A Kernel Approach to the Stinespring and Kraus Representations

James Tian · 2025

Read A Kernel Approach to the Stinespring and Kraus Representations by James Tian (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-00256.txt.

Introduction

Completely positive maps describe how quantum systems change while probabilities stay sensible. James Tian gives a self-contained derivation of two structure theorems for such maps in 2025, the Stinespring and Kraus representations. The derivation uses only scalar positive-definite kernels.

Textbook proofs of these theorems often rely on heavier machinery from operator algebras. A kernel-only route makes the results reachable for readers with a more modest background. The theorems underpin quantum information theory, where Kraus operators are the standard way to model noise in quantum devices. Kraus operators are the form in which noise models are usually handed to engineers. Stinespring's theorem explains why those operators must exist in the first place.

Catalog record

CREATORJames Tian
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE14 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-00256

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Kernel Approach to the Stinespring and Kraus Representations James Tian Abstract . We give a self-contained derivation of the Stinespring and Kraus structure theorems for completely positive maps using only scalar positive-definite kernels. 2000 Mathematics Subject Classification Primary: 46E22; Secondary: 47A20, 47L25, 1. Introduction Completely positive (CP) maps describe a fundamental class of maps between algebras of bounded operators, specifically those that preserve positivity at every matrix level. Two classical theorems capture their structure: Stinespring’s dilation theorem, which realizes any CP map as a compression of a representation on a larger Hilbert space, and Kraus’s theorem, which expresses the same map as a finite or countable sum of conjugations by bounded operators (see, e.g., [ Sti55 , Kra71 , Pau02 ] ). These results, although standard, are often proved using heavy machinery, including C^{*} -algebra representation theory, the Gelfand-Naimark-Segal construction, or matrix algebra arguments. This paper shows that both theorems follow directly from an elementary geometric idea involving two scalar positive definite kernels. The approach proceeds in two stages: 1. From a CP map, we define a scalar kernel whose positivity follows directly from complete positivity. The associated reproducing kernel Hilbert space automatically contains the data of the Stinespring dilation, including the dilation space and the * -representation. 2. A second kernel, built from the matrix units of the underlying Hilbert space, reveals a canonical unitary identification of the dilation space with a concrete tensor product. Under this identification, the abstract Stinespring form becomes the familiar tensor-product expression, and the Kraus operators appear immediately by expanding along an orthonormal basis of the auxiliary space. The argument requires only elementary facts about positive definite kernels and Hilbert space geometry. It parallels familiar constructions in classical analysis, such as the reproducing kernel of a Hilbert space of functions or the Kolmogorov factorization of a covariance function. In this way the kernel perspective clarifies the geometry behind complete positivity: it shows how the dilation space is generated by the “columns” of the map, how tensor structure emerges naturally, and how the Kraus operators reflect orthogonal components of this structure. Throughout the paper, let H be a complex Hilbert space and \mathscr{B}\left(H\right) the space of bounded operators on H . All inner products are linear in the second variable. 2. The Stinespring Dilation We begin by building the dilation space from a scalar kernel defined directly by the CP map. The kernel will be shown to be positive definite (p.d.), and its reproducing kernel Hilbert space (RKHS) will serve as the dilation space. Theorem 2.1 (Stinespring dilation) . Let \varphi:\mathscr{B}\left(H\right)\rightarrow\mathscr{B}\left(H\right) be a CP map. There exists a minimal Stinespring dilation \left(\pi,H_{K},V\right) such that \varphi\left(S\right)=V^{*}\pi\left(S\right)V where: (1) H_{K} is the RKHS associated with the \mathbb{C} -valued kernel K\left(\left(S,a\right),\left(T,b\right)\right):=\left\langle a,\varphi\left(S^{*}T\right)b\right\rangle. (2) V:H\rightarrow H_{K} is the bounded linear operator, Vb:=K_{\left(I,b\right)}=K\left(\cdot,\left(I,b\right)\right). (3) \pi:\mathscr{B}\left(H\right)\rightarrow\mathscr{B}\left(H_{K}\right) is the unital * -representation defined on the dense span of sections by \pi\left(S\right)K_{\left(T,b\right)}:=K_{\left(ST,b\right)}. Proof. Complete positivity of \varphi ensures K is a scalar-valued positive definite (p.d.) kernel on \left(\mathscr{B}\left(H\right)\times H\right)\times\left(\mathscr{B}\left(H\right)\times H\right) . Indeed, for any finite family \left\{\left(S_{i},a_{i}\right)\right\}^{n}_{i=1} in \mathscr{B}\left(H\right)\times H , and \left\{c_{i}\right\}^{n}_{i=1} in \mathbb{C} , setting b_{i}=c_{i}a_{i} , then \displaystyle\sum\nolimits_{i,j}\overline{c_{i}}c_{j}K\left(\left(S_{i},a_{i}\right),\left(S_{j},a_{j}\right)\right) \displaystyle=\sum\nolimits_{i,j}\overline{c_{i}}c_{j}\left\langle a_{i},\varphi\left(S^{*}_{i}S_{j}\right)a_{j}\right\rangle \displaystyle=\sum\nolimits_{i,j}\left\langle b_{i},\varphi\left(S^{*}_{i}S_{j}\right)b_{j}\right\rangle \displaystyle=\left\langle\left(b_{1},\dots,b_{n}\right),\left[\varphi\left(S^{*}_{i}S_{j}\right)\right]\left(b_{1},\dots,b_{n}\right)\right\rangle\geq 0 because the block matrix \left[\varphi\left(S^{*}_{i}S_{j}\right)\right] is positive in M_{n}\left(\mathscr{B}\left(H\right)\right) . Let H_{K} be the corresponding RKHS with sections K_{\left(T,b\right)}:=\left\langle\cdot,K\left(\cdot,T\right)b\right\rangle. By definition, H_{K}=\overline{span}\left\{K_{\left(T,b\right)}:T\in\mathscr{B}\left(H\right),\>b\in H\right\}. In particular, it consists of complex-valued functions on \mathscr{B}\left(H\right)\times H with reproducing kernel K . See, e.g., [ Aro50 , Ped58 , Sza21 ] . Define V:H\rightarrow H_{K} by Vb=K_{\left(I,b\right)} , and \pi:\mathscr{B}\left(H\right)\rightarrow\mathscr{B}\left(H_{K}\right) by \pi\left(S\right)K_{\left(T,b\right)}=K_{\left(ST,b\right)} . To see \pi\left(S\right) is well defined and bounded, take a finite sum u=\sum_{k}c_{k}K_{\left(T_{k},b_{k}\right)} . Then \displaystyle\left\|\pi\left(S\right)u\right\|^{2} \displaystyle=\Big\|\sum\nolimits_{k}c_{k}K_{\left(ST_{k},b_{k}\right)}\Big\|^{2} \displaystyle=\sum\nolimits_{k,l}\overline{c_{k}}c_{l}K\left(\left(ST_{k},b_{k}\right),\left(ST_{l},b_{l}\right)\right) \displaystyle=\sum\nolimits_{k,l}\overline{c_{k}}c_{l}\left\langle b_{k},\varphi\left(T^{*}_{k}S^{*}ST_{l}\right)b_{l}\right\rangle. Since S^{*}S\leq\left\|S\right\|^{2}I in \mathscr{B}\left(H\right) and the map A\mapsto T^{*}_{k}AT_{k} is positive, we have \left[T^{*}_{k}S^{*}ST_{l}\right]\leq\left\|S\right\|^{2}\left[T^{*}_{k}T_{l}\right] in M_{n}\left(\mathscr{B}\left(H\right)\right) . Applying the CP map id_{M_{n}}\otimes\varphi and pairing with \left(b_{k}\right)\in H^{\oplus n} yields \sum\nolimits_{k,l}\overline{c_{k}}c_{l}\left\langle b_{k},\varphi\left(T^{*}_{k}S^{*}ST_{l}\right)b_{l}\right\rangle\leq\left\|S\right\|^{2}\sum\nolimits_{k,l}\overline{c_{k}}c_{l}\left\langle b_{k},\varphi\left(T^{*}_{k}T_{l}\right)b_{l}\right\rangle=\left\|S\right\|^{2}\left\|u\right\|^{2}. Hence \left\|\pi\left(S\right)\right\|\leq\left\|S\right\| . One checks that \pi\left(S_{1}\right)\pi\left(S_{2}\right)=\pi\left(S_{1}S_{2}\right) on the dense set of sections. Since K is Hermitian, \pi\left(S\right)^{*}=\pi\left(S^{*}\right) . Hence \pi is a unital * -representation of \mathscr{B}\left(H\right) . For all S\in\mathscr{B}\left(H\right) and a,b\in H , \left\langle a,\varphi\left(S\right)b\right\rangle=K\left(\left(I,a\right),\left(S,b\right)\right)=\left\langle Va,\pi\left(S\right)Vb\right\rangle_{H\_\{K\}}, that is, \varphi\left(S\right)=V^{*}\pi\left(S\right)V . Finally, the dilation is minimal by construction, as the spanning set for H_{K} is precisely \left\{K_{(S,a)}\right\}=\left\{\pi\left(S\right)Va\right\} . ∎ 3. The Kraus Representation The dilation obtained in the previous section is still abstract: it lives inside the reproducing kernel Hilbert space generated by the first kernel. Our next goal is to uncover the concrete tensor product structure hidden inside that space. This structure is responsible for the familiar form of a completely positive map as a sum of operator conjugations. To reveal it, we introduce a second scalar positive definite kernel, built from the matrix units of the underlying Hilbert space. This auxiliary kernel generates a new reproducing kernel Hilbert space which plays the role of the “ancillary” tensor factor. A simple computation then shows that the original dilation space H_{K} is unitarily equivalent to an honest tensor product of Hilbert spaces. Once this identification is made, the Kraus representation follows immediately by expanding the dilation operator along an orthonormal basis of the auxiliary space. Theorem 3.1 (canonical tensor structure) . Let \left(\pi,H_{K},V\right) be the minimal Stinespring dilation from Theorem 2.1 . Fix an orthonormal basis (ONB) \left\{e_{i}\right\} in H , and let E_{ij}=\left|e_{i}\left\rangle\right\langle e_{j}\right| . Let H_{\hat{K}} be the RKHS of the kernel \hat{K}\left(\left(i,a\right),\left(j,b\right)\right):=\left\langle a,\varphi\left(E_{ij}\right)b\right\rangle. Then the Stinespring space H_{K} is canonically unitary to H\otimes H_{\hat{K}} via an isomorphism U that identifies: (1) The representation \pi\left(S\right) with S\otimes I_{H_{\hat{K}}} . (2) The operator V with W:H\rightarrow H\otimes H_{\hat{K}} , where Wa:=\sum\nolimits_{i}e_{i}\otimes\hat{K}_{\left(i,a\right)}. This identification transforms the abstract Stinespring form \varphi(S)=V^{*}\pi(S)V into the canonical tensor product form \varphi(S)=W^{*}\left(S\otimes I_{H_{\hat{K}}}\right)W. Furthermore, expanding W in an ONB \{f_{\alpha}\} of H_{\hat{K}} yields the Kraus form \varphi\left(S\right)=\sum V^{*}_{\alpha}SV_{\alpha}, where V_{\alpha}:=\left(I_{H}\otimes\left\langle f_{\alpha},\cdot\right\rangle\right)W. Proof. The proof proceeds in two steps. First, we establish the unitary isomorphism U . Second, we transport the representation (\pi,V) along U to derive the tensor product and Kraus forms. See the diagram below. \textstyle{H\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces} \scriptstyle{\varphi\left(S\right)} \scriptstyle{V} \scriptstyle{W} \textstyle{H_{K}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces} \scriptstyle{U} \scriptstyle{\pi\left(S\right)} \textstyle{H\otimes H_{\hat{K}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces} \scriptstyle{S\otimes I_{H_{\hat{K}}}} \textstyle{H\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces} \scriptstyle{V} \scriptstyle{W} \textstyle{H_{K}\ignorespaces\ignorespaces\ignorespaces\ignorespaces} \scriptstyle{U} \textstyle{H\otimes H_{\hat{K}}} Step 1. Fix an ONB \left\{e_{i}\right\}_{i\in I} in H , and let E_{ij}=\left|e_{i}\left\rangle\right\langle e_{j}\right| . The block matrix [\varphi(E_{ij})] is positive by complete positivity, so \hat{K} is a scalar-valued p.d. kernel defined on \left(I\times H\right)\times\left(I\times H\right) . Let H_{\hat{K}} be the associated RKHS with sections \hat{K}_{\left(i,a\right)} , so that H_{\hat{K}}=\overline{span}\mathop{\{}\hat{K}_{\left(i,a\right)}:i\in I,\>a\in H\mathop{\}}. We now relate the kernel K from Theorem 2.1 to this new kernel \hat{K} . For any \left(S,a\right) and \left(T,b\right) , \displaystyle K\left(\left(S,a\right),\left(T,b\right)\right) \displaystyle=\left\langle a,\varphi\left(S^{*}T\right)b\right\rangle \displaystyle=\sum\nolimits_{i,j}\left\langle e_{i},S^{*}Te_{j}\right\rangle\left\langle a,\varphi\left(E_{ij}\right)b\right\rangle \displaystyle=\sum\nolimits_{i,j}\left\langle Se_{i},Te_{j}\right\rangle\left\langle\hat{K}_{\left(i,a\right)},\hat{K}_{\left(j,b\right)}\right\rangle_{H_{\hat{K}}} \displaystyle=\left\langle\sum\nolimits_{i}Se_{i}\otimes\hat{K}_{\left(i,a\right)},\sum\nolimits_{j}Te_{j}\otimes\hat{K}_{\left(j,b\right)}\right\rangle_{H\otimes H_{\hat{K}}}. This calculation shows that the linear map U:H_{K}\rightarrow H\otimes H_{\hat{K}} , defined on the spanning set of sections by U\left(K_{\left(T,b\right)}\right):=\sum\nolimits_{i}Te_{i}\otimes\hat{K}_{\left(i,b\right)} is isometric, with dense range in H\otimes H_{\hat{K}} . Thus, U is unitary, and H_{K}\simeq H\otimes H_{\hat{K}}. Step 2. We now transport the Stinespring representation (\pi,V) from H_{K} to H\otimes H_{\hat{K}} . Using U , we get \displaystyle U\pi\left(S\right)U^{*}\left(\sum\nolimits_{i}Te_{i}\otimes\hat{K}_{\left(i,b\right)}\right) \displaystyle=U\left(\pi\left(S\right)K_{\left(T,b\right)}\right) \displaystyle=U\left(K_{\left(ST,b\right)}\right) \displaystyle=\sum\nolimits_{i}STe_{i}\otimes\hat{K}_{\left(i,b\right)} \displaystyle=\left(S\otimes I\right)\left(\sum\nolimits_{i}Te_{i}\otimes\hat{K}_{\left(i,b\right)}\right). So \pi\left(S\right) is unitarily equivalent to S\otimes I_{H_{\hat{K}}} . Also, the operator V becomes W:=UV : Wa:=U\left(Va\right)=U\left(K_{\left(I,a\right)}\right)=\sum\nolimits_{i}e_{i}\otimes\hat{K}_{\left(i,a\right)}. Therefore, \displaystyle\varphi\left(S\right) \displaystyle=V^{*}\pi\left(S\right)V=\left(UV\right)^{*}\left(U\pi\left(S\right)U^{*}\right)\left(UV\right) \displaystyle=W^{*}\left(S\otimes I\right)W. Lastly, let \left\{f_{\alpha}\right\}_{\alpha\in\Lambda} be an ONB of H_{\hat{K}} . Define the operators V_{\alpha}:H\rightarrow H by V_{\alpha}:=\left(I_{H}\otimes\left\langle f_{\alpha},\right\rangle_{H_{\hat{K}}}\right)W. This means V_{\alpha}a is the vector

Read the complete text