Kraus Operator-Sum Representation

• Kraus operator-sum representation is a formalism that models open quantum systems via a set of operators ensuring complete positivity and trace preservation.
• It encodes the effects of system-environment interactions, including decoherence, dissipation, and conditional evolution, through discrete operator branches.
• Its derivation from unitary dilation and partial trace provides a practical framework for quantum process tomography and noise modeling in diverse quantum channels.

The Kraus operator-sum representation is the standard reduced description of open-system quantum dynamics obtained when a system interacts unitarily with an environment and the environment is subsequently discarded. For a system state $\rho$, the evolved state is written as

$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$

with the completeness condition $\sum_k K_k^\dagger K_k=I$ for trace-preserving evolution. In this form, the operators $K_k$ act only on the system Hilbert space, yet encode the effect of joint system-environment dynamics, decoherence, dissipation, and conditional branches of evolution. The representation is the natural discrete-time generalization of unitary dynamics for non-isolated systems and is the fixed-time counterpart of continuous-time generator descriptions such as the Gorini–Kossakowski–Sudarshan–Lindblad form (Maziero, 2015, Kuramochi, 2024).

1. Definition, channel structure, and operational meaning

For a completely positive trace-preserving map in Schrödinger picture, the operator-sum form is

$$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$

The same condition appears in several equivalent contexts: as the trace-preserving constraint for learned process models, as the unitality condition in the Heisenberg picture, and as the channel normalization inherited from a Stinespring dilation (Ahmed et al., 2022, Andersson, 2015). If the normalization is relaxed to

$\sum_k K_k^\dagger K_k\le I,$

the map is trace-non-increasing rather than trace-preserving, corresponding to a conditional process or selected outcome (Maziero, 2015).

The physical interpretation is that the $K_k$ label environmentally conditioned or measurement-like branches of the reduced evolution. A single unitary acting only on the system generally cannot represent irreversible processes such as spontaneous emission, whereas the operator-sum form can do so by summing over branches generated by system-environment correlations and partial trace (Maziero, 2015). In the Heisenberg picture the same channel is written as

$\Phi(A)=\sum_k K_k^* A K_k,$

with $\sum_k K_k^*K_k=\mathbf 1$ expressing unitality of $\Phi$ and trace preservation of the dual Schrödinger-picture map (Andersson, 2015).

The representation is not merely formal. It is used as an explicit model class for tomography, for continuous measurement theory, for Gaussian channels, and for data-driven noise learning, precisely because complete positivity is automatic once the map is written in this quadratic form (Ahmed et al., 2022, Sakai et al., 22 Apr 2026).

2. Derivation from unitary dilation and partial trace

A standard derivation begins by assuming that system and environment together form a closed system undergoing unitary dynamics, while the initial state is factorized: $\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$0 The reduced dynamics is then

$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$1

or, in the pure-environment simplification,

$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$2

Expanding in a product basis and carrying out the partial trace yields matrix elements

$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$3

from which one obtains directly

$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$4

This construction also proves that the $\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$5 are linear operators on the system Hilbert space and that the map is positive and trace-preserving (Maziero, 2015).

This derivation is intentionally restricted to factorized initial states. The same source remarks that initially correlated system-environment states are tied to the issue of maps that may fail to be completely positive, so the elementary derivation avoids that complication. A mixed initial environment may be purified, enlarging the environment while keeping the same reduced dynamics (Maziero, 2015).

Microscopic constructions need not stop at the elementary dilation argument. For a bosonic reservoir with no initial system-reservoir correlations, the reduced non-Markovian dynamics can be written exactly as a Kraus map involving an infinite family of time-dependent Kraus matrices, with coefficients determined by reservoir pair-correlation functions (Wonderen et al., 2018). In Keldysh real-time theory, the same operator-sum structure emerges by grouping diagrams into measurement-conditioned completely positive pieces and obtaining Kraus operators by “cutting” suitable diagram groups in half (Reimer et al., 2018).

3. Structural properties: nonuniqueness, rank, and operator bases

A Kraus decomposition is not unique. If

$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$6

for a unitary matrix $\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$7, then $\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$8 represents the same physical map. In the dilation picture, this freedom corresponds to an additional unitary rotation on the environment after the joint evolution (Maziero, 2015). The nonuniqueness is also operationally important in tomography and channel comparison, where one often passes to Choi form precisely because the Choi matrix is unique even when the Kraus set is not (Sakai et al., 22 Apr 2026).

For a system of dimension $\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,$9, the minimum number of Kraus operators is the Choi rank $\sum_k K_k^\dagger K_k=I$0, and the maximum required is $\sum_k K_k^\dagger K_k=I$1 when $\sum_k K_k^\dagger K_k=I$2. Equivalently, no more than $\sum_k K_k^\dagger K_k=I$3 Kraus operators are needed for a $\sum_k K_k^\dagger K_k=I$4-dimensional system, while unitary channels have $\sum_k K_k^\dagger K_k=I$5 (Maziero, 2015, Ahmed et al., 2022). This low-rank viewpoint is central in compressed or learned process models, where one chooses a small Kraus rank to avoid direct estimation of a full $\sum_k K_k^\dagger K_k=I$6 Choi matrix (Ahmed et al., 2022).

The representation also admits concrete operator bases. On $\sum_k K_k^\dagger K_k=I$7, the Weyl shift-phase operators

$\sum_k K_k^\dagger K_k=I$8

form a basis for the full operator space. In an explicit system-environment derivation, these same operators arise naturally on the system side of the unitary evolution, and the paper identifies

$\sum_k K_k^\dagger K_k=I$9

as a possible normalized set of operation elements in an operator-sum model (Wilmott, 2011).

For single-mode Bosonic Gaussian channels, the operator-sum representation can likewise be developed from a system-ancilla unitary dilation. The Kraus operators are obtained from ancilla Fock-basis matrix elements

$K_k$0

and discrete linearly independent Kraus sets can be constructed for noisy Gaussian channels. In the entanglement-breaking case, rank-one Kraus operators emerge naturally, whereas for the quantum-limited amplifier and attenuator families the paper states that there is not even one finite-rank operator in the entire linear span of the Kraus operators (Ivan et al., 2010).

4. Relation to continuous-time generators and master equations

A fixed Kraus map describes a single completely positive map at a given time, whereas continuous-time Markovian evolution is described by a generator $K_k$1. The relation between the two becomes explicit by examining the Kraus operators of an infinitesimal channel $K_k$2. In the finite-dimensional semigroup setting, one writes

$K_k$3

decomposes each infinitesimal Kraus operator into a scalar trace part and a traceless part $K_k$4, and finds the decisive scaling

$K_k$5

Taking standard parts of $K_k$6 and $K_k$7 yields jump operators $K_k$8 and Hamiltonian $K_k$9, and the generator assumes the Gorini–Kossakowski–Sudarshan–Lindblad form

$$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$0

This derivation makes explicit that the dissipative terms survive at first order only because the nontrivial Kraus components scale as $$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$1, not as $$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$2 (Kuramochi, 2024).

The converse direction is also constructive. Starting from a generator already in GKSL form, one can build small-time Kraus operators

$$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$3

and thereby recover complete positivity and unitality of the semigroup (Kuramochi, 2024).

Exact model-specific master-equation reconstructions are also available. For an exactly solvable spin-star model with layered environment, the reduced qubit map has matrix representation

$$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$4

leading to an explicit time-dependent Pauli-channel Kraus representation and, from the same map, exact time-local and memory-kernel master equations of TCL and Nakajima–Zwanzig type (Mahdian et al., 2013). More recently, a closed-form approximate Kraus-map solution for general GKSL dynamics has been derived by treating the Hamiltonian exactly while remaining linear in the dissipator; the resulting Krausized approximation is not exactly trace-preserving at finite truncation, with

$$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$5

but it extends the Kraus description to regimes with arbitrarily strong coherent driving and weak dissipation (Chishti et al., 11 Mar 2026).

5. Representative channels and explicit formulas

A canonical example is the amplitude-damping channel describing a two-level atom interacting with the vacuum electromagnetic field. With system basis $$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$6 and environment initially in $$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$7, the joint phenomenological evolution is

$$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$8

where $$\mathcal{E}(\rho)=\sum_k K_k\rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k=I.$$9 and $\sum_k K_k^\dagger K_k\le I,$0. The resulting Kraus operators are

$\sum_k K_k^\dagger K_k\le I,$1

with $\sum_k K_k^\dagger K_k\le I,$2 and $\sum_k K_k^\dagger K_k\le I,$3. For a qubit Bloch vector $\sum_k K_k^\dagger K_k\le I,$4, the evolved vector is

$\sum_k K_k^\dagger K_k\le I,$5

and the density matrix transforms as

$\sum_k K_k^\dagger K_k\le I,$6

Thus the excited-state population decays as $\sum_k K_k^\dagger K_k\le I,$7, coherence decays as $\sum_k K_k^\dagger K_k\le I,$8, and the asymptotic state is $\sum_k K_k^\dagger K_k\le I,$9. With the $K_k$0-type coherence measure $K_k$1, one has

$K_k$2

so coherence decays monotonically to zero (Maziero, 2015).

The operator-sum framework also captures unital channels. In the spin-star model, the exact reduced qubit dynamics is a Pauli-diagonal unital channel whose Kraus operators are proportional to $K_k$3, with coefficients determined by the exact bath-dependent functions $K_k$4 and $K_k$5 (Mahdian et al., 2013).

A different microscopic example is the generalized one-qubit depolarizing channel derived from a Hamiltonian model with three bosonic bath sectors coupled through $K_k$6. The resulting master equation yields a unital but generally anisotropic channel whose action on the Pauli basis is

$K_k$7

and whose Kraus operators reduce to the standard depolarizing-channel Kraus operators only in the special regime $K_k$8 and $K_k$9 (Arsenijevic et al., 2015).

6. Methodological and applied roles

In quantum process tomography, the Kraus representation can be used as the primary learnable model rather than as a post hoc channel description. A process is parameterized directly as

$\Phi(A)=\sum_k K_k^* A K_k,$0

with the $\Phi(A)=\sum_k K_k^* A K_k,$1 stacked into a single matrix

$\Phi(A)=\sum_k K_k^* A K_k,$2

so that trace preservation becomes the Stiefel-manifold constraint $\Phi(A)=\sum_k K_k^* A K_k,$3. This converts physicality into a constrained optimization problem: complete positivity is automatic from the operator-sum form, and trace preservation is enforced exactly throughout gradient descent (Ahmed et al., 2022).

A closely related idea appears in hardware noise learning. Each noise channel is represented by Kraus operators extracted from a Stinespring unitary

$\Phi(A)=\sum_k K_k^* A K_k,$4

which guarantees complete positivity and trace preservation by construction. In the reported framework, such channels are attached to native gate types, crosstalk interactions, state preparation, and measurement, and are optimized end-to-end against observed output distributions while the noisy circuits are simulated as matrix product density operators (Sakai et al., 22 Apr 2026).

The representation is also central in continuous-measurement theory. For continuous isotropic monitoring of $\Phi(A)=\sum_k K_k^* A K_k,$5 in a spin-$\Phi(A)=\sum_k K_k^* A K_k,$6 system, each measurement record determines a time-dependent Kraus operator $\Phi(A)=\sum_k K_k^* A K_k,$7. These operators lie on a submanifold diffeomorphic to $\Phi(A)=\sum_k K_k^* A K_k,$8, while the associated POVM elements $\Phi(A)=\sum_k K_k^* A K_k,$9 belong to the symmetric space $\sum_k K_k^*K_k=\mathbf 1$0, identified in the paper as the 3-hyperboloid. The resulting POVM converges rapidly to the spin-coherent-state POVM, so the generalized $\sum_k K_k^*K_k=\mathbf 1$1-function becomes the outcome distribution of a concrete continuous measurement process (Jackson et al., 2021).

These applications illustrate a recurrent pattern: the operator-sum representation is useful not only because it classifies CPTP maps, but because it exposes a parameterization in which complete positivity is algebraically built in and physically interpretable.

7. Generalizations, variants, and conceptual boundaries

The basic operator-sum theorem admits several structured generalizations. One line of work imposes additional algebraic symmetry. For a distinguished state $\sum_k K_k^*K_k=\mathbf 1$2, a $\sum_k K_k^*K_k=\mathbf 1$3-adapted minimal Kraus family can satisfy a $\sum_k K_k^*K_k=\mathbf 1$4-sphere condition,

$\sum_k K_k^*K_k=\mathbf 1$5

and this condition generates a KMS state on the Kraus algebra and a detailed-balance structure connected to compact quantum groups (Andersson, 2015).

Another line constructs explicit channel families from group representations. For diagonal density matrices on $\sum_k K_k^*K_k=\mathbf 1$6, the defining permutation representation of the symmetric group yields channels with Kraus operators

$\sum_k K_k^*K_k=\mathbf 1$7

leading to

$\sum_k K_k^*K_k=\mathbf 1$8

These maps are CPTP, unital, and in the cyclic-subgroup case form explicit semigroups with closed-form orbit and limit-set structure (Cattabriga et al., 2021).

There are also settings in which the usual internal Kraus form is obstructed. For semigroups $\sum_k K_k^*K_k=\mathbf 1$9 on finite group algebras $\Phi$0, the paper on Kraus-like decompositions proves that a standard Kraus decomposition with Kraus operators inside $\Phi$1 generally fails for strict length functions, motivating instead a character-induced decomposition

$\Phi$2

with nonnegative coefficients in the conditionally negative-definite case (Boretsky et al., 2022). A different workaround appears in the operator sum-difference representation, where a Choi matrix is decomposed into simpler Hermitian pieces, producing “positive” and “negative” Kraus operators and an exact formula

$\Phi$3

The construction keeps the overall Choi matrix positive but may require as many as $\Phi$4 operators rather than $\Phi$5 (Omkar et al., 2012).

A frequent terminological confusion concerns the “Kraus-Cirac” decomposition of two-qubit unitaries. That decomposition concerns canonical factorization of a two-qubit unitary into local terms and a nonlocal exponential, and is not the standard Kraus representation of a general quantum channel (Soeda et al., 2014). By contrast, some recent work outside quantum information adopts the genuine operator-sum structure in new domains: for example, relation operators in knowledge-graph embedding are required to satisfy linearity, trace preservation, and complete positivity, leading to the real-valued Kraus form

$\Phi$6

with Kraus rank interpreted as a per-relation complexity measure (Chaki, 11 May 2026).

Taken together, these developments mark the operator-sum representation as both a foundational theorem about open quantum dynamics and a flexible algebraic language whose precise implementation depends on the ambient operator algebra, the microscopic model, and the structural constraints one wishes to preserve.