--- title: Quasi-Unitary Embedding url: https://www.emergentmind.com/topics/quasi-unitary-embedding type: topic --- # Quasi-Unitary Embedding “Quasi-unitary embedding” does not denote a single universally fixed notion across the arXiv literature. Instead, it labels a family of constructions in which an embedding preserves unitary structure exactly, approximately, or in a weakened algebraic sense. In one formal usage, it means comparison of operators or quadratic forms on different Hilbert spaces by identification maps that are unitary up to a controlled error parameter \(\delta\) [2503.22611]. In coding theory, it refers to matrices \(B\in \mathcal M(\mathbb F_{q^2},k\times k)\) satisfying \(BB^\dagger\) diagonal over \(\mathbb F_q^\ast\), which is sufficient to preserve Hermitian dual-containing structure under matrix-product code constructions [2012.15691]. In dynamical systems, the phrase is used in connection with embedding a quasi-periodic cocycle as the Poincaré map of a nearby analytic quasi-periodic linear flow [1202.2911]. In quantum-information-inspired geometry, it describes attempts to embed spaces of quregisters into \(\mathrm{GL}(2^n)\) or \(\mathrm{SU}(2^n)\), a program that succeeds for qubits and fails globally for \(n\ge 2\) [1604.07498]. This variety suggests that the expression functions less as a single definition than as a recurrent structural motif. ## 1. Terminological scope and delimitations Several papers explicitly separate “quasi-unitary embedding” from nearby but distinct notions. In manifold learning, the relevant concept is a **quasi-isometric embedding**: the projection \(p\in\mathcal G(m,k)\) is chosen to minimize the worst secant distortion \[ D_\Sigma(p)=\max_{\sigma\in\Sigma}\left|1-\|p^T\sigma\|_2^2\right|, \] with \(\Sigma\) the set of normalized secants of the sampled point cloud. The paper states that this is “approximate isometry under dimension reduction, not approximate unitarity,” even though both ideas concern controlled deviation from ideal length preservation [1709.01972]. A different, unrelated use of “embedding” occurs in geometric group theory. There, a map between finitely generated groups is a quasi-isometric embedding if there exist constants \(A\ge 1\), \(B\ge 0\) such that \[ \frac{1}{A}d(g,h)-B \le d(f(g),f(h))\le A\,d(g,h)+B, \] and the theory is applied, for example, to embeddings \(T_n\hookrightarrow T_2\) between generalized Thompson groups [1803.00866]. This is a coarse metric notion rather than a unitary or almost-unitary one. Other literatures replace the phrase by more precise terms. One paper studies **quasi-unitary equivalence** of operators and forms [2503.22611]; another studies **embedding** of selfadjoint operator spaces via complete isometry of Werner unitisations and notes that it does **not** use the exact phrase “quasi-unitary embedding” as its main term [2510.24326]. A plausible implication is that encyclopedia treatment is best organized by mathematical mechanism rather than by a single lexical definition. ## 2. Quasi-unitary equivalence on varying Hilbert spaces The most explicit formalization appears in the theory of non-negative self-adjoint operators \(A\) on \(\mathcal H\) and \(\widetilde A\) on \(\widetilde{\mathcal H}\), compared by identification operators \[ J:\mathcal H\to \widetilde{\mathcal H},\qquad J':\widetilde{\mathcal H}\to \mathcal H. \] The paper says that \(J\) is \(\delta\)-quasi-unitary with \(\delta\)-quasi-adjoint \(J'\) if \[ |Jf| \le (1+\delta)|f|,\qquad |(Jf,u)-(f,J'u)| \le \delta |f|\,|u| \] and \[ |f-J'Jf| \le \delta |f|_2,\qquad |u-JJ'u| \le \delta |u|_2. \] The operators are then \(\delta\)-quasi-unitarily equivalent if, in addition, \[ \bigl|\widetilde A Jf - JAf, u\bigr| \le \delta\, |f|_2\,|u|_2. \] When \(\delta=0\), \(J\) is genuinely unitary, \(J'=J^*\), and the construction reduces to ordinary unitary equivalence [2503.22611]. The same framework is formulated at the level of closed, non-negative quadratic forms \(\mathfrak a\) and \(\widetilde{\mathfrak a}\). Besides \(J,J'\), one introduces \[ J_1:\mathcal H_1\to \widetilde{\mathcal H}_1,\qquad J_1':\widetilde{\mathcal H}_1\to \mathcal H_1, \] with \(\mathcal H_1=\operatorname{dom}\mathfrak a\) and \(\|f\|_1^2=\|f\|^2+\mathfrak a(f)\). The forms are \(\delta\)-close if \[ \bigl|\mathfrak a(Jf,u)-\widetilde{\mathfrak a}(f,J'u)\bigr| \le \delta\,\|f\|_1\,\|u\|_1. \] This form-level notion is the paper’s fundamental abstract definition, because form domains are often easier to control than operator domains [2503.22611]. Its principal consequence is generalized norm resolvent convergence on varying Hilbert spaces. If \(\mathfrak a\) and \(\widetilde{\mathfrak a}\) are \(\delta\)-quasi-unitarily equivalent, then \[ \bigl\|\widetilde R(z)J - JR(z)\bigr\| \le C(z)\,\delta, \] for \(R(z)=(A-z)^{-1}\) and \(\widetilde R(z)=(\widetilde A-z)^{-1}\), with \[ C(z)=4\Bigl(1+\frac{|z+1|}{d(z,\sigma(A)\cup \sigma(\widetilde A))}\Bigr). \] The theory further yields convergence of spectral projections, heat operators, eigenvalues, and eigenfunctions, and it proves spectral exactness on compact spectral sets. The framework is stable under composition: if \(A\) and \(\widetilde A\) are \(\delta\)-quasi-unitarily equivalent and \(\widetilde A\) and \(\widehat A\) are \(\varepsilon\)-quasi-unitarily equivalent, then \(A\) and \(\widehat A\) are \(5(\delta+\varepsilon)\)-quasi-unitarily equivalent at operator level, while the form-level transitivity estimate is \(14(\delta+\varepsilon)\) [2503.22611]. Two example classes illustrate the abstraction. For graph approximations of the unit interval, the graph energy forms and the continuum energy form are \(\delta_m\)-quasi-unitarily equivalent with \[ \delta_m=(1+\sqrt2)\,2^{-m}. \] For the Sierpiński gasket, the corresponding rate is \[ \delta_m=\frac{(1+1/3)\sqrt2}{5^{m/2}}. \] The paper also treats manifolds with small obstacles, where the Neumann Laplacian on \(X\setminus B\) is compared to the Laplacian on \(X\) by natural restriction and extension maps [2503.22611]. ## 3. Finite-field quasi-unitary matrices and matrix-product codes In coding theory, a matrix \(B\in \mathcal M(\mathbb F_{q^2},k\times k)\) is **quasi-unitary** if \[ BB^\dagger = \operatorname{diag}(\lambda_1,\ldots,\lambda_k), \qquad \lambda_i \in \mathbb{F}_q^\ast. \] If \(BB^\dagger=I_k\), then \(B\) is unitary. The essential point is that the Hermitian Gram matrix is diagonal over the base field, not necessarily the identity. The paper emphasizes that this diagonal form is enough to preserve Hermitian dual-containing structure in matrix-product codes [2012.15691]. The matrix-product construction is \[ C(A)=[C_1,C_2,\ldots,C_k]A, \] where the constituent codes have the same length and \(A\in\mathcal M(\mathbb F_q,k\times s)\). If \(A\) has full rank and \(C_i\) are \([n,t_i,d_i]_q\), then \[ C(A) \text{ is } [sn,\sum_{i=1}^k t_i,\ge d]_q, \qquad d = \min_{1\le i\le k}\{D_i(A)d_i\}. \] In the Hermitian setting, the dual satisfies \[ ([C_1,\ldots,C_k]A)^{\perp_H} = [C_1^{\perp_H},\ldots,C_k^{\perp_H}](A^{-1})^\dagger. \] If \(A\) is quasi-unitary and \(C_i^{\perp_H}\subseteq C_i\), then \(C(A)\) is Hermitian dual-containing, hence yields a quantum code with parameters \[ [[kn,2\sum_{i=1}^k t_i-kn,\ge d]]_q. \] This is the Hermitian analogue of the quasi-orthogonal Euclidean construction [2012.15691]. The mechanism is algebraically transparent. If \[ AA^\dagger = \operatorname{diag}(\lambda_1,\ldots,\lambda_k), \] then \[ (A^{-1})^\dagger = (AA^\dagger)^{-1}A. \] Thus the Hermitian dual of the matrix-product code is described by the same defining matrix up to diagonal scaling. Since multiplication by nonzero field elements preserves linear-code containment, dual containment survives the embedding effected by \(A\) [2012.15691]. A stronger condition is **NSC** (“non-singular by columns”). For square \(k\times k\) NSC matrices, \[ D_i(A)=k+1-i, \] so the minimum-distance bound becomes \[ d=\min_{1\le i\le k}\{(k+1-i)d_i\}. \] The paper proves that if all leading principal minors of \(AA^\dagger\) are nonzero, then there exists a lower unitriangular matrix \(L\) such that \(LA\) is quasi-unitary; if \(A\) is NSC, then \(LA\) is NSC quasi-unitary. It also proves that for every \(k(1+\alpha)B(\alpha), \] for irrational \(\alpha\) with finite Liouville exponent \(B(\alpha)\). For the almost Mathieu operator, the corresponding condition is \[ \lambda < c\,e^{-B(\alpha)}. \] This use of embedding is thus not merely representational; it serves as a transfer principle between discrete and continuous quasi-periodic dynamics [1202.2911]. ## 5. Quregisters, special linear groups, and entanglement-sensitive failure of unitarity A more geometric and quantum-information-oriented formulation studies embeddings of the unit sphere of \(H_n\cong \mathbb C^{2^n}\) into matrix groups. For \(n=1\), the central map is \[ \Psi_1:S_1(\mathbb C)\to \mathrm{SU}(2),\qquad \begin{bmatrix}x_0\\x_1\end{bmatrix}\mapsto \begin{bmatrix} x_0 & -\overline{x_1}\\ x_1 & \overline{x_0} \end{bmatrix}. \] Because \(|x_0|^2+|x_1|^2=1\), the columns are orthonormal and \[ \det \Psi_1(x)=x_0\overline{x_0}+x_1\overline{x_1}=1, \] so \(\Psi_1\) is a bijection \(S_1(\mathbb C)\to \mathrm{SU}(2)\). The group law of \(\mathrm{SU}(2)\) is therefore transported to the qubit sphere by \[ x\star_1 y := \Psi_1^{-1}\big(\Psi_1(x)\Psi_1(y)\big) \] [1604.07498]. The paper emphasizes that this picture is special to \(n=1\). For \(n=2\), one has \[ H_2=H_1\otimes H_1,\qquad \dim H_2=4, \] and the natural candidate is a tensor-product construction based on \(\Psi_1(c_0)\otimes\Psi_1(c_1)\in \mathrm{SU}(4)\). For a \(2\)-quregister \(x=[x_0,x_1,x_2,x_3]^T\), separability is characterized by \[ x_0x_3=x_1x_2. \] On the separable locus, the chartwise maps \(\Phi_{2ku}\) land in \(\mathrm{SU}(4)\), and the tensor-product construction behaves correctly. Globally, however, the paper states that the embedding obtained by tensoring \(\Psi_1\) fails to determine a bijection between \(S_3(\mathbb C)\) and \(\mathrm{SU}(4)\) [1604.07498]. This failure is quantified by entanglement. Writing \[ t(x)=x_0x_3-x_1x_2,\qquad s(x)=\sqrt{1-4|t(x)|^2}, \] the reduced density matrices have eigenvalues \[ \lambda_0=\frac12(1-s(x)),\qquad \lambda_1=\frac12(1+s(x)), \] so the reduced von Neumann entropy is \[ E(\rho(x)) = -\lambda_0\log_2\lambda_0-\lambda_1\log_2\lambda_1. \] Moreover, the paper defines \[ \nu_k(x)=\|\Phi_{2ku}(x)\|_2-1 \] and derives \[ \nu_k(x)=\sqrt{1+2|t(x)|}-1. \] Thus \[ \nu_k(x)=0 \iff x \text{ is separable}, \] while Bell states are maximally entangled and give \(\nu_k=\sqrt2-1\). In this setting, quasi-unitary embedding is exact on the separable submanifold and fails on entangled states in a way that becomes an entanglement measure consistent with von Neumann entropy [1604.07498]. ## 6. Operator-algebraic and representation-theoretic analogues In selfadjoint operator-space theory, the nearest formal analogue is an embedding compatible with **Werner’s unitisation**. For a completely isometric complete order embedding \[ \phi:E\to F, \] the paper calls \(\phi\) an embedding if the induced map on unitisations \[ \phi^\#:E^\#\to F^\#,\qquad (x,a)\mapsto (\phi(x),a), \] is completely isometric, equivalently a unital complete order embedding. This condition is characterized by several extension properties: every bounded positive functional on each matrix level extends with the same norm; quasistates extend to quasistates; every completely contractive completely positive map into \(M_k(\mathbb C)\) or \(B(K)\) extends with the same cb-norm; and \(\phi\) is a gauge maximal isometry in Russell’s sense. The key distance formula is \[ \operatorname{dist}(x,M_n(E)_+) = \sup_{\psi\in \mathrm{CCP}(M_n(E),\mathbb C)} -\psi(x), \qquad x\in M_n(E)_{sa}. \] The paper also notes explicitly that it does **not** use “quasi-unitary embedding” as its main term, although the unitisation-compatible embedding property is the closest related notion [2510.24326]. A different operator-algebraic direction is given by **unitary correlation sets**. With \[ \mathcal U_{nc}(n) \] the universal C\(^*\)-algebra generated by entries of an abstract unitary matrix, and \[ \mathcal V_n=\operatorname{span}\{1,u_{ij},u_{ij}^*:1\le i,j\le n\}, \] the paper studies tensor-product models \(UC_{qa}(n,m)\), \(UC_{qc}(n,m)\), and compressed sets \(B_t(n,m)\). Its central equivalence is that Connes’ embedding problem has a positive answer if and only if \[ UC_{qa}(n,m)=UC_{qc}(n,m)\ \text{for all } n,m\ge 2, \] equivalently if and only if \[ B_{qa}(n,m)=B_{qc}(n,m)\ \text{for all } n,m\ge 2, \] equivalently if and only if \[ M_n\otimes_{qa} M_n = M_n\otimes_{qc} M_n \] isometrically for all \(n\ge 2\). Here the embedding-type question is encoded in finite-dimensional tensor norms rather than in a single explicit map between Hilbert spaces [1612.02791]. In automorphic representation theory, the analogue is an \(L\)-group embedding for quasi-split unitary groups. The paper defines \[ {}^LU_{E/F}(N)=\mathrm{GL}_N(\mathbb C)\rtimes W_F \] and \[ {}^LG_{E/F}(N) = (\mathrm{GL}_N(\mathbb C)\times \mathrm{GL}_N(\mathbb C))\rtimes W_F, \] and constructs canonical \(L\)-embeddings \[ \widehat{U}_{E/F}(N)\hookrightarrow {}^LG_{E/F}(N), \] depending on a character \(X_k\in Z_E\). The image is identified with the conjugate self-dual parameters of parity \[ n = (-1)^{N-1}k. \] Although the paper does not use the phrase “quasi-unitary embedding” explicitly, the entire endoscopic framework is organized around these structural embeddings and the transfer of parameters, packets, and automorphic representations between unitary groups and twisted forms of \(\mathrm{Res}_{E/F}\mathrm{GL}_N\) [1206.0882]. Taken together, these operator-theoretic, coding-theoretic, dynamical, geometric, and representation-theoretic usages show that “quasi-unitary embedding” is best understood as a family of constructions in which exact unitarity is replaced by a controlled defect, a diagonal Gram form, a unitisation-compatible order structure, or a transfer map into a larger unitary framework. The common theme is not literal uniformity of definition, but the preservation of a unitary-type invariant under embedding. Source: https://www.emergentmind.com/topics/quasi-unitary-embedding