Covariant Qubit Channels: The Mixture Family, Unitality and Choi Rank

Introduction

This article closes the structural account. Three of its ingredients are already proved elsewhere in the corpus, and it uses them without re-deriving them. Unitality is the algebraic fixing of the centre, so that a unital map leaves the maximally mixed state where it is. Non-invertibility is the only route across the level sets of the quaternionic norm, so that no reversible motion is a thermal motion. And the semigroup law along the thermal path holds exactly, because its parameter multiplies rather than adds.

What remained looked like a classification, and it is not one. The algebra $\mathbb{B}\cong M_2(\mathbb{C})$ admits every completely positive trace-preserving map, because the channel theory of $M_2(\mathbb{C})$ is the channel theory of a qubit and the algebra is that algebra. A classification written in quaternionic notation would restate Choi's theorem and the Fujiwara–Algoet range criterion with different letters. That is not a mathematical error and it is not worth an article.

The questions with content are two. What forces a channel to leave the algebra at all, and what selects the thermal channel from the set that remains. The first is answered by simplicity, and the answer is exact: a channel is a map that has abandoned the product, and no such map is an endomorphism of the algebra. The second is answered by covariance together with complete positivity, and the answer is a one-parameter family whose physically admissible segment is forced.

The plan. Admission is refused as a goal, and the imported classification is stated once with its two boundaries. Simplicity is then used to show that no channel is internal. Covariance selects a family, complete positivity and the mixture reading cut it to a segment, and unitality gives it its arrow. The Choi rank of that family is computed, and it places the thermal channel one rung above every dilation the corpus has so far needed. The two senses in which the word ceiling has been used are then separated, and what is refused is named.

Admission Is Not the Question

The algebra is eight-dimensional over $\mathbb{R}$ and four over $\mathbb{C}$. $\mathbb{B}=\mathbb{B}_{\Phi}\oplus\mathbb{B}_{\Psi}$, each sector of four real dimensions; the coefficients are complex, so $\dim_{\mathbb{C}}\mathbb{B}=4$ and $\dim_{\mathbb{R}}\mathbb{B}=8$. The real dimension is the one that matters for dilation counting below.

A state map is an affine map of the ball. Write a state as $\tilde{Ϛ}_{\Psi}=\tfrac12(e_t+i\vec s_{\Psi})$; a channel acts by $\vec s_{\Psi}\mapsto M\vec s_{\Psi}+\vec t$ with $M$ real $3\times3$ and $\vec t\in\mathbb{R}^{3}$, and complete positivity is a condition on the Choi matrix, the $4\times4$ matrix $\mathsf C=\sum_l\mathrm{vec}(\tilde K_l)\mathrm{vec}(\tilde K_l)^{\dagger}$ assembled from the Kraus operators. Everything in this paragraph is standard, it holds for $M_2(\mathbb{C})$, and it holds therefore for $\mathbb{B}$ with no change of statement.

Two conditions run alongside trace preservation, and they bound the affine data.

  • Trace preservation, $\sum_l\tilde K_l^{*}\tilde K_l=e_t$, keeps a state a state.
  • Unitality, $\sum_l\tilde K_l\tilde K_l^{*}=e_t$, keeps the centre fixed.

The second is not implied by the first, and the pair is the dividing line the rest of the article uses. Dephasing satisfies both; amplitude damping satisfies only the first. Verified on the canonical Kraus sets: dephasing with $\{\sqrt{\tfrac12}e_t,\sqrt{\tfrac12}\vec e_z\}$ is trace-preserving and unital; amplitude damping is trace-preserving and not unital.

The classification is imported whole, and the article does not improve it. What the article adds is which sub-family survives three requirements imposed from the algebra's own structure, and at what rank.

Simplicity: Why No Channel Is Internal

$\mathbb{B}$ is simple, and that is the whole reason a channel cannot live inside it. A nonzero algebra endomorphism of a simple algebra is injective, because its kernel is a two-sided ideal; an injective endomorphism of a finite-dimensional algebra is surjective; and a unital automorphism of $M_2(\mathbb{C})$ is inner by Skolem–Noether. So every unital multiplicative map is a conjugation.

A conjugation has one Kraus operator, and trace preservation then forces it to be unitary. With $\mathcal E(\tilde Q)=\tilde U\tilde Q\tilde U^{*}$ the Kraus sum has the single term $\tilde K=\tilde U$, and $\tilde K^{*}\tilde K=e_t$ is exactly the unit-norm condition. A completely positive map that is multiplicative is a $*$-homomorphism, so the chain closes in both directions:

$$ \text{multiplicative}\iff\text{Choi rank }1\iff\text{conjugation by a unit}\iff N\text{ preserved}. $$

Verified on a random inner endomorphism. For $\mathcal G$ of unit determinant and $\mathcal E(\tilde Q)=\mathcal G\tilde Q\mathcal G^{-1}$, multiplicativity holds to $5.0\times10^{-16}$, unitality to $2.6\times10^{-16}$, and the Choi rank of the single Kraus operator is $1$.

So a channel is exactly a map that has given up the product. Irreversibility and non-multiplicativity are one statement here, and the statement is not about size. It is that $\mathbb{B}$ is simple, its only two-sided ideals being $0$ and itself, so there is no internal way to lose information: any map that contracts the ball must carry what it destroyed somewhere the algebra does not reach.

One distinction this forces, and the corpus has been careful about it elsewhere. A Lorentz boost is a congruence with a single factor and preserves $N$ of the element, but it does not preserve $2\,\mathrm{Sc}$, so it is not a state map at all: it moves the slice. For $\tilde{Ϛ}_{\Psi}=\tfrac12(e_t+0.4\vec e_z)$ and a boost of rapidity $0.9$ along $\vec e_x$, the scalar part goes from $0.500000$ to $0.716543$, and $2\,\mathrm{Sc}$ from $1.000000$ to $1.433086$. The reversible group acting on states is therefore the unitary subgroup, and every use of the word rotor below means a unit element of it.

Covariance Selects the Family

The algebra admits every channel, so the thermal one must be selected, and the selector is symmetry. Require covariance under the action the algebra itself supplies:

$$ \mathcal E(\tilde U\tilde Q\tilde U^{*})=\tilde U\,\mathcal E(\tilde Q)\,\tilde U^{*} \qquad\text{for every unit rotor }\tilde U . $$

Twirling collapses the affine data to two numbers, one of which is zero. Averaging $RMR^{\mathsf T}$ over the rotation group gives $\tfrac13\operatorname{tr}(M)I$, because the average of $R_{ia}R_{jb}$ over the orthogonal group is $\tfrac13\delta_{ij}\delta_{ab}$, and averaging $R\vec t$ gives $\vec 0$. A Monte-Carlo twirl over the Haar measure shows this to the sampling error: with $10^{5}$ draws the off-diagonal entries of the averaged $M$ are $2.4\times10^{-3}$ and the averaged $\vec t$ is $1.9\times10^{-3}$, both of which shrink with the sample count and neither of which is a residual.

What survives is a one-parameter family.

$$ \mathcal E_\lambda(\tilde Q)=(1-\lambda)\tilde Q+\lambda\,\mathrm{Sc}(\tilde Q)\,e_t, \qquad \vec s_{\Psi}\mapsto(1-\lambda)\vec s_{\Psi}, $$

which is the mixture with the centre at weight $\lambda$. It is covariant, and covariance holds to $3.3\times10^{-16}$ over three hundred random unitaries and states. It is trace-preserving, it fixes the centre, and it is the channel the companion The Gibbs Path Is Not a Group Orbit already uses.

Three boundaries cut the real line, and they are not the same boundary. With $c=1-\lambda$ the family is $\Phi_c(\tilde{Ϛ})=c\,\tilde{Ϛ}+(1-c)\mathrm{Sc}(\tilde{Ϛ})e_t$, whose Choi eigenvalues are

$$ \frac{1+3c}{2}\ \text{once},\qquad \frac{1-c}{2}\ \text{three times}. $$

condition range of $c$ range of $\lambda$ what fails outside
complete positivity $c\ge-\tfrac13$ $\lambda\le\tfrac43$ the Choi matrix has a negative eigenvalue
positivity, so states map to states $-1\le c\le1$ $0\le\lambda\le2$ the image leaves the ball
a convex mixture of the state and the centre $0\le c\le1$ $0\le\lambda\le1$ the map reflects the Bloch vector instead of shrinking it

All three follow from two eigenvalue formulas, and none of them is approximate. For a pure input the eigenvalues of the image are $\tfrac12(1+c)$ and $\tfrac12(1-c)$, both non-negative exactly for $\lvert c\rvert\le1$. For $c=-\tfrac13$ the Choi eigenvalues are $0,\tfrac23,\tfrac23,\tfrac23$; for $c=1$ they are $0,0,0,2$. The window $-1\le c<-\tfrac13$ is the standard one in which the map is positive but not completely positive, and it is where the corpus's exercise on positive-but-not-CP maps sits.

The physical segment is the last row, and it is forced rather than chosen. A mixture with weight in $[0,1]$ is the only reading of $\mathcal E_\lambda$ under which $\lambda$ is a probability. Covariance gives the family; complete positivity permits more of it than physics uses; and the mixture reading cuts it to the segment the thermal path needs. Nothing about the value of $\lambda$ is selected — that is the bath, and §What Is Refused returns to it.

Unitality Is the Arrow

Unitality is not a technical side condition; it is the direction of time. For a unital channel the Bloch radius cannot grow, and $N=\tfrac14(1-r^{2})$ therefore cannot fall:

$$ r'=\lvert M\vec s_{\Psi}+\vec t\rvert\le r \quad\Longrightarrow\quad N'\ge N \quad\Longrightarrow\quad \text{purity non-increasing, entropy non-decreasing.} $$

Verified over four hundred random unital channels. Each is an average of conjugations by Haar unitaries, hence unital and trace-preserving by construction; the largest change in the radius over all draws is $r'-r=-1.8\times10^{-4}$, and it is never positive. The inequality is the contraction of the ball toward the centre, and on the thermal curve it is the statement that a unital operation heats: it is why the companion calls the depolarising channel a heating channel and not a cooling one.

The centre is the only state that every unital channel fixes. It is fixed by each conjugation and therefore by every average of them, to $1.6\times10^{-16}$ over two hundred draws. Conversely, if a state is fixed by every unital channel it is fixed by the fully dephasing map along any axis: a state at $0.7\vec e_z$, dephased along $\vec e_x$, has radius $0.0$ exactly, so it must already have been the centre.

The sign of the escape is the physics. A non-unital channel can lower $N$, and that is refrigeration and not a channel's failure: damping with Kraus weight $g=0.447$ takes $r$ from $0.2000$ to $0.8400$ and $N$ from $0.240000$ to $0.073600$. The coarse-graining article already says that such a map is not a coarse-graining, because it exports entropy to an environment and keeps no ledger of the export. Unitality is precisely the condition that no such export is hidden.

The semigroup law is inherited, not reproved. $\mathcal E_{\lambda_1}\circ\mathcal E_{\lambda_2}=\mathcal E_{\lambda_1+\lambda_2-\lambda_1\lambda_2}$ because $(1-\lambda_1)(1-\lambda_2)=1-(\lambda_1+\lambda_2-\lambda_1\lambda_2)$, and the companion computes the residual as exactly zero. So the family is a semigroup, unital throughout, with the centre as its unique common fixed point and $N$ monotone along it.

The Staircase of Choi Ranks

The price of abandoning multiplicativity is counted by the Choi rank, and the count is the size of the environment. The Choi rank is the minimal number of Kraus operators, and by Stinespring it is also the minimal environment dimension in a dilation $\mathcal E(\tilde{Ϛ})=\operatorname{tr}_E\bigl(\tilde W(\tilde{Ϛ}\otimes\tilde\Pi_0)\tilde W^{*}\bigr)$.

channel Choi rank minimal dilated algebra real dimension
unitary conjugation $1$ $\mathbb{B}\cong M_2(\mathbb{C})$ $8$
dephasing, damping $2$ $\mathbb{B}\otimes\mathbb{B}\cong M_4(\mathbb{C})$ $32$
mixture, depolarising $4$ $M_2(\mathbb{C})\otimes M_4(\mathbb{C})\cong M_8(\mathbb{C})$ $128$

The rank does not interpolate, and that is the sharpest fact in this section. At $\lambda=0$ the channel is the identity and the rank is $1$; for every $\lambda\in(0,1]$ the rank is $4$. Verified at $\lambda=0.05$, $0.5$ and $1$. There is no gradual increase, because the rank is an integer and the family is a straight line through the identity: the channel leaves the reversible class at the first step and lands at full rank.

The selected channel therefore sits one rung above everything the corpus has dilated. $\mathbb{B}\otimes\mathbb{B}$ covers the rank-$\le2$ channels, which is exactly dephasing and damping and the one-qubit environment they need. The covariant channel needs an environment of dimension four, so its minimal carrier is $M_2\otimes M_4\cong M_8$, of $128$ real dimensions. And at the boundary of complete positivity the rank is $3$, not $4$: the rank itself changes where the Choi matrix acquires its zero eigenvalue, so the boundary of positivity is visible in the rank and not only in the spectrum.

The Two Ceilings

The word has been used for two different statements, and they should be separated. Both are true, and conflating them would make each look weaker than it is.

The fixed ceiling is the centre as the limit of contraction: every unital channel sends the ball inward, the fixed points of a channel are where its contraction stops, and for a unital family the stopping place is the centre. This is the sense in which the ledger's phrase the centre as the ceiling of every channel is meant, and it is a statement about the arrow.

The dimensional ceiling is the Stinespring staircase: a channel of Choi rank $k$ requires an environment of dimension $k$, so the algebra that carries it is $\mathbb{B}\otimes M_k$, and the rank is the count of how much multiplicativity was abandoned. This is a statement about structure, and it is a corollary of §Simplicity rather than an independent fact — dimensions follow from rank, and rank follows from the loss of the product.

The same staircase is the corpus's gauge ceiling read on another carrier. Grand Unification in the Biquaternion Framework: Where the Embedding Fails passes that ceiling by moving from $\mathbb{B}$ to the matrix algebra over it, and the move here is the same enlargement: the matrix factor is the environment and its size is the Choi rank. The two share the shape of the enlargement and not the subject — one counts the factors a gauge group needs, the other the dimensions a discarded system needs.

The two are related by one sentence. The centre is where a contractive channel stops, and the rank is what the channel had to leave the algebra to do; a channel that stopped nowhere would have rank $1$ and be a rotation.

What the Algebra Does Not Give

The value of $\lambda$ is not derived, and neither is the identity of the bath. Covariance, complete positivity and the mixture reading force the shape of the family; they say nothing about which member a given physical situation realises. That is the temperature, or equivalently the fixed point, and it is data supplied from outside the algebra — exactly as the Hamiltonian is data under O18 and the outcome basis is data under O19.

Irreversibility is never internal, and this is a theorem rather than a remark. A completely positive multiplicative map on $\mathbb{B}$ is a $*$-homomorphism, hence inner, hence unitary conjugation, hence Choi rank $1$. So no channel with an environment can be written inside $\mathbb{B}$, and demanding one would be demanding a non-injective endomorphism of a simple algebra. The refusal is structural, and it is the reason the framework needs an environment at all.

What is claimed is smaller and precise. The algebra supplies the geometry, the family that covariance selects, the arrow that unitality gives it, the rank it costs, and the two ceilings. It does not supply the temperature at which the bath sits, and it does not supply the bath.

Summary

Channels are not classified here; they are selected. $\mathbb{B}\cong M_2(\mathbb{C})$ admits every completely positive trace-preserving map, so the classification is imported from Choi theory and named as such, with its Choi-matrix condition and its two boundaries: complete positivity at $c\ge-\tfrac13$ and positivity at $\lvert c\rvert\le1$.

Simplicity is the structural fact. A nonzero endomorphism of $\mathbb{B}$ is injective, a unital one is an automorphism, an automorphism of $M_2(\mathbb{C})$ is inner, and an inner one has Choi rank $1$. So multiplicative, rank $1$, conjugation by a unit, and $N$ preserved are one condition, and a channel is precisely a map that has abandoned the product. That is why no channel is internal, and why a boost does not count as one: it preserves $N$ but not $2\,\mathrm{Sc}$, so it moves the slice.

Covariance under unit rotor conjugation selects the mixture family $\mathcal E_\lambda(\tilde Q)=(1-\lambda)\tilde Q+\lambda\,\mathrm{Sc}(\tilde Q)e_t$, verified to $3.3\times10^{-16}$; the twirl sends $M$ to $\tfrac13\operatorname{tr}(M)I$ and $\vec t$ to zero. Its Choi eigenvalues are $\tfrac12(1+3c)$ once and $\tfrac12(1-c)$ three times with $c=1-\lambda$, which places complete positivity at $c\ge-\tfrac13$, positivity at $\lvert c\rvert\le1$, and the mixture reading at $0\le\lambda\le1$.

Unitality is the arrow: $r'\le r$, so $N'\ge N$, purity non-increasing and entropy non-decreasing, verified with the worst change $r'-r=-1.8\times10^{-4}$ over four hundred channels. The centre is the unique common fixed point of all unital channels, and a non-unital channel that lowers $N$ from $0.240000$ to $0.073600$ is refrigeration rather than an error.

The Choi rank is the price: $1$ for a conjugation and $8$ real dimensions, $2$ for dephasing and damping with $\mathbb{B}\otimes\mathbb{B}$ and $32$, and $4$ for the selected family with $M_8$ and $128$ — reached at $\lambda=0.05$ already, since the rank jumps from $1$ to $4$ at the identity and never interpolates. Two ceilings follow: the centre as the limit of contraction, and the Stinespring staircase as the cost of leaving the algebra.

Summary of Notation

Symbol Meaning
$\mathbb{B}\cong M_2(\mathbb{C})$ Biquaternion algebra, $\dim_{\mathbb{R}}=8$, $\dim_{\mathbb{C}}=4$
$\tilde{Ϛ}_{\Psi}=\tfrac12(e_t+i\vec s_{\Psi})$ State, with Bloch vector $\vec s_{\Psi}$
$N=\tfrac14(1-\lvert\vec s_{\Psi}\rvert^{2})$ Quaternionic norm of the state
$M,\vec t$ Affine data of a channel, $\vec s_{\Psi}\mapsto M\vec s_{\Psi}+\vec t$
$\mathsf C=\sum_l\mathrm{vec}(\tilde K_l)\mathrm{vec}(\tilde K_l)^{\dagger}$ Choi matrix
$\operatorname{rank}\mathsf C$ Choi rank: minimal Kraus number, minimal environment dimension
$\sum_l\tilde K_l^{*}\tilde K_l=e_t$ Trace preservation
$\sum_l\tilde K_l\tilde K_l^{*}=e_t$ Unitality: the centre is fixed
$\mathcal E_\lambda(\tilde Q)=(1-\lambda)\tilde Q+\lambda\,\mathrm{Sc}(\tilde Q)e_t$ The covariant family; $\vec s_{\Psi}\mapsto(1-\lambda)\vec s_{\Psi}$
$\Phi_c$ with $c=1-\lambda$ The same family in depolarising form
$\tfrac12(1+3c)$, $\tfrac12(1-c)$ Choi eigenvalues of $\Phi_c$, once and three times
$c\ge-\tfrac13$ Complete positivity
$\lvert c\rvert\le1$ Positivity: states map to states
$0\le\lambda\le1$ The mixture reading

Further Reading

  • M.-D. Choi, Completely positive linear maps on complex matrices, Linear Algebra Appl. 10, 285 (1975), for the Choi matrix and the equivalence of complete positivity with its positivity.
  • A. Fujiwara and P. Algoet, One-to-one parametrization of quantum channels, Phys. Rev. A 59, 3290 (1999), for the affine parametrisation of qubit channels used in §Admission.
  • W. F. Stinespring, Positive functions on $C^{*}$-algebras, Proc. Amer. Math. Soc. 6, 211 (1955), for the dilation whose environment dimension is the Choi rank.
  • I. Bengtsson and K. Życzkowski, Geometry of Quantum States (Cambridge, 2006), for the Bloch ball, the affine action of channels on it, and the contraction of its radius.
  • M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information (Cambridge, 2000), for the Kraus representation and the canonical channels used as examples.
  • G. Lindblad, On the generators of quantum dynamical semigroups, Commun. Math. Phys. 48, 119 (1976), for the semigroup whose contraction is the continuous form of the arrow in §Unitality.