Decoherence-Free Subspaces

A decoherence-free subspace is a subspace whose states are not distinguished by a specified noise process. The environment may still interact with the physical degrees of freedom, but it acts identically on every state in the protected subspace, so it cannot learn which logical state was stored there.

Protection is specific to a noise model. Independent local noise can destroy an encoding that is perfectly protected against a collective fluctuation, and a Hamiltonian can move the state out of the protected subspace.

Helpful background. Lindblad operators describe Markovian noise; tensor-product spaces describe an encoded system.

Conditions for a decoherence-free subspace

Consider a finite-dimensional, time-independent generator

$$\dot\rho=-\frac{i}{\hbar}[H,\rho]+\sum_\mu\mathcal D[L_\mu]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\frac12\{L^\dagger L,\rho\}.$$

Rates are absorbed into $L_\mu$, which has units of inverse square-root time. A subspace $\mathcal C$ with projector $P$ is decoherence free if all its states undergo a unitary evolution that stays inside $\mathcal C$.

A particularly simple sufficient condition is

$$L_\mu P=\lambda_\mu P, \qquad L_\mu^\dagger P=\lambda_\mu^*P, \qquad (I-P)HP=0.$$

Each noise operator and its adjoint acts as the same scalar on every code vector, and $H$ preserves the code. For $\rho=P\rho P$, these conditions give

$$L_\mu\rho L_\mu^\dagger=|\lambda_\mu|^2\rho, \qquad L_\mu^\dagger L_\mu\rho=|\lambda_\mu|^2\rho,$$

so every dissipator vanishes. For normal $L_\mu$, the adjoint condition follows from the eigenvector condition. For a dark subspace, $L_\mu P=0$, the dissipator also vanishes without an adjoint condition.

For nonnormal operators and nonzero $\lambda_\mu$, the condition $L_\mu P=\lambda_\mu P$ alone does not make the dissipator vanish. On supported states its action instead equals a commutator:

$$\sum_\mu\mathcal D[L_\mu]\rho =-\frac{i}{\hbar}[H_c,\rho], \qquad H_c=\frac{i\hbar}{2}\sum_\mu (\lambda_\mu^*L_\mu-\lambda_\mu L_\mu^\dagger).$$

Thus the more general sufficient condition is that $H+H_c$ preserves $\mathcal C$, together with the common scalar eigenvalue condition. This extra condition prevents initially pure code states from being driven into a region where the noise distinguishes them (Karasik et al., 2008). The counterexample in the exercises shows why it matters.

Physical Meaning

For a pure initially uncorrelated environment, a protected encoding leaves no logical-state information in the environment. For two encoded states $\lvert0_L\rangle$ and $\lvert1_L\rangle$, a protected interaction has the schematic action

$$\lvert\psi_L\rangle\lvert E_0\rangle \longmapsto U_L\lvert\psi_L\rangle \lvert E_{\mathcal C}\rangle$$

for every encoded state

$$\lvert\psi_L\rangle = \alpha\lvert0_L\rangle + \beta\lvert1_L\rangle.$$

The environment state $\lvert E_{\mathcal C}\rangle$ is independent of $\alpha$ and $\beta$. Tracing out the environment therefore does not suppress coherence between $\lvert0_L\rangle$ and $\lvert1_L\rangle$.

By contrast, if

$$\lvert0_L\rangle\lvert E_0\rangle \longmapsto \lvert0_L\rangle\lvert E_0'\rangle, \qquad \lvert1_L\rangle\lvert E_0\rangle \longmapsto \lvert1_L\rangle\lvert E_1'\rangle$$

with $\langle E_1'|E_0'\rangle$ small, the environment has distinguished the logical states and the subspace is not decoherence free for that process.

This is the subspace version of the pointer-state principle. A pointer sector can preserve coherence inside a degenerate monitored sector even while coherences between sectors are suppressed.

Example: Collective Dephasing of Two Qubits

Consider two qubits subject to collective dephasing with one Lindblad operator

$$L=Z_1+Z_2.$$

In the computational basis,

$$\begin{aligned} L\lvert00\rangle&=2\lvert00\rangle,\\ L\lvert01\rangle&=0,\\ L\lvert10\rangle&=0,\\ L\lvert11\rangle&=-2\lvert11\rangle. \end{aligned}$$

The subspace

$$\mathcal C = \operatorname{span} \{ \lvert01\rangle, \lvert10\rangle \}$$

is decoherence free for this noise, because $L$ has the same eigenvalue $0$ on every vector in $\mathcal C$:

$$L\lvert\psi\rangle=0 \qquad \lvert\psi\rangle\in\mathcal C.$$

One may encode a logical qubit as

$$\lvert0_L\rangle=\lvert01\rangle, \qquad \lvert1_L\rangle=\lvert10\rangle.$$

Collective phase noise cannot distinguish the two code states. It sees only the total $Z$ value, and both states have the same value.

Independent dephasing is different. If the noise operators are $Z_1$ and $Z_2$ separately, then

$$Z_1\lvert01\rangle=\lvert01\rangle, \qquad Z_1\lvert10\rangle=-\lvert10\rangle.$$

The environment can now distinguish the two encoded states, so the same subspace is no longer decoherence free.

Collective Noise and Symmetry

Decoherence-free subspaces usually arise from symmetry. If the bath couples only to collective operators,

$$S_\alpha = \sum_{j=1}^N s_\alpha^{(j)},$$

then common scalar eigenspaces or representation multiplicity spaces can carry protected information, provided the remaining dynamics preserve the encoding.

For collective dephasing,

$$S_z=\sum_{j=1}^N Z_j,$$

each fixed-eigenvalue sector of $S_z$ is protected against the dephasing generated by $S_z$ itself. Coherence inside the fixed-eigenvalue sector survives; coherence between different sectors decays.

For fully collective $SU(2)$ noise with

$$S_x, \qquad S_y, \qquad S_z,$$

singlet sectors and representation multiplicity spaces can be protected. This is where angular-momentum decomposition becomes operational: the environment may couple to total spin, while information stored in multiplicity labels is invisible to that coupling.

The symmetry statement is the main point. Protection follows from how the noise algebra acts, not from the physical qubits being far apart or from the state having small energy.

Dark States and Decay

A dark state is annihilated by a jump operator:

$$J\lvert\psi\rangle=0.$$

Dark states are common sources of decoherence-free structure. For collective spontaneous emission from two identical two-level systems, the jump operator is

$$J_-=\sigma_-^{(1)}+\sigma_-^{(2)}.$$

The singlet state

$$\lvert\psi_-\rangle = \frac{ \lvert ge\rangle-\lvert eg\rangle } {\sqrt2}$$

is dark:

$$J_-\lvert\psi_-\rangle=0.$$

The ground state $\lvert gg\rangle$ is also dark. Their span is protected when the Hamiltonian preserves it and there are no additional independent decay or excitation operators. Being dark is not the same as being uniquely prepared. A large dark subspace can preserve information; reservoir engineering asks how to make a desired dark state or subspace also attracting.

For the jump-operator language, see Lindblad Operators. For designed attraction into dark states, see Reservoir Engineering.

Noiseless Subsystems

A decoherence-free subspace is the simplest case. More generally, the protected information may live in a tensor factor inside a sector:

$$\mathcal H \cong \bigoplus_J \left( \mathcal H_{A_J}\otimes\mathcal H_{B_J} \right) \oplus \mathcal K.$$

A noiseless subsystem is a factor $\mathcal H_{A_J}$ on which the noise acts trivially while it may act nontrivially on $\mathcal H_{B_J}$:

$$E_a \big|_{\mathcal H_{A_J}\otimes\mathcal H_{B_J}} = I_{A_J}\otimes M_{a,J}.$$

Then a state encoded in $A_J$ is protected even if $B_J$ changes:

$$\rho_{A_J}\otimes\rho_{B_J} \longmapsto \rho_{A_J}\otimes \Phi_{B_J}(\rho_{B_J}).$$

This is stronger and more flexible than looking only for one subspace on which every error operator is a scalar. It is especially natural when symmetry decomposes the Hilbert space into irreducible representations and multiplicity spaces. The noise may act on the representation factor while leaving the multiplicity factor untouched.

Relation to Fixed Points

A decoherence-free subspace is related to fixed points, but the concepts are not identical.

If every state supported on $\mathcal C$ is unchanged by the dissipative part of the generator, then the dissipator has a large fixed set. But the full dynamics on $\mathcal C$ may still include a Hamiltonian:

$$\rho_{\mathcal C}(t) = U_{\mathcal C}(t) \rho_{\mathcal C}(0) U_{\mathcal C}^\dagger(t).$$

Thus the encoded information can evolve unitarily while remaining protected from the specified dissipator. A stationary state is fixed in time; a decoherence-free subspace can support nontrivial logical unitary dynamics.

For the channel-level fixed-point language, see Channel Composition and Fixed Points.

Relation to Quantum Error Correction

Decoherence-free subspaces are passive error-avoidance structures. They do not require measuring syndromes and applying corrections if the actual noise respects the required symmetry.

Quantum error correction is broader. A code projector $P$ corrects a set of errors $\{E_a\}$ when the Knill–Laflamme condition holds:

$$P E_a^\dagger E_b P = c_{ab}P.$$

A decoherence-free subspace is a special, more restrictive case in which the errors act as scalars on the code:

$$E_a P = c_a P$$

up to an allowed logical unitary convention. No syndrome is needed because the encoded state was never distinguished by the error.

The tradeoff is clear:

Method	Strength	Limitation
decoherence-free subspace	passive protection against symmetric noise	fails when symmetry-breaking noise dominates
noiseless subsystem	protects a logical factor under a noise algebra	requires the correct algebraic decomposition
dynamical decoupling	averages unwanted couplings with control pulses	requires fast, accurate control
active error correction	corrects broader error sets	needs syndrome extraction, redundancy, and fault-tolerant design

The quantum-error-correction introduction explains how active correction can handle errors outside this symmetry-protected setting.

Imperfectly collective noise: a residual rate

Let the actual Hermitian jump operator be

$$L=\sqrt\gamma\,[Z_1+Z_2+\epsilon(Z_1-Z_2)], \qquad \gamma\ge0.$$

On $|01\rangle$ and $|10\rangle$ it has eigenvalues $2\sqrt\gamma\epsilon$ and $-2\sqrt\gamma\epsilon$. Therefore the logical off-diagonal element obeys

$$\dot\rho_{01,10}=-\frac12(4\sqrt\gamma\epsilon)^2\rho_{01,10} =-8\gamma\epsilon^2\rho_{01,10}.$$

For $\gamma=100\,\mathrm{s}^{-1}$ and $\epsilon=0.01$, the coherence time is $12.5\,\mathrm s$. This quadratic suppression depends on the stated single-operator model. Independent local channels with fixed rates, Hamiltonian leakage, or control errors require their own calculation; they are not reduced merely by making this $\epsilon$ small.

Exercises

1. Collective dephasing code. Let $L=Z_1+Z_2$. Show that every state in $\operatorname{span}\{\lvert01\rangle,\lvert10\rangle\}$ is annihilated by $L$.

Solution

Using $Z\lvert0\rangle=\lvert0\rangle$ and $Z\lvert1\rangle=-\lvert1\rangle$,

$$L\lvert01\rangle = (1-1)\lvert01\rangle = 0,$$

and

$$L\lvert10\rangle = (-1+1)\lvert10\rangle = 0.$$

For any

$$\lvert\psi\rangle = \alpha\lvert01\rangle+\beta\lvert10\rangle,$$

linearity gives

$$L\lvert\psi\rangle=0.$$

Thus the whole two-dimensional subspace is decoherence free for this collective dephasing operator.

2. Independent dephasing failure. For the same encoded states, show that $Z_1$ distinguishes $\lvert01\rangle$ and $\lvert10\rangle$.

Solution

The first qubit is $\lvert0\rangle$ in $\lvert01\rangle$ and $\lvert1\rangle$ in $\lvert10\rangle$. Therefore

$$Z_1\lvert01\rangle=\lvert01\rangle, \qquad Z_1\lvert10\rangle=-\lvert10\rangle.$$

The two states have different eigenvalues of $Z_1$. A bath that monitors $Z_1$ can distinguish them, so the collective-dephasing code is not protected against independent dephasing on qubit 1.

3. Why the adjoint condition matters. Take $H=0$ and $L=\sqrt\kappa\begin{pmatrix}1&1\\0&0\end{pmatrix}$, with $\kappa>0$. Although $L|0\rangle=\sqrt\kappa|0\rangle$, show that $\operatorname{span}\{|0\rangle\}$ is not invariant under the master equation.

Solution

For $\rho=|0\rangle\langle0|$,

$$L\rho L^\dagger=\kappa\begin{pmatrix}1&0\\0&0\end{pmatrix}, \qquad \frac12\{L^\dagger L,\rho\} =\kappa\begin{pmatrix}1&1/2\\1/2&0\end{pmatrix}.$$

Consequently

$$\dot\rho=-\frac\kappa2\begin{pmatrix}0&1\\1&0\end{pmatrix}\ne0.$$

The derivative develops matrix elements outside the proposed one-dimensional subspace. Here $L^\dagger|0\rangle=\sqrt\kappa(|0\rangle+|1\rangle)$, so the stronger scalar-adjoint condition fails; $H+H_c$ also fails to preserve the subspace. Zero instantaneous purity loss at $t=0$ is not a guarantee of a protected finite-time evolution.

4. Dark singlet. Let $J_-=\sigma_-^{(1)}+\sigma_-^{(2)}$ and

$$\lvert\psi_-\rangle = \frac{\lvert ge\rangle-\lvert eg\rangle}{\sqrt2}.$$

Show that $J_-\lvert\psi_-\rangle=0$.

Solution

The lowering operator gives

$$J_-\lvert ge\rangle=\lvert gg\rangle, \qquad J_-\lvert eg\rangle=\lvert gg\rangle.$$

Therefore

$$J_-\lvert\psi_-\rangle = \frac{\lvert gg\rangle-\lvert gg\rangle}{\sqrt2} = 0.$$

The minus sign produces destructive interference in the emission amplitude.

5. DFS versus QEC. Explain why a decoherence-free subspace is a special case of quantum error correction but does not replace active error correction.

Solution

For a DFS, the relevant errors act as scalars on the code:

$$E_aP=c_aP.$$

Then

$$P E_a^\dagger E_b P = c_a^*c_b P,$$

which satisfies the Knill–Laflamme condition. No syndrome is needed because the error did not distinguish the encoded state.

Active error correction is broader. It can correct errors that move the state into distinguishable syndrome subspaces, provided the syndrome can be measured and corrected without learning the logical state. A DFS protects only against noise with the required symmetry or algebraic action; symmetry-breaking noise generally requires other methods.

References

• D. Bacon, D. A. Lidar, and K. B. Whaley, “Robustness of decoherence-free subspaces for quantum computation,” Physical Review A 60, 1944–1955 (1999). DOI.
• R. I. Karasik, K.-P. Marzlin, B. C. Sanders, and K. B. Whaley, “Criteria for dynamically stable decoherence-free subspaces and incoherently generated coherences,” Physical Review A 77, 052301 (2008). DOI.
• E. Knill, R. Laflamme, and L. Viola, “Theory of quantum error correction for general noise,” Physical Review Letters 84, 2525–2528 (2000). DOI.
• D. A. Lidar, I. L. Chuang, and K. B. Whaley, “Decoherence-free subspaces for quantum computation,” Physical Review Letters 81, 2594–2597 (1998). DOI.
• D. A. Lidar and K. B. Whaley, “Decoherence-free subspaces and subsystems,” in Irreversible Quantum Dynamics, Lecture Notes in Physics 622, Springer (2003). Author manuscript.
• P. Zanardi and M. Rasetti, “Noiseless quantum codes,” Physical Review Letters 79, 3306–3309 (1997). DOI.