)$1, but fails for $\operatorname{Ran}(V^*V)$2 where non-circularity and failure of genericity appear (He et al., 2022). The critical conjecture is that if $\operatorname{Ran}(V^*V)$3 is a disk, it must be centered at $\operatorname{Ran}(V^*V)$4 for any partial isometry—a statement proved for $\operatorname{Ran}(V^*V)$5 and all rank $\operatorname{Ran}(V^*V)$6, but open in general.

Classification under similarity and unitary similarity leverages the SVD, Jordan canonical form, and Livšic characteristic functions. Notably, every partial isometry splits into completely non-unitary and unitary summands, and unitary equivalence for defect-one operators reduces to characteristic polynomial equality (Garcia et al., 2019). Trace invariants and Livšic functions classify in higher dimensions.

4. Partial Isometries in Geometric and Metric Contexts

The framework extends beyond operator theory and matrix analysis. In metric geometry, a partial isometry between finite metric spaces is a bijection between subsets that preserves distances (Hubička et al., 2018). The extension property for partial isometries (EPPA) asserts that every partial isometry of a finite metric space extends to a global isometry of a larger space. This is realized through combinatorial constructions involving edge-labeled graphs, Möbius strip gadgets, and shortest-path completions.

In sub-Riemannian geometry, a $\operatorname{Ran}(V^*V)$7-map $\operatorname{Ran}(V^*V)$8 is called a partial isometry if $\operatorname{Ran}(V^*V)$9 is isometric, i.e., $\operatorname{Ran}(VV^*)$0. The generalization of the Nash-Kuiper theorem asserts that every strictly $\operatorname{Ran}(VV^*)$1-short $\operatorname{Ran}(VV^*)$2-immersion $\operatorname{Ran}(VV^*)$3 into a Riemannian manifold of dimension $\operatorname{Ran}(VV^*)$4 is $\operatorname{Ran}(VV^*)$5-homotopic to a partial isometry; if $\operatorname{Ran}(VV^*)$6, every sub-Riemannian manifold admits a partial isometry into Euclidean space (Datta, 2010).

This unifies operator-theoretic and geometric perspectives, linking path-isometric embeddings, leafwise isometries in foliated manifolds, and convex integration techniques.

5. Semigroup Theory, Inverse Semigroups, and Quantum Generalizations

The set of partial isometries is closed under composition only in special circumstances. The Halmos-Wallen theorem characterizes power partial isometries by decompositions into unitary, unilateral shift, backward shift, and finite Jordan (truncated shift) blocks (Huef et al., 2017). In self-adjoint semigroups of partial isometries, these coincide with faithful *-representations of abstract inverse semigroups. The structure theorem for such semigroups expresses each partial isometry as a generalized weighted composition operator on $\operatorname{Ran}(VV^*)$7 spaces, exhibiting both measure-class transport and measurable fields of unitaries (Popov et al., 2013).

Self-adjoint closure of semigroups preserves partial isometry structure precisely when the range projections generate an abelian von Neumann algebra of uniform finite multiplicity (Bernik et al., 2014). Failure of this condition leads to semigroups whose self-adjoint closure contains non-partial isometries.

Quantum analogues generalize classical semigroups to "liberated" and "half-liberated" versions, governed by universal $\operatorname{Ran}(VV^*)$8-algebras on generators subject to partial isometry relations and higher commutation constraints. These interpolate between classical and free probabilistic laws via the Bercovici-Pata bijection and permit a detailed Weingarten calculus (Banica, 2014).

6. Differential-Geometric and Metric Aspects

The space of partial isometries $\operatorname{Ran}(VV^*)$9 on a Hilbert space inherits a homogeneous Finsler geometry, realized as a quotient of $A\in M_n(\mathbb{C})$0 by the stabilizer of a reference partial isometry (Andruchow, 2021). The tangent space at each point consists of elements of the form $A\in M_n(\mathbb{C})$1 with $A\in M_n(\mathbb{C})$2 self-adjoint. The associated Finsler norm informs geodesic distance and the geometry of minimal-length paths, which have explicit descriptions via two-sided exponentials. This theory relates the geometry of $A\in M_n(\mathbb{C})$3 to unitary orbits and the Grassmannian, providing tools for isometric analysis in operator and projection manifolds.

In pseudo-Hilbert (Loynes $A\in M_n(\mathbb{C})$4-) spaces, partial gramian isometries generalize the classical notion. Two accessible subspaces $A\in M_n(\mathbb{C})$5 are initial and final spaces of a partial gramian isometry provided the norm of the difference of their associated gramian projections is strictly less than $A\in M_n(\mathbb{C})$6 (Gaşpar et al., 2013). This sufficient condition enables explicit construction and relates to perturbational analysis of projections even in the indefinite inner product case.

7. Applications, Open Problems, and Further Directions

Partial isometries underpin dilation theory, model spaces, extension and completion problems, and have function-theoretic manifestations through Livšic characteristic functions in de Branges–Rovnyak spaces. Their role in combinatorial geometry is fundamental to the automorphism groups of universal spaces, such as the Urysohn space (Hubička et al., 2018). Active research areas include:

• Structure and classification of products and powers of partial isometries;
• Quantitative geometry of numerical ranges and field of values for higher-dimensional partial isometries;
• Robustness of semigroup closure and stability under non-uniformity in von Neumann algebra multiplicities;
• Connections between free and classical probabilities via liberation of quantum partial isometries (Banica, 2014);
• Existence and explicit construction of partial isometries under perturbation in indefinite metric spaces.

Major conjectures remain, such as conditions for numerical ranges to be disks centered at zero in all dimensions and the extension property for other finite structures beyond metric spaces. These underline the central, unifying role of partial isometries across operator theory, geometry, combinatorics, and noncommutative probability.