betaFeedback

WorksPhysics and Astronomy

Universal Quantum Computation in a Wigner's Friend Framework

Abstract

The Wigner's Friend thought experiment illustrates a core tension in quantum mechanics: an observer inside an isolated laboratory (the Friend) measures a definite outcome, while an external observer (Wigner) describes the laboratory as a coherent superposition. Here we show that this discrepancy, usually studied as a conflict between descriptions, has a direct computational consequence. We namely propose a protocol by which Wigner can turn the Friend's laboratory into a universal quantum computer, whereby the Friend is restricted to apply simple classical operations within the laboratory. We also experimentally implement a proof-of-principle demonstration of this protocol for single-qubit gates using single photons. By recasting the coherence between incompatible descriptions as a computational resource, we thus link the foundations of quantum measurement to quantum computation.

Introduction

The Wigner’s Friend thought experiment, proposed in 1961 to highlight the measurement problem [1], has recently come back into focus, through the use of modern quantum information tools, which allow for a more rigorous analysis [2, 3, 4, 5, 6, 7, 8, 9, 10]. In the original thought experiment, a Friend performs a measurement of a quantum system inside an isolated laboratory, recording a definite outcome (from her point of view). However, from the perspective of an external agent, Wigner, who is powerful enough to treat the Friend in her laboratory as a closed quantum system, the situation looks different. Wigner describes the Friend’s state and the measured system as becoming entangled via the measurement interaction [11]. As such, the measurement outcome (and the Friend herself) are left in a superposition, entangled with the quantum system. Therefore, Wigner does not describe the laboratory as hosting a definite outcome, in seeming conflict with the Friend’s observations. This conflict, while implied by ordinary quantum theory, is surely puzzling, and is accordingly made sense of in different ways by different interpretations of the theory [6, 7, 12, 13].

A range of extensions to the above thought experiment have since been developed [2], including nested observers [3], no-go theorems on observer-independent facts [4, 14], on local friendliness [10, 15], and experimental tests of local observer independence [9, 10, 16]. In all of these, the Wigner’s Friend scenario is used to make the disagreement between the agents’ descriptions explicit. Here, instead, we propose a computational consequence of the thought experiment. We show that

Wigner can turn the Friend’s laboratory—which contains a rudimentary classical computer—into a universal quantum computer even when the Friend is restricted to applying simple bit permutations that cannot even achieve universal classical computation. Wigner achieves this by sending the Friend messages that are superpositions of instructions to apply different permutations. The Friend only ever reads a definite classical instruction and permutes bits; she never learns about the superposed message, which Wigner prepares outside the laboratory. The resource allowing this computational power stems from Wigner’s quantum-coherent manipulation of the entire lab, including the Friend and her environment. Universal quantum computation would not be possible if information about the Friend’s operations were to leak outside of the lab.

Our result builds on recent work showing that quantum-coherently controlling a restricted set of classical operations, which by themselves are not universal for classical computation, is sufficient to achieve universal quantum computation [17], building on a broader line of work on the coherent control of quantum operations [18, 19, 20]. While that construction was formulated in an abstract control-target setting, here we embed it into a Wigner’s Friend scenario where the Friend acts as an observing mediator between Wigner’s control message and a classical target register inside her laboratory. More precisely, the Friend is assumed to be located within a laboratory, where she has access to the aforementioned target system, on which computations are to be applied. Wigner is located outside of the laboratory, but still allowed to communicate with the Friend, by exchanging messages through a quantum channel, taking the place of the control qubit. Our protocol then consists of the following steps:

  1. Wigner prepares a message to send into the labora-

tory, naming some of the classical operations that the Friend is instructed to apply to her register of bits. This message may be prepared in a coherent superposition of different instructions (Figure 1-Step 1).

Simplified sketch of the protocol steps

Figure 1. Simplified sketch of the protocol steps. Step 1: Wigner (W) prepares the message ∣ψ⟩|\psi\rangle in a superposition state, and sends it to the laboratory, where the Friend (F) waits, with her memory initially being in the “ready state” R. The only operations that the Friend can do on her target register - illustrated as a 2 × 2 grid - are the identity operation or a permutation. Step 2: The Friend reads the message - with her memory recording its content - and applies the instructed operation on the target register. Step 3: Wigner reverses the state of the whole lab, except for the target register: the Friend’s memory is thus again in state R. In Step 4, Wigner applies a unitary operation on the message and sends it back into the lab. Iterations of this procedure effectively generate unitary transformations on the target register.

  1. The Friend receives the message, reads it, memorizes its content, and applies the corresponding operation on her target (Figure 1-Step 2).

  1. The message is returned to Wigner, who then applies an “erasure unitary operation” to the whole laboratory except for the target system, thereby reversing the Friend’s state –including her memory – to what it was before she had received the message (Figure 1-Steps 3 & 4).

  1. After this, Wigner proceeds by either reading the message or applying a unitary operation on it (Figure 1-Steps 3 & 4), before again sending it back into the laboratory and repeating steps 1 to 3.

Remarkably, iterating this procedure allows Wigner to execute universal quantum computation on the target system, even though the Friend only ever performs deterministic, non-universal classical operations. Because the Friend’s operations alone cannot even generate universal classical computation, the protocol’s computational power stems predominantly from Wigner’s coherent control of the lab, including the Friend and her memory.

Notice that the technologically infeasible point of our protocol is step 3, which requires the possibility of quantum-coherently manipulating a macroscopic laboratory that contains a human being embedded in a complex environment. Despite its (current) technological infeasibility, this possibility—which has also been appealed to in previous work on Wigner’s Friend [10, 21]—is nevertheless a theoretically legitimate one, at least under the usual assumption that unitary quantum theory applies to arbitrarily complex physical systems [22]. Without this assumption our protocol could not work, since records within the laboratory would essentially decohere the state of the target system.

While the macroscopic implementation of our protocol can not currently be experimentally tested, here we implement a microscopic analogy thereof, using smaller quantum systems to demonstrate the use of our protocol to implement single-qubit gates. In particular, we use a photonic system to implement the logical gate RX(θ)=e−iθ2X^R_X(\theta) = e^{-\frac{i\theta}{2}\hat{X}} at sixteen values of θ\theta covering a full rotation about the logical XX axis, with average fidelity ⟨F⟩=0.9885±0.0012\langle F\rangle= 0.9885\pm0.0012 and average purity ⟨P⟩=0.9775±0.0023\langle P\rangle= 0.9775\pm0.0023 obtained by maximum-likelihood quantum state tomography [23]. In our experiment, the target system and the Friend are encoded in different degrees of freedom of a single photon [24], and are thus not physically separate

systems. Nevertheless, our experiment does demonstrate a realization of the formal protocol structure sketched above. Whether this structure can be scaled to progressively larger systems, however, remains an open experimental challenge.

Results

Framework

Aside from Wigner, three systems are needed for our protocol: a target system made out of 2n2n qubits (bits from the Friend’s perspective), a quantum message, and the Friend. The latter three will enter the quantummechanical descriptions. Wigner will, on the contrary, not be modeled explicitly, as he is an external agent that conducts experiments on the laboratory and on the message.

State-spaces

Target system and classical operations. The Friend is assumed to be located inside a laboratory, where she has access to a target system made out of 2n2n qubits. She is, however, only allowed to execute a limited range of operations on the target, amounting to mere permutations in a fixed basis. This makes said qubits effectively behave as classical bits. More precisely, let the so-called “classical basis” be spanned by vectors {∣x⃗⟩≡∣x1⋯x2n⟩, xi∈{0,1}}\{|\vec{x}\rangle\equiv|x_1\cdots x_{2n}\rangle,\ x_i \in\{0,1\}\}. We assume that the Friend is only capable of implementing operations drawn from the following set

G={1}∪⋃l=1n{G1(l),G2(l)},(1)\mathcal{G}=\{1\}\cup\bigcup_{l=1}^{n}\left\{G_1^{(l)},G_2^{(l)}\right\}, \tag*{(1)}

where Gi(l)≡1⊗(2l−2)⊗Gi⊗1⊗(2n−2l)G_i^{(l)}\equiv1^{\otimes(2l-2)}\otimes G_i\otimes1^{\otimes(2n-2l)} acts non-trivially only on the ll-th qubit pair as

G1∣xlxl+1⟩=∣(xl⊕1)xl+1⟩,G_1|x_lx_{l+1}\rangle=|(x_l\oplus1)x_{l+1}\rangle,
G2∣xlxl+1⟩=∣xl(xl+1⊕xl)⟩.(2)G_2|x_lx_{l+1}\rangle=|x_l(x_{l+1}\oplus x_l)\rangle. \tag*{(2)}

Each element of G\mathcal{G} is a permutation of the classical basis: either a NOT on the first qubit pertaining to a pair, or a CNOT on the pair, or the identity. Relative to the Friend’s operations, the 2n2n qubits thus behave as if they were 2n2n classical bits. Note that NOTs and CNOTs alone cannot generate even the AND operation. The Friend can accordingly not even implement universal classical computation, let alone quantum computation. Therefore, all of the computational power of our protocol will have to come from outside the laboratory.

Message. The message is a quantum system that Wigner uses to instruct the Friend which operation she is to execute within the laboratory. The Friend will read the message by measuring it in the computational basis and from her point of view observe a definite outcome. The state space HM\mathcal{H}_M associated to the message will accordingly correspond to the space of possible operations that the Friend is capable of carrying out in the laboratory:

HM=span⁡{∣g⟩M:g∈G},(3)\mathcal{H}_M=\operatorname{span}\{|g\rangle_M:g\in\mathcal{G}\}, \tag*{(3)}

Each state ∣Gi(l)⟩M|G_i^{(l)}\rangle_M carries the instruction to execute operation Gi(l)G_i^{(l)}, whereas ∣1⟩M|1\rangle_M carries the instruction to do nothing at all. Note that no restriction is placed on the quantum operations that Wigner is allowed to implement on the message outside of the laboratory, though only a few limited single-qubit gates will already suffice.

Friend. The Friend remains inside the lab, and reads messages sent by Wigner, telling her which operation to perform on the target system. As usual in Wigner-Friend scenarios, we will model the Friend as a memory that keeps a record of the operations that she has implemented. We thus choose to model the state-space HF\mathcal{H}_F associated to the Friend in analogy with the state-space associated to the message:

HF=span⁡({∣R⟩F}∪{∣g⟩F:g∈G}).(4)\mathcal{H}_F=\operatorname{span}\left(\{|R\rangle_F\}\cup\{|g\rangle_F:g\in\mathcal{G}\}\right). \tag*{(4)}

Each state ∣Gi(l)⟩F|G_i^{(l)}\rangle_F represents the Friend after having read and executed instruction Gi(l)G_i^{(l)}, ∣1⟩F|1\rangle_F represents the Friend after having read the instruction to do nothing, and having done nothing; finally, the ready state ∣R⟩F|R\rangle_F represents the Friend before having received any message at all.

Operations

Read and execute. In our protocol, Wigner sends quantum messages into the laboratory (step 1), where the Friend reads it and follows the instructions (step 2). We model this latter step as a sequence of two transformations. First, the reading is described by a coupling unitary CC, which transitions the Friend’s memory from its ready-state into a record of the read instruction:

C∣g⟩M∣R⟩F=∣g⟩M∣g⟩F,g∈G.(5)C|g\rangle_M|R\rangle_F=|g\rangle_M|g\rangle_F,\qquad g\in\mathcal{G}. \tag*{(5)}

Next, the Friend executes the instruction by applying the corresponding gate to the target system: if she reads that operation gg is to be executed, she applies this gg on the target system. The net effect of this two-step procedure is

∣g⟩M∣R⟩F∣ϕ⟩⟼∣g⟩M∣g⟩Fg∣ϕ⟩,g∈G,(6)|g\rangle_M|R\rangle_F|\phi\rangle\longmapsto|g\rangle_M|g\rangle_Fg|\phi\rangle,\qquad g\in\mathcal{G}, \tag*{(6)}

where ∣ϕ⟩|\phi\rangle is an arbitrary state of the target system. Note that Eq. (5) specifies CC only on a specific list of input states—those in which the Friend is in a ready-state—thereby underspecifying the total unitary.

The erasure unitary. After step 2, the message and the Friend’s memory are correlated. If this correlation were to survive, the Friend’s memory would effectively amount to a which-path record that destroys the interference between different instruction branches. In order to maintain coherence, Wigner must therefore reset the Friend’s memory to its ready-state (step 3). More precisely, after the Friend reads the message, executes its instruction, and sends the message outside of the laboratory, Wigner implements an “erasure-unitary” EE that acts as follows:

E∣g⟩M∣g⟩F=∣g⟩M∣R⟩F,g∈G.(7)E|g\rangle_M|g\rangle_F=|g\rangle_M|R\rangle_F,\qquad g\in\mathcal{G}. \tag*{(7)}

This requires coherent quantum control over the entire laboratory, similar to the additional elements of the extended Wigner’s Friend scenarios [9, 10, 22]; we will return to this in § 7.

Logical Qubit encoding

As introduced in § 2 A 1, the Friend’s target system consists of 2n2n target qubits. In our protocol, these will effectively encode nn “logical qubits” with basis vectors given by:

∣0⟩L≡12(∣10⟩−∣11⟩),∣1⟩L≡12(∣00⟩−∣01⟩),(8)\lvert0\rangle_L \equiv\frac{1}{\sqrt{2}}(\lvert10\rangle-\lvert11\rangle), \qquad\lvert1\rangle_L \equiv\frac{1}{\sqrt{2}}(\lvert00\rangle-\lvert01\rangle), \tag*{(8)}

where the classical basis vectors ∣xy⟩\lvert xy\rangle correspond to pairs of adjacent target qubits. The ll-th logical qubit will accordingly be encoded in the above two-dimensional subspace of target qubits (2l−1,2l)(2l-1,2l), for any l=1,…,nl=1,\ldots,n. The logical qubit states cannot be prepared by the Friend alone, since they are given by superpositions of classical basis states. Nevertheless, as we show in § 8 A, they can be prepared with Wigner’s help, via a procedure similar to the general protocol presented below.

Note that the Friend’s gates—which amount to mere permutations in the classical basis—effectively act as Pauli gates in the logical basis:

G1(l)∣y⟩L=∣y⊕1⟩L,G2(l)∣y⟩L=−(−1)y∣y⟩L,(9)G^{(l)}_1\lvert y\rangle_L=\lvert y\oplus1\rangle_L,\qquad G^{(l)}_2\lvert y\rangle_L=-(-1)^y\lvert y\rangle_L, \tag*{(9)}

Thus, in the logical basis: G1(l)≡X^G^{(l)}_1\equiv\hat{X} and G2(l)≡−Z^G^{(l)}_2\equiv-\hat{Z}.

The protocol

We will now show how combining the elements introduced above enables the implementation of single-qubit rotations e−iθ2Xe^{-i\frac{\theta}{2}X} and eiθ2Ze^{i\frac{\theta}{2}Z}, as well as two-qubit entangling gates e−iθ2X⊗Xe^{-i\frac{\theta}{2}X\otimes X} and e−iθ2Z⊗Ze^{-i\frac{\theta}{2}Z\otimes Z}, for arbitrary angles θ\theta. Since these gates form a universal set, this will imply that our protocol allows Wigner to execute any quantum gate on the nn-logical-qubit subspace of the target system [28]. Implementing each basis gate requires one full iteration of the four-step scheme announced in § 1, followed by a second repetition of Steps 2 and 3, as will be detailed in what follows - see also Figure 2 for the circuit that represents the protocol.

Quantum circuit of the protocol

Figure 2. Quantum circuit of the protocol. a) Complete quantum circuit on the message (m), memory (f) and target (t), whose two wires are the pair of bits the Friend always acts on in the classical basis. Initially, the state of the message is an arbitrary ∣m⟩M\lvert m\rangle_M state, the memory is in ∣R⟩F\lvert R\rangle_F, and the target qubit is already in a given logical state ∣φ⟩L\lvert\varphi\rangle_L. b) Step 1: Wigner prepares the message ∣+⟩M≡(∣1⟩M+∣G⟩M)/2\lvert+\rangle_M\equiv(\lvert1\rangle_M+\lvert G\rangle_M)/\sqrt{2}, by acting on it with a unitary, UU. c) Step 2: the coupling CC takes her memory from ∣R⟩F\lvert R\rangle_F to a record of what she read, ∣g⟩F\lvert g\rangle_F. In one branch it becomes ∣1⟩F\lvert1\rangle_F and she leaves the register alone; in the other it becomes ∣G⟩F\lvert G\rangle_F and applies the classical operation. d) Steps 3 and 4: the erasure EE resets the Friend’s memory to ∣R⟩F\lvert R\rangle_F. Wigner then applies RX(θ)R_X(\theta) to the message, encoding the desired rotation angle. e) Steps 2 and 3 are repeated, leaving the message in the original state after Wigner’s initialisation, ∣+⟩M\lvert+\rangle_M, the memory in state ∣R⟩F\lvert R\rangle_F, and having applied e−iθG/2e^{-i\theta G/2} to the input target state, which for G=G1G=G_1 is the logical rotation RX(θ)R_X(\theta).

Single-qubit gates

Let ∣ψ⟩L\lvert\psi\rangle_L be an arbitrary state of the ll-th logical qubit. To implement a rotation generated by Gi(l)G^{(l)}_i (i=1,2i=1,2), Wigner initializes the message in the superposition ∣+(i)⟩M≡12(∣1⟩M+∣Gi(l)⟩M)\lvert+^{(i)}\rangle_M\equiv\frac{1}{\sqrt{2}}(\lvert1\rangle_M+\lvert G^{(l)}_i\rangle_M) and sends it into the laboratory (Step 1). The Friend then reads the message and applies the corresponding classical permutation on

her target register (Step 2), after which Wigner applies the erasure unitary EE, resetting the Friend’s memory to ∣R⟩F\lvert R\rangle_F (Step 3). Up to this point, the joint state of the message, the Friend, and the logical target evolves as

Step 1⟶∣+(i)⟩M∣R⟩F∣ψ⟩LStep 2⟶12(∣1⟩M∣1⟩F∣ψ⟩L+∣Gi(l)⟩M∣Gi(l)⟩FGi(l)∣ψ⟩L)Step 3⟶12(∣1⟩M∣ψ⟩L+∣Gi(l)⟩MGi(l)∣ψ⟩L)∣R⟩F.(10)\begin{aligned} \text{Step 1}\quad&\longrightarrow\lvert+^{(i)}\rangle_M\lvert R\rangle_F\lvert\psi\rangle_L\\ \text{Step 2}\quad&\longrightarrow\frac{1}{\sqrt{2}}\left(\lvert1\rangle_M\lvert1\rangle_F\lvert\psi\rangle_L+\lvert G^{(l)}_i\rangle_M\lvert G^{(l)}_i\rangle_FG^{(l)}_i\lvert\psi\rangle_L\right)\\ \text{Step 3}\quad&\longrightarrow\frac{1}{\sqrt{2}}\left(\lvert1\rangle_M\lvert\psi\rangle_L+\lvert G^{(l)}_i\rangle_MG^{(l)}_i\lvert\psi\rangle_L\right)\lvert R\rangle_F. \tag*{(10)} \end{aligned}

With the Friend’s memory factored out, the message and the target register remain entangled. In Step 4, Wigner applies the single-qubit rotation Ri(l)(θ)=e−iθ2XMR_i^{(l)}(\theta) = e^{-i\frac{\theta}{2}X_M} to the returned message, corresponding to a rotation by angle θ\theta around the XX axis of the two-dimensional message subspace spanned by {∣1⟩M,∣Gi(l)⟩M}\{|1\rangle_M, |G_i^{(l)}\rangle_M\}. He then sends it back into the laboratory for a second read-execute-erase round (repeating Steps 2 and 3). As it can be easily verified, this second round decouples the message and leaves the joint system in the final state

∣+i(l)⟩M∣R⟩Fe−iθ2Gi(l)∣ψ⟩L.(11)|+_i^{(l)}\rangle_M |R\rangle_F e^{-i\frac{\theta}{2}G_i^{(l)}} |\psi\rangle_L . \tag*{(11)}

Since G1(l)≡XG_1^{(l)} \equiv X and G2(l)≡−ZG_2^{(l)} \equiv-Z on the ll-th logical subspace (Eq. (9)), choosing i=1i = 1 implements e−iθ2Xe^{-i\frac{\theta}{2}X} while choosing i=2i = 2 implements eiθ2Ze^{i\frac{\theta}{2}Z} on the logical qubit, leaving both the message and the Friend’s memory restored to their initial states.

Two logical qubits and universality

It is simple to notice that, by replacing gate Gi(l)G_i^{(l)} by any other gate GG that satisfies G2=1G^2 = \mathbb{1}, our protocol can implement the according rotation eiθ2Ge^{i\frac{\theta}{2}G}. Now, since (Gi(l)Gi(k))2=1\left(G_i^{(l)}G_i^{(k)}\right)^2 = \mathbb{1}, an analogous protocol can also be used to generate two-qubit gates e−iθ2Gi(l)⊗Gi(k)e^{-i\frac{\theta}{2}G_i^{(l)}\otimes G_i^{(k)}} on the logical qubit-pair (l,k)(l,k). To implement these the state-spaces of the message and the Friend simply need to be expanded to include states that correspond to pairs of operations: e.g. ∣Gi(l),Gi(k)⟩M|G_i^{(l)},G_i^{(k)}\rangle_M is the state of the message that contains the instruction to implement Gi(l)Gi(k)G_i^{(l)}G_i^{(k)}, whereas ∣Gi(l),Gi(k)⟩F|G_i^{(l)},G_i^{(k)}\rangle_F is the state of the Friend after reading this message and executing its instruction.

Running the steps from before, with the message instead initially prepared in ∣+i(l,k)⟩M≡(∣1(l),1(k)⟩M+∣Gi(l),Gi(k)⟩M)/2|+_i^{(l,k)}\rangle_M \equiv(|1^{(l)},1^{(k)}\rangle_M + |G_i^{(l)},G_i^{(k)}\rangle_M)/\sqrt{2}, thus yields output state

∣+i(l,k)⟩M∣R⟩Fe−iθ2Gi(l)⊗Gi(k)∣ψ⟩L,(12)|+_i^{(l,k)}\rangle_M |R\rangle_F e^{-i\frac{\theta}{2}G_i^{(l)}\otimes G_i^{(k)}}|\psi\rangle_L , \tag*{(12)}

thereby implementing e−iθ2X⊗Xe^{-i\frac{\theta}{2}X\otimes X} or e−iθ2Z⊗Ze^{-i\frac{\theta}{2}Z\otimes Z} on the logical qubit-pair (l,k)(l,k). These entangling gates, together with the prior single-qubit gates, make universal quantum computing possible.

Experimental demonstration

We will now experimentally implement the single-qubit gate e−iθ2X^e^{-i\frac{\theta}{2}\hat{X}} using our protocol. That is, we set n=1n=1 and G=G1(1)G=G_1^{(1)}. Our experiment has two stages, shown in Figure 3: Wigner, following the steps in § 8 A, first prepares ∣0⟩L|0\rangle_L from a classical bit string (orange), and then runs steps 1-4 followed by another round of step 2 and step 3. Step 1 is shown in the green region, steps 2 and 3 are represented in the blue areas, while step 4 is shown in the pink area. All of the required state spaces are encoded in a single heralded photon at 1550 nm.

Illustration of the experimental setup

Figure 3. Illustration of the experimental setup. The experiment uses heralded single photons at 1550 nm. Polarisation encodes the message MM, a 2×22 \times2 grid of spatial modes encodes the classical basis, and the path relative to the double polarising beam splitters (dPBS) encodes the Friend’s memory. Wigner prepares the message by using a half-waveplate to modify the polarisation (green). State preparation (orange) converts the input into the logical state ∣0⟩L|0\rangle_L using a calcite beam displacer, waveplates and a polariser. In pink we have the action corresponding to step 4. Each gate stage (blue) implement steps 2 and 3: the entry dPBS is the coupling CC, a pair of cylindrical lens telescope applies the permutation G1=X^G_1 = \hat{X} to the reflected arm, carrying the record ∣G1⟩F|G_1\rangle_F, and the exit dPBS is the erasure EE. Between the two rounds Wigner (pink) applies the rotation RX(θ)R_X(\theta) to the message using a quarter-, half-, quarter-waveplate gadget. The output is analysed on the readout stage (grey) by performing single-photon tomography. The legend at right identifies the optical components.

Encoding. Each system is encoded in a different degree of freedom (DOF) of the same photon. The “classical” two-bit target system is encoded in a 2×22 \times2 grid of spatial modes. A single photon in each mode thus represents a two-bit string as labeled in Figure 4a. The Friend’s allowed classical operations are permutations of these spatial modes. Experimentally, gates G1G_1 and G2G_2 are implemented with cylindrical-lens telescopes to permute modes as illustrated in Figure 4c and d, respectively. For the state preparation protocol (§ 8 A), the Friend implements G2G_2. However, since for this step the photon enters the experiment in a well-defined spatial mode (i.e. in the “classical state” ∣10⟩|10\rangle), we use a different implementation of G2G_2 shown in Figure 4b. In particular, we use a calcite beam displacer to shift the mode upwards by 2.7 mm. This allows us to precisely set the row spacing, which will be essential for our final verification of the protocol, wherein the spatial modes must be coherently recombined (readout stage of Figure 3).

Experimental encoding and gates

Figure 4. Experimental encoding and gates. Optical implementation of the classical gates on the spatially encoded target. a) Bit encoding: the classical basis {∣00⟩,∣01⟩,∣10⟩,∣11⟩}\{|00\rangle, |01\rangle, |10\rangle, |11\rangle\} is mapped onto the four spatial modes of a 2×22 \times2 grid. b) The G2G_2 (−Z^-\hat{Z}) gate used during state preparation, implemented with a calcite beam displacer that shifts the input ∣10⟩|10\rangle mode vertically onto ∣11⟩|11\rangle. c) The G1G_1 (X^\hat{X}) gate used in the logical gate stage, implemented with a 1:1 cylindrical-lens imaging system whose optical axis is aligned vertically with the centre of the grid, exchanging the two columns and mapping ∣x1x2⟩↦∣(x1⊕1)x2⟩|x_1x_2\rangle\mapsto|(x_1 \oplus1)x_2\rangle. d) The same 1:1 imaging system with its axis rotated to horizontal, and placed so that only the first column passes through the lenses and the bits ∣10⟩↔∣11⟩|10\rangle\leftrightarrow|11\rangle (the rows in the first-column) are swapped, giving G2G_2 (−Z^-\hat{Z}) on the full grid.

Wigner’s message is encoded in the polarisation of the same photon with ∣H⟩↔∣1⟩M|H\rangle\leftrightarrow|1\rangle_M and ∣V⟩↔∣G⟩M|V\rangle\leftrightarrow|G\rangle_M. To read the message, the Friend must couple it to her memory. This is accomplished with a double polarising beam splitters (dPBS) as follows. The Friend’s memory is another path degree of freedom of the photon. Before the dPBS, the memory is in the ready state ∣R⟩F|R\rangle_F, as there is only a single path and no record of any instruction. Upon entering the dPBS ∣H⟩|H\rangle is transmitted and ∣V⟩|V\rangle reflected. Thus the entire grid is transmitted or reflected. This is the coupling CC of Eq. (5). The transmitted arm encodes ∣1⟩F|1\rangle_F. Thus, this arm is empty and does nothing. On the other hand, the reflected arm represents ∣G⟩F|G\rangle_F, in which the permutation GG is placed. The absence or presence of GG thus implements the controlled-∣G⟩|G\rangle gates of Figure 2. After this, a second dPBS recombines the two arms into one path without measuring which one the photon took, which implements the erasure EE of Eq. (7). Unlike a standard PBS, a dPBS sends its transmitted and reflected outputs out parallel to each other, allowing the interferometer to close with fewer mirrors than a conventional one. This matters because mirrors can introduce incoherence between the two branches that the protocol must keep coherent. In practice the setup remains stable without active stabilisation for at least 30 minutes, which is longer than a full characterization measurement (Supplementary Information, SI A).

In summary, Wigner only acts on the polarization of the photon (the message). The Friend reads the polarization by coupling it to spatial modes using dPBSs. She cannot otherwise act on the polarization, she is only able to implement permutations on the spatial modes.

State preparation. The photon starts in ∣H⟩∣10⟩|H\rangle|10\rangle, with the message reading ∣1⟩M|1\rangle_M. Wigner applies a Hadamard to the message with a HWP at 22.5∘22.5^\circ and sends the photon through an interferometer whose reflected arm holds the beam displacer, where the Friend executes G2G_2 as ∣10⟩↦∣11⟩|10\rangle\mapsto|11\rangle (Figure 4b). A second Hadamard and a polariser transmitting ∣V⟩\lvert V\rangle then leave the target in ∣0⟩L\lvert0\rangle_L, with success probability 1/21/2. The coherence between ∣10⟩\lvert10\rangle and ∣11⟩\lvert11\rangle comes directly from the coherence of Wigner’s message: the Friend simply coupled the message to her additional spatial modes (memory) by displacing the photon based on its polarization.

Gate. After the state preparation, Wigner acts on the message state again, using a HWP at 67.5∘67.5^\circ to rotate ∣V⟩\lvert V\rangle to ∣D⟩\lvert D\rangle and prepares ∣+⟩M∣0⟩L\lvert+\rangle_M\lvert0\rangle_L. In the first gate stage, the reflected arm holds two cylindrical lenses of 80 mm focal length in a 1:1 inverting telescope, axes vertical and centred on the grid. The telescope swaps the two columns, ∣x1x2⟩↦∣(x1⊕1)x2⟩\lvert x_1x_2\rangle\mapsto\lvert(x_1\oplus1)x_2\rangle, which is G1G_1 (Figure 4c). On the logical qubit it takes ∣0⟩L\lvert0\rangle_L to ∣1⟩L\lvert1\rangle_L, so after the first exit dPBS

∣D⟩∣0⟩L⟶12(∣H⟩∣0⟩L+∣V⟩∣1⟩L).(13)\lvert D\rangle\lvert0\rangle_L \longrightarrow\frac{1}{\sqrt{2}}\left(\lvert H\rangle\lvert0\rangle_L+\lvert V\rangle\lvert1\rangle_L\right). \tag*{(13)}

Between the two gate stages, Wigner rotates the polarisation with a quarter-, half-, quarter-wave plate sequence, RX(θ)=QWP⁡(90∘)HWP⁡(θ/4)QWP⁡(90∘)R_X(\theta) = \operatorname{QWP}(90^\circ)\operatorname{HWP}(\theta/4)\operatorname{QWP}(90^\circ). The angle θ\theta is set by the angle of a single HWP. With c=cos⁡θ2c = \cos\frac{\theta}{2} and s=sin⁡θ2s = \sin\frac{\theta}{2}, the state becomes

12[∣H⟩(c∣0⟩L−is∣1⟩L)+∣V⟩(c∣1⟩L−is∣0⟩L)].(14)\frac{1}{\sqrt{2}}\left[\lvert H\rangle\left(c\lvert0\rangle_L-is\lvert1\rangle_L\right)+\lvert V\rangle\left(c\lvert1\rangle_L-is\lvert0\rangle_L\right)\right]. \tag*{(14)}

The second gate stage is a copy of the first. It flips the logical qubit in the ∣V⟩\lvert V\rangle branch only, which makes the two branches equal and leaves

∣D⟩⊗(c∣0⟩L−is∣1⟩L)=∣+⟩Me−iθ2X^∣0⟩L,(15)\lvert D\rangle\otimes\left(c\lvert0\rangle_L-is\lvert1\rangle_L\right)=\lvert+\rangle_M e^{-\frac{i\theta}{2}\hat{X}}\lvert0\rangle_L, \tag*{(15)}

as predicted by Eq. (11). No measurement appears anywhere in this stage, so it is deterministic. The Friend’s optics never see θ\theta. Both gate stages are identical at every angle, and the only element that changes is Wigner’s HWP.

Readout. To verify the action of the gate, we step outside of the framework, and perform tomography on the logical qubit space as follows. To read out the logical qubit we map it onto polarisation, ∣0⟩L→∣H⟩\lvert0\rangle_L \to\lvert H\rangle and ∣1⟩L→∣V⟩\lvert1\rangle_L \to\lvert V\rangle. A HWP across the whole grid takes ∣D⟩\lvert D\rangle to ∣H⟩\lvert H\rangle, a second HWP rotates the right column alone to ∣V⟩\lvert V\rangle, and a calcite beam displacer merges the two columns. Each row then carries the logical state in its polarisation, and the two rows carry the same state up to a global phase (Sec. IVB). We analyse the top row with a QWP, a HWP and a PBS. Both PBS outputs are coupled into single-mode fibres and detected with superconducting nanowire single-photon detectors (PhotonSpot), in coincidence with the herald; these elements are not shown in Figure 3. The photon pairs come from type-0 spontaneous parametric down-conversion in a 30 mm PPKTP crystal inside a Sagnac loop, pumped in one direction only by a 775 nm continuous-wave laser so that the pair carries no polarisation entanglement. For each tomography setting we acquire data 10 s, resulting in approximately 1550 coincidences per measurement. The density matrices are reconstructed by maximum-likelihood estimation [23] with Monte Carlo simulations used to estimate uncertainties. Note that the readout writes the logical qubit onto the message’s polarisation, and it assumes the message has returned to ∣D⟩\lvert D\rangle, as in Eq. (15). Without resetting the message, errors would appear in the reconstructed logical state, so the fidelities below jointly test the gate and the resetting of the message.

Results. Our tomography results for different gates are shown in Figure 5. In Figure 5a we display the reconstructed density matrix of the full two-qubit target system in the classical basis gives F=0.9932±0.0012\mathcal{F}=0.9932\pm0.0012 when θ=0\theta=0. In this case, the expected output state is ∣0⟩L\lvert0\rangle_L, which is apparent from the off-diagonal terms between ∣10⟩\lvert10\rangle and ∣11⟩\lvert11\rangle. In Figure 5b we set θ/2=π/4\theta/2=\pi/4, and again perform full two-qubit tomography, observing coherence between the logical states ∣0⟩L\lvert0\rangle_L and ∣1⟩L\lvert1\rangle_L, as expected. The fidelity with ∣θ/2⟩=cos⁡(θ/2)∣0⟩L−isin⁡(θ/2)∣1⟩L\lvert\theta/2\rangle=\cos(\theta/2)\lvert0\rangle_L-i\sin(\theta/2)\lvert1\rangle_L is F=0.986±0.002\mathcal{F}=0.986\pm0.002. To visualize this more easily, in panel c we show the logical states directly. Scanning θ/2∈[0,π)\theta/2\in[0,\pi) in steps of π/16\pi/16, and performing a tomography at each step. The Bloch vectors follow the great circle that RX(θ)R_X(\theta) traces through the YY–ZZ plane (Figure 5c). Averaged over all sixteen angles, the state fidelity is ⟨F⟩=0.9885±0.0012\langle\mathcal{F}\rangle=0.9885\pm0.0012 and the purity is ⟨P⟩=0.9775±0.0023\langle\mathcal{P}\rangle=0.9775\pm0.0023. A check that the bottom row reproduces the top-row state is given in § 8 B and SI A. Two losses are built into the scheme. The preparation passes half of the photons, and the readout uses one row out of two, so the overall success probability is at most 1/41/4.

Reconstructed experimental density matrices and Bloch sphere evolution

Figure 5. Reconstructed experimental density matrices and Bloch sphere evolution. a) Reconstructed density matrix in the classical basis of the input state ∣0⟩L\lvert0\rangle_L, after state preparation and prior to the protocol implementation, with F=0.9932(12)\mathcal{F} = 0.9932(12). b) Reconstructed target qubit density matrix in the classical basis for the target state at θ/2=π/4\theta/2 = \pi/4, the reconstruction yields a fidelity of F=0.986(2)\mathcal{F} = 0.986(2). c) Bloch sphere evolution of target logical qubit under different angles. For each angle we conducted a tomography, where we mapped the logical basis to polarisation basis. Three of the tomographies are shown for angles: θ/2={0, π/4, 3π/8}\theta/2 = \{0,\ \pi/4,\ 3\pi/8\}, with respective fidelities of F={0.9894(19),0.986(2),0.988(7)}\mathcal{F} = \{0.9894(19), 0.986(2), 0.988(7)\}.

The residual infidelity is consistent with technical imperfections of the setup, such as polarisation mixing from the finite extinction ratio of the dPBS and spatial-mode aberrations introduced by the cylindrical-lens telescope. Imperfections of this kind can leave the two arms of the interferometer partially distinguishable after the exit dPBS, so that some information about which operation was performed survives the erasure. In SI B we introduce a simple model of this situation, in which such information leaks into an environment. The environment is then left in one of two states, where VV is the real part of their overlap and V=1V=1 for a perfect erasure. Within this model, the applied operation becomes an incoherent mixture of the target gate (e−iθ2X^∣0⟩L)(e^{-i\frac{\theta}{2}\hat{X}}\lvert0\rangle_L) and a rotation by the same angle with opposite sign, e+iθ2X^∣0⟩Le^{+i\frac{\theta}{2}\hat{X}}\lvert0\rangle_L. Imperfections in the tomography, such as imperfect projective measurements, affect the reconstructed states in the same way as a surviving record, and since the individual contributions of these imperfections were not quantified, the measured fidelities cannot be used to extract VV.

Discussion

We have shown that a gate set of classical permutations, which is not universal even for classical computation, suffices for universal quantum computation when the instruction specifying which permutation the Friend applies is encoded in a quantum message prepared by Wigner. The permutation determines the generator, and hence the rotation axis, of each logical gate. The rotation angle is set by Wigner’s operation on the message outside the laboratory. Since the Friend reads the message only in the instruction basis, she does not have any access to the angle, and can thus gain no knowledge of it. The erasure EE is required because the Friend’s memory stores which-path information about the executed instruction, which would otherwise suppress the interference between the instruction branches. For a perfect erasure the message and the Friend’s memory return to their initial states after each gate, and only the target system is transformed. We experimentally implemented the single-qubit sector of the protocol, for the gate family generated by G1G_1, with heralded single photons. In particular, we used two bits to encode one logical qubit and showed that we can perform coherent rotations of this qubit around the XX axis.

Universal quantum computation in this framework is possible only as long as the laboratory does not leak information about which permutation the Friend applied, implying that the computational power arises from the quantum coherence of the Friend’s operations, i.e. from the coherence that underlies the different descriptions that Wigner and the Friend hold of one and the same process. Our results thus establish an operational link between the foundations of quantum measurement and quantum computation.

Our implementation is microscopic, with the Friend’s memory carried by a path degree of freedom of a single photon; extending it to more complex observers would require coherent control over correspondingly larger systems. Quantum computers may eventually allow these ideas to be tested. For example, it has been proposed to realise the Friend as a human-level artificial intelligence running on a quantum computer, in the context of Local Friendliness tests [15]. Extended Wigner’s Friend circuits have already been run on quantum hardware, with friends of increasing observer-likeness quantified by a branching measure [26] and with friends implemented as reversible circuits executing simple agent-like decision rules [16], both reporting Local Friendliness violations. Our results suggest that such experiments may also have computational consequences, since the coherent control over a friend’s record that they require is the ingredient our protocol turns into universal quantum computation. Whether this connection can be made quantitative for more complex friends remains open, as does whether the measurement of a genuine observer can be coherently reversed, which extended Wigner’s Friend arguments also require.

Methods

State preparation

The logical input ∣0⟩L=(∣10⟩−∣11⟩)/2|0\rangle_L = (|10\rangle- |11\rangle)/\sqrt{2} is prepared from the classical state ∣10⟩|10\rangle using only the framework’s allowed operations with the active instruction pair {1,G2}\{1, G_2\}, given that the Friend’s action alone on the target qubits cannot generate superposition of qubits in the classical basis.

State Preparation circuit

Figure 6. State Preparation circuit. The quantum circuit producing the logical zero ∣0⟩L|0\rangle_L in the dual-rail encoding. The Friend state, initialised in the ready state, ∣R⟩F|R\rangle_F, is disentangled at each stage.

  1. The computer is initialised in the classical state ∣10⟩|10\rangle; the source produces ∣H⟩|H\rangle, i.e. the message state ∣1⟩M|1\rangle_M.

  1. A half-wave plate (HWP) at 22.5∘22.5^\circ implements a Hadamard on the message, preparing (∣1⟩M+∣G2⟩M)/2≡∣D⟩(|1\rangle_M + |G_2\rangle_M)/\sqrt{2} \equiv|D\rangle.

  1. The photon enters the interferometer. The entry PBS transmits the ∣H⟩|H\rangle (= ∣1⟩M|1\rangle_M) component and reflects the ∣V⟩|V\rangle (= ∣G2⟩M|G_2\rangle_M) component, taking the Friend’s memory from ∣R⟩F|R\rangle_F to ∣1⟩F|1\rangle_F and ∣G2⟩F|G_2\rangle_F respectively. In the reflected branch a calcite beam displacer shifts the occupied spatial mode vertically, ∣10⟩↦∣11⟩|10\rangle\mapsto|11\rangle, executing G2G_2 (see Figure 4b) for the experimental implementation); the transmitted path contains no optics and executes the identity. Entry is the coupling unitary CC of Eq. (5); together with the conditional optics it realises the read–execute round of Eq. (6).

  1. The photon exits through the second PBS, which merges the two branches into a single outside path: the Friend’s memory returns from ∣g⟩F|g\rangle_F to ∣R⟩F|R\rangle_F in both branches without being read. This is the same operator applied in its erasing direction, the EE of Eq. (7), realised optically.

  1. A further Hadamard (HWP) followed by a polariser at 90∘90^\circ implements the projective measurement ΠV=∣V⟩⟨V∣\Pi_V = |V\rangle\langle V| on the message, leaving the state in ∣G⟩M|G\rangle_M.

  1. An optional bit-flip X^\hat{X} on the message returns it to ∣1⟩M|1\rangle_M.

Conditioned on the projection in step 5, the computer is left in ∣0⟩L|0\rangle_L. The projection succeeds with probability p=1/2p = 1/2, and this cost originates in step 5 alone, not in the conditional gate. In principle the success probability can be increased by routing failed events back through the same preparation stage [17]; this recycling was not implemented here due to the architecture of the experiment.

Tomography

After the protocol, the logical qubit encoded in the spatial modes is mapped onto polarisation for tomographic reconstruction, with ∣0⟩L≡∣H⟩|0\rangle_L \equiv|H\rangle and ∣1⟩L≡∣V⟩|1\rangle_L \equiv|V\rangle. At the gate output the message is in ∣D⟩|D\rangle, so the whole grid carries diagonal polarisation. A first HWP rotates the polarisation of the entire grid to ∣H⟩|H\rangle, and a second HWP acting only on the second column rotates that column to ∣V⟩|V\rangle:

∣D⟩⊗(α∣0⟩L+β∣1⟩L)→∣H⟩⊗(α∣0⟩L+β∣1⟩L)→α∣0⟩L⊗∣H⟩+β∣1⟩L⊗∣V⟩.(16)|D\rangle\otimes(\alpha|0\rangle_L + \beta|1\rangle_L) \to|H\rangle\otimes(\alpha|0\rangle_L + \beta|1\rangle_L) \to\alpha|0\rangle_L \otimes|H\rangle+ \beta|1\rangle_L \otimes|V\rangle. \tag*{(16)}

Writing the state in terms of the top (tt) and bottom (bb) rows of the grid,

12((α∣H⟩t+β∣V⟩t)−(α∣H⟩b+β∣V⟩b)),(17)\frac{1}{\sqrt{2}}\left((\alpha|H\rangle_t + \beta|V\rangle_t) - (\alpha|H\rangle_b + \beta|V\rangle_b)\right), \tag*{(17)}

the two rows carry identical tomographic information up to a global phase. The columns are recombined using a calcite beam displacer. The tomography reported in the main text is performed on the top row alone, accepting the factor-of-two loss in count rate from the 1/21/\sqrt{2} prefactor; the bottom-row cross-check below confirms the choice is ideally unbiased. Each tomographic basis setting is recorded for 10 s, yielding ≈1550\approx1550 coincidence counts. Density matrices ρ~\tilde{\rho} are reconstructed by maximum-likelihood estimation [23], with uncertainties obtained by standard Monte Carlo resampling of the coincidence counts assuming Poissonian statistics. The uncertainty quoted on angle-averaged quantities is the propagated per-angle Monte Carlo uncertainty of the mean. The fidelity to the pure target σ=∣θ/2⟩⟨θ/2∣\sigma= |\theta/2\rangle\langle\theta/2|, with ∣θ/2⟩=cos⁡(θ/2)∣0⟩L−isin⁡(θ/2)∣1⟩L|\theta/2\rangle= \cos(\theta/2)|0\rangle_L - i\sin(\theta/2)|1\rangle_L, and the Bloch vector are

F=Tr⁡(σ~ρ),r⃗=Tr⁡(ρ~σ⃗),σ⃗=(X^,Y^,Z^).(18)\mathcal{F} = \operatorname{Tr}(\tilde{\sigma}\rho), \qquad\vec{r} = \operatorname{Tr}(\tilde{\rho}\vec{\sigma}), \qquad\vec{\sigma} = (\hat{X},\hat{Y},\hat{Z}). \tag*{(18)}

This tomography lies outside the framework of the main text. It is performed by the external agent directly on the Friend’s computer, to verify that the final state is a coherent superposition of logical basis states, which the Friend, restricted to permutations, could not have produced. From the Friend’s perspective the computer holds a pair of classical bits at all times.

Acknowledgments

We thank Clemens Lindner for helpful discussions and work on an early version of the protocol.

The financial support by the Austrian Federal Ministry of Labour and Economy, the National Foundation for Research, Technology and Development and the Christian Doppler Research Association is gratefully acknowledged. This research was funded in whole or in part by the Austrian Science Fund (FWF) [10.55776/COE1] (Quantum Science Austria), [10.55776/F71] (BeyondC) and [10.55776/DOC162]. For open access purposes, the author has applied a CC BY public copyright license to any author accepted manuscript version arising from this submission. Funded (Co-funded) by the European Union (HORIZON Research and Innovation Programme, EPIQUE, No 101135288). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Commission-EU. Neither the European Union nor the granting authority can be held responsible for them.

Data and Code Availability

All the data that are necessary to replicate, verify, falsify and/or reuse this research are available online at Ref. [27].

Competing Interests

The authors declare no competing interests.

Additional Data

Tomography comparison

As discussed above, after the logical-to-polarisation mapping the top and bottom rows of the 2×22 \times2 grid carry ideally the same tomographic information up to a global phase, and the tomography reported in § 2 C uses only the top row. To check that this choice does not bias the result, we performed an independent tomography of the bottom row at θ/2=3π/4\theta/2 = 3\pi/4 (corresponding to ∣L⟩=1/2(∣H⟩+i∣V⟩)|L\rangle= 1/\sqrt{2}(|H\rangle+ i|V\rangle), up to a global phase), where the cylindrical-lens imaging system is expected to have a large effect because the output is in a maximal superposition of ∣0⟩L|0\rangle_{\mathrm{L}} and ∣1⟩L|1\rangle_{\mathrm{L}}. Figure 7 shows the reconstructed density matrices for both rows. The fidelity of the top row is Ft=0.9852(21)\mathcal{F}_{t} = 0.9852(21), and the bottom is Fb=0.9853(22)\mathcal{F}_{b} = 0.9853(22). The two values agree within their uncertainties, supporting the use of the top row as representative of the output at the tested setting. The top row was chosen solely for experimental convenience.

Top and bottom tomography comparison

Figure 7. Top and Bottom Tomography Comparison. Comparison between the tomography of the top and bottom rows for ∣θ/2=3π/4⟩|\theta/2 = 3\pi/4\rangle. In a) we have the tomography for the top row with {P,F}={0.9711(41),0.9852(21)}\{\mathcal{P},\mathcal{F}\} = \{0.9711(41), 0.9852(21)\}. In b) we have the tomography of the same state using instead the bottom row. The purity and fidelity of the state is {P,F}={0.9714(43),0.9853(22)}\{\mathcal{P},\mathcal{F}\} = \{0.9714(43), 0.9853(22)\}.

Single state Fidelity

The purity P=Tr⁡(ρ~2)\mathcal{P} = \operatorname{Tr}(\tilde{\rho}^{2}) and the fidelity F\mathcal{F} with to the ideal target state ∣θ/2⟩|\theta/2\rangle, are computed from the same MLE-reconstructed density matrices used for the Bloch trajectory of (Figure 5). Figure 8 shows the fidelity of the sixteen measured half-angles individually (plus the input state and the tomography implemented in the bottom row, discussed in § 8 B), rather than the angle-averaged values quoted in § 2 C. The error bars come from Monte Carlo resampling of the Poissonian counts.

Fidelity versus angle

Figure 8. Additional experimental data. Fidelity F\mathcal{F} with respect to the ideal target state ∣θ/2⟩\lvert\theta/2\rangle (bottom), as defined in Eq. (18), for each of the sixteen measured half-angles, plus the input state ∣0⟩L\lvert0\rangle_L, labeled as Logical, and the state measured from the bottom row (last set of data, Bottom 3π/43\pi/4). Dashed lines mark the ideal value F=1\mathcal{F}=1. The average fidelity and purity values over all transformed states are ⟨F⟩=0.9885±0.0012\langle\mathcal{F}\rangle= 0.9885 \pm0.0012 and ⟨P⟩=0.9775±0.0023\langle P\rangle= 0.9775 \pm0.0023.

Stability

To check that the optical setup is passively stable over the timescale of a full tomographic acquisition, we recorded the Stokes parameters and power of the output beam for 30 minutes using classical (laser) light, with the waveplates set to the configuration that prepares the state ∣R⟩=1/2(∣H⟩−i∣V⟩)|R\rangle= 1/\sqrt{2}(|H\rangle- i|V\rangle) (θ/2=π/4\theta/2 = \pi/4). Figure 9 shows that all three Stokes parameters remain flat to within the measurement noise, with no systematic drift. Furthermore, the power also shows very small drifts over the recorded time. The dPBS paths and the cylindrical-lens 4f4f imaging system therefore remain stable for longer than a full tomographic run, and no active stabilisation is required.

Stability data

Figure 9. Stability data. Taken with classical light for the waveplate configuration θ/2=π/4\theta/2=\pi/4, corresponding to ∣R⟩\lvert R\rangle. For the Polarimeter PAX1000, the convention is that s3=+1s_3=+1 for ∣R⟩\lvert R\rangle. The plot shows the normalised Stokes parameters s1s_1, s2s_2, s3s_3, left axis and continuous lines, and the optical power, right axis and dashed line, recorded over 30 minutes. The mean recorded power was 0.1031 mW, showing a slow drift of 2.6% peak-to-peak from the laser source. The polarisation-encoded message is insensitive to this drift: the Stokes parameters remain flat to within the measurement noise, confirming that the dPBS loops and the cylindrical-lens imaging system are passively stable over a timescale longer than a full tomographic acquisition.

Quantifying the erased record

In the ideal protocol nothing is consumed. Equation (11) returns the message to ∣+(i)⟩M\lvert+^{(i)}\rangle_M and the memory to ∣R⟩F\lvert R\rangle_F exactly. Each round, consisting of coupling, execution and erasure, writes a record of the executed instruction into the Friend’s memory and then removes it. Here we ask what happens when the removal is incomplete.

We model this with an environment onto which information about the executed instruction can leak, so that a record of the branch survives the erasure. Before each erasure the environment acquires a partial copy of the record, ∣g⟩F∣0⟩E→∣g⟩F(cos⁡η∣0⟩+sin⁡η∣g⟩)E\lvert g\rangle_F\lvert0\rangle_E \to\lvert g\rangle_F(\cos\eta\lvert0\rangle+\sin\eta\lvert g\rangle)_E for g∈{1,G}g\in\{1,G\}, with ∣R⟩F\lvert R\rangle_F unaffected. For η=0\eta=0 nothing leaks, and for η=π/2\eta=\pi/2 a perfect copy of the record survives. In this model each environment has the Hilbert space

HE(i)=span⁡({∣0⟩E}∪{∣g⟩E:g∈{1,G}}),(19)\mathcal{H}^{(i)}_E=\operatorname{span}\left(\{\lvert0\rangle_E\}\cup\{\lvert g\rangle_E:g\in\{1,G\}\}\right), \tag*{(19)}

where the superscript labels the leak, i=1i=1 before the first erasure and i=2i=2 before the second. We will see that the second leak does not affect the target. We write ∣⋅⟩E1\lvert\cdot\rangle_{E_1} for states in HE(1)\mathcal{H}^{(1)}_E and ∣⋅⟩E2\lvert\cdot\rangle_{E_2} for states in HE(2)\mathcal{H}^{(2)}_E.

The model assumes that the leak acts only on the Friend’s memory, between execution and erasure, and leaves the message and the target untouched, and that both branches leak with the same strength η\eta into orthogonal environment states. It only covers photons that are detected. Photons that leave through the unused output port of the exit dPBS are lost before detection, so they lower the count rate but not the measured fidelity. We abbreviate the environment states as ∣eg⟩E=cos⁡η∣0⟩E+sin⁡η∣g⟩E|e_g\rangle_E = \cos\eta|0\rangle_E + \sin\eta|g\rangle_E and denote by VV the real part of their overlap,

V=Re⁡⟨e1∣eG⟩E=cos⁡2η.(20)V = \operatorname{Re}\langle e_1|e_G\rangle_E = \cos^2\eta. \tag*{(20)}

In this model V=1V = 1 means that the two records are identical and V=0V = 0 that they are perfectly distinguishable, so 1−V1 - V quantifies how distinguishable the records are. Including the environment, the protocol runs as follows.

  1. The initial state is

12(∣1⟩M+∣G⟩M)∣R⟩F∣ψ⟩L⊗∣0⟩E∣0⟩E2.\frac{1}{\sqrt{2}}\left(|1\rangle_M + |G\rangle_M\right)|R\rangle_F|\psi\rangle_L \otimes|0\rangle_E|0\rangle_{E_2}.
  1. The coupling and execution step gives

12(∣1⟩M∣1⟩F∣ψ⟩L+∣G⟩M∣G⟩FG∣ψ⟩L)⊗∣0⟩E∣0⟩E2.\frac{1}{\sqrt{2}}\left(|1\rangle_M|1\rangle_F|\psi\rangle_L + |G\rangle_M|G\rangle_F G|\psi\rangle_L\right)\otimes|0\rangle_E|0\rangle_{E_2}.
  1. The first leak copies part of the record into the environment,

12(∣1⟩M∣1⟩F∣e1⟩E∣ψ⟩L+∣G⟩M∣G⟩F∣eG⟩EG∣ψ⟩L)⊗∣0⟩E2.\frac{1}{\sqrt{2}}\left(|1\rangle_M|1\rangle_F|e_1\rangle_E|\psi\rangle_L + |G\rangle_M|G\rangle_F|e_G\rangle_E G|\psi\rangle_L\right)\otimes|0\rangle_{E_2}.
  1. Wigner applies the erasure,

12(∣1⟩M∣e1⟩E∣ψ⟩L+∣G⟩M∣eG⟩EG∣ψ⟩L)⊗∣R⟩F∣0⟩E2.\frac{1}{\sqrt{2}}\left(|1\rangle_M|e_1\rangle_E|\psi\rangle_L + |G\rangle_M|e_G\rangle_E G|\psi\rangle_L\right)\otimes|R\rangle_F|0\rangle_{E_2}.
  1. Wigner rotates the message. With cθ=cos⁡θ2c_\theta= \cos\frac{\theta}{2} and sθ=sin⁡θ2s_\theta= \sin\frac{\theta}{2}, the state becomes

12[(∣1⟩M(cθ∣e1⟩E∣ψ⟩L−isθ∣eG⟩EG∣ψ⟩L)+∣G⟩M(cθ∣eG⟩EG∣ψ⟩L−isθ∣e1⟩E∣ψ⟩L))]⊗∣R⟩F∣0⟩E2.\frac{1}{\sqrt{2}}\left[\left(|1\rangle_M\left(c_\theta|e_1\rangle_E|\psi\rangle_L - i s_\theta|e_G\rangle_E G|\psi\rangle_L\right) + |G\rangle_M\left(c_\theta|e_G\rangle_E G|\psi\rangle_L - i s_\theta|e_1\rangle_E|\psi\rangle_L\right)\right)\right]\otimes|R\rangle_F|0\rangle_{E_2}.
  1. The second coupling and execution follow. Writing gˉ\bar{g} for the label of the other branch and defining the combined environment–target states ∣wg⟩EL=cθ∣eg⟩E∣ψ⟩L−isθ∣egˉ⟩EG∣ψ⟩L|w_g\rangle_{EL} = c_\theta|e_g\rangle_E|\psi\rangle_L - i s_\theta|e_{\bar{g}}\rangle_E G|\psi\rangle_L, the state is

12(∣1⟩M∣1⟩F∣w1⟩EL+∣G⟩M∣G⟩F∣wG⟩EL)⊗∣0⟩E2.\frac{1}{\sqrt{2}}\left(|1\rangle_M|1\rangle_F|w_1\rangle_{EL} + |G\rangle_M|G\rangle_F|w_G\rangle_{EL}\right)\otimes|0\rangle_{E_2}.
  1. The second leak gives

12(∣1⟩M∣1⟩F∣e1⟩E2∣w1⟩EL+∣G⟩M∣G⟩F∣eG⟩E2∣wG⟩EL).\frac{1}{\sqrt{2}}\left(|1\rangle_M|1\rangle_F|e_1\rangle_{E_2}|w_1\rangle_{EL} + |G\rangle_M|G\rangle_F|e_G\rangle_{E_2}|w_G\rangle_{EL}\right).
  1. Wigner applies the second erasure, and the final state is

12(∣1⟩M∣e1⟩E2∣w1⟩EL+∣G⟩M∣eG⟩E2∣wG⟩EL)⊗∣R⟩F.\frac{1}{\sqrt{2}}\left(|1\rangle_M|e_1\rangle_{E_2}|w_1\rangle_{EL} + |G\rangle_M|e_G\rangle_{E_2}|w_G\rangle_{EL}\right)\otimes|R\rangle_F.

To obtain the state of the target we trace out the message, both environments and the Friend. Since ∣1⟩M|1\rangle_M and ∣G⟩M|G\rangle_M are orthogonal, the trace over MM removes every cross-term between the two branches. Tracing out MM, and dropping the decoupled memory FF, gives

ρLE,E2=12(∣e1⟩E2⟨e1∣⊗∣w1⟩EL⟨w1∣+∣eG⟩E2⟨eG∣⊗∣wG⟩EL⟨wG∣).\rho_{LE,E_2} = \frac{1}{2}\left(|e_1\rangle_{E_2}\langle e_1| \otimes|w_1\rangle_{EL}\langle w_1| + |e_G\rangle_{E_2}\langle e_G| \otimes|w_G\rangle_{EL}\langle w_G|\right).

The trace over the second environment only uses the normalisation ⟨eg∣eg⟩E2=1\langle e_g|e_g\rangle_{E_2} = 1. With the cross-terms already gone, the overlap ⟨e1∣eG⟩E2\langle e_1|e_G\rangle_{E_2} never appears, and the second leak drops out of the target’s dynamics,

ρLE=12(∣w1⟩EL⟨w1∣+∣wG⟩EL⟨wG∣).\rho_{LE}=\frac{1}{2}\left(|w_1\rangle_{EL}\langle w_1|+|w_G\rangle_{EL}\langle w_G|\right).

Tracing over the first environment gives ρL=12(T1+TG)\rho_L=\frac{1}{2}(T_1+T_G) with Tg=Tr⁡E[∣wg⟩⟨wg∣EL]T_g=\operatorname{Tr}_{E}\left[|w_g\rangle\langle w_g|_{EL}\right]. Expanding ∣wg⟩EL|w_g\rangle_{EL}, with ρ=∣ψ⟩L⟨ψ∣\rho=|\psi\rangle_L\langle\psi|,

T1=cθ2ρ+sθ2GρG+icθsθ⟨eG∣e1⟩EρG−icθsθ⟨e1∣eG⟩EGρ,T_1=c_\theta^2\rho+s_\theta^2G\rho G+ic_\theta s_\theta\langle e_G|e_1\rangle_E\rho G-ic_\theta s_\theta\langle e_1|e_G\rangle_EG\rho,
TG=cθ2ρ+sθ2GρG+icθsθ⟨e1∣eG⟩EρG−icθsθ⟨eG∣e1⟩EGρ.T_G=c_\theta^2\rho+s_\theta^2G\rho G+ic_\theta s_\theta\langle e_1|e_G\rangle_E\rho G-ic_\theta s_\theta\langle e_G|e_1\rangle_EG\rho.

In the sum each overlap appears together with its complex conjugate, so the imaginary parts cancel and only the real part survives,

ρL=cθ2ρ+sθ2GρG+icθsθV(ρG−Gρ).(21)\rho_L=c_\theta^2\rho+s_\theta^2G\rho G+ic_\theta s_\theta V\left(\rho G-G\rho\right). \tag*{(21)}

The derivation uses the two environment states only through their overlap, so Eq. (21) holds for any pure records ∣e1⟩E|e_1\rangle_E and ∣eG⟩E|e_G\rangle_E, not only for the model of Eq. (19). In particular it covers a record that carries a relative phase, ⟨e1∣eG⟩E=∣⟨e1∣eG⟩E∣eiχ\langle e_1|e_G\rangle_E=|\langle e_1|e_G\rangle_E|e^{i\chi}, for which V=∣⟨e1∣eG⟩E∣cos⁡χV=|\langle e_1|e_G\rangle_E|\cos\chi, and a dephased record and a partially copied record act identically on the target whenever they share the same VV.

To identify the channel in Eq. (21), we compare it with a mixture of the two rotation signs. Writing U±θ=cθ1∓isθGU_{\pm\theta}=c_\theta\mathbf{1}\mp is_\theta G and expanding,

UθρUθ†=cθ2ρ+sθ2GρG+icθsθ(ρG−Gρ),U−θρU−θ†=cθ2ρ+sθ2GρG−icθsθ(ρG−Gρ),U_\theta\rho U_\theta^\dagger=c_\theta^2\rho+s_\theta^2G\rho G+ic_\theta s_\theta\left(\rho G-G\rho\right),\qquad U_{-\theta}\rho U_{-\theta}^\dagger=c_\theta^2\rho+s_\theta^2G\rho G-ic_\theta s_\theta\left(\rho G-G\rho\right),

so that for any weight pp

(1−p)UθρUθ†+p U−θρU−θ†=cθ2ρ+sθ2GρG+icθsθ(1−2p)(ρG−Gρ).(1-p)U_\theta\rho U_\theta^\dagger+p\,U_{-\theta}\rho U_{-\theta}^\dagger=c_\theta^2\rho+s_\theta^2G\rho G+ic_\theta s_\theta(1-2p)\left(\rho G-G\rho\right).

This reproduces Eq. (21) term by term if and only if 1−2p=V1-2p=V, and the target therefore evolves as

Λ(ρ)=(1−p)UθρUθ†+p U−θρU−θ†,p=1−V2.(22)\Lambda(\rho)=(1-p)U_\theta\rho U_\theta^\dagger+p\,U_{-\theta}\rho U_{-\theta}^\dagger,\qquad p=\frac{1-V}{2}. \tag*{(22)}

The overlap multiplies only the commutator term ρG−Gρ\rho G-G\rho, which is the part of ρL\rho_L that changes sign with θ\theta. The coherence of the record therefore carries the phase relation that distinguishes UθU_\theta from U−θU_{-\theta}, and losing it randomises the sense of the rotation while leaving its size intact. For V=0V=0, reached for instance by a perfect copy (η=π/2)(\eta=\pi/2), the channel is an equal mixture of UθU_\theta and U−θU_{-\theta}.

For an input ∣ψ⟩|\psi\rangle with pure target ∣ϕ⟩=Uθ∣ψ⟩|\phi\rangle=U_\theta|\psi\rangle, the fidelity of the output is F=⟨ϕ∣Λ(∣ψ⟩⟨ψ∣)∣ϕ⟩\mathcal{F}=\langle\phi|\Lambda(|\psi\rangle\langle\psi|)|\phi\rangle. With Eq. (22) this gives F=(1−p)+p∣⟨ψ∣Uθ†U−θ∣ψ⟩∣2\mathcal{F}=(1-p)+p|\langle\psi|U_\theta^\dagger U_{-\theta}|\psi\rangle|^2, where the second term is the squared overlap between the ideal output and the output rotated in the opposite direction. Using G2=1G^2=1, Uθ†U−θ=(cθ1+isθG)2=cos⁡θ1+isin⁡θG=eiθGU_\theta^\dagger U_{-\theta}=(c_\theta\mathbf{1}+is_\theta G)^2=\cos\theta\mathbf{1}+i\sin\theta G=e^{i\theta G}. With g~=⟨ψ∣G∣ψ⟩\tilde{g}=\langle\psi|G|\psi\rangle, which is real since GG is Hermitian, ∣⟨ψ∣eiθG∣ψ⟩∣2=cos⁡2θ+g~2sin⁡2θ|\langle\psi|e^{i\theta G}|\psi\rangle|^2=\cos^2\theta+\tilde{g}^2\sin^2\theta, so that

1−F=p(1−g~2)sin⁡2θ.(23)1-\mathcal{F}=p(1-\tilde{g}^2)\sin^2\theta. \tag*{(23)}

The input ∣0⟩L|0\rangle_L used in the experiment has g~=⟨0∣X^∣0⟩=0\tilde{g}=\langle0|\hat{X}|0\rangle=0, giving 1−F(θ)=1−V2sin⁡2θ1-\mathcal{F}(\theta)=\frac{1-V}{2}\sin^2\theta. Eigenstates of GG, with g~=±1\tilde{g}=\pm1, are taken to the same state by both rotations up to a global phase, so this mechanism leaves their fidelity unchanged. An input orthogonal to the rotation axis, such as ∣0⟩L|0\rangle_L, is instead maximally sensitive, since the two rotation signs send it to the most distant pair of states.

To average Eq. (23) over input states it is enough to average g~2\tilde{g}^2. Since GG is Hermitian and unitary with G2=1G^2=\mathbf{1}, on a single logical qubit it has eigenvalues ±1\pm1 and vanishing trace, so it is a Pauli operator along some axis, G=u^⋅σ⃗G=\hat{u}\cdot\vec{\sigma}, with u^\hat{u} a real unit vector and σ⃗=(X^,Y^,Z^)\vec{\sigma}=(\hat{X},\hat{Y},\hat{Z}). Writing a pure input as ∣ψ⟩⟨ψ∣=12(1+n^⋅σ⃗)|\psi\rangle\langle\psi|=\frac{1}{2}(\mathbf{1}+\hat{n}\cdot\vec{\sigma}) with Bloch vector n^\hat{n}, and using Tr⁡(σiσj)=2δij\operatorname{Tr}(\sigma_i\sigma_j)=2\delta_{ij},

g~=⟨ψ∣G∣ψ⟩=12Tr⁡[(1+n^⋅σ⃗)(u^⋅σ⃗)]=n^⋅u^=cos⁡ϑ.(24)\tilde{g}=\langle\psi|G|\psi\rangle=\frac{1}{2}\operatorname{Tr}\left[(\mathbf{1}+\hat{n}\cdot\vec{\sigma})(\hat{u}\cdot\vec{\sigma})\right]=\hat{n}\cdot\hat{u}=\cos\vartheta. \tag*{(24)}

where ϑ\vartheta is the angle between the input Bloch vector and the axis of GG. The overlap g~\tilde{g} is thus the projection of the input onto the axis of the operation the Friend performs. Haar-random pure states of a qubit have Bloch vectors distributed uniformly on the sphere, with normalised measure dμ=sin⁡ϑ dϑ dϕ/4πd\mu= \sin\vartheta\,d\vartheta\,d\phi/4\pi, so with u=cos⁡ϑu = \cos\vartheta

⟨q^2⟩=14π∫02πdϕ∫0πcos⁡2ϑsin⁡ϑ dϑ=12∫−11u2 du=13,(25)\langle\hat{q}^{2}\rangle= \frac{1}{4\pi}\int_{0}^{2\pi}d\phi\int_{0}^{\pi}\cos^{2}\vartheta\sin\vartheta\,d\vartheta= \frac{1}{2}\int_{-1}^{1}u^{2}\,du = \frac{1}{3}, \tag*{(25)}

independently of the axis u^\hat{u}. Substituting into Eq. (23) gives 1−Fˉ(θ)=1−V3sin⁡2θ1-\bar{F}(\theta)=\frac{1-V}{3}\sin^{2}\theta, the input-independent statement about the channel. The experiment measures the state-specific law with the coefficient 12\frac{1}{2}, since it prepares ∣0⟩L|0\rangle_L rather than an ensemble of inputs. The factor sin⁡2θ\sin^{2}\theta vanishes at θ=0\theta=0 and θ=π\theta=\pi, where the gate is a permutation up to a global phase and needs no interference between the instruction branches, and is largest at θ=π/2\theta=\pi/2 (Figure 10).

State infidelity of the gate as a function of gate angle $\theta$ for three values of $1-V$

Figure 10. The erasure–fidelity law. State infidelity of the gate as a function of θ\theta for three values of 1−V1-V. Lines are the closed form 1−V2sin⁡2θ\frac{1-V}{2}\sin^{2}\theta of Eq. (23) for the input ∣0⟩L\lvert0\rangle_L; markers are the exact simulation of the six-step protocol with the leak model. Dotted lines mark θ=0,π,2π\theta=0,\pi,2\pi, where the infidelity vanishes because the gate is a permutation up to a global phase.

The second leak attaches its record to the branch label, so the overlap ⟨e1∣eG⟩E2\langle e_1|e_G\rangle_{E_2} can only multiply cross-terms between the two branches, and those have already been removed by the orthogonality of the untouched message states. The first leak behaves differently because Wigner’s rotation in step 5 mixes the two branches, placing both ∣e1⟩E|e_1\rangle_E and ∣eG⟩E|e_G\rangle_E inside each ∣wg⟩EL|w_g\rangle_{EL}, where they are correlated with the target rather than with the branch label. The role of the second erasure is to return the message and the memory to their initial states, so an imperfect second erasure degrades subsequent rounds rather than the present one. In the ideal case, with no leak, the second read–execute round maps the state to 12(∣1,1⟩+∣G,G⟩)⊗Uθ∣ψ⟩\frac{1}{\sqrt{2}}(|1,1\rangle+|G,G\rangle)\otimes U_\theta|\psi\rangle, using G2=1G^2=1, and the computer has already factored out with the gate complete before the erasure is applied.

Experimental imperfections. In the experiment the Friend’s record is the path of the photon through the interferometer, so any degree of freedom that still distinguishes the two arms after the exit dPBS is a surviving record in the sense of Eq. (21). Aberrations of the cylindrical-lens telescope, which sits only in the reflected arm, can leave the two arms in slightly different transverse modes, so that the path remains partly distinguishable after the exit dPBS. Polarisation mixing in the dPBS makes the coupling and erasure themselves imperfect, so that the two arms are not recombined exactly. The tomography contributes separately through the finite contrast of each measurement basis, which, as shown below, affects the reconstructed states in the same way as a surviving record.

Comparison with the data. The measured infidelities of § 2 C also contain contributions from technical imperfections, and the angular dependence alone cannot separate them from a surviving record. To see this, write the ideal output as ∣θ/2⟩=cos⁡θ2∣0⟩L−isin⁡θ2∣1⟩L|\theta/2\rangle=\cos\frac{\theta}{2}|0\rangle_L-i\sin\frac{\theta}{2}|1\rangle_L, whose Bloch vector is

t⃗(θ)=(0,−sin⁡θ,cos⁡θ).(26)\vec{t}(\theta)=(0,-\sin\theta,\cos\theta). \tag*{(26)}

For any reconstructed state ρ~=12(1+r⃗⋅σ⃗)\tilde{\rho}=\frac{1}{2}(1+\vec{r}\cdot\vec{\sigma}) the fidelity to this pure target is F=12(1+t⃗⋅r⃗)\mathcal{F}=\frac{1}{2}(1+\vec{t}\cdot\vec{r}), so

1−F=12(1+rYsin⁡θ−rZcos⁡θ),(27)1-\mathcal{F}=\frac{1}{2}(1+r_Y\sin\theta-r_Z\cos\theta), \tag*{(27)}

and only the YY and ZZ components of the output enter. Applied to ∣0⟩L|0\rangle_L, the channel of Eq. (22) mixes the Bloch vectors (0,±sin⁡θ,cos⁡θ)(0,\pm\sin\theta,\cos\theta) with weights 1−p1-p and pp, giving r⃗=(0,−Vsin⁡θ,cos⁡θ)\vec{r}=(0,-V\sin\theta,\cos\theta): the leak leaves the ZZ component unchanged and simply multiplies the YY component by VV. A tomography with finite contrast ck≤1c_k\leq1 in the basis kk has measurement operators 12(1±ckσk)\frac{1}{2}(1\pm c_k\sigma_k) and returns ckrkc_kr_k in place of each Bloch component, as long as the reconstruction lies inside the Bloch ball. With both effects present the reconstructed state has r⃗=(0,−cYVsin⁡θ,cZcos⁡θ)\vec{r}=(0,-c_YV\sin\theta,c_Z\cos\theta), and Eq. (27) gives

1−F=1−cZ2+cZ−cYV2sin⁡2θ.(28)1-\mathcal{F}=\frac{1-c_Z}{2}+\frac{c_Z-c_YV}{2}\sin^{2}\theta. \tag*{(28)}

For cY=cZ=1c_Y=c_Z=1 this reduces to Eq. (23), but a readout that is slightly worse in the YY basis than in the ZZ basis produces the same sin⁡2θ\sin^{2}\theta dependence with no surviving record at all. More generally, VV and cYc_Y enter the reconstructed state only through their product cYVc_YV, so the pairs (cY,V)(c_Y,V) and (cYV,1)(c_YV,1) give identical density matrices at every angle. A fit of the per-angle infidelities fixes cZc_Z and cYVc_YV but cannot separate the two factors, and we therefore treat Eq. (23) as a prediction of the framework rather than extracting VV from the data.

References

  1. [1]Wigner, E. P. (1961). Remarks on the mind-body question. In The Scientist Speculates (pp. 284–302).
  2. [2]Schmid, D., Yīng, Y., & Leifer, M. (2024). A review and analysis of six extended Wigner’s friend arguments. https://arxiv.org/abs/2308.16220
  3. [3]Frauchiger, D. & Renner, R. Quantum theory cannot consistently describe the use of itself. Nat. Commun. 9, 3711 (2018).DOI
  4. [4]Brukner, Č. A no-go theorem for observer-independent facts. Entropy 20, 350 (2018).DOI
  5. [5]Brukner, Č. On the quantum measurement problem. In Quantum [Un]Speakables II: Half a Century of Bell’s Theorem (eds Bertlmann, R. & Zeilinger, A.) 95–117 (Springer, Cham, 2017).arxiv.org/abs/1507.05255
  6. [6]Baumann, V. & Wolf, S. On formalisms and interpretations. Quantum 2, 99 (2018).DOI
  7. [7]Healey, R. (2018). Quantum theory and the limits of objectivity. Foundations of Physics, 48, 1568–1589.
  8. [8]Żukowski, M. & Markiewicz, M. Physics and metaphysics of Wigner’s friends: even performed premeasurements have no results. Phys. Rev. Lett. 126, 130402 (2021).DOI
  9. [9]Proietti, M. et al. Experimental test of local observer independence. Sci. Adv. 5, eaaw9832 (2019).DOI
  10. [10]Bong, K.-W. et al. A strong no-go theorem on the Wigner’s friend paradox. Nat. Phys. 16, 1199–1205 (2020).DOI
  11. [11]Zurek, W. H. Decoherence, einselection, and the quantum origins of the classical. Rev. Mod. Phys. 75, 715–775 (2003).DOI
  12. [12]Nurgalieva, N. & Renner, R. Testing quantum theory with thought experiments. Contemp. Phys. 61, 193–216 (2020).DOI
  13. [13]Rovelli, C. Relational quantum mechanics. Int. J. Theor. Phys. 35, 1637–1678 (1996).DOI
  14. [14]Allard Guérin, P., Baumann, V., Del Santo, F. & Brukner, Č. A no-go theorem for the persistent reality of Wigner’s friend’s perception. Commun. Phys. 4, 93 (2021).DOI
  15. [15]Wiseman, H. M., Cavalcanti, E. G. & Rieffel, E. G. A “thoughtful” Local Friendliness no-go theorem: a prospective experiment with new assumptions to suit. Quantum 7, 1112 (2023).DOI
  16. [16]Laux, J. & Cavalcanti, E. G. Extended Wigner’s friend scenarios with agent-like observers on quantum computers. Preprint at https://arxiv.org/abs/2609.12527 (2026).
  17. [17]Horvat, S., Gao, X. & Dakić, B. Universal quantum computation via quantum controlled classical operations. J. Phys. A: Math. Theor. 55, 075301 (2022).DOI
  18. [18]Chiribella, G., D’Ariano, G. M., Perinotti, P. & Valiron, B. Quantum computations without definite causal structure. Phys. Rev. A 88, 022318 (2013).DOI
  19. [19]Araújo, M., Feix, A., Costa, F. & Brukner, Č. Quantum circuits cannot control unknown operations. New J. Phys. 16, 093026 (2014).DOI
  20. [20]Procopio, L. M. et al. Experimental superposition of orders of quantum gates. Nat. Commun. 6, 7913 (2015).DOI
  21. [21]Evans, P. W. What is it like to be unitarily reversed? Eur. J. Philos. Sci. 14, 56 (2024).DOI
  22. [22]Deutsch, D. Quantum theory as a universal physical theory. Int. J. Theor. Phys. 24, 1–41 (1985).DOI
  23. [23]Hradil, Z. Quantum-state estimation. Phys. Rev. A 55, R1561–R1564 (1997).DOI
  24. [24]Englert, B.-G., Kurtsiefer, C. & Weinfurter, H. Universal unitary gate for single-photon two-qubit states. Phys. Rev. A 63, 032303 (2001).DOI
  25. [26]Zeng, W. J., Labib, F. & Russo, V. Towards violations of Local Friendliness with quantum computers. Quantum 9, 1851 (2025).DOI
  26. [27]Cladera-Rosselló, M., Horvat, S., Dakić, B., Rozema, L. A., & Walther, P. Universal Quantum Computation in a Wigner’s Friend Framework. Zenodo https://doi.org/10.5281/zenodo.24167209 (2026).
  27. [28]It is to be noted that our protocol structure resembles the exponential-swap construction of Ref. [25].. (n.d.). It is to be noted that our protocol structure resembles the exponential-swap construction of Ref. [25]..

Paper details

vlnc:W_39pd5dnu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents