Bosonic Gaussian Purification Channel

• Bosonic Gaussian Purification Channel is a framework employing pure-environment dilations, heralded maps, and random operations to conditionally reduce Gaussian noise in continuous-variable quantum systems.
• It utilizes affine transformations on first and second moments with parameters (τ, ν) to model loss, amplification, and classical noise, establishing criteria for effective purification.
• The concept unifies operational, state-theoretic, and extremality notions to enhance transmitted entanglement and to quantify minimal environmental complexity in quantum channels.

In bosonic continuous-variable quantum information, purification language is used for several technically distinct constructions that all live within the phase-space formalism of Gaussian states and channels. A bosonic Gaussian channel is characterized by affine action on first and second moments,

$$m \longmapsto Xm+x,\qquad V \longmapsto XVX^\intercal+Y,$$

with complete-positivity condition

$Y+i\Omega \ge iX\Omega X^\intercal$

(Rodriguez et al., 27 Mar 2025). In the frequently used single-mode phase-insensitive parametrization one writes

$X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$

so that loss, amplification, additive classical noise, and the identity channel are distinguished by the pair $(\tau,\nu)$ (Shajilal et al., 2024). Within this setting, a “bosonic Gaussian purification channel” may refer to a pure-environment Gaussian dilation and its complementary output, an exact random channel that outputs Gaussian purifications of a restricted class of Gaussian states, or a heralded map that conditionally suppresses Gaussian noise. These notions are related, but they are not equivalent.

1. Formal setting and principal meanings

The most conservative use of purification language is the Stinespring one: a channel is represented by a Gaussian unitary coupling between system and environment, and a pure environment state turns the joint output into a purification of the reduced channel output. A second use is state-theoretic: given a mixed Gaussian state, one seeks a channel that outputs copies of one of its Gaussian purifications, randomized over the allowed ancilla-unitary freedom. A third use is operational and error-corrective: a conditional map is said to have purification-like action if it reduces effective added noise, increases transmitted entanglement, or drives an effective channel toward the identity on accepted events (Caruso et al., 2010).

For single-mode phase-insensitive channels, the standard covariance-matrix law

$$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$

provides the natural language for purification-like noise suppression (Shajilal et al., 2024). The same $(\tau,\nu)$ plane organizes ordinary loss channels,

$0<\tau<1,\qquad \nu=(1-\tau)\chi,$

amplifier channels,

$\tau>1,\qquad \nu=(\tau-1)\chi,$

additive classical noise channels,

$\tau=1,\qquad \nu>0,$

and the identity channel,

$\tau=1,\qquad \nu=0$

(Shajilal et al., 2024). In this sense, “purification” most often means decreasing $Y+i\Omega \ge iX\Omega X^\intercal$0, increasing transmitted correlations, or realizing a pure-environment dilation, rather than universally mapping mixed Gaussian inputs to pure outputs.

2. Pure-environment Gaussian dilation and complementary structure

For an arbitrary multimode bosonic Gaussian channel $Y+i\Omega \ge iX\Omega X^\intercal$1, the constructive Gaussian dilation problem was solved by the explicit unitary-dilation theory of Holevo and Werner. In that framework,

$Y+i\Omega \ge iX\Omega X^\intercal$2

where $Y+i\Omega \ge iX\Omega X^\intercal$3 is Gaussian and $Y+i\Omega \ge iX\Omega X^\intercal$4 is Gaussian; when $Y+i\Omega \ge iX\Omega X^\intercal$5 is pure this is the Gaussian Stinespring representation (Caruso et al., 2010). Writing the channel in characteristic-function form with matrices $Y+i\Omega \ge iX\Omega X^\intercal$6 and

$Y+i\Omega \ge iX\Omega X^\intercal$7

the minimum number of environmental modes required for a pure Gaussian dilation is

$Y+i\Omega \ge iX\Omega X^\intercal$8

(Caruso et al., 2010). This gives a sharp structural measure of the environmental complexity of a bosonic Gaussian channel and, equivalently, of the smallest pure Gaussian purification space needed to realize it.

The same purification picture underlies extremality. Under the nondegeneracy assumption $Y+i\Omega \ge iX\Omega X^\intercal$9, a Bosonic linear channel is extreme iff it has a complementary channel with trivial kernel, and in the finite-mode Gaussian case this becomes the statement that a bosonic Gaussian channel is extreme iff it has minimal noise (Holevo, 2011). In the Gaussian parametrization

$X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$0

complete positivity requires

$X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$1

and minimality of $X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$2 is equivalent to purity of the environmental Gaussian state in the dilation (Holevo, 2011). In this precise sense, pure-environment realization, minimal noise, complementary-channel faithfulness, and extremality coincide.

When the environment is mixed, purification is indispensable if one wants the true complementary channel rather than only a weakly complementary one. For attenuators and amplifiers with mixed Gaussian environment $X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$3, one introduces a purifier $X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$4 and a pure state $X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$5 such that

$X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$6

and then defines the complementary channel on the enlarged dilation by

$X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$7

(Jeong, 2020). This distinction is operationally important in private-capacity bounds, where the conditional quantum entropy power inequality is applied precisely after purifying the Gaussian environment (Jeong, 2020).

The purification/complementarity structure changes in singular cases. For Bosonic linear channels with noise-free variables $X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$8 or rank-deficient $X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1,$9, the corresponding weak complementary channel exhibits dual singularities, and the noiseless sector induces direct-sum decompositions and explicit reversing channels (Shirokov, 2013). At the level of dilation theory, this identifies which canonical variables survive noiselessly and which sectors reduce to discrete classical-quantum behavior.

Beyond exact Gaussian channels, the dilation viewpoint extends approximately to all phase-space linear bosonic channels. Every linear bosonic channel is in the strong-operator closure of the Gaussian-dilatable channels, meaning that it can be approximated by channels realized through a Gaussian unitary and an ancillary state, with the non-Gaussianity shifted into the ancilla (Lami et al., 2018). Exact Gaussian dilation can fail for specific linear bosonic channels, such as the binary displacement channel, but approximate Gaussian-dilation realizability is universal in the linear bosonic class (Lami et al., 2018).

3. Heralded Gaussian noise suppression and purification-like channel engineering

A more operational notion of bosonic Gaussian purification appears in heralded continuous-variable teleportation. The phase-insensitive single-mode Gaussian map

$(\tau,\nu)$0

can be engineered by finite-squeezing teleportation, and augmented by a measurement-based noiseless linear amplifier (MBNLA) to access lower-noise effective channels than deterministic teleportation can reach (Shajilal et al., 2024). The ideal NLA is associated with the non-unitary operator

$(\tau,\nu)$1

acting approximately on coherent states as

$(\tau,\nu)$2

and the measurement-based implementation postselects dual-homodyne outcomes $(\tau,\nu)$3 through the filter

$(\tau,\nu)$4

Accepted outcomes are rescaled according to

$(\tau,\nu)$5

and, unlike earlier heralded teleportation restricted to $(\tau,\nu)$6, the non-unity-gain regime allows

$(\tau,\nu)$7

(Shajilal et al., 2024).

The resulting map is trace-decreasing and conditional, not deterministic CPTP, but it can suppress effective Gaussian noise and move simulated channels toward the identity. The reported reach includes

$(\tau,\nu)$8

interpreted as substantially closer to the identity than the deterministic baseline from the same physical resources (Shajilal et al., 2024). Increasing $(\tau,\nu)$9 pushes loss-channel points toward larger $$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$0 and smaller $$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$1, increases transmitted entanglement as quantified through the entanglement of formation of the Choi state, and can even realize effective points in the $$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$2 plane that the authors call “non-physical Gaussian channels,” meaning channels with no deterministic physical Gaussian equivalent (Shajilal et al., 2024). The crucial caveat is explicit: there is no violation of Gaussian no-go theorems because the MBNLA is probabilistic.

A second heralded purification-like mechanism is passive linear-optical unitary averaging against stochastic phase noise. In that protocol, one uses multiple parallel applications of a phase-noisy bosonic channel, balanced interferometric encoding/decoding, and vacuum postselection on $$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$3 output modes (Swain et al., 2023). For a fixed realization of the random phases, the successful output is again Gaussian; after averaging over the phase ensemble, the state is generally non-Gaussian but approximately Gaussian for weak noise. In that weak-noise regime the protocol improves purity, squeezing, and entanglement, typically succeeds with high probability, and already for $$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$4 improves output squeezing by about $$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$5–$$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$6 dB (Swain et al., 2023). This is therefore a probabilistic bosonic purification protocol, but not a deterministic Gaussian channel in the strict CPTP Gaussian sense.

4. Exact random Gaussian purification channels

A distinct and exact use of the term appears in recent work on random purification channels. For passive Gaussian bosons, an exact channel was constructed that, given $$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$7 copies of an unknown $$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$8-mode passive Gaussian state $$V_{\rm out}=X V_{\rm in}X^T+Y,\qquad X=\sqrt{\tau}\,\mathbb 1,\qquad Y=\nu\,\mathbb 1$$9, outputs $(\tau,\nu)$0 copies of one randomly chosen Gaussian purification of $(\tau,\nu)$1 (Mele et al., 18 Dec 2025). If $(\tau,\nu)$2 denotes the standard Gaussian purification and $(\tau,\nu)$3 is a passive Gaussian unitary indexed by $(\tau,\nu)$4, the channel satisfies

$(\tau,\nu)$5

(Mele et al., 18 Dec 2025). Each output purification is Gaussian, and each has mean photon number exactly twice that of the initial passive Gaussian state (Mele et al., 18 Dec 2025). The construction is exact, completely positive, and trace preserving, and relies on the characterization of the commutant of passive Gaussian unitaries through bosonic Howe duality.

A parallel but more general symmetry-theoretic construction establishes the existence of bosonic Gaussian purification channels for gauge-invariant Gaussian states. For every number of modes $(\tau,\nu)$6 and copies $(\tau,\nu)$7, there exists a channel

$(\tau,\nu)$8

from $(\tau,\nu)$9 to $0<\tau<1,\qquad \nu=(1-\tau)\chi,$0 such that, for every gauge-invariant bosonic Gaussian state $0<\tau<1,\qquad \nu=(1-\tau)\chi,$1,

$0<\tau<1,\qquad \nu=(1-\tau)\chi,$2

(Walter et al., 17 Dec 2025). The randomization is over passive Gaussian unitaries, i.e. the compact subgroup $0<\tau<1,\qquad \nu=(1-\tau)\chi,$3. The restriction to gauge-invariant states is essential: the full bosonic Gaussian group is noncompact, there is no Haar probability measure for a uniform twirl over all Gaussian unitaries, and the infinite-dimensional irreducible sectors do not admit maximally mixed states (Walter et al., 17 Dec 2025).

These random purification channels are exact purification channels in the strict state-theoretic sense. They do not suppress noise on a transmitted signal; rather, they map copies of a mixed Gaussian state to copies of a random Gaussian purification of that state. They therefore address a different problem from Gaussian error mitigation, even though both are described by purification language.

5. Least-noisy outputs, extremal preparations, and adjacent notions

Purification should be distinguished from extremal-output theorems. For every single-mode phase-insensitive bosonic Gaussian channel $0<\tau<1,\qquad \nu=(1-\tau)\chi,$4, the output generated by any coherent state majorizes the output generated by any other input: $0<\tau<1,\qquad \nu=(1-\tau)\chi,$5 (Mari et al., 2013). Equivalently, coherent states uniquely minimize every nonnegative strictly concave unitary-invariant output functional, including the von Neumann entropy and the relevant Rényi entropies (Mari et al., 2013). This identifies the least noisy output permitted by the channel, but it does not furnish a purification protocol or a channel that purifies arbitrary mixed states.

The same distinction appears in Gaussian thermodynamics. For a Gaussian channel $0<\tau<1,\qquad \nu=(1-\tau)\chi,$6 and a quadratic Hamiltonian

$0<\tau<1,\qquad \nu=(1-\tau)\chi,$7

the minimum output energy over all input states is

$0<\tau<1,\qquad \nu=(1-\tau)\chi,$8

when $0<\tau<1,\qquad \nu=(1-\tau)\chi,$9 is invertible (Rodriguez et al., 27 Mar 2025). An optimal input is Gaussian with

$\tau>1,\qquad \nu=(\tau-1)\chi,$0

where $\tau>1,\qquad \nu=(\tau-1)\chi,$1 Williamson-diagonalizes $\tau>1,\qquad \nu=(\tau-1)\chi,$2 (Rodriguez et al., 27 Mar 2025). This is an exact minimal-output-energy theorem, and it is closely related to cooling or excitation suppression, but it is not an entropy-purification theorem. It characterizes the pure squeezed-displaced Gaussian preparation that yields the coldest output relative to a chosen quadratic observable, with the irreducible energetic penalty $\tau>1,\qquad \nu=(\tau-1)\chi,$3 fixed by channel-added noise (Rodriguez et al., 27 Mar 2025).

Accordingly, least-noisy outputs, minimal-output-energy preparations, and pure-environment dilations are adjacent but distinct concepts. They all identify extremal structure of bosonic Gaussian channels, yet only the dilation and random-purification constructions are purification channels in the literal sense.

6. Limitations, misconceptions, and non-Gaussian optima

The principal misconception is to identify every purification-like effect with an ordinary deterministic Gaussian CPTP map. The heralded teleportation scheme of Gaussian noise suppression is explicitly conditional and probabilistic, and its ability to reach near-identity behavior or “non-physical Gaussian channels” relies on postselection approximating the NLA $\tau>1,\qquad \nu=(\tau-1)\chi,$4 (Shajilal et al., 2024). The unitary-averaging protocol against phase noise is likewise heralded and only approximately Gaussian after averaging over the stochastic phase ensemble (Swain et al., 2023). These are purification-like channels in an operational sense, but not ordinary Gaussian channels in the Holevo-Werner classification.

A second misconception is to assume that an optimal purification task between Gaussian input and Gaussian target should itself be achieved by a Gaussian channel. For coherent thermal states, the asymptotically optimal phase-insensitive purification protocol is not Gaussian (Yadavalli et al., 2024). In that setting, the task is to convert many copies of

$\tau>1,\qquad \nu=(\tau-1)\chi,$5

into a single mode close to $\tau>1,\qquad \nu=(\tau-1)\chi,$6. Among Gaussian phase-insensitive channels, the optimal asymptotic infidelity factor is

$\tau>1,\qquad \nu=(\tau-1)\chi,$7

and the optimal Gaussian protocol is a linear-optical beam-splitter scheme with vacuum ancilla (Yadavalli et al., 2024). But the true optimum over all phase-insensitive channels is

$\tau>1,\qquad \nu=(\tau-1)\chi,$8

so the globally optimal purification channel must be non-Gaussian (Yadavalli et al., 2024). The same work identifies the inverse purity of coherence, derived from the RLD Fisher information, as the asymptotic error coefficient governing optimal purification (Yadavalli et al., 2024).

Finally, pure-environment implementation does not by itself guarantee Gaussianity of the resulting channel. One may keep the standard Gaussian dilating unitaries $\tau>1,\qquad \nu=(\tau-1)\chi,$9 or $\tau=1,\qquad \nu>0,$0 and replace the Gaussian environment by a pure superposition of two maximally distinguishable Gaussian states, obtaining linear bosonic non-Gaussian attenuator- or amplifier-type channels (Volkoff, 2017). Such constructions remain purification-based at the level of dilation, but they lie outside the Gaussian channel class.

Taken together, the literature supports a precise taxonomy. A bosonic Gaussian purification channel may be a pure-environment Gaussian dilation and complementary map; an exact random channel that outputs Gaussian purifications of passive or gauge-invariant Gaussian states; or a heralded, non-Gaussian-assisted map that conditionally suppresses effective Gaussian noise. What it is not, in general, is a universal deterministic Gaussian CPTP transformer that purifies arbitrary mixed bosonic Gaussian states.