Quantum information cannot be copied [1 ] , and neither can an unknown quantum operation: no finite quantum program can store an arbitrary unitary for later execution [2 ] , and copies of a gate are necessarily imperfect [3 ] . What can be stored is a specific gate, frozen into an entangled resource and released later by local measurement and classical feedforward [4 , 5 , 6 ] . Gate teleportation of this kind is the elementary operation of distributed quantum computing and of a quantum internet [7 , 8 ] : it applies a non-local entangling gate between data held at different nodes without moving the data, and it has been demonstrated between logical qubits, between separated trapped ions and across optical network links [9 , 10 , 11 ] . In every such protocol, however, the resource is bound to its destination when it is created. The half of the entangled pair that completes the gate is sent to one node, and whoever holds it holds the gate.
Secret sharing offers a different way to hold a resource. A threshold scheme splits a secret among n n n parties so that only authorized subsets recover it and all other sets learn nothing [12 ] . Its quantum analog shares an unknown quantum state in the same way [13 , 14 , 15 ] , with constructions based on graph states and on absolutely maximally entangled (AME) states [16 , 17 ] , states that are maximally entangled across every bipartition, and demonstrations with photons and optical modes, recently in eleven dimensions [18 , 19 ] . In all of these, the shared object is a state. Authority over an operation has also been shared, in quantum operation sharing [20 ] , controlled remote implementation of operations [21 ] and threshold-controlled teleportation [22 ] . There, however, the recipient is fixed in advance, and every controller must measure and announce an outcome, so each can veto the operation and all of them observe it. Here, we ask: can the capability to act , namely, applying an entangling gate to data that never enter the scheme, be secret-shared so that the node exercising it is chosen after distribution and the unselected sharers take no part?
A gate shared in this way would meet three requirements of an entanglement-based network: joint authorization, anonymity and early distribution. Authority to use a gate can be split among independent custodians, none of whom individually holds anything of value, as threshold cryptography does for classical keys. Moreover, nodes that are not chosen should learn neither the data nor whether anyone was chosen, a requirement shared with anonymous quantum communication [23 ] and with secure multiparty and delegated quantum computation [24 , 25 ] . Because remote entanglement is slow to establish [11 ] , the user’s entanglement should be in place before the destination is known. However, the obvious ways of deferring the destination fail at least one of these requirements. For example, a provider that shares a Bell pair with the user and later swaps its half onto the chosen node defers the choice, but the chooser already holds the gate. A Greenberger–Horne–Zeilinger (GHZ) state spread over the user and all candidates, as in quantum information secret sharing [13 , 15 ] , requires every unselected node to measure and announce its outcome, so any one of them can veto the gate, and it leaves each share classically correlated with the user (Supplementary Section S13).
Encrypted cloning [26 ] supplies the missing ingredient. It splits a qubit into n n n signal copies, each indistinguishable from noise, any one of which can be decrypted once by whoever holds an entangled key. Qudit versions of encrypted cloning have recently been developed [27 , 28 ] , with Ref. [28 ] relating the construction to AME states and secret sharing. Here we apply encrypted cloning not to data but to one half of a gate-teleportation resource, and show that the result is a threshold quantum secret-sharing scheme for a non-local two-qudit gate. Specifically, a provider distributes encrypted shares to n n n cancandidate nodes. Releasing n n n key qudits later lets exactly one node complete an exact CSUM \operatorname{CSUM} CSUM with the user, in every measurement branch, for any dimension d d d and for data entangled with arbitrary external systems, while every coalition of fewer than n n n candidates, or the key register alone, holds nothing. We derive the access structure of this encoder in closed form; for two candidates the five distributed registers form a ( ( 3 , 5 ) ) ((3,5)) (( 3 , 5 )) threshold scheme, as anticipated in Ref. [28 ] . The user alone decides which power CSUM k \operatorname{CSUM}^{k} CSUM k is applied, and k k k is not revealed in any transmitted register or public message. Encoder and decoder are explicit circuits of O ( n ) O(n) O ( n ) gates, independent of d d d , and every claim is proved in closed form. We work with qudits because high-dimensional carriers are increasingly available [29 , 30 , 31 ] and bring higher capacity and noise tolerance to cryptographic tasks [32 ] . The construction is novel already for d = 2 d=2 d = 2 ; for d > 2 d>2 d > 2 it relies on two features of the arithmetic of the qudit Clifford group [33 , 34 ] that are not present for qubits: a flat phase angle that depends on the parity of d d d , which we show is necessary for the quadratic-phase encoder, and sign bookkeeping in the teleportation stage that makes every power of CSUM \operatorname{CSUM} CSUM available.
Protocol
The protocol runs in three stages as depicted in Figure 1 .
Figure 1. Deferred-recipient teleportation of a two-qudit gate. In stage 1, a maximally entangled pair ∣ Φ ⟩ B A B C |\Phi\rangle_{B_A B_C} ∣Φ ⟩ B A B C is shared between Alice, who holds B A B_A B A , and the provider, who holds B C B_C B C ; it may be prepared by either party. The provider also prepares n n n Bell pairs ∣ Φ ⟩ S j K j |\Phi\rangle_{S_j K_j} ∣Φ ⟩ S j K j and applies the encrypted-cloning encoder U e n c U_{\mathrm{enc}} U enc to B C B_C B C and the signal qudits S 1 , … , S n S_1,\ldots,S_n S 1 , … , S n , then sends one signal to each of n n n candidate sites and withholds the key qudits K 1 , … , K n K_1,\ldots,K_n K 1 , … , K n . Each share alone carries no information, and the provider keeps nothing that can complete the gate. Alice’s data qudit D A D_A D A is untouched at this stage. The recipient (Bob) is chosen only after distribution in the next stage: releasing the key register to one site lets it apply U d e c U_{\mathrm{dec}} U dec and recover B C B_C B C , with its entanglement with B A B_A B A , at unit fidelity. All other shares (dashed, crossed) stay maximally mixed, their holders cannot tell that a selection was made, and the spent key cannot decrypt them. In the final stage, Alice applies CSUM ( D A → B A ) \operatorname{CSUM}(D_A \to B_A) CSUM ( D A → B A ) and Bob applies CSUB ( S B → D B ) \operatorname{CSUB}(S_B \to D_B) CSUB ( S B → D B ) . One computational-basis and one Fourier-basis measurement, two publicly exchanged dits and the Weyl corrections X d m B X_d^{m_B} X d m B on D B D_B D B and Z d m A Z_d^{m_A} Z d m A on D A D_A D A complete an exact CSUM ( D A → D B ) \operatorname{CSUM}(D_A \to D_B) CSUM ( D A → D B ) in every measurement branch, although D A D_A D A and D B D_B D B never interact directly. Replacing the two gates by CSUM k 1 \operatorname{CSUM}^{k_1} CSUM k 1 and CSUB k 2 \operatorname{CSUB}^{k_2} CSUB k 2 yields CSUM k 1 k 2 \operatorname{CSUM}^{k_1k_2} CSUM k 1 k 2 .
Stage 1 — distributing the resource. A provider, who may be Alice herself, prepares a maximally entangled pair ∣ Φ ⟩ B A B C |\Phi\rangle_{B_A B_C} ∣Φ ⟩ B A B C and gives B A B_A B A to Alice. By applying U e n c U_{\mathrm{enc}} U enc it encodes B C B_C B C into n n n signal qudits S 1 , … , S n S_1,\ldots,S_n S 1 , … , S n , each initially Bell-paired with a key qudit K j K_j K j , sends S j S_j S j to candidate site j j j and retains the key register ( K 1 , … , K n ) (K_1,\ldots,K_n) ( K 1 , … , K n ) . Alice’s data qudit D A D_A D A plays no part and need not yet exist.
Stage 2 — one site redeems the key. At any later time and by any criterion external to the protocol, the provider selects a site, Bob, and releases the key register to that site alone. Bob applies a decoder U d e c U_{\mathrm{dec}} U dec to ( S B , K 1 , … , K n ) (S_B,K_1,\ldots,K_n) ( S B , K 1 , … , K n ) and recovers B C B_C B C with its full entanglement with B A B_A B A , at fidelity exactly one: ( B A , S B ) (B_A,S_B) ( B A , S B ) is then a perfect Bell pair between Alice and Bob. Every other share is left untouched and maximally mixed. Operational requirement: the key register must reach Bob intact over an authenticated quantum channel [35 ] ; the keys alone carry no information, but an unselected site that obtained them could redeem the gate in Bob’s place.
Stage 3 — local operations complete the gate. Alice and Bob complete a non-local gate by local operations and one round of classical communication. Alice applies CSUM ( D A → B A ) \operatorname{CSUM}(D_A \to B_A) CSUM ( D A → B A ) and measures B A B_A B A in the computational basis (outcome m A m_A m A ); Bob applies CSUB ( S B → D B ) \operatorname{CSUB}(S_B \to D_B) CSUB ( S B → D B ) and measures S B S_B S B in the Fourier basis { F † ∣ m ⟩ } \{F^\dagger|m\rangle\} { F † ∣ m ⟩} , i.e. applies F F F (defined below) and reads out m B m_B m B in the computational basis; after exchanging the two dits, Bob applies X d m A X_d^{m_A} X d m A to D B D_B D B and Alice applies Z d m B Z_d^{m_B} Z d m B to D A D_A D A . Only Stage 1 requires quantum communication with all candidates, and the two dits of Stage 3 may be announced publicly (Methods). Operational requirement: the dits need integrity but no secrecy [36 ] ; altering them by ( δ A , δ B ) (\delta_A,\delta_B) ( δ A , δ B ) replaces CSUM by ( Z d δ B ⊗ X d δ A ) (Z_d^{\delta_B}\otimes X_d^{\delta_A}) ( Z d δ B ⊗ X d δ A ) CSUM, a definite Weyl error rather than a randomization (Methods).
Trust assumptions and anonymity. The provider is trusted to follow the protocol and knows the recipient, as does Alice. Candidate sites are colluding semi-honest: they may store, measure and pool shares but do not deviate. Anonymity refers to these unselected sites: their registers and the public transcript ( m A , m B ) (m_A,m_B) ( m A , m B ) are independent of whether, and which, site was selected (Supplementary Section S13).
Encoding and decoding circuits
The construction uses the shift and clock operators X d ∣ j ⟩ = ∣ j + 1 ⟩ X_d\lvert j\rangle=\lvert j+1\rangle X d ∣ j ⟩ = ∣ j + 1 ⟩ and Z d ∣ j ⟩ = ω j ∣ j ⟩ Z_d\lvert j\rangle=\omega^j\lvert j\rangle Z d ∣ j ⟩ = ω j ∣ j ⟩ with ω = e 2 π i / d \omega=e^{2\pi i/d} ω = e 2 π i / d and all labels taken modulo d d d ; the discrete Fourier transform F k j = ω j k / d F_{kj}=\omega^{jk}/\sqrt{d} F k j = ω j k / d ; the sum gate CSUM ∣ c , t ⟩ = ∣ c , c + t ⟩ \lvert c,t\rangle=\lvert c,c+t\rangle ∣ c , t ⟩ = ∣ c , c + t ⟩ and its inverse CSUB; and the quadratic phase gate
P θ = diag ( e i θ k 2 ) k = 0 d − 1 . (1) P_\theta=\operatorname{diag}\left(e^{i\theta k^2}\right)_{k=0}^{d-1}. \tag*{(1)} P θ = diag ( e i θ k 2 ) k = 0 d − 1 . (1)
With L \mathcal{L} L the ladder CSUM( B C → S 1 ) (B_C\to S_1) ( B C → S 1 ) , CSUM( S 1 → S 2 ) (S_1\to S_2) ( S 1 → S 2 ) , … \ldots … , CSUM( S n − 1 → S n ) (S_{n-1}\to S_n) ( S n − 1 → S n ) and F = F ⊗ ( n + 1 ) \mathbb{F}=F^{\otimes(n+1)} F = F ⊗ ( n + 1 ) on ( B C , S 1 , … , S n ) (B_C,S_1,\ldots,S_n) ( B C , S 1 , … , S n ) , the encoder is
U e n c = U X U Z , U Z = L † ( P θ on S n ) L , U X = F † U Z F , (2) U_{\mathrm{enc}}=U_XU_Z,\qquad U_Z=\mathcal{L}^{\dagger}(P_\theta\text{ on }S_n)\mathcal{L},\qquad U_X=\mathbb{F}^{\dagger}U_Z\mathbb{F}, \tag*{(2)} U enc = U X U Z , U Z = L † ( P θ on S n ) L , U X = F † U Z F , (2)
which costs 4 n 4n 4 n two-qudit gates, two phase gates and two Fourier layers (Figure 2 a). For d = 2 d=2 d = 2 it reduces to the qubit encoder of Ref. [26 ] ; the Pauli-exponential form used there has no qudit counterpart, which is why we build the encoder as a circuit (Methods). Each S j S_j S j starts in the Bell state ∣ Φ ⟩ S j K j = d − 1 / 2 ∑ k ∣ k k ⟩ \lvert\Phi\rangle_{S_jK_j}=d^{-1/2}\sum_k\lvert kk\rangle ∣ Φ ⟩ S j K j = d − 1/2 ∑ k ∣ k k ⟩ . Privacy of the unselected shares holds for every θ \theta θ , but exact recovery requires θ \theta θ to be a flat angle (below), the smallest of which is
Figure 2. Explicit encoding and decoding circuits. a, The encoder U e n c U_{\mathrm{enc}} U enc of Eq. (2) : a CSUM ladder accumulates the total register sum onto S n S_n S n , a single quadratic phase gate P θ ∗ P_{\theta^*} P θ ∗ is applied, and the ladder is inverted; the whole block is then repeated conjugated by F ⊗ ( n + 1 ) F^{\otimes(n+1)} F ⊗ ( n + 1 ) . Key registers K j K_j K j are spectators throughout. b, The decoder U d e c ( B ) U_{\mathrm{dec}}^{(B)} U dec ( B ) of Eq. (5) , acting only on ( S B , K 1 , … , K n ) (S_B,K_1,\ldots,K_n) ( S B , K 1 , … , K n ) : CSUB( K B → S B ) (K_B\to S_B) ( K B → S B ) extracts the shift label onto S B S_B S B ; CSUM( S B → K j ) (S_B\to K_j) ( S B → K j ) restores each unselected pair to ∣ Φ ⟩ \lvert\Phi\rangle ∣ Φ ⟩ ; CSUM( K j → K B ) (K_j\to K_B) ( K j → K B ) and CSUB( S B → K B ) n (S_B\to K_B)^n ( S B → K B ) n collect the phase label onto K B K_B K B ; the single-qudit gates Λ \Lambda Λ and Γ \Gamma Γ cancel the two Gauss-sum chirps; and J J J followed by CSUM( K B → S B ) (K_B\to S_B) ( K B → S B ) recombines. Total: 3 n 3n 3 n two-qudit and three single-qudit gates, independent of the d 2 ( n + 1 ) d^{2(n+1)} d 2 ( n + 1 ) Hilbert-space dimension a generic recovery map would act on.
θ ∗ ( d ) = 2 π d gcd ( 2 , d ) = { 2 π / d , d odd , π / d , d even . (3) \theta^*(d)=\frac{2\pi}{d\operatorname{gcd}(2,d)}=
\begin{cases}
2\pi/d, & d\ \text{odd},\\
\pi/d, & d\ \text{even}.
\end{cases} \tag*{(3)} θ ∗ ( d ) = d gcd ( 2 , d ) 2 π = { 2 π / d , π / d , d odd , d even . (3)
The decoder U d e c U_{\mathrm{dec}} U dec acts only on ( S B , K 1 , … , K n ) (S_B,K_1,\ldots,K_n) ( S B , K 1 , … , K n ) and never touches B A B_A B A , D A D_A D A , D B D_B D B or any unselected share. It is an explicit circuit, obtained by inverting the Weyl-sum form of the encoder term by term (Methods). With ν ( d ) = gcd ( 2 , d ) \nu(d)=\gcd(2,d) ν ( d ) = g cd( 2 , d ) , ν i n v = 2 / ν \nu_{\mathrm{inv}}=2/\nu ν inv = 2/ ν (always invertible modulo d d d ) and
λ ( d ) = { ω 4 − 1 , d odd , e i π / d , d even , (4) \lambda(d)=\begin{cases}\omega^{4^{-1}}, & d\ \text{odd},\\ e^{i\pi/d}, & d\ \text{even},\end{cases} \tag*{(4)} λ ( d ) = { ω 4 − 1 , e iπ / d , d odd , d even , (4)
where 4 − 1 4^{-1} 4 − 1 is the inverse of 4 modulo d d d , it is the ordered product, read right to left,
U d e c ( B ) = CSUM ( K B → S B ) ⋅ J S B ⋅ Γ K B ⋅ Λ S B U_{\mathrm{dec}}^{(B)}=\operatorname{CSUM}(K_B\to S_B)\cdot J_{S_B}\cdot\Gamma_{K_B}\cdot\Lambda_{S_B} U dec ( B ) = CSUM ( K B → S B ) ⋅ J S B ⋅ Γ K B ⋅ Λ S B
⋅ CSUB ( S B → K B ) n ⋅ ∏ j ≠ B CSUM ( K j → K B ) \qquad\cdot\operatorname{CSUB}(S_B\to K_B)^n\cdot\prod_{j\ne B}\operatorname{CSUM}(K_j\to K_B) ⋅ CSUB ( S B → K B ) n ⋅ j = B ∏ CSUM ( K j → K B )
⋅ ∏ j ≠ B CSUM ( S B → K j ) ⋅ CSUB ( K B → S B ) , (5) \qquad\cdot\prod_{j\ne B}\operatorname{CSUM}(S_B\to K_j)\cdot\operatorname{CSUB}(K_B\to S_B), \tag*{(5)} ⋅ j = B ∏ CSUM ( S B → K j ) ⋅ CSUB ( K B → S B ) , (5)
with single-qudit gates Λ = diag ( λ k 2 ) \Lambda=\operatorname{diag}(\lambda^{k^2}) Λ = diag ( λ k 2 ) , J ∣ x ⟩ = ∣ − x ⟩ J|x\rangle=|-x\rangle J ∣ x ⟩ = ∣ − x ⟩ and Γ = P θ † R F P θ † \Gamma=P_\theta^\dagger R F P_\theta^\dagger Γ = P θ † R F P θ † , where R ∣ u ⟩ = ∣ ν i n v − 1 u ⟩ R|u\rangle=| \nu_{\mathrm{inv}}^{-1}u\rangle R ∣ u ⟩ = ∣ ν inv − 1 u ⟩ . The decoder uses 3 n 3n 3 n two-qudit and three single-qudit gates (Figure 2 b), independent of d d d .
Exact gate, privacy and single use
The central result is that the protocol implements the gate exactly, for any input and in every measurement branch. Take any d ≥ 2 d\ge2 d ≥ 2 and n ≥ 2 n\ge2 n ≥ 2 , an encoder with θ = θ ∗ \theta=\theta^* θ = θ ∗ (or any flat angle, with the decoder adjusted as in Methods), and any site B B B . Let Alice apply CSUM k 1 ( D A → B A ) \operatorname{CSUM}^{k_1}(D_A\to B_A) CSUM k 1 ( D A → B A ) and Bob CSUB k 2 ( S B → D B ) \operatorname{CSUB}^{k_2}(S_B\to D_B) CSUB k 2 ( S B → D B ) , with corrections X D k 2 m A X_D^{k_2m_A} X D k 2 m A on D B D_B D B and Z D k 1 m B Z_D^{k_1m_B} Z D k 1 m B on D A D_A D A , and set k ≡ k 1 k 2 ( m o d d ) k\equiv k_1k_2\pmod d k ≡ k 1 k 2 ( mod d ) (Figure 1 is k 1 = k 2 = 1 k_1=k_2=1 k 1 = k 2 = 1 ). The input is an arbitrary density operator ρ \rho ρ on D A D B R D_AD_BR D A D B R , where R R R is a reference system of any dimension, held by anyone, on which the protocol does not act. With ideal operations, delivery of the intact key register to B B B and faithful exchange of ( m A , m B ) (m_A,m_B) ( m A , m B ) , every branch ( m A , m B ) (m_A,m_B) ( m A , m B ) occurs with probability 1 / d 2 1/d^2 1/ d 2 , independently of ρ \rho ρ , and the corrected state conditioned on that branch is
ρ o u t ( m A , m B ) = ( CSUM k ⊗ I R ) ρ ( CSUM k ⊗ I R ) † , (6) \rho_{\mathrm{out}}^{(m_A,m_B)}=(\operatorname{CSUM}^{k}\otimes I_R)\rho(\operatorname{CSUM}^{k}\otimes I_R)^\dagger, \tag*{(6)} ρ out ( m A , m B ) = ( CSUM k ⊗ I R ) ρ ( CSUM k ⊗ I R ) † , (6)
with CSUM k \operatorname{CSUM}^{k} CSUM k acting as D A → D B D_A\to D_B D A → D B . All other registers end in a state that does not depend on ρ \rho ρ and is uncorrelated with D A D B R D_AD_BR D A D B R .
For a pure input ∣ χ ⟩ = ∑ a , b c a , b ∣ a ⟩ D A ∣ b ⟩ D B |\chi\rangle=\sum_{a,b}c_{a,b}|a\rangle_{D_A}|b\rangle_{D_B} ∣ χ ⟩ = ∑ a , b c a , b ∣ a ⟩ D A ∣ b ⟩ D B and k = 1 k=1 k = 1 , (6) reads CSUM ∣ χ ⟩ = ∑ a , b c a , b ∣ a ⟩ D A ∣ b + a ⟩ D B \operatorname{CSUM}|\chi\rangle=\sum_{a,b}c_{a,b}|a\rangle_{D_A}|b+a\rangle_{D_B} CSUM ∣ χ ⟩ = ∑ a , b c a , b ∣ a ⟩ D A ∣ b + a ⟩ D B . (6) equates branch-conditioned outputs, not branch averages: nothing is post-selected, every branch is corrected to the same target, and because R R R is arbitrary it covers mixed inputs and inputs entangled with systems outside anyone’s control. The derivation (Methods) combines two exact identities: the decoder returns a perfect Bell pair between Alice and Bob, (21) , and that pair teleports CSUM k \operatorname{CSUM}^{k} CSUM k deterministically, (22) .
Privacy is equally exact. Any set of unselected shares—one site, or all n − 1 n-1 n − 1 non-recipients pooling their qudits—is left in I / d ∣ C ∣ ⊗ ρ E I/d^{|C|}\otimes\rho_E I / d ∣ C ∣ ⊗ ρ E , maximally mixed and uncorrelated with any external system E E E , for every d d d , every n ≥ 2 n\ge2 n ≥ 2 and every phase angle (Methods). An unselected site therefore learns nothing about Alice’s data and cannot distinguish a run in which a different site was selected from one in which no site was selected at all. The key is single-use. After decoding, the full register is the product of maximally entangled pairs
∣ Φ ⟩ B A S B ⊗ ∣ Φ ⟩ B C K B ⊗ ⨂ j ≠ B ∣ Φ ⟩ S j K j , (7) |\Phi\rangle_{B_AS_B}\otimes|\Phi\rangle_{B_CK_B}\otimes\bigotimes_{j\ne B}|\Phi\rangle_{S_jK_j}, \tag*{(7)} ∣Φ ⟩ B A S B ⊗ ∣Φ ⟩ B C K B ⊗ j = B ⨂ ∣Φ ⟩ S j K j , (7)
in which every unselected share is re-paired with its own key and retains no correlation with B A B_A B A . A second site that applies the decoder with the spent key obtains entanglement fidelity 1 / d 2 1/d^2 1/ d 2 with B A B_A B A , the value of a random guess. Only the coalition of all n n n candidates, which before selection includes the future recipient, can be correlated with B A B_A B A . It too is private if n n n is chosen with gcd ( n − η 2 , d ) = 1 \gcd(n-\eta^2,d)=1 g cd( n − η 2 , d ) = 1 , where η = 1 \eta=1 η = 1 for even d d d and η = 2 − 1 \eta=2^{-1} η = 2 − 1 modulo d d d for odd d d d ; for n = 2 n=2 n = 2 this holds unless 7 ∣ d 7\mid d 7 ∣ d (Supplementary Section S8).
Secret sharing of a gate
The encoded register is a quantum secret-sharing scheme. The secret is B C B_C B C ’s entanglement with B A B_A B A , and the 2 n + 1 2n+1 2 n + 1 shareholders are B C B_C B C , the n n n signal qudits and the n n n key qudits. At a flat angle the encoder is a Clifford circuit, so the encoded register is a stabilizer state, and its access structure follows exactly from counting solutions of linear congruences modulo d d d (Supplementary Section S9). A set of registers that misses at least one signal–key pair entirely holds the secret if and only if it contains B C B_C B C and at least one complete pair ( S j , K j ) (S_j,K_j) ( S j , K j ) ; a set that touches every pair holds it if it contains any complete pair; and the only remaining sets, which hold exactly one member of each pair, carry an amount of information fixed by a single gcd condition extending earlier classifications of informative subsets in encrypted cloning [37 , 38 , 39 ] .
Three consequences give the scheme its operational meaning. First, the key register on its own carries no information for any d d d and n n n . Handing the n n n key qudits to n n n independent custodians therefore makes the gate available only if every custodian releases its key, while the custodians together, without a share, learn nothing—joint authorization of a quantum operation rather than of a message. Second, the recipient’s set ( S B , K 1 , … , K n ) (S_B,K_1,\ldots,K_n) ( S B , K 1 , … , K n ) is authorized without B C B_C B C , so the provider may discard B C B_C B C once the shares are sent and never itself holds the gate. Third, for n = 2 n=2 n = 2 the five registers form a ( ( 3 , 5 ) ) ((3,5)) (( 3 , 5 )) threshold scheme, the qudit analog of the scheme that Ref. [14 ] obtains from the five-qubit code: any three recover the secret, and any two may be lost or intercepted without leaking anything or destroying the gate. Lim and Lo showed that encrypted cloning with two pairs is equivalent to such a scheme in every dimension [28 ] ; for the present encoder the access structure shows that this holds at θ ∗ \theta^* θ ∗ whenever gcd ( 7 , d ) = 1 \gcd(7,d)=1 g cd( 7 , d ) = 1 , and for d = 7 d=7 d = 7 another flat angle restores it. For n ≥ 3 n\ge3 n ≥ 3 the scheme is no longer threshold: some unauthorized sets hold exactly half of the secret, as in ramp schemes [40 ] . Unlike in secret sharing of a state, the reconstructed secret is an entangled resource that the recipient spends on a gate with data that never entered the scheme: the protocol shares a gate.
Two features of qudit arithmetic
Two elements of the construction are invisible at d = 2 d=2 d = 2 and must be handled explicitly for $d>2 ({xref}Fig. 3). Both reflect known features of the qudit Clifford group {citep}[33 , 34 ] . The first is a genuine constraint: for the quadratic-phase encoder of {xref}Eq. (2)`, we show that no other angle gives exact recovery by any decoder. The second is a bookkeeping rule that, once made explicit, turns into a resource (Methods).
The flat angle splits by parity. Because B C B_C B C never reaches the recipient, exact recovery requires the provider’s pair to be left maximally mixed, which forces the Gauss sum g ^ ( p ) = ∑ k e i θ k 2 ω − p k \hat{g}(p)=\sum_k e^{i\theta k^2}\omega^{-pk} g ^ ( p ) = ∑ k e i θ k 2 ω − p k to be flat, ∣ g ^ ( p ) ∣ = d |\hat{g}(p)|=\sqrt{d} ∣ g ^ ( p ) ∣ = d for every p p p (Methods). This holds exactly on the lattice θ = π s / d \theta=\pi s/d θ = π s / d with gcd ( s , d ) = 1 \gcd(s,d)=1 g cd( s , d ) = 1 and s d sd s d even (see Figure 3 b), whose smallest positive member is Eq. (3) . For odd d d d it is reached by completing the square with the inverse of 2 modulo d d d ; for even d d d that inverse does not exist, and a 2 d 2d 2 d -th root of unity is needed instead. Rescal-
Figure 3. Two features hidden at d = 2 d = 2 d = 2 . a, The flatness spread max p ∣ g ^ ( p ) ∣ − min p ∣ g ^ ( p ) ∣ \max_p |\hat{g}(p)|-\min_p |\hat{g}(p)| max p ∣ g ^ ( p ) ∣ − min p ∣ g ^ ( p ) ∣ , in units of d \sqrt{d} d , as a function of the phase angle θ \theta θ for one odd (d = 3 d = 3 d = 3 ) and one even (d = 4 d = 4 d = 4 ) dimension; filled circles mark its exact zeros. The naive rescaling of the qubit angle π / 2 \pi/2 π /2 to π / d \pi/d π / d (×) is flat for d = 4 d = 4 d = 4 but not for d = 3 d = 3 d = 3 , where ∣ g ^ ( p ) ∣ = ( 1 , 7 , 1 ) |\hat{g}(p)| = (1,\sqrt{7},1) ∣ g ^ ( p ) ∣ = ( 1 , 7 , 1 ) and the spread is ( 7 − 1 ) / 3 ≈ 0.95 (\sqrt{7}-1)/\sqrt{3}\approx0.95 ( 7 − 1 ) / 3 ≈ 0.95 . b, The zero set is the lattice Θ d = { π s / d : gcd ( s , d ) = 1 , s d even } \Theta_d=\{\pi s/d:\gcd(s,d)=1,sd\text{ even}\} Θ d = { π s / d : g cd( s , d ) = 1 , s d even } , shown for d = 2 , … , 9 d=2,\ldots,9 d = 2 , … , 9 , with open rings on its smallest positive member θ ∗ \theta^* θ ∗ : 2 π / d 2\pi/d 2 π / d for odd d d d and π / d \pi/d π / d for even d d d . c, Fidelity to CSUM ∣ c , t ⟩ |c,t\rangle ∣ c , t ⟩ at d = 3 d=3 d = 3 of the literal transcription of the qubit circuit, in which both parties apply CSUM and the qubit corrections X d m A X_d^{m_A} X d m A , Z d m B Z_d^{m_B} Z d m B are used, for every input ( c , t ) (c,t) ( c , t ) and every measurement branch ( m A , m B ) (m_A,m_B) ( m A , m B ) . This circuit teleports CSUB, so it reproduces CSUM only on the ν ( d ) / d \nu(d)/d ν ( d ) / d fraction with m A ≡ c ( m o d d / ν ( d ) ) m_A\equiv c\pmod{d/\nu(d)} m A ≡ c ( mod d / ν ( d )) (the diagonal blocks), a coincidence with the unmeasurable logical value c c c . Choosing opposite signs for the two gates, k 1 k 2 ≡ 1 k_1k_2\equiv1 k 1 k 2 ≡ 1 in Eq. (8) , makes every entry equal to 1 (not shown).
ing the qubit angle π / 2 \pi/2 π /2 to π / d \pi/d π / d therefore fails for every odd d d d ; at d = 3 d=3 d = 3 it gives ∣ g ^ ( p ) ∣ = ( 1 , 7 , 1 ) |\hat{g}(p)|=(1,\sqrt{7},1) ∣ g ^ ( p ) ∣ = ( 1 , 7 , 1 ) (Figure 3 a,Figure 3 b).
The teleported power is a product of local powers. If Alice applies C S U M k 1 ( D A → B A ) \mathrm{CSUM}^{k_1}(D_A\to B_A) CSUM k 1 ( D A → B A ) and Bob applies C S U M k 2 ( S B → D B ) \mathrm{CSUM}^{k_2}(S_B\to D_B) CSUM k 2 ( S B → D B ) , then with corrections X d k 2 m A X_d^{k_2m_A} X d k 2 m A on D B D_B D B and Z d k 1 m B Z_d^{k_1m_B} Z d k 1 m B on D A D_A D A the protocol applies
C S U M k ( D A → D B ) , k ≡ k 1 k 2 ( m o d d ) , (8) \mathrm{CSUM}^{k}(D_A\to D_B),\qquad k\equiv k_1k_2\pmod d, \tag*{(8)} CSUM k ( D A → D B ) , k ≡ k 1 k 2 ( mod d ) , (8)
exactly in every branch; the protocol of Figure 1 is k 1 = k 2 = 1 k_1=k_2=1 k 1 = k 2 = 1 . Because CSUM is its own inverse only for qubits, the literal transcription of the qubit circuit, in which both parties apply CSUM (k 1 = 1 k_1=1 k 1 = 1 , k 2 = − 1 k_2=-1 k 2 = − 1 ), teleports CSUB; used with the qubit corrections it leaves D B D_B D B displaced by an amount that depends on the unknown logical value of D A D_A D A (Figure 3 c). Any choice with k 1 k 2 ≡ 1 k_1k_2\equiv1 k 1 k 2 ≡ 1 gives CSUM; which party carries the inverse is immaterial. More usefully, Eq. (8) makes every power of CSUM reachable with one Bell pair and two dits. If Bob fixes k 2 = 1 k_2=1 k 2 = 1 , Alice alone chooses k = k 1 k=k_1 k = k 1 , possibly after the recipient has been selected. Bob’s operations, the key register and the public dits ( m A , m B ) (m_A,m_B) ( m A , m B ) , which are uniformly distributed for every k k k , are then all independent of k k k : the protocol reveals nothing about the power applied beyond what the gate’s own action on Bob’s data qudit can reveal (Methods).
Effect of noise on the distributed registers
To move beyond ideal operations we apply the qudit depolarizing channel D p ( ρ ) = ( 1 − p ) ρ + p I Tr [ ρ ] / d \mathcal{D}_p(\rho)=(1-p)\rho+p\,\mathbb{I}\operatorname{Tr}[\rho]/d D p ( ρ ) = ( 1 − p ) ρ + p I Tr [ ρ ] / d to the qudits that travel or are stored between encoding and decoding, and compute the process fidelity F \mathcal{F} F of the completed gate exactly (Methods). The model is deliberately limited to this storage and transmission stage: state preparation, the encoder and decoder gates, the Stage 3 operations and both measurements are taken to be ideal. The results therefore quantify how errors on the distributed registers propagate to the gate. They are not an end-to-end fault-tolerance analysis, which would require encoding the full register (Discussion). Noise on the redeemed share (S B S_B S B ) gives
F ( p ) = 1 − ( 1 − 1 d 2 ) p , (9) \mathcal{F}(p)=1-\left(1-\frac{1}{d^2}\right)p, \tag*{(9)} F ( p ) = 1 − ( 1 − d 2 1 ) p , (9)
linear from 1 to the random-unitary value 1 / d 2 1/d^2 1/ d 2 , with no threshold (Figure 4 a). Noise on the unselected shares has no effect at all, whereas the key qudits are as sensitive as the redeemed share (Figure 4 b); for K B K_B K B this holds whenever gcd ( n − η 2 , d ) = 1 \gcd(n-\eta^2,d)=1 g cd( n − η 2 , d ) = 1 , so only the redeemed share and the key register need protection. Three aspects of the noise response should be kept apart. The floor does not compound: depolarizing both S B S_B S B and K B K_B K B completely still leaves 1 / d 2 1/d^2 1/ d 2 rather than the 1 / d 4 1/d^4 1/ d 4 that independent errors would give (Figure 4 c), because part of the joint errors cancel in the decoder. The full p p p -dependence does compound: with both registers noisy, F ( p ) = 1 − ( 1 − 1 / d 2 ) [ 1 − ( 1 − p ) 2 ] \mathcal{F}(p)=1-(1-1/d^2)[1-(1-p)^2] F ( p ) = 1 − ( 1 − 1/ d 2 ) [ 1 − ( 1 − p ) 2 ] , so the infidelity roughly doubles at small p p p . And the scaling with n n n is set by the number of exposed key qudits. Since F \mathcal{F} F is multilinear in the channels, its first-order term is additive, and with S B S_B S B and all n n n keys each depolarized with probability p p p , 1 − F ≃ ( n + 1 ) ( 1 − 1 / d 2 ) p 1-\mathcal{F}\simeq(n+1)(1-1/d^2)p 1 − F ≃ ( n + 1 ) ( 1 − 1/ d 2 ) p to first order in p p p (slightly less when gcd ( n − η 2 , d ) > 1 \gcd(n-\eta^2,d)>1 g cd( n − η 2 , d ) > 1 ; Methods). Protecting the key register is thus a cost that grows linearly with the number of candidates, even though no single error is worse than the loss of one link.
Figure 4. Performance under depolarizing noise. Each panel is computed exactly: the channel D p ( ρ ) = ( 1 − p ) ρ + p I Tr [ ρ ] / d \mathcal{D}_p(\rho)=(1-p)\rho+p\,\mathbb{I}\operatorname{Tr}[\rho]/d D p ( ρ ) = ( 1 − p ) ρ + p I Tr [ ρ ] / d is expanded in Weyl operators and every branch propagated as a pure state, so no sampling is involved. The figure of merit is the process fidelity of the completed non-local CSUM, with D A D_A D A and D B D_B D B each maximally entangled with a reference qudit, summed over all d 2 d^2 d 2 measurement branches. a, Process fidelity versus the depolarizing probability p p p on the redeemed share S B S_B S B , for d = 2 , 3 , 4 d=2,3,4 d = 2 , 3 , 4 at n = 2 n=2 n = 2 . The response is exactly linear, Eq. (9) , from 1 at p = 0 p=0 p = 0 to 1 / d 2 1/d^2 1/ d 2 at p = 1 p=1 p = 1 . b, The same, resolved by which register is exposed (d = 3 d=3 d = 3 , n = 2 n=2 n = 2 ). Noise on the redeemed share S B S_B S B , on the released key K B K_B K B and on an unselected key K j K_j K j act identically; noise on an unselected share S j S_j S j leaves the gate untouched. c, The floor does not compound for errors on S B S_B S B and K B K_B K B : depolarizing both (red) keeps the 1 / d 2 1/d^2 1/ d 2 floor rather than falling to the 1 / d 4 1/d^4 1/ d 4 that independent errors would give (gray dashed).
Discussion
This protocol can be viewed as encrypted cloning of a gate rather than of a state. While an unknown unitary cannot be copied or programmed [2 , 3 ] , our scheme offers the ability to apply one fixed gate to be distributed among n n n sites, with exactly one site able to exercise it at a later time. For every unselected site, the scheme functions as a private quantum channel [41 ] . It is closely related to quantum one-time programs [42 ] , copy protection [43 ] and unclonable encryption [44 ] , which typically require computational or hardware assumptions. In contrast, our protocol achieves single use and privacy through information-theoretic means, though it is limited to a single Clifford gate and assumes parties do not deviate from the protocol. Furthermore, it extends the concept of teleporting a unitary operation [6 , 45 ] by allowing the destination of the operation, as well as its timing, to be decided after the resource is created. Because the power k k k of the CSUM is chosen locally by the user, the provider, the candidates and the public channel remain ignorant of which gate was applied, a limited form of the operation hiding sought in blind and delegated quantum computation [25 ] .
The construction yields two results of independent interest. First, removing the data qudits and the teleportation stage yields qudit encrypted cloning with an explicit 3 n 3n 3 n -gate decoder, exact transfer of entanglement with an arbitrary external system, and flat angles that are both necessary and sufficient. This complements recent qudit constructions whose decoders scale with d d d or are given as operator sums [27 , 28 ] . Second, deleting the cloning stage leaves deterministic non-local CSUMk ^k k for every k ∈ Z d k \in\mathbb{Z}_d k ∈ Z d (Supplementary Sections S7 and S11). The decoder structure also has practical consequences, discussed in Methods: all n n n auxiliary Bell pairs survive decoding and can be reused, a lost key can be replaced by a declined share, and Alice may apply her gate at the last moment.
Our privacy guarantee applies to parties that observe without deviating from the protocol. However, if an unselected site obtains the entire key register, it holds an authorized set, so the key must be transmitted over an authenticated quantum channel. Active attacks on the key or on the classical channel are not covered by this protocol (Supplementary Section S13). Extending the scheme to verifiable secret sharing [24 ] against active adversaries, to non-Clifford or general two-qudit gates, and to fault-tolerant encodings of the full register are natural directions for future work. The resource requirements are modest—2 n + 4 2n+4 2 n + 4 qudits, O ( n ) O(n) O ( n ) two-qudit gates, and one round of classical communication—and are within reach of current qudit processors [29 , 31 , 46 ] and photonic platforms where high-dimensional secret sharing has already been demonstrated [19 ] .
Methods
Qudit conventions
For a qudit with dimension d ≥ 2 d \ge2 d ≥ 2 and computational basis { ∣ 0 ⟩ , … , ∣ d − 1 ⟩ } \{|0\rangle,\ldots,|d-1\rangle\} { ∣0 ⟩ , … , ∣ d − 1 ⟩} , the Weyl–Heisenberg (generalized Pauli) operators [33 ] are X d ∣ j ⟩ = ∣ j + 1 m o d d ⟩ X_d|j\rangle=|j+1\bmod d\rangle X d ∣ j ⟩ = ∣ j + 1 mod d ⟩ and Z d ∣ j ⟩ = ω j ∣ j ⟩ Z_d|j\rangle=\omega^j|j\rangle Z d ∣ j ⟩ = ω j ∣ j ⟩ , which satisfy
Z d X d = ω X d Z d , Z d a X d b = ω a b X d b Z d a ( a , b ∈ Z d ) . (10) Z_dX_d=\omega X_dZ_d,\qquad Z_d^aX_d^b=\omega^{ab}X_d^bZ_d^a\qquad(a,b\in\mathbb{Z}_d). \tag*{(10)} Z d X d = ω X d Z d , Z d a X d b = ω ab X d b Z d a ( a , b ∈ Z d ) . (10)
where ω = e 2 π i / d \omega=e^{2\pi i/d} ω = e 2 π i / d . For p , q ∈ Z d p,q\in\mathbb{Z}_d p , q ∈ Z d , W p , q : = X d p Z d q W_{p,q}:=X_d^pZ_d^q W p , q := X d p Z d q ; the d 2 d^2 d 2 operators { W p , q } \{W_{p,q}\} { W p , q } form an orthogonal operator basis,
Tr [ W p , q † W p ′ , q ′ ] = d δ p p ′ δ q q ′ . (11) \operatorname{Tr}\left[W_{p,q}^{\dagger}W_{p',q'}\right]=d\,\delta_{pp'}\delta_{qq'}. \tag*{(11)} Tr [ W p , q † W p ′ , q ′ ] = d δ p p ′ δ q q ′ . (11)
The discrete Fourier transform F F F (F k j = ω j k / d F_{kj}=\omega^{jk}/\sqrt{d} F k j = ω j k / d ) satisfies F − 1 Z d F = X d F^{-1}Z_dF=X_d F − 1 Z d F = X d , and the maximally entangled state ∣ Φ ⟩ = d − 1 / 2 ∑ j ∣ j j ⟩ |\Phi\rangle=d^{-1/2}\sum_j|jj\rangle ∣Φ ⟩ = d − 1/2 ∑ j ∣ j j ⟩ obeys, for any operator M M M , ⟨ Φ ∣ ( M ⊗ I ) ∣ Φ ⟩ = Tr [ M ] / d \langle\Phi|(M\otimes I)|\Phi\rangle=\operatorname{Tr}[M]/d ⟨ Φ∣ ( M ⊗ I ) ∣Φ ⟩ = Tr [ M ] / d . Together with CSUM, CSUB and P θ P_\theta P θ as defined in the main text, these are the only objects the constructions below require; all arithmetic on qudit labels is modulo d d d throughout.
Why the encoder is built as a circuit
The qubit encrypted-cloning encoder [26 ] ,
U e n c ( q u b i t ) = e − i π 4 X ( B C ) ⊗ ⨂ j = 1 n X ( S j ) e − i π 4 Z ( B C ) ⊗ ⨂ j = 1 n Z ( S j ) , (12) U_{\mathrm{enc}}^{(\mathrm{qubit})}=e^{-i\frac{\pi}{4}X^{(B_C)}\otimes\bigotimes_{j=1}^{n}X^{(S_j)}}e^{-i\frac{\pi}{4}Z^{(B_C)}\otimes\bigotimes_{j=1}^{n}Z^{(S_j)}}, \tag*{(12)} U enc ( qubit ) = e − i 4 π X ( B C ) ⊗ ⨂ j = 1 n X ( S j ) e − i 4 π Z ( B C ) ⊗ ⨂ j = 1 n Z ( S j ) , (12)
is a product of two exponentials of Pauli strings, a form that relies on X 2 = Z 2 = I 2 X^2=Z^2=I_2 X 2 = Z 2 = I 2 and on the single-qubit Pauli group closing on four operators. Neither fact survives for d > 2 d>2 d > 2 : X d X_d X d and Z d Z_d Z d have order d d d , and { W p , q } \{W_{p,q}\} { W p , q } is a set of d 2 d^2 d 2 distinct operators, so the two-exponential form has nothing to generalize to. We therefore define the encoder by the circuit of Eq. (2) and derive its properties from the operator structure below (Supplementary Section S2).
The component U Z U_Z U Z in Eq. (2) is diagonal in the computational basis of ( B C , S 1 , … , S n ) (B_C,S_1,\ldots,S_n) ( B C , S 1 , … , S n ) . This is because the forward CSUM chain L \mathcal{L} L accumulates the running sum onto S n S_n S n , the phase gate P θ P_\theta P θ multiplies by e i θ T 2 e^{i\theta T^2} e i θ T 2 with T = ( a + s 1 + ⋯ + s n ) m o d d T=(a+s_1+\cdots+s_n)\bmod d T = ( a + s 1 + ⋯ + s n ) mod d , and the reverse chain restores every register, so U Z ∣ a , s 1 , … , s n ⟩ = e i θ T 2 ∣ a , s 1 , … , s n ⟩ U_Z|a,s_1,\ldots,s_n\rangle=e^{i\theta T^2}|a,s_1,\ldots,s_n\rangle U Z ∣ a , s 1 , … , s n ⟩ = e i θ T 2 ∣ a , s 1 , … , s n ⟩ . Writing g ( k ) = e i θ k 2 g(k)=e^{i\theta k^2} g ( k ) = e i θ k 2 in terms of its discrete Fourier coefficients g ^ ( p ) = ∑ k g ( k ) ω − p k \hat{g}(p)=\sum_k g(k)\,\omega^{-pk} g ^ ( p ) = ∑ k g ( k ) ω − p k (because g ( k ) = g ( − k ) g(k)=g(-k) g ( k ) = g ( − k ) , g ^ ( p ) = g ^ ( − p ) \hat{g}(p)=\hat{g}(-p) g ^ ( p ) = g ^ ( − p ) , so the opposite sign convention gives identical coefficients) gives U Z = 1 d ∑ p g ^ ( p ) ( Z d p ) ⊗ ( n + 1 ) U_Z=\frac{1}{d}\sum_p\hat{g}(p)(Z_d^p)^{\otimes(n+1)} U Z = d 1 ∑ p g ^ ( p ) ( Z d p ) ⊗ ( n + 1 ) , and conjugating by F = F ⊗ ( n + 1 ) F=F^{\otimes(n+1)} F = F ⊗ ( n + 1 ) turns each Z d p Z_d^p Z d p into X d p X_d^p X d p . Multiplying the two factors and reducing with Eq. (10) collects the result into the Weyl basis,
U e n c = 1 d 2 ∑ p , q = 0 d − 1 μ p , q W p , q ⊗ ( n + 1 ) , μ p , q : = g ^ ( p ) g ^ ( q ) , (13) U_{\mathrm{enc}}=\frac{1}{d^2}\sum_{p,q=0}^{d-1}\mu_{p,q}W_{p,q}^{\otimes(n+1)},\qquad\mu_{p,q}:=\hat{g}(p)\hat{g}(q), \tag*{(13)} U enc = d 2 1 p , q = 0 ∑ d − 1 μ p , q W p , q ⊗ ( n + 1 ) , μ p , q := g ^ ( p ) g ^ ( q ) , (13)
where W p , q ⊗ ( n + 1 ) W_{p,q}^{\otimes(n+1)} W p , q ⊗ ( n + 1 ) acts as W p , q = X d p Z d q W_{p,q}=X_d^pZ_d^q W p , q = X d p Z d q on each of B C , S 1 , … , S n B_C,S_1,\ldots,S_n B C , S 1 , … , S n and as the identity on every key qudit K j K_j K j ; with this ordering of X d X_d X d and Z d Z_d Z d the reordering phases generated by Eq. (10) cancel identically and no residual factor survives. Everything that follows—the flat angle, single-share privacy, and the explicit decoder—is read off this one closed-form expansion. For d = 2 d=2 d = 2 it reduces to the qubit encoder of Ref. [26 ] , up to the global phase i i i (checked numerically for n ≤ 3 n\leq3 n ≤ 3 ).
Flat angles and the chirp identity
Tracing out every signal–key pair in Eq. (17) below (the pair states ( W p , q ⊗ I ) ∣ Φ ⟩ (W_{p,q}\otimes I)|\Phi\rangle ( W p , q ⊗ I ) ∣Φ ⟩ are orthonormal in ( p , q ) (p,q) ( p , q ) ) leaves the provider’s pair in the Bell-diagonal state ρ B A B C = ∑ p , q ∣ μ p , q ∣ 2 d 4 ( W p , q ) B C ∣ Φ ⟩ ⟨ Φ ∣ ( W p , q † ) B C \rho_{B_AB_C}=\sum_{p,q}\frac{|\mu_{p,q}|^2}{d^4}(W_{p,q})_{B_C}|\Phi\rangle\langle\Phi|(W_{p,q}^{\dagger})_{B_C} ρ B A B C = ∑ p , q d 4 ∣ μ p , q ∣ 2 ( W p , q ) B C ∣Φ ⟩ ⟨ Φ∣ ( W p , q † ) B C . The decoder never touches B C B_C B C , so exact recovery of B A B_A B A ’s partner at S B S_B S B requires B A B_A B A to be decoupled from B C B_C B C , i.e. ρ B A B C = I / d 2 \rho_{B_AB_C}=I/d^2 ρ B A B C = I / d 2 , i.e. all d 2 d^2 d 2 weights equal. Recovery therefore requires ∣ g ^ ( p ) ∣ |\hat{g}(p)| ∣ g ^ ( p ) ∣ to be independent of p p p , which by Parseval’s identity ( ∑ p ∣ g ^ ( p ) ∣ 2 = d ∑ k ∣ g ( k ) ∣ 2 = d 2 ) (\sum_p|\hat{g}(p)|^2=d\sum_k|g(k)|^2=d^2) ( ∑ p ∣ g ^ ( p ) ∣ 2 = d ∑ k ∣ g ( k ) ∣ 2 = d 2 ) forces ∣ g ^ ( p ) ∣ = d |\hat{g}(p)|=\sqrt{d} ∣ g ^ ( p ) ∣ = d for every p p p . The scope of this necessity statement is precise: for the quadratic-phase encoder U e n c ( θ ) U_{\mathrm{enc}}(\theta) U enc ( θ ) of Eq. (2) , exact recovery by any decoder acting on ( S B , K 1 , … , K n ) (S_B,K_1,\ldots,K_n) ( S B , K 1 , … , K n ) implies flatness of ∣ g ^ ∣ |\hat{g}| ∣ g ^ ∣ , which pins θ \theta θ to the set Θ d \Theta_d Θ d derived below. It does not exclude encoders outside this one-parameter family. (Privacy, by contrast, needs no condition on θ \theta θ ; see below.) Flatness is not an approximate numerical fact to be checked dimension by dimension: at θ ∗ ( d ) \theta^*(d) θ ∗ ( d ) the Gauss sum g ^ ( p ) \hat{g}(p) g ^ ( p ) is a pure chirp ,
g ^ ( p ) = G λ − p 2 , G = ∑ k ∈ Z d λ k 2 , ∣ G ∣ = d , (14) \hat{g}(p)=G\lambda^{-p^2},\qquad G=\sum_{k\in\mathbb{Z}_d}\lambda^{k^2},\qquad|G|=\sqrt{d}, \tag*{(14)} g ^ ( p ) = G λ − p 2 , G = k ∈ Z d ∑ λ k 2 , ∣ G ∣ = d , (14)
with λ \lambda λ as in Eq. (4) , so that ∣ g ^ ( p ) ∣ = ∣ G ∣ = d |\hat{g}(p)|=|G|=\sqrt{d} ∣ g ^ ( p ) ∣ = ∣ G ∣ = d identically, for every p p p and every d d d . The proof is completion of the square. For odd d d d , θ ∗ = 2 π / d \theta^*=2\pi/d θ ∗ = 2 π / d gives g ( k ) = ω k 2 g(k)=\omega^{k^2} g ( k ) = ω k 2 and 2 2 2 is invertible modulo d d d , so k 2 − p k = ( k − p ⋅ 2 − 1 ) 2 − p 2 ⋅ 4 − 1 k^2-pk=(k-p\cdot2^{-1})^2-p^2\cdot4^{-1} k 2 − p k = ( k − p ⋅ 2 − 1 ) 2 − p 2 ⋅ 4 − 1 and re-indexing k ↦ k + p ⋅ 2 − 1 k\mapsto k+p\cdot2^{-1} k ↦ k + p ⋅ 2 − 1 yields Eq. (14) with λ = ω 4 − 1 \lambda=\omega^{4^{-1}} λ = ω 4 − 1 . For even d d d no such inverse exists; instead, at λ = ζ = e i π / d \lambda=\zeta=e^{i\pi/d} λ = ζ = e iπ / d (a 2 d 2d 2 d -th root of unity with ζ 2 = ω \zeta^2=\omega ζ 2 = ω ) one has g ^ ( p ) = ∑ k ζ k 2 − 2 p k = ζ − p 2 ∑ k ζ ( k − p ) 2 \hat{g}(p)=\sum_k\zeta^{k^2-2pk}=\zeta^{-p^2}\sum_k\zeta^{(k-p)^2} g ^ ( p ) = ∑ k ζ k 2 − 2 p k = ζ − p 2 ∑ k ζ ( k − p ) 2 , and ζ ( k + d ) 2 = ζ k 2 ω k d e i π d = ζ k 2 \zeta^{(k+d)^2}=\zeta^{k^2}\omega^{kd}e^{i\pi d}=\zeta^{k^2} ζ ( k + d ) 2 = ζ k 2 ω k d e iπ d = ζ k 2 because d d d is even, so the shifted sum is the unshifted one and Eq. (14) follows at θ ∗ = π / d \theta^*=\pi/d θ ∗ = π / d , half the odd-d d d value. In both cases ∣ G ∣ = d |G|=\sqrt{d} ∣ G ∣ = d is the classical evaluation of the quadratic Gauss sum. The identity Eq. (14) is the single structural fact the decoder inverts: all of the p p p -dependence of g ^ \hat{g} g ^ sits in a quadratic phase, which one single-qudit diagonal gate removes.
The same argument identifies every flat angle, not just the smallest. Writing θ = π s / d \theta=\pi s/d θ = π s / d , the periodicity step requires e i π s d = 1 e^{i\pi sd}=1 e iπ s d = 1 , i.e. s d sd s d even, and the resulting Gauss sum G s = ∑ k e i π s k 2 / d G_s=\sum_k e^{i\pi s k^2/d} G s = ∑ k e iπ s k 2 / d has modulus d \sqrt{d} d exactly when gcd ( s , d ) = 1 \gcd(s,d)=1 g cd( s , d ) = 1 ; when gcd ( s , d ) = m > 1 \gcd(s,d)=m>1 g cd( s , d ) = m > 1 the sum degenerates and ∣ g ^ ( p ) ∣ |\hat{g}(p)| ∣ g ^ ( p ) ∣ takes the value m d \sqrt{md} m d on a sublattice of p p p and zero elsewhere. Angles off the lattice π Z d / d \pi\mathbb{Z}_d/d π Z d / d are irrational multiples of the relevant root of unity and never flat. Hence the flat set is exactly { θ = π s / d : gcd ( s , d ) = 1 , s d even } \{\theta=\pi s/d:\gcd(s,d)=1,\ sd\text{ even}\} { θ = π s / d : g cd( s , d ) = 1 , s d even } and θ ∗ ( d ) \theta^*(d) θ ∗ ( d ) of Eq. (3) is its smallest positive member. At every flat angle completing the square again gives a pure chirp, g ^ ( p ) = G s λ s − p 2 \hat{g}(p)=G_s\lambda_s^{-p^2} g ^ ( p ) = G s λ s − p 2 , now with
λ s = { ω ( 2 s ) − 1 , d odd , e i π s / d s ˉ , d even , (15) \lambda_s=
\begin{cases}
\omega^{(2s)^{-1}}, & d\ \text{odd},\\
e^{i\pi s/d}\bar{s}, & d\ \text{even},
\end{cases} \tag*{(15)} λ s = { ω ( 2 s ) − 1 , e iπ s / d s ˉ , d odd , d even , (15)
where ( 2 s ) − 1 (2s)^{-1} ( 2 s ) − 1 is taken modulo d d d and s ˉ \bar{s} s ˉ is the inverse of s s s modulo 2 d 2d 2 d ; at θ ∗ \theta^* θ ∗ this reduces to Eq. (4) . All flat angles are therefore equivalent for the protocol: the decoder is unchanged except that Λ = diag ( λ s k 2 ) \Lambda=\operatorname{diag}(\lambda_s^{k^2}) Λ = diag ( λ s k 2 ) , P θ ∗ P_{\theta^*} P θ ∗ in Γ \Gamma Γ is replaced by P θ P_\theta P θ , and R ∣ u ⟩ = ∣ s − 1 u ⟩ R|u\rangle=|s^{-1}u\rangle R ∣ u ⟩ = ∣ s − 1 u ⟩ with s − 1 s^{-1} s − 1 taken modulo d d d .
Privacy of unselected shares
Acting with U e n c U_{\mathrm{enc}} U enc on an input ∣ ψ ⟩ B C E |\psi\rangle_{BCE} ∣ ψ ⟩ B C E (entangled with an external system E E E of any dimension) tensored with the key Bell pairs ∣ Φ ⟩ S j K j |\Phi\rangle_{S_jK_j} ∣Φ ⟩ S j K j gives, by Eq. (13) , U e n c ∣ ψ ⟩ ⊗ ⨂ j ∣ Φ ⟩ S j K j = 1 d 2 ∑ p , q μ p , q ∣ χ p , q ⟩ B C E ⊗ ∣ ϕ p , q ⟩ S j K j U_{\mathrm{enc}}|\psi\rangle\otimes\bigotimes_j|\Phi\rangle_{S_jK_j}=\frac{1}{d^2}\sum_{p,q}\mu_{p,q}|\chi_{p,q}\rangle_{BCE}\otimes|\phi_{p,q}\rangle_{S_jK_j} U enc ∣ ψ ⟩ ⊗ ⨂ j ∣Φ ⟩ S j K j = d 2 1 ∑ p , q μ p , q ∣ χ p , q ⟩ B C E ⊗ ∣ ϕ p , q ⟩ S j K j , with ∣ χ p , q ⟩ = ( W p , q ⊗ I E ) ∣ ψ ⟩ |\chi_{p,q}\rangle=(W_{p,q}\otimes I_E)|\psi\rangle ∣ χ p , q ⟩ = ( W p , q ⊗ I E ) ∣ ψ ⟩ and ∣ ϕ p , q ⟩ = ( W p , q ⊗ I ) ∣ Φ ⟩ S j K j |\phi_{p,q}\rangle=(W_{p,q}\otimes I)|\Phi\rangle_{S_jK_j} ∣ ϕ p , q ⟩ = ( W p , q ⊗ I ) ∣Φ ⟩ S j K j . By Eq. (11) , ⟨ ϕ p ′ , q ′ ∣ ϕ p , q ⟩ = δ p p ′ δ q q ′ \langle\phi_{p',q'}|\phi_{p,q}\rangle=\delta_{pp'}\delta_{qq'} ⟨ ϕ p ′ , q ′ ∣ ϕ p , q ⟩ = δ p p ′ δ q q ′ on each pair. Let C ⊊ { 1 , … , n } C\subsetneq\{1,\ldots,n\} C ⊊ { 1 , … , n } be any coalition of signal qudits that is not the whole set, and let j 0 ∈ C c j_0\in C^c j 0 ∈ C c be a spectator. Because ∣ ϕ p , q ⟩ S j 0 K j 0 |\phi_{p,q}\rangle_{S_{j_0}K_{j_0}} ∣ ϕ p , q ⟩ S j 0 K j 0 is retained in the partial trace and is orthogonal for distinct ( p , q ) (p,q) ( p , q ) , every cross term ( p , q ) ≠ ( p ′ , q ′ ) (p,q)\ne(p',q') ( p , q ) = ( p ′ , q ′ ) is annihilated, and the reduced state of C C C together with E E E collapses to the incoherent mixture
ρ C E = ∑ p , q ∣ μ p , q ∣ 2 d 4 Tr B [ ∣ χ p , q ⟩ ⟨ χ p , q ∣ ] (16) \rho_{CE}=\sum_{p,q}\frac{|\mu_{p,q}|^2}{d^4}\operatorname{Tr}_{B}\left[|\chi_{p,q}\rangle\langle\chi_{p,q}|\right] \tag*{(16)} ρ C E = p , q ∑ d 4 ∣ μ p , q ∣ 2 Tr B [ ∣ χ p , q ⟩ ⟨ χ p , q ∣ ] (16)
⊗ ⨂ j ∈ C Tr K j [ ∣ ϕ p , q ⟩ ⟨ ϕ p , q ∣ ] = I ⊗ ∣ C ∣ d ∣ C ∣ ⊗ ρ E , \otimes\bigotimes_{j\in C}\operatorname{Tr}_{K_j}\left[|\phi_{p,q}\rangle\langle\phi_{p,q}|\right]=\frac{\mathbb{I}^{\otimes|C|}}{d^{|C|}}\otimes\rho_E, ⊗ j ∈ C ⨂ Tr K j [ ∣ ϕ p , q ⟩ ⟨ ϕ p , q ∣ ] = d ∣ C ∣ I ⊗ ∣ C ∣ ⊗ ρ E ,
using Tr B [ ∣ χ p , q ⟩ ⟨ χ p , q ∣ ] = ρ E \operatorname{Tr}_{B} [|\chi_{p,q}\rangle\langle\chi_{p,q}|]=\rho_E Tr B [ ∣ χ p , q ⟩ ⟨ χ p , q ∣ ] = ρ E (the Weyl operator acts on B C BC B C alone and is removed by the partial trace, for any E E E ), Tr K j [ ∣ ϕ p , q ⟩ ⟨ ϕ p , q ∣ ] = I / d \operatorname{Tr}_{K_j} [|\phi_{p,q}\rangle\langle\phi_{p,q}|]=\mathbb{I}/d Tr K j [ ∣ ϕ p , q ⟩ ⟨ ϕ p , q ∣ ] = I / d for every ( p , q ) (p,q) ( p , q ) , and normalization ∑ p , q ∣ μ p , q ∣ 2 = d 4 \sum_{p,q}|\mu_{p,q}|^2=d^4 ∑ p , q ∣ μ p , q ∣ 2 = d 4 . No property of θ \theta θ enters: privacy of proper coalitions holds at every phase angle, and the flat angle is required only for recovery. Every proper coalition therefore holds a state that is maximally mixed and uncorrelated with E E E , for any E E E , every d d d and every n ≥ 2 n\ge2 n ≥ 2 . The argument requires n ≥ 2 n\ge2 n ≥ 2 : for n = 1 n=1 n = 1 the single share has no spectator, so it is the all-shares coalition of Supplementary Section S8, and it is private only when gcd ( 1 − η 2 , d ) = 1 \gcd(1-\eta^2,d)=1 g cd( 1 − η 2 , d ) = 1 . For d = 2 , 3 , 4 , 6 d=2,3,4,6 d = 2 , 3 , 4 , 6 it is correlated with E E E (the trace distance 1 / 2 1/2 1/2 from I / d ⊗ ρ E \mathbb{I}/d\otimes\rho_E I / d ⊗ ρ E at d = 2 d=2 d = 2 ), whereas for d = 5 , 7 d=5,7 d = 5 , 7 it happens to be private. Recovery at n = 1 n=1 n = 1 remains exact.
The decoder is obtained by inverting Eq. (13) term by term. Because U e n c U_{\mathrm{enc}} U enc is a sum over Weyl operators applied identically to B C B_C B C and to every signal qudit, the post-encoding state is a coherent superposition of d 2 d^2 d 2 branches labeled by ( p , q ) (p,q) ( p , q ) ,
U e n c ∣ Φ ⟩ B A B C ⨂ j ∣ Φ ⟩ S j K j = 1 d 2 ∑ p , q μ p , q ( W p , q ) B C ∣ Φ ⟩ B A B C ⨂ j = 1 n ( W p , q ⊗ I ) ∣ Φ ⟩ S j K j , (17) U_{\mathrm{enc}}|\Phi\rangle_{B_AB_C}\bigotimes_j|\Phi\rangle_{S_jK_j}
=
\frac{1}{d^2}\sum_{p,q}\mu_{p,q}(W_{p,q})_{B_C}|\Phi\rangle_{B_AB_C}\bigotimes_{j=1}^{n}(W_{p,q}\otimes\mathbb{I})|\Phi\rangle_{S_jK_j}, \tag*{(17)} U enc ∣Φ ⟩ B A B C j ⨂ ∣Φ ⟩ S j K j = d 2 1 p , q ∑ μ p , q ( W p , q ) B C ∣Φ ⟩ B A B C j = 1 ⨂ n ( W p , q ⊗ I ) ∣Φ ⟩ S j K j , (17)
in which the same label ( p , q ) (p,q) ( p , q ) is shared by every register. The decoder’s task is therefore not to identify an abstract recovery map on a d 2 ( n + 1 ) d^{2(n+1)} d 2 ( n + 1 ) -dimensional space, but to coherently transfer the shared label from the n n n signal–key pairs onto the two registers ( S B , K B ) (S_B,K_B) ( S B , K B ) and then undo it. Each step below is an exact operator identity; no approximation or search enters.
Step 1 — label extraction (CSUB( K B → S B ) (K_B\to S_B) ( K B → S B ) ). On the pair ( S B , K B ) (S_B,K_B) ( S B , K B ) the branch state is ( W p , q ⊗ I ) ∣ Φ ⟩ = d − 1 / 2 ∑ k ω q k ∣ k + p ⟩ ∣ k ⟩ (W_{p,q}\otimes\mathbb{I})|\Phi\rangle=d^{-1/2}\sum_k\omega^{qk}|k+p\rangle|k\rangle ( W p , q ⊗ I ) ∣Φ ⟩ = d − 1/2 ∑ k ω q k ∣ k + p ⟩ ∣ k ⟩ . Subtracting the key value from the signal register maps ∣ k + p ⟩ ∣ k ⟩ ↦ ∣ p ⟩ ∣ k ⟩ |k+p\rangle|k\rangle\mapsto|p\rangle|k\rangle ∣ k + p ⟩ ∣ k ⟩ ↦ ∣ p ⟩ ∣ k ⟩ , factorizing the pair into ∣ p ⟩ S B ⊗ F ∣ q ⟩ K B |p\rangle_{S_B}\otimes F|q\rangle_{K_B} ∣ p ⟩ S B ⊗ F ∣ q ⟩ K B : the shift label p p p is now a computational-basis value on S B S_B S B , and the phase label q q q a Fourier state on K B K_B K B .
Step 2 — spectator restoration (CSUM( S B → K j ) (S_B\to K_j) ( S B → K j ) , j ≠ B j\ne B j = B ). Since S B S_B S B holds ∣ p ⟩ |p\rangle ∣ p ⟩ , adding it into each spectator key register maps ∣ k + p ⟩ S j ∣ k ⟩ K j ↦ ∣ k + p ⟩ S j ∣ k + p ⟩ K j |k+p\rangle_{S_j}|k\rangle_{K_j}\mapsto|k+p\rangle_{S_j}|k+p\rangle_{K_j} ∣ k + p ⟩ S j ∣ k ⟩ K j ↦ ∣ k + p ⟩ S j ∣ k + p ⟩ K j , i.e. every unselected pair returns to ∣ Φ ⟩ S j K j |\Phi\rangle_{S_jK_j} ∣Φ ⟩ S j K j up to the residual phase ω q k \omega^{qk} ω q k .
Step 3 — phase collection (CSUM(K j → K B K_j \to K_B K j → K B ) then CSUB(S B → K B S_B \to K_B S B → K B )n ^n n ). Because K B K_B K B holds a Fourier state, CSUM(c → K B c \to K_B c → K B ) acts on the control as Z d q Z_d^q Z d q and leaves K B K_B K B untouched, while CSUB(c → K B c \to K_B c → K B ) acts as Z d + q Z_d^{+q} Z d + q ; both identities follow from F − 1 Z d F = X d F^{-1}Z_dF=X_d F − 1 Z d F = X d . Applying the first to each of the n − 1 n-1 n − 1 spectator keys strips the phase ω − q ( k + p ) \omega^{-q(k+p)} ω − q ( k + p ) off each, canceling the residue of Step 2 and leaving all n − 1 n-1 n − 1 spectator pairs in exactly ∣ Φ ⟩ S j K j |\Phi\rangle_{S_jK_j} ∣Φ ⟩ S j K j ; applying the second n n n times to S B S_B S B removes the accumulated ω q p \omega^{qp} ω q p bookkeeping. Steps 1–3 use 3 n − 1 3n-1 3 n − 1 two-qudit gates and together implement the exact map
( W p , q ⊗ I ) ⊗ n ⨂ j ∣ Φ ⟩ S j K j ⟼ ω p q ∣ p ⟩ S B ⊗ F ∣ q ⟩ K B ⊗ ⨂ j ≠ B ∣ Φ ⟩ S j K j , (18) (W_{p,q}\otimes\mathbb{I})^{\otimes n}\bigotimes_j|\Phi\rangle_{S_jK_j}
\longmapsto\omega^{pq}|p\rangle_{S_B}\otimes F|q\rangle_{K_B}\otimes\bigotimes_{j\ne B}|\Phi\rangle_{S_jK_j}, \tag*{(18)} ( W p , q ⊗ I ) ⊗ n j ⨂ ∣Φ ⟩ S j K j ⟼ ω pq ∣ p ⟩ S B ⊗ F ∣ q ⟩ K B ⊗ j = B ⨂ ∣Φ ⟩ S j K j , (18)
independently of n n n — the unselected sites are returned to their original Bell pairs and never touched again.
Step 4 — chirp cancellation (Λ \Lambda Λ on S B S_B S B , Γ \Gamma Γ on K B K_B K B ). What remains is the branch weight μ p , q = g ^ ( p ) g ^ ( q ) = G 2 λ − p 2 − q 2 \mu_{p,q}=\hat{g}(p)\hat{g}(q)=G^2\lambda^{-p^2-q^2} μ p , q = g ^ ( p ) g ^ ( q ) = G 2 λ − p 2 − q 2 from Eq. (14) , a product of two independent chirps. This is precisely why a single-qudit diagonal gate suffices on each register: Λ = diag ( λ k 2 ) \Lambda=\operatorname{diag}(\lambda^{k^2}) Λ = diag ( λ k 2 ) on S B S_B S B removes λ − p 2 \lambda^{-p^2} λ − p 2 , while Γ = P θ ∗ † R F † P θ ∗ † \Gamma=P_{\theta^*}^{\dagger}RF^{\dagger}P_{\theta^*}^{\dagger} Γ = P θ ∗ † R F † P θ ∗ † on K B K_B K B obeys the identity
Γ F ∣ q ⟩ = G ‾ d λ q 2 F ∣ − q ⟩ , (19) \Gamma F|q\rangle=\frac{\overline{G}}{\sqrt{d}}\lambda^{q^2}F|-q\rangle, \tag*{(19)} Γ F ∣ q ⟩ = d G λ q 2 F ∣ − q ⟩ , (19)
where G ‾ \overline{G} G is the complex conjugate of the Gauss sum G G G of Eq. (14) , removing λ − q 2 \lambda^{-q^2} λ − q 2 and simultaneously conjugating the phase label. Both gates are q q q - and p p p -independent, so they act identically on every branch and the superposition is preserved. All d 2 d^2 d 2 branch weights are now equal in modulus and in phase, up to the global constant G 2 G ‾ / d G^2\overline{G}/\sqrt{d} G 2 G / d ; had the flatness of ∣ g ^ ∣ |\hat{g}| ∣ g ^ ∣ been only approximate, no such cancellation would exist, which is the sense in which Eq. (14) is the structural fact the whole construction rests on.
Step 5 — recombination (J J J on S B S_B S B , CSUM(K B → S B K_B \to S_B K B → S B )). The state is now ∝ ∑ p , q ω p q ( W p , q ) B C ∣ Φ ⟩ B A B C ⊗ ∣ p ⟩ S B ⊗ F ∣ − q ⟩ K B \propto\sum_{p,q}\omega^{pq}(W_{p,q})_{BC}|\Phi\rangle_{B_A B_C}\otimes|p\rangle_{S_B}\otimes F|-q\rangle_{K_B} ∝ ∑ p , q ω pq ( W p , q ) B C ∣Φ ⟩ B A B C ⊗ ∣ p ⟩ S B ⊗ F ∣ − q ⟩ K B . Applying J ∣ x ⟩ = ∣ − x ⟩ J|x\rangle=|-x\rangle J ∣ x ⟩ = ∣ − x ⟩ to S B S_B S B and then CSUM(K B → S B K_B\to S_B K B → S B ) collapses this double sum: the p p p -sum acts as a completeness relation transferring B C B_C B C ’s shift onto S B S_B S B , and the q q q -sum does the same for the phase. The result is exactly
∣ Φ ⟩ B A S B ⊗ ∣ Φ ⟩ B C K B ⊗ ⨂ j ≠ B ∣ Φ ⟩ S j K j , (20) |\Phi\rangle_{B_A S_B}\otimes|\Phi\rangle_{B_CK_B}\otimes\bigotimes_{j\ne B}|\Phi\rangle_{S_jK_j}, \tag*{(20)} ∣Φ ⟩ B A S B ⊗ ∣Φ ⟩ B C K B ⊗ j = B ⨂ ∣Φ ⟩ S j K j , (20)
which is Eq. (7) : Alice’s entanglement has been transferred to the selected site with fidelity exactly one, and B C B_C B C has been swapped into the key register K B K_B K B .
Counting gates, Steps 1–3 contribute 3 n − 1 3n-1 3 n − 1 two-qudit gates, Step 4 two single-qudit gates, and Step 5 one single-qudit and one two-qudit gate: 3 n 3n 3 n two-qudit and three single-qudit gates in total, matching Eq. (5) and O ( n ) O(n) O ( n ) in the number of sites; the ( d , n ) = ( 3 , 2 ) (d,n)=(3,2) ( d , n ) = ( 3 , 2 ) case is unrolled explicitly in Supplementary Section S4, and the full protocol resource count is tabulated in Supplementary Section S5. The derivation uses no property of B A B_A B A , so it extends to an arbitrary partner of B C B_C B C . For every d ≥ 2 d\ge2 d ≥ 2 , n ≥ 1 n\ge1 n ≥ 1 , flat angle θ \theta θ , site B B B and pure state ∣ ψ ⟩ B C E |\psi\rangle_{B_CE} ∣ ψ ⟩ B C E with E E E of any dimension,
U d e c ( B ) U e n c ∣ ψ ⟩ B C E ⨂ j ∣ Φ ⟩ S j K j = ∣ ψ ⟩ S B E ⊗ ∣ Φ ⟩ B C K B ⊗ ⨂ j ≠ B ∣ Φ ⟩ S j K j . (21) U_{\mathrm{dec}}^{(B)}U_{\mathrm{enc}}|\psi\rangle_{B_CE}\bigotimes_j|\Phi\rangle_{S_jK_j}
=
|\psi\rangle_{S_BE}\otimes|\Phi\rangle_{B_CK_B}\otimes\bigotimes_{j\ne B}|\Phi\rangle_{S_jK_j}. \tag*{(21)} U dec ( B ) U enc ∣ ψ ⟩ B C E j ⨂ ∣Φ ⟩ S j K j = ∣ ψ ⟩ S B E ⊗ ∣Φ ⟩ B C K B ⊗ j = B ⨂ ∣Φ ⟩ S j K j . (21)
Steps 1–5 establish this for ∣ ψ ⟩ = ∣ Φ ⟩ B C B ′ |\psi\rangle=|\Phi\rangle_{B_CB'} ∣ ψ ⟩ = ∣Φ ⟩ B C B ′ . A general ∣ ψ ⟩ B C E |\psi\rangle_{B_CE} ∣ ψ ⟩ B C E equals d ( I ⊗ V ) ∣ Φ ⟩ B C B ′ \sqrt{d}(\mathbb{I}\otimes V)|\Phi\rangle_{B_CB'} d ( I ⊗ V ) ∣Φ ⟩ B C B ′ for a linear map V : B ′ → E V:B'\to E V : B ′ → E , which commutes with the encoder and decoder because they never act on B ′ B' B ′ (Supplementary Section S7). Eq. (7) also settles key reuse: the keys are left in Bell pairs with their own signal qudits and retain no correlation with B A B_A B A , so a second site applying the same decoder obtains entanglement fidelity 1 / d 2 1/d^2 1/ d 2 .
Deterministic non-local CSUMk ^k k
We use the labels of the main text: data qudits D A D_A D A (Alice) and D B D_B D B (Bob), and a resource pair ∣ Φ ⟩ B A S B |\Phi\rangle_{B_AS_B} ∣Φ ⟩ B A S B . Alice applies CSUMk 1 ( D A → B A ) ^{k_1}(D_A\to B_A) k 1 ( D A → B A ) and Bob applies CSUBk 2 ( S B → D B ) ^{k_2}(S_B\to D_B) k 2 ( S B → D B ) . Alice then measures B A B_A B A in the computational basis (outcome m A m_A m A ), and Bob applies F F F to S B S_B S B and measures it in the computational basis (outcome m B m_B m B ). For all k 1 , k 2 ∈ Z d k_1,k_2\in\mathbb{Z}_d k 1 , k 2 ∈ Z d and every branch ( m A , m B ) (m_A,m_B) ( m A , m B ) , the corrected branch operator on D A D B D_AD_B D A D B is
M m A , m B : = ( Z d k 1 m B ⊗ X d k 2 m A ) ( ⟨ m A ∣ B A ⟨ m B ∣ S B F S B ) ( CSUB S B → D B k 2 CSUM D A → B A k 1 ∣ Φ ⟩ B A S B ) = ω m A m B d CSUM D A → D B k 1 k 2 . (22) M_{m_A,m_B}:=\left(Z_d^{k_1m_B}\otimes X_d^{k_2m_A}\right)\left(\langle m_A|_{B_A}\langle m_B|_{S_B}F_{S_B}\right)
\left(\operatorname{CSUB}_{S_B\to D_B}^{k_2}\operatorname{CSUM}_{D_A\to B_A}^{k_1}|\Phi\rangle_{B_AS_B}\right)
=\frac{\omega^{m_Am_B}}{d}\operatorname{CSUM}_{D_A\to D_B}^{k_1k_2}. \tag*{(22)} M m A , m B := ( Z d k 1 m B ⊗ X d k 2 m A ) ( ⟨ m A ∣ B A ⟨ m B ∣ S B F S B ) ( CSUB S B → D B k 2 CSUM D A → B A k 1 ∣Φ ⟩ B A S B ) = d ω m A m B CSUM D A → D B k 1 k 2 . (22)
To see this, track a basis input ∣ c ⟩ D A ∣ t ⟩ D B |c\rangle_{D_A}|t\rangle_{D_B} ∣ c ⟩ D A ∣ t ⟩ D B ; by linearity this fixes the operator. After the two gates the state is d − 1 / 2 ∑ j ∣ c ⟩ D A ∣ t + k 1 c ⟩ B A ∣ j ⟩ S B ∣ t − k 2 j ⟩ D B d^{-1/2}\sum_j|c\rangle_{D_A}|t+k_1c\rangle_{B_A}|j\rangle_{S_B}|t-k_2j\rangle_{D_B} d − 1/2 ∑ j ∣ c ⟩ D A ∣ t + k 1 c ⟩ B A ∣ j ⟩ S B ∣ t − k 2 j ⟩ D B . The outcome m A m_A m A fixes j = m A − k 1 c j=m_A-k_1c j = m A − k 1 c , which leaves D B D_B D B in ∣ t − k 2 m A + k 1 k 2 c ⟩ |t-k_2m_A+k_1k_2c\rangle ∣ t − k 2 m A + k 1 k 2 c ⟩ and S B S_B S B in ∣ m A − k 1 c ⟩ |m_A-k_1c\rangle ∣ m A − k 1 c ⟩ . Measuring S B S_B S B after F F F gives each m B m_B m B with probability 1 / d 1/d 1/ d and the phase ω ( m A − k 1 c ) m B \omega^{(m_A-k_1c)m_B} ω ( m A − k 1 c ) m B . The corrections X d k 2 m A X_d^{k_2m_A} X d k 2 m A on D B D_B D B and Z d k 1 m B Z_d^{k_1m_B} Z d k 1 m B on D A D_A D A remove the m A m_A m A shift and the c c c -dependent phase, and leave
∣ c ⟩ D A ∣ t + k c ⟩ D B , k ≡ k 1 k 2 ( m o d d ) , (23) |c\rangle_{D_A}|t+k c\rangle_{D_B},\qquad k\equiv k_1k_2\pmod d, \tag*{(23)} ∣ c ⟩ D A ∣ t + k c ⟩ D B , k ≡ k 1 k 2 ( mod d ) , (23)
up to the global phase ω m A m B \omega^{m_Am_B} ω m A m B , in each of the d 2 d^2 d 2 equiprobable branches. The amplitude is 1 / d 1/d 1/ d for every ( c , t ) (c,t) ( c , t ) , which establishes Eq. (22) and hence Eq. (8) . The protocol of the main text is k 1 = k 2 = 1 k_1=k_2=1 k 1 = k 2 = 1 , with corrections X d m A X_d^{m_A} X d m A and Z d m B Z_d^{m_B} Z d m B .
Altered classical messages. Because the corrections are applied after the measurements and act on D A D_A D A and D B D_B D B directly, an alteration of the transcript composes with Eq. (22) from the left. If Bob receives m A + δ A m_A+\delta_A m A + δ A and Alice receives m B + δ B m_B+\delta_B m B + δ B , branch ( m A , m B ) (m_A,m_B) ( m A , m B ) implements
( Z d k 1 δ B ⊗ X d k 2 δ A ) CSUM k ( D A → D B ) (24) \left(Z_d^{k_1\delta_B}\otimes X_d^{k_2\delta_A}\right)\operatorname{CSUM}^{k}(D_A\to D_B) \tag*{(24)} ( Z d k 1 δ B ⊗ X d k 2 δ A ) CSUM k ( D A → D B ) (24)
exactly, still with probability 1 / d 2 1/d^2 1/ d 2 and for every input. The transcript is uniform and independent of the data, so a party controlling the channel learns nothing from it, and the most it can do is append a Weyl error of its choosing, determined by ( δ A , δ B ) (\delta_A,\delta_B) ( δ A , δ B ) . That error is undone if the true transcript is later authenticated. Authenticating the two dits reduces the adversarial case to Eq. (6) .
Only the product k 1 k 2 k_1k_2 k 1 k 2 matters, so CSUM is obtained whenever k 1 k 2 ≡ 1 k_1k_2\equiv1 k 1 k 2 ≡ 1 : Bob may apply CSUM ( S B → D B ) \operatorname{CSUM}(S_B\to D_B) CSUM ( S B → D B ) (k 2 = − 1 k_2=-1 k 2 = − 1 ) provided Alice applies CSUB ( D A → B A ) \operatorname{CSUB}(D_A\to B_A) CSUB ( D A → B A ) (k 1 = − 1 k_1=-1 k 1 = − 1 ) and the corrections are X d − m A X_d^{-m_A} X d − m A and Z d − m B Z_d^{-m_B} Z d − m B . What fails is the literal transcription of the qubit circuit, in which both parties apply CSUM (k 1 = 1 k_1=1 k 1 = 1 , k 2 = − 1 k_2=-1 k 2 = − 1 ): it teleports CSUB, not CSUM. Applying the qubit correction X d m A X_d^{m_A} X d m A anyway leaves D B D_B D B in ∣ t − c + 2 m A ⟩ \lvert t-c+2m_A\rangle ∣ t − c + 2 m A ⟩ , displaced from the target ∣ t + c ⟩ \lvert t+c\rangle ∣ t + c ⟩ by $2(m_A-c)$. This shift depends on the logical value c c c , which is in general an unknown superposition, and no function of ( m A , m B ) (m_A,m_B) ( m A , m B ) can supply it without measuring D A D_A D A . At d = 2 d=2 d = 2 the displacement is 0 for every c c c , which is why the inversion is invisible for qubits. For d > 2 d>2 d > 2 it vanishes only when m A ≡ c ( m o d d / ν ( d ) ) m_A\equiv c\pmod{d/\nu(d)} m A ≡ c ( mod d / ν ( d )) , i.e. on a fraction ν ( d ) / d \nu(d)/d ν ( d ) / d of (input, branch) pairs: 1/3 at d = 3 d=3 d = 3 , 1/2 at d = 4 d=4 d = 4 , 1/5 at $d=5 (Figure 3 c, Supplementary Section S12). Which branches succeed depends on the unknown c c c , so the parties can neither identify nor post-select them; the remedy is simply to fix the signs so that k 1 k 2 ≡ 1 k_1k_2\equiv1 k 1 k 2 ≡ 1 . The complete analysis, including the resource count, is given in Supplementary § 1.6.11 .
Hiding the power of CSUM
With k 2 = 1 k_2=1 k 2 = 1 , Bob’s gate CSUB ( S B → D B ) \operatorname{CSUB}(S_B\to D_B) CSUB ( S B → D B ) , his Fourier-basis measurement and his correction X d m A X_d^{m_A} X d m A are the same for every k k k , and the encoder, the decoder and the key register never involve k k k , which enters only through Alice’s local gate CSUM k ( D A → B A ) \operatorname{CSUM}^{k}(D_A\to B_A) CSUM k ( D A → B A ) and her correction Z d k m B Z_d^{k m_B} Z d k m B . The public outcomes are uniform for every k k k : in the derivation above m A = j + k 1 c m_A=j+k_1c m A = j + k 1 c with j j j uniform over Z d \mathbb{Z}_d Z d , and m B m_B m B is uniform because S B S_B S B is in a computational-basis state when it is measured in the Fourier basis. The joint distribution of everything that Bob, the provider or an unselected site receives from the protocol is therefore independent of k k k , and Alice may fix k k k at any time before her measurement. What the protocol cannot remove is the action of the gate itself on Bob’s data qudit. For a product input, with Alice’s computational-basis distribution p ( c ) = ⟨ c ∣ ρ D A ∣ c ⟩ p(c)=\langle c\vert\rho_{D_A}\vert c\rangle p ( c ) = ⟨ c ∣ ρ D A ∣ c ⟩ , the gate acts on D B D_B D B as the channel
ρ ↦ ∑ c p ( c ) X d k c ρ X d − k c , \rho\mapsto\sum_c p(c)X_d^{kc}\rho X_d^{-kc}, ρ ↦ c ∑ p ( c ) X d k c ρ X d − k c ,
whatever Bob prepares. Even with D B D_B D B entangled with a reference he keeps, the output is a mixture of the orthogonal states ( X d k c ⊗ I ) ∣ Φ ⟩ (X_d^{kc}\otimes I)\lvert\Phi\rangle ( X d k c ⊗ I ) ∣ Φ ⟩ , so the most Bob can extract is a single sample of the product k c kc k c , drawn from Alice’s input distribution. Because that input is in general unknown to him, this sample does not separate k k k from c c c . If p ( c ) p(c) p ( c ) is uniform, as for D A D_A D A in any Fourier-basis state or maximally entangled with an external system, and gcd ( k , d ) = 1 \gcd(k,d)=1 g cd( k , d ) = 1 , the sample is uniform and D B D_B D B ’s output is independent of k k k altogether. In general, repeated use can reveal at most which subgroup gcd ( k , d ) Z d \gcd(k,d)\mathbb{Z}_d g cd( k , d ) Z d the products k c kc k c fall in. This is exactly the leakage of an ideal local CSUM k \operatorname{CSUM}^{k} CSUM k , so the protocol hides k k k as well as any implementation of the gate could.
Derivation of the exact gate and timing of Alice’s operation
Purification. Let ∣ ρ ⟩ D A D B R R ′ \lvert\rho\rangle_{D_A D_B R R'} ∣ ρ ⟩ D A D B R R ′ purify ρ \rho ρ with an additional reference R ′ R' R ′ . Every operation of the protocol acts as the identity on R ′ R' R ′ , so it suffices to show that each branch maps ∣ ρ ⟩ \lvert\rho\rangle ∣ ρ ⟩ to ω m A m B d ( CSUM k ⊗ I R R ′ ) ∣ ρ ⟩ \frac{\omega^{m_A m_B}}{d}(\operatorname{CSUM}^{k}\otimes I_{R R'})\lvert\rho\rangle d ω m A m B ( CSUM k ⊗ I R R ′ ) ∣ ρ ⟩ times a fixed state of the remaining registers; tracing out R ′ R' R ′ then gives Eq. (6) for ρ \rho ρ .
Recovery. Before Stage 3, the registers D A , D B , R , R ′ D_A,D_B,R,R' D A , D B , R , R ′ have not been acted on, and the joint state is ∣ ρ ⟩ ⊗ U d e c ( B ) U e n c ∣ Φ ⟩ B A B C ⊗ ⨂ j ∣ Φ ⟩ S j K j \lvert\rho\rangle\otimes U_{\mathrm{dec}}^{(B)}U_{\mathrm{enc}}\lvert\Phi\rangle_{B_A B_C}\otimes\bigotimes_j\lvert\Phi\rangle_{S_jK_j} ∣ ρ ⟩ ⊗ U dec ( B ) U enc ∣ Φ ⟩ B A B C ⊗ ⨂ j ∣ Φ ⟩ S j K j . Eq. (21) with E = B A E=B_A E = B A turns the second factor into ∣ Φ ⟩ B A S B ⊗ ∣ Φ ⟩ B C K B ⊗ ⨂ j ≠ B ∣ Φ ⟩ S j K j \lvert\Phi\rangle_{B_A S_B}\otimes\lvert\Phi\rangle_{B_C K_B}\otimes\bigotimes_{j\ne B}\lvert\Phi\rangle_{S_jK_j} ∣ Φ ⟩ B A S B ⊗ ∣ Φ ⟩ B C K B ⊗ ⨂ j = B ∣ Φ ⟩ S j K j , a fixed state independent of ρ \rho ρ .
Teleportation. Stage 3 acts only on D A , D B , B A , S B D_A,D_B,B_A,S_B D A , D B , B A , S B . By Eq. (22) , branch ( m A , m B ) (m_A,m_B) ( m A , m B ) applies M m A , m B ⊗ I R R ′ = ω m A m B d CSUM k ⊗ I R R ′ M_{m_A,m_B}\otimes I_{R R'}=\frac{\omega^{m_A m_B}}{d}\operatorname{CSUM}^{k}\otimes I_{R R'} M m A , m B ⊗ I R R ′ = d ω m A m B CSUM k ⊗ I R R ′ to ∣ ρ ⟩ \lvert\rho\rangle ∣ ρ ⟩ and leaves the other registers in ∣ Φ ⟩ B C K B ⊗ ⨂ j ≠ B ∣ Φ ⟩ S j K j \lvert\Phi\rangle_{B_C K_B}\otimes\bigotimes_{j\ne B}\lvert\Phi\rangle_{S_jK_j} ∣ Φ ⟩ B C K B ⊗ ⨂ j = B ∣ Φ ⟩ S j K j . The branch probability is ∥ M m A , m B ∣ ρ ⟩ ∥ 2 = 1 / d 2 \lVert M_{m_A,m_B}\lvert\rho\rangle\rVert^2=1/d^2 ∥ M m A , m B ∣ ρ ⟩ ∥ 2 = 1/ d 2 because CSUM k \operatorname{CSUM}^{k} CSUM k is unitary, and the normalized output is ( CSUM k ⊗ I ) ∣ ρ ⟩ (\operatorname{CSUM}^{k}\otimes I)\lvert\rho\rangle ( CSUM k ⊗ I ) ∣ ρ ⟩ , up to a global phase, which is Eq. (6) . Eq. (22) is an operator identity on D A D B D_A D_B D A D B , so it applies unchanged to inputs entangled with R ′ R' R ′ ; nothing about R R R is used beyond the fact that the protocol does not act on it.
Timing. Alice’s gate CSUM k 1 ( D A → B A ) \operatorname{CSUM}^{k_1}(D_A\to B_A) CSUM k 1 ( D A → B A ) acts only on ( D A , B A ) (D_A,B_A) ( D A , B A ) , while the encoder and decoder act only on ( B C , { S j , K j } ) (B_C,\{S_j,K_j\}) ( B C , { S j , K j }) . The two commute, so Alice may apply her gate in Stage 1, before encoding, or in Stage 3, as in the main text, with the same joint state; in the first case Eq. (21) is applied with E = B A D A D B R R ′ E=B_A D_A D_B R R' E = B A D A D B R R ′ . The same disjointness carries the privacy and single-use guarantees over to the composed protocol unchanged. It also makes m A m_A m A and m B m_B m B safe to announce openly: each is a function of a register the encoder never touched, or of S B S_B S B after decoding, so neither carries information about any unselected share.
Practical consequences of the decoder
Reuse of entanglement. Only the provider’s original pair is consumed. After decoding, every unselected pair ( S j , K j ) (S_j,K_j) ( S j , K j ) is back in ∣ Φ ⟩ |\Phi\rangle ∣Φ ⟩ , now shared between site j j j and the recipient, and ( B C , K B ) (B_C,K_B) ( B C , K B ) is a fresh Bell pair between provider and recipient (7) ; all n n n can be reused in later rounds. The key is single-use, but the entanglement is not.
Replacing lost keys. For j ≠ B j \ne B j = B the decoder uses K j K_j K j only as the partner of S j S_j S j in a Bell pair, so a key lost in transit can be replaced by the share S j S_j S j of a site that has declined the gate, provided it is still coherent. Because the two halves of a Bell pair are interchangeable only up to a transpose, the decoder must be adjusted (Supplementary Section S7); with the adjustment recovery is exact, and without it the fidelity falls to 1 / d 2 1/d^2 1/ d 2 .
Late binding of the data. Alice’s entangling gate commutes with the encoder and decoder, which act on disjoint registers, so she may apply it at the last moment, and D A D_A D A need not exist when the resource is distributed.
Each exposed register is acted on, between encoding and decoding, by D p ( ρ ) = ( 1 − p ) ρ + p I Tr [ ρ ] / d = ( 1 − p + p / d 2 ) ρ + ( p / d 2 ) ∑ ( a , b ) ≠ ( 0 , 0 ) W a , b ρ W a , b † \mathcal{D}_p(\rho)=(1-p)\rho+p\mathbb{I}\operatorname{Tr}[\rho]/d=(1-p+p/d^2)\rho+(p/d^2)\sum_{(a,b)\ne(0,0)}W_{a,b}\rho W_{a,b}^{\dagger} D p ( ρ ) = ( 1 − p ) ρ + p I Tr [ ρ ] / d = ( 1 − p + p / d 2 ) ρ + ( p / d 2 ) ∑ ( a , b ) = ( 0 , 0 ) W a , b ρ W a , b † . The figure of merit is the process fidelity of the completed gate: D A D_A D A and D B D_B D B start maximally entangled with reference qudits R A , R B R_A,R_B R A , R B (unrelated to the reference R R R of Eq. (6) ), and the corrected output of each branch is compared, via the Uhlmann–Jozsa fidelity, with the Choi state ∣ C ⟩ = ( I R A R B ⊗ C S U M ) ∣ Φ ⟩ R A D A ∣ Φ ⟩ D B R B |C\rangle=(\mathbb{I}_{R_AR_B}\otimes C_{\mathrm{SUM}})|\Phi\rangle_{R_AD_A}|\Phi\rangle_{D_BR_B} ∣ C ⟩ = ( I R A R B ⊗ C SUM ) ∣Φ ⟩ R A D A ∣Φ ⟩ D B R B ,
F = ∑ m A , m B P ( m A , m B ) ⟨ C ∣ ρ R A D A D B R B ( m A , m B ) ∣ C ⟩ (25) \mathcal{F}=\sum_{m_A,m_B}P(m_A,m_B)\langle C|\rho_{R_AD_AD_BR_B}^{(m_A,m_B)}|C\rangle \tag*{(25)} F = m A , m B ∑ P ( m A , m B ) ⟨ C ∣ ρ R A D A D B R B ( m A , m B ) ∣ C ⟩ (25)
Unlike a fidelity averaged over computational-basis inputs, this is sensitive to phase errors, and a fully random gate scores 1 / d 2 1/d^2 1/ d 2 rather than 1 / d 1/d 1/ d . Because F \mathcal{F} F is linear in each channel, noise on m m m registers gives F ( p ) = ∑ T ( 1 − p ) m − ∣ T ∣ p ∣ T ∣ F T \mathcal{F}(p)=\sum_{T}(1-p)^{m-|T|}p^{|T|}\mathcal{F}_T F ( p ) = ∑ T ( 1 − p ) m − ∣ T ∣ p ∣ T ∣ F T , where F T \mathcal{F}_T F T is the fidelity with the registers in T T T fully depolarized, i.e. averaged uniformly over the d 2 ∣ T ∣ d^{2|T|} d 2∣ T ∣ Weyl errors. Each error is propagated as a pure state, so every curve in Figure 4 is exact. For a single register F T \mathcal{F}_T F T is 1 / d 2 1/d^2 1/ d 2 for S B S_B S B and every K j ≠ B K_{j\ne B} K j = B (each non-trivial Weyl error there reaches the data as a non-trivial Weyl error), 1 1 1 for every S j ≠ B S_{j\ne B} S j = B , and gcd ( n − η 2 , d ) / d 2 \gcd(n-\eta^2,d)/d^2 g cd( n − η 2 , d ) / d 2 for K B K_B K B , with η \eta η as defined in the main text. The released key is special because the decoder folds the other n − 1 n-1 n − 1 key labels into it: a Weyl error W a , b W_{a,b} W a , b on K B K_B K B is harmless exactly when a ≡ η b a\equiv\eta b a ≡ η b and ( n − η 2 ) b ≡ 0 (n-\eta^2)b\equiv0 ( n − η 2 ) b ≡ 0 , which has gcd ( n − η 2 , d ) \gcd(n-\eta^2,d) g cd( n − η 2 , d ) solutions. This is the same arithmetic condition that governs the all-shares coalition. It equals 1 / d 2 1/d^2 1/ d 2 for every case in Figure 4 (n = 2 n=2 n = 2 , d ≤ 4 d\le4 d ≤ 4 ), but not, for example, at ( d , n ) = ( 2 , 3 ) (d,n)=(2,3) ( d , n ) = ( 2 , 3 ) or ( 7 , 2 ) (7,2) ( 7 , 2 ) . Eq. (9) and its two-register analog follow from these values. We verified them in the full protocol for d = 2 d=2 d = 2 (n ≤ 5 n\le5 n ≤ 5 ), d = 3 d=3 d = 3 (n ≤ 4 n\le4 n ≤ 4 ), d = 4 d=4 d = 4 (n ≤ 3 n\le3 n ≤ 3 ) and ( d , n ) = ( 5 , 2 ) (d,n)=(5,2) ( d , n ) = ( 5 , 2 ) , and the K B K_B K B rule additionally in the cloning layer for d = 5 , 6 , 7 d=5,6,7 d = 5 , 6 , 7 with n ≤ 4 n\le4 n ≤ 4 where feasible, including cases with gcd ( n − η 2 , d ) > 1 \gcd(n-\eta^2,d)>1 g cd( n − η 2 , d ) > 1 .
Which registers matter. After encoding, no operation of the protocol acts on an unselected share S j ≠ B S_{j\ne B} S j = B , so noise on it commutes with everything that follows and disappears when it is traced out; a provider need not protect the n − 1 n-1 n − 1 shares that will not be redeemed. The keys K j ≠ B K_{j\ne B} K j = B , by contrast, are decoder inputs, and depolarizing any one of them is as damaging as depolarizing S B S_B S B . If a lost key has been replaced by an unselected share (below), that share joins the protected set. For two registers, of the d 4 d^4 d 4 joint Weyl errors on ( S B , K B ) (S_B,K_B) ( S B , K B ) , d 2 d^2 d 2 cancel exactly in the decoder, giving F ( p ) = 1 − ( 1 − 1 / d 2 ) [ 1 − ( 1 − p ) 2 ] \mathcal{F}(p)=1-(1-1/d^2)[1-(1-p)^2] F ( p ) = 1 − ( 1 − 1/ d 2 ) [ 1 − ( 1 − p ) 2 ] with floor 1 / d 2 1/d^2 1/ d 2 . For general sets of noisy registers, multilinearity makes the first-order term additive, 1 − F ≃ p ∑ j ( 1 − F j ) 1-\mathcal{F}\simeq p\sum_j(1-\mathcal{F}_j) 1 − F ≃ p ∑ j ( 1 − F j ) , so exposing S B S_B S B and all n n n keys gives 1 − F ≃ [ n ( 1 − 1 / d 2 ) + 1 − gcd ( n − η 2 , d ) / d 2 ] p 1-\mathcal{F}\simeq[n(1-1/d^2)+1-\gcd(n-\eta^2,d)/d^2]p 1 − F ≃ [ n ( 1 − 1/ d 2 ) + 1 − g cd( n − η 2 , d ) / d 2 ] p at small p p p , which is ( n + 1 ) ( 1 − 1 / d 2 ) p (n+1)(1-1/d^2)p ( n + 1 ) ( 1 − 1/ d 2 ) p when the gcd is 1 1 1 . The pair floor of 1 / d 2 1/d^2 1/ d 2 for ( S B , K B ) (S_B,K_B) ( S B , K B ) holds for every ( d , n ) (d,n) ( d , n ) we tested, including those with gcd ( n − η 2 , d ) > 1 \gcd(n-\eta^2,d)>1 g cd( n − η 2 , d ) > 1 . Gates and measurements are ideal throughout this analysis.
Independent numerical cross-check
Every claim above is proved in closed form. As a check against algebraic slips we also evaluated the construction numerically with an exact state-vector implementation in NumPy, in which every gate, measurement and correction acts directly on the qudit tensor. The checks reproduce, to < 10 − 9 <10^{-9} < 1 0 − 9 and typically ∼ 10 − 15 \sim10^{-15} ∼ 1 0 − 15 : Eqs. (14) , (13) , (18) and (19) ; the flat-angle lattice, including a fine scan of θ \theta θ ; exact recovery for every site B B B at d ≤ 9 d\le9 d ≤ 9 (n = 2 n=2 n = 2 ), d ≤ 6 d\le6 d ≤ 6 (n = 3 n=3 n = 3 ) and d ≤ 4 d\le4 d ≤ 4 (n = 4 n=4 n = 4 ), and on every flat angle for d ≤ 9 d\le9 d ≤ 9 , with the post-decoding state Eq. (7) ; privacy of every proper coalition (d ≤ 7 d\le7 d ≤ 7 , n ≤ 4 n\le4 n ≤ 4 ); the full ( 2 n + 4 ) (2n+4) ( 2 n + 4 ) -qudit protocol, including measurements and feed-forward, in every branch and for every power k k k ; the success fraction of the literal qubit transcription, Eq. (36) ; the access structure of Supplementary Section S9, including the symplectic map Eq. (32) , the stabilizer group and the counting formula Eq. (34) ; key replacement; the k k k -independence of the public outcomes; the effect of altered classical messages, Eq. (24) , for every ( k 1 , k 2 ) (k_1,k_2) ( k 1 , k 2 ) and alteration at d ≤ 5 d\le5 d ≤ 5 ; and the noise coefficients F T \mathcal{F}_T F T , including the first-order scaling with the number of exposed keys.
Use of AI tools. Large language models (Claude Opus, Anthropic) were used to help write and refactor the numerical code and to draft the manuscript text. After using this tool, the author reviewed and edited the content as needed and take full responsibility for the content of the publication. Access to the code is described under Code availability.
Data availability
No experimental datasets were generated or analyzed during this study. The results reported here are analytic and follow from the constructions given in Methods; the verification outputs quoted as cross-checks (truth-table fidelities, coalition trace distances, and the angle scan) are generated by the code described below and are available on the same terms (Code availability).
Code availability
The code used for the numerical cross-checks is not publicly released because the protocol is the subject of a pending patent application [47 ] (Competing interests). It is registered as FERMILAB-CODE-2026-45-SQMS and will be made available to editors and referees for peer review, and to other researchers under a research license on request to the author or to Fermilab’s Office of Partnerships and Technology Transfer (OPTT@fnal.gov). It consists of a NumPy module implementing the encoder Eq. (2) , the closed-form decoder Eq. (5) , the full protocol and the noise model as exact state-vector operations, and two Jupyter notebooks with their outputs. The first checks every quantitative statement of this paper with an explicit assertion. The second regenerates Figure 3 , Figure 4 and Figure S2 , so every plotted quantity is recomputed.
Figure S2. Privacy, recovery and the access structure of the shared gate. a, Mutual information I ( R : A ) I(R:A) I ( R : A ) with the reference R R R , in units of its maximum 2 log d 2\log d 2 log d , as a function of the encoder’s phase angle θ \theta θ , for a single unselected share (A = S j A=S_j A = S j , dashed) and for the recipient’s set (A = { S B , K 1 , K 2 } A=\{S_B,K_1,K_2\} A = { S B , K 1 , K 2 } , solid), at d = 3 d=3 d = 3 , n = 2 n=2 n = 2 . The share carries nothing at every angle; the recipient recovers the reference exactly only on the flat angles (open circles), the smallest of which is θ ∗ \theta^* θ ∗ . b, I ( R : A ) I(R:A) I ( R : A ) for every subset A A A of the five registers { B C , S 1 , S 2 , K 1 , K 2 } \{B_C,S_1,S_2,K_1,K_2\} { B C , S 1 , S 2 , K 1 , K 2 } (n = 2 n=2 n = 2 ), grouped by size ∣ A ∣ |A| ∣ A ∣ ; numbers give how many subsets share each value, and marker area scales with that count. At d = 3 d=3 d = 3 , θ ∗ \theta^* θ ∗ (circles) and at d = 7 d=7 d = 7 , θ = 4 π / 7 \theta=4\pi/7 θ = 4 π /7 (triangles) every set of at most two registers holds nothing and every set of at least three holds everything: a ( ( 3 , 5 ) ) ((3,5)) (( 3 , 5 )) threshold scheme. At d = 7 d=7 d = 7 , θ ∗ \theta^* θ ∗ (squares) six sets hold exactly half, as predicted by Eq. (31) . c, Fidelity obtained by a second site attempting to decrypt with the spent key: state fidelity 1 / d 1/d 1/ d and entanglement fidelity 1 / d 2 1/d^2 1/ d 2 , both the values of a random guess (black bars), against fidelity exactly 1 for the legitimate recipient.
Acknowledgments
We acknowledge helpful feedback from Henry Lamm. This work was supported by the U.S. Department of Energy, Office of Science, National Quantum Information Science Research Centers, Superconducting Quantum Materials and Systems Center (SQMS), under Contract No. 89243024CSC000002. Fermilab is operated by Fermi Forward Discovery Group, LLC under Contract No. 89243024CSC000002 with the U.S. Department of Energy, Office of Science, Office of High Energy Physics.
Author contributions
T.R. conceived the project, developed the theory, carried out the analytical derivations and numerical verification, and wrote the manuscript.
Competing interests
T.R. is the inventor on a pending patent application covering the protocol described in this work, U.S. Provisional Patent Application No. 64/182,988 [47 ] , filed by Fermi Forward Discovery Group, LLC.
References
Extended qudit gate identities
Beyond the commutation relation Eq. (10) (Methods) and the Weyl orthogonality relation Eq. (11) (Methods) used throughout the main text, the following identities are used implicitly at various points and are collected here for reference. With W p , q = X d p Z d q W_{p,q}=X_d^pZ_d^q W p , q = X d p Z d q , the Weyl operators compose as
W p , q W p ′ , q ′ = ω q p ′ W p + p ′ , q + q ′ (26) W_{p,q}\,W_{p',q'}=\omega^{qp'}W_{p+p',q+q'} \tag*{(26)} W p , q W p ′ , q ′ = ω q p ′ W p + p ′ , q + q ′ (26)
obtained by commuting Z d q Z_d^q Z d q through X d p ′ X_d^{p'} X d p ′ using Eq. (10) (Methods). The Fourier transform exchanges the two generators in both directions, F † Z d F = X d F^\dagger Z_dF=X_d F † Z d F = X d and F X d F † = Z d FX_dF^\dagger=Z_d F X d F † = Z d , consistent with F 4 = I F^4=I F 4 = I . For any operator M M M on a single qudit and the maximally entangled state ∣ Φ ⟩ = d − 1 / 2 ∑ j ∣ j j ⟩ |\Phi\rangle=d^{-1/2}\sum_j|jj\rangle ∣Φ ⟩ = d − 1/2 ∑ j ∣ j j ⟩ , the transpose trick ( M ⊗ I ) ∣ Φ ⟩ = ( I ⊗ M T ) ∣ Φ ⟩ (M\otimes I)|\Phi\rangle=(I\otimes M^T)|\Phi\rangle ( M ⊗ I ) ∣Φ ⟩ = ( I ⊗ M T ) ∣Φ ⟩ holds in the computational basis, together with ⟨ Φ ∣ ( M ⊗ I ) ∣ Φ ⟩ = Tr [ M ] / d \langle\Phi|(M\otimes I)|\Phi\rangle=\operatorname{Tr}[M]/d ⟨ Φ∣ ( M ⊗ I ) ∣Φ ⟩ = Tr [ M ] / d ; both are used repeatedly when moving an operator from one half of a Bell pair to the other, as in the extension of exact recovery to arbitrary inputs (Methods, Eq. (21) , and Section S7). Finally, C S U B : = C S U M − 1 = C S U M † \mathrm{CSUB}:=\mathrm{CSUM}^{-1}=\mathrm{CSUM}^{\dagger} CSUB := CSUM − 1 = CSUM † since CSUM is a computational-basis permutation matrix (real and orthogonal), so inverting it costs nothing beyond relabelling the gate.
Why the qubit encoder does not generalize algebraically
For d = 2 d=2 d = 2 , X d 2 = Z d 2 = I X_d^2=Z_d^2=I X d 2 = Z d 2 = I , so the Weyl group collapses to just four distinct operators, { I , X , Z , X Z } \{I,X,Z,XZ\} { I , X , Z , X Z } (with X Z ∝ Y XZ\propto Y X Z ∝ Y up to phase), and every one of them is again a single-qubit Pauli string. This is what lets the qubit encoder be written compactly as a product of two exponentials of Pauli strings: any diagonal phase built from squares of sums of Z Z Z ’s expands, using Z 2 = I Z^2=I Z 2 = I , into a sum of purely pairwise Z i Z j Z_iZ_j Z i Z j terms with fixed coefficients, which exponentiates directly into a product of two-qubit phase gates. For d > 2 d>2 d > 2 , X d d = Z d d = I X_d^d=Z_d^d=I X d d = Z d d = I but no smaller power vanishes, so { W p , q } p , q ∈ Z d \{W_{p,q}\}_{p,q\in\mathbb{Z}_d} { W p , q } p , q ∈ Z d is a full set of d 2 d^2 d 2 distinct operators rather than a group of order 4: for d = 3 d=3 d = 3 , for instance, direct computation confirms X d Z d X_dZ_d X d Z d is neither proportional to I I I , X d X_d X d , Z d Z_d Z d , X d 2 X_d^2 X d 2 , Z d 2 Z_d^2 Z d 2 nor to any other single generator, and the analogous expansion of e i θ T 2 e^{i\theta T^2} e i θ T 2 (with T T T a sum of d d d -level registers) does not collapse to pairwise terms with d d d -independent coefficients, because Z d 2 ≠ I Z_d^2\ne I Z d 2 = I leaves genuinely higher powers Z d 2 k Z_d^{2k} Z d 2 k in the expansion for k = 1 , … , ⌊ ( d − 1 ) / 2 ⌋ k=1,\ldots,\lfloor(d-1)/2\rfloor k = 1 , … , ⌊( d − 1 ) /2 ⌋ . This is why the construction here abandons the two-exponential route entirely in favor of the explicit CSUM-ladder circuit of Eq. (2) , whose cost is O ( n ) O(n) O ( n ) for every d d d by direct construction rather than by algebraic collapse.
Full unrolled protocol circuit
Figure S1 unrolls the entire protocol into elementary qudit gates on all 2 n + 4 2n+4 2 n + 4 registers, with the encoder and decoder shown as the blocks U e n c U_{\mathrm{enc}} U enc and U d e c U_{\mathrm{dec}} U dec whose interiors are given explicitly in Figure 2 . The data qudits D A D_A D A and D B D_B D B are acted on by arbitrary unitaries U A U_A U A and U B U_B U B that entangle them with external systems, drawn as extra wires; these unitaries are never referenced by the protocol, which is what makes the gate act on arbitrary entangled inputs rather than only on product inputs. The unselected shares S j ≠ B S_{j\ne B} S j = B , the retained keys K j ≠ B K_{j\ne B} K j = B , and the provider’s copy B C B_C B C are discarded, marked by the bins: no part of the protocol conditions on them, and by Eq. (16) they carry nothing to condition on. The equivalence sign asserts that the whole left-hand side reduces to a single CSUM between D A D_A D A and D B D_B D B acting through the same U A , U B U_A,U_B U A , U B .
Figure S1. The protocol unrolled into elementary qudit gates. a, The complete deferred-recipient construction on all 2 n + 4 2n+4 2 n + 4 registers. Vertical wavy connectors joining two wires denote a Bell pair shared between those two registers, prepared by the F d F_d F d -then-CSUM gadget shown at the right: the provider’s pair ( B A , B C ) (B_A,B_C) ( B A , B C ) , and one pair ( S j , K j ) (S_j,K_j) ( S j , K j ) for each candidate site. The data qudits D A D_A D A and D B D_B D B enter through arbitrary unitaries U A U_A U A , U B U_B U B (red) that entangle them with external wires; the protocol never references those wires, which is why the gate is exact on arbitrarily entangled inputs. The provider encodes B C B_C B C into the shares with U e n c U_{\mathrm{enc}} U enc ; site B redeems the released key and applies U d e c U_{\mathrm{dec}} U dec ; B C B_C B C , the unselected shares and the retained keys are then discarded (bins). A computational-basis measurement on B A B_A B A and a Fourier-basis measurement on S B S_B S B return dits m A m_A m A , m B m_B m B , which drive Z d k 1 m B Z_d^{k_1m_B} Z d k 1 m B on D A D_A D A and X d k 2 m A X_d^{k_2m_A} X d k 2 m A on D B D_B D B . The result equals a bare CSUM ( D A → D B ) \operatorname{CSUM}(D_A\to D_B) CSUM ( D A → D B ) acting through the same U A U_A U A , U B U_B U B , in every one of the d 2 d^2 d 2 branches. b, The cloning layer alone: U d e c U_{\mathrm{dec}} U dec inverts U e n c U_{\mathrm{enc}} U enc on the redeeming site’s share and key, returning ∣ B C ⟩ \lvert B_C\rangle ∣ B C ⟩ exactly. c, The teleportation layer alone, with the resource pair supplied directly. Bob’s entangling gate is CSUB; running it as CSUM while keeping Alice’s gate and the corrections unchanged succeeds on only a ν ( d ) / d \nu(d)/d ν ( d ) / d fraction of entries (Section S12).
Panels b and c isolate the two restrictions discussed in the main text, and are analyzed in Sections S7 and § 1.6.11 . Panel b keeps only the cloning layer: U e n c U_{\mathrm{enc}} U enc followed by U d e c U_{\mathrm{dec}} U dec returns ∣ B C ⟩ |B_C\rangle ∣ B C ⟩ to the redeeming site exactly, which is the qudit encrypted-cloning statement. Panel c keeps only the teleportation layer, with the resource pair ( B A , B B ) (B_A,B_B) ( B A , B B ) supplied directly rather than recovered from a share. Bob’s entangling gate is CSUB( B B → D B ) (B_B\to D_B) ( B B → D B ) (k 1 = k 2 = 1 k_1=k_2=1 k 1 = k 2 = 1 in Eq. (8) ); the role of this sign is discussed in Section S12.
Decoder: worked example
The decoder derivation in Methods is complete as stated; this section gives the ( d , n ) = ( 3 , 2 ) (d,n)=(3,2) ( d , n ) = ( 3 , 2 ) case explicitly as a check on the index conventions. At d = 3 d=3 d = 3 (odd), ν ( 3 ) = gcd ( 2 , 3 ) = 1 \nu(3)=\operatorname{gcd}(2,3)=1 ν ( 3 ) = gcd ( 2 , 3 ) = 1 and ν i n v ( 3 ) = 2 \nu_{\mathrm{inv}}(3)=2 ν inv ( 3 ) = 2 ; θ ∗ = 2 π / 3 \theta^*=2\pi/3 θ ∗ = 2 π /3 , ν i n v − 1 = 2 m o d 3 \nu_{\mathrm{inv}}^{-1}=2\bmod3 ν inv − 1 = 2 mod 3 , and λ = ω 4 − 1 = ω \lambda=\omega^{4^{-1}}=\omega λ = ω 4 − 1 = ω since 4 ≡ 1 m o d 3 4\equiv1\bmod3 4 ≡ 1 mod 3 . Hence Λ = diag ( 1 , ω , ω ) \Lambda=\operatorname{diag}(1,\omega,\omega) Λ = diag ( 1 , ω , ω ) (from λ k 2 \lambda^{k^2} λ k 2 with k 2 = 0 , 1 , 1 k^2=0,1,1 k 2 = 0 , 1 , 1 ), R = diag - perm ∣ u ⟩ ↦ ∣ 2 u ⟩ R=\operatorname{diag}\text{-}\operatorname{perm}|u\rangle\mapsto|2u\rangle R = diag - perm ∣ u ⟩ ↦ ∣2 u ⟩ , i.e. the transposition ( 1 2 ) (1\,2) ( 1 2 ) , and Γ = P † R P † \Gamma=P^\dagger R P^\dagger Γ = P † R P † with P = diag ( 1 , ω , ω ) P=\operatorname{diag}(1,\omega,\omega) P = diag ( 1 , ω , ω ) . Taking B = 1 B=1 B = 1 , Eq. (5) unrolls to the six two-qudit gates
CSUB( K 1 → S 1 ) (K_1\to S_1) ( K 1 → S 1 ) , CSUM( S 1 → K 2 ) (S_1\to K_2) ( S 1 → K 2 ) , CSUM( K 2 → K 1 ) (K_2\to K_1) ( K 2 → K 1 ) ,
CSUB( S 1 → K 1 ) 2 (S_1\to K_1)^2 ( S 1 → K 1 ) 2 , CSUM( K 1 → S 1 ) (K_1\to S_1) ( K 1 → S 1 ) , Eq. (S2)
interleaved with Λ S 1 \Lambda_{S_1} Λ S 1 , Γ K 1 \Gamma_{K_1} Γ K 1 and J S 1 J_{S_1} J S 1 as in Eq. (5) , matching 3 n = 6 3n=6 3 n = 6 two-qudit and three single-qudit gates. Applying this to Eq. (17) reproduces Eq. (7) exactly. Gate-level circuit diagrams in the style of Figure 2 b for ( d , n ) = ( 2 , 2 ) (d,n)=(2,2) ( d , n ) = ( 2 , 2 ) and ( 3 , 2 ) (3,2) ( 3 , 2 ) follow mechanically from the same unrolling.
Resource count vs. n n n and d d d
The complete protocol (Figure S1 ) uses 2 n + 4 2n+4 2 n + 4 qudits: D A , D B , B A D_A,D_B,B_A D A , D B , B A , the encoded qudit (relabelled S 1 , … , S n S_1,\ldots,S_n S 1 , … , S n after encoding), and the n n n key qudits K 1 , … , K n K_1,\ldots,K_n K 1 , … , K n . Table S1 accounts for gate and measurement cost as a function of n n n . Every entry is independent of d d d , including the decoder: with Eq. (5) in hand its cost is a gate count, O ( n ) O(n) O ( n ) , rather than the d 2 ( n + 1 ) d^{2(n+1)} d 2 ( n + 1 ) Hilbert-space dimension on which a generic recovery map would have to be synthesised. The whole protocol is therefore linear in the number of sites and constant in the qudit dimension, so the only quantity that grows with d d d is the cost of implementing a single elementary qudit gate on hardware.
Table S1. Resource count vs. n n n , independent of d d d throughout. “CSUM-type” counts elementary CSUM or CSUB gates; “Fourier layers” counts applications of F ⊗ ( n + 1 ) F^{\otimes(n+1)} F ⊗ ( n + 1 ) or F − ⊗ ( n + 1 ) F^{-\otimes(n+1)} F −⊗ ( n + 1 ) .
The flat-angle lattice
Methods shows that at θ ⋆ ( d ) \theta^\star(d) θ ⋆ ( d ) the Gauss sum is the pure chirp g ^ ( p ) = G λ − p 2 \hat{g}(p)=G\lambda^{-p^2} g ^ ( p ) = G λ − p 2 , so ∣ g ^ ( p ) ∣ = d |\hat{g}(p)|=\sqrt{d} ∣ g ^ ( p ) ∣ = d is an identity rather than a coincidence of small d d d . Here we record which other angles share that property, since θ ⋆ \theta^\star θ ⋆ is the smallest flat angle but not the only one.
Write θ = π s / d \theta=\pi s/d θ = π s / d with s ∈ Z s\in\mathbb{Z} s ∈ Z , so that g ( k ) = e i π s k 2 / d g(k)=e^{i\pi sk^2/d} g ( k ) = e iπ s k 2 / d and g ^ ( p ) = ∑ k e i π ( s k 2 − 2 p k ) / d \hat{g}(p)=\sum_k e^{i\pi(sk^2-2pk)/d} g ^ ( p ) = ∑ k e iπ ( s k 2 − 2 p k ) / d . Re-indexing k ↦ k + d k\mapsto k+d k ↦ k + d leaves g g g invariant iff e i π s d = 1 e^{i\pi sd}=1 e iπ s d = 1 , i.e. iff s d sd s d is even. Given that, and gcd ( s , d ) = 1 \gcd(s,d)=1 g cd( s , d ) = 1 , the square can be completed. For even d d d , s s s is odd and hence invertible modulo 2 d 2d 2 d , with inverse s ˉ \bar{s} s ˉ , and s k 2 − 2 p k ≡ s ( k − s ˉ p ) 2 − s ˉ p 2 ( m o d 2 d ) sk^2-2pk\equiv s(k-\bar{s}p)^2-\bar{s}p^2\pmod{2d} s k 2 − 2 p k ≡ s ( k − s ˉ p ) 2 − s ˉ p 2 ( mod 2 d ) . For odd d d d , s s s is even and 2 s 2s 2 s is invertible modulo d d d , and the same step runs in Z d \mathbb{Z}_d Z d with ω \omega ω . Either way g ^ ( p ) = G s λ s − p 2 \hat{g}(p)=G_s\lambda_s^{-p^2} g ^ ( p ) = G s λ s − p 2 with G s = ∑ k e i π s k 2 / d G_s=\sum_k e^{i\pi sk^2/d} G s = ∑ k e iπ s k 2 / d and λ s \lambda_s λ s as in Eq. (15) . Note that λ s ≠ e i π s / d \lambda_s\ne e^{i\pi s/d} λ s = e iπ s / d in general: at θ ⋆ \theta^\star θ ⋆ for odd d d d , λ = ω 4 = 1 \lambda=\omega^4=1 λ = ω 4 = 1 , not ω \omega ω . The classical evaluation of the quadratic Gauss sum gives ∣ G s ∣ = d |G_s|=\sqrt{d} ∣ G s ∣ = d exactly when gcd ( s , d ) = 1 \gcd(s,d)=1 g cd( s , d ) = 1 . When m = gcd ( s , d ) > 1 m=\gcd(s,d)>1 m = g cd( s , d ) > 1 the sum degenerates: ∣ g ^ ( p ) ∣ |\hat{g}(p)| ∣ g ^ ( p ) ∣ equals m d \sqrt{md} m d on the sublattice p ∈ m Z d p\in m\mathbb{Z}_d p ∈ m Z d (or a coset of it) and vanishes elsewhere, so flatness fails maximally — e.g. at d = 8 d=8 d = 8 , s = 2 s=2 s = 2 the spectrum is ( 4 , 0 , 4 , 0 , 4 , 0 , 4 , 0 ) (4,0,4,0,4,0,4,0) ( 4 , 0 , 4 , 0 , 4 , 0 , 4 , 0 ) , and at d = 12 d=12 d = 12 , s = 4 s=4 s = 4 it is ( 0 , 0 , 6.93 , 0 , … ) (0,0,6.93,0,\ldots) ( 0 , 0 , 6.93 , 0 , … ) . For θ \theta θ off the lattice π Z / d \pi\mathbb{Z}/d π Z / d , λ \lambda λ is not a 2 d 2d 2 d -th root of unity, no re-indexing is available, and flatness fails; a fine scan of θ ∈ [ 0 , 2 π ) \theta\in[0,2\pi) θ ∈ [ 0 , 2 π ) at 10 5 10^5 1 0 5 points per dimension for d = 3 , … , 8 d=3,\ldots,8 d = 3 , … , 8 finds no flat angle outside { π s / d } \{\pi s/d\} { π s / d } .
The flat set is therefore exactly
Θ d = { π s / d : gcd ( s , d ) = 1 , s d even } , (27) \Theta_d=\{\pi s/d:\gcd(s,d)=1,\ sd\ \text{even}\}, \tag*{(27)} Θ d = { π s / d : g cd( s , d ) = 1 , s d even } , (27)
which has φ ( d ) \varphi(d) φ ( d ) members in [ 0 , 2 π ) [0,2\pi) [ 0 , 2 π ) for odd d d d and 2 φ ( d ) 2\varphi(d) 2 φ ( d ) for even d d d (φ \varphi φ Euler’s totient): two members for d = 2 , 3 d=2,3 d = 2 , 3 , four for d = 4 , 5 d=4,5 d = 4 , 5 , eight for d = 8 d=8 d = 8 . Every member works identically in the protocol, with the decoder modified as stated after Eq. (15) (λ ↦ λ s \lambda\mapsto\lambda_s λ ↦ λ s in Λ \Lambda Λ , P θ ⋆ ↦ P θ P_{\theta^\star}\mapsto P_\theta P θ ⋆ ↦ P θ and R ∣ u ⟩ = ∣ s − 1 u ⟩ R|u\rangle=|s^{-1}u\rangle R ∣ u ⟩ = ∣ s − 1 u ⟩ in Γ \Gamma Γ ); we confirmed exact recovery on every flat angle for d ≤ 9 d\le9 d ≤ 9 ; θ ⋆ ( d ) \theta^\star(d) θ ⋆ ( d ) of Eq. (3) is simply the smallest positive element, s = 1 s=1 s = 1 for even d d d and s = 2 s=2 s = 2 for odd d d d . This also explains why no single parity-independent formula exists: s = 1 s=1 s = 1 is disallowed for odd d d d because s d sd s d is then odd, which is precisely the failure of 2 2 2 to be invertible seen in Methods.
Restriction I: qudit encrypted cloning
Removing the data qudits and the teleportation stage from the protocol leaves encrypted cloning of a single qudit, the circuit of Figure S1 . Here we collect what the main text and Methods establish about this restriction on its own. The input is the provider’s register B C B_C B C , which may be entangled with an external system E E E of any dimension, so that the joint state is an arbitrary ∣ ψ ⟩ B C E |\psi\rangle_{B_CE} ∣ ψ ⟩ B C E . In the full protocol E E E is Alice’s qudit B A B_A B A together with anything B A B_A B A is entangled with. The encoder Eq. (2) acts on ( B C , S 1 , … , S n ) (B_C,S_1,\ldots,S_n) ( B C , S 1 , … , S n ) with each S j S_j S j initially Bell-paired with a key qudit K j K_j K j , and the decoder Eq. (5) acts on ( S B , K 1 , … , K n ) (S_B,K_1,\ldots,K_n) ( S B , K 1 , … , K n ) .
Privacy holds at every angle. Methods, Eq. (16) , shows that any set of shares short of all n n n is left in I / d ∣ C ∣ ⊗ ρ E \mathbb{I}/d^{|C|}\otimes\rho_E I / d ∣ C ∣ ⊗ ρ E . That derivation uses only the Weyl expansion Eq. (13) and the normalization ∑ p , q ∣ μ p , q ∣ 2 = d 4 \sum_{p,q}|\mu_{p,q}|^2=d^4 ∑ p , q ∣ μ p , q ∣ 2 = d 4 , both valid for any θ \theta θ , so a losing site learns nothing whether or not the angle is tuned. This requires n ≥ 2 n\ge2 n ≥ 2 . A single share has no spectator pair to suppress the cross terms, so for n = 1 n=1 n = 1 it is the all-shares coalition of Section S8 and is private only when gcd ( 1 − η 2 , d ) = 1 \gcd(1-\eta^2,d)=1 g cd( 1 − η 2 , d ) = 1 ; at d = 2 d=2 d = 2 , for example, its trace distance from I / d ⊗ ρ E \mathbb{I}/d\otimes\rho_E I / d ⊗ ρ E is 1 / 2 1/2 1/2 , even though recovery still works.
Recovery holds on the flat angles. For θ ∈ Θ d \theta\in\Theta_d θ ∈ Θ d (Section S6), and with the decoder modified as stated after Eq. (15) , the decoder at any site B B B returns
∣ ψ ⟩ S B E ⊗ ∣ Φ ⟩ B C K B ⊗ ⨂ j ≠ B ∣ Φ ⟩ S j K j , (28) |\psi\rangle_{S_BE}\otimes|\Phi\rangle_{B_CK_B}\otimes\bigotimes_{j\ne B}|\Phi\rangle_{S_jK_j}, \tag*{(28)} ∣ ψ ⟩ S B E ⊗ ∣Φ ⟩ B C K B ⊗ j = B ⨂ ∣Φ ⟩ S j K j , (28)
which is the equality drawn in Figure S1 . The Methods derivation (Steps 1–5) treats the case where B C B_C B C is maximally entangled with a reference B ′ B' B ′ , and it uses nothing beyond the chirp from Eq. (14) . Any other input follows, because every ∣ ψ ⟩ B C E |\psi\rangle_{B_CE} ∣ ψ ⟩ B C E can be written as d ( I ⊗ V ) ∣ Φ ⟩ B C B ′ \sqrt{d}(\mathbb{I}\otimes V)|\Phi\rangle_{B_CB'} d ( I ⊗ V ) ∣Φ ⟩ B C B ′ for some linear map V : B ′ → E V:B'\to E V : B ′ → E . Since the decoder never touches B ′ B' B ′ , applying V V V afterwards commutes with it and turns ∣ Φ ⟩ B ′ S B |\Phi\rangle_{B'S_B} ∣Φ ⟩ B ′ S B into ∣ ψ ⟩ S B E |\psi\rangle_{S_BE} ∣ ψ ⟩ S B E . Entanglement between B C B_C B C and the outside world is therefore transferred to the recipient intact, not only the local state of B C B_C B C .
Off the flat angles, no decoder exists. The statement concerns the encoder family U e n c ( θ ) U_{\mathrm{enc}}(\theta) U enc ( θ ) of Eq. (2) , and excludes every decoder, not only that of Eq. (5) , acting on the recipient’s registers. Take B C B_C B C maximally entangled with B ′ B' B ′ . After exact recovery ( B ′ , S B ) (B',S_B) ( B ′ , S B ) is in a pure maximally entangled state, so it cannot be correlated with anything else, in particular with B C B_C B C , which never reaches the recipient. Hence the encoded state must already have ρ B ′ B C = I / d 2 \rho_{B'B_C}=\mathbb{I}/d^2 ρ B ′ B C = I / d 2 . Tracing all signal–key pairs out of Eq. (17) gives the Bell-diagonal state ρ B ′ B C = ∑ p , q ∣ μ p , q ∣ 2 ( W p , q ) B C ∣ Φ ⟩ ⟨ Φ ∣ ( W p , q ) B C † \rho_{B'B_C}=\sum_{p,q}|\mu_{p,q}|^2(W_{p,q})_{B_C}|\Phi\rangle\langle\Phi|(W_{p,q})_{B_C}^{\dagger} ρ B ′ B C = ∑ p , q ∣ μ p , q ∣ 2 ( W p , q ) B C ∣Φ ⟩ ⟨ Φ∣ ( W p , q ) B C † , which is maximally mixed only if all d 2 d^2 d 2 weights are equal, i.e. ∣ g ^ ( p ) ∣ = d |\hat{g}(p)|=\sqrt{d} ∣ g ^ ( p ) ∣ = d for every p p p . By Section S6 this is exactly the condition θ ∈ Θ d \theta\in\Theta_d θ ∈ Θ d . Numerically, at off-lattice angles the decoder fidelity drops well below one (to 0.15 0.15 0.15 at d = 3 d=3 d = 3 and 0.065 0.065 0.065 at d = 4 d=4 d = 4 in our tests), while the unselected shares remain exactly maximally mixed.
Cost. The encoder uses 4 n 4n 4 n two-qudit gates, two phase gates and two Fourier layers, and the decoder uses 3 n 3n 3 n two-qudit gates and three single-qudit gates. Neither count depends on d d d .
Replacing lost keys. Suppose the keys K j K_j K j for j j j in some set L L L of unselected sites are lost, while the corresponding shares S j S_j S j are still coherent and can be handed to the recipient. The two halves of a Bell pair are interchangeable only up to a transpose: ( W p , q ⊗ I ) ∣ Φ ⟩ = ( I ⊗ W p , q T ) ∣ Φ ⟩ (W_{p,q}\otimes\mathbb{I})|\Phi\rangle=(\mathbb{I}\otimes W_{p,q}^{T})|\Phi\rangle ( W p , q ⊗ I ) ∣Φ ⟩ = ( I ⊗ W p , q T ) ∣Φ ⟩ with W p , q T = Z d q X d − p = ω − p q W − p , q W_{p,q}^{T}=Z_{d}^{q}X_{d}^{-p}=\omega^{-pq}W_{-p,q} W p , q T = Z d q X d − p = ω − pq W − p , q . Seen from S j S_j S j , the shift label therefore appears with the opposite sign, the phase label with the same sign, and each exchanged pair carries an extra ω − p q \omega^{-pq} ω − pq . The decoder compensates with three changes for every j ∈ L j\in L j ∈ L : CSUM ( S B → K j ) \operatorname{CSUM}(S_B\to K_j) CSUM ( S B → K j ) becomes CSUB ( S B → S j ) \operatorname{CSUB}(S_B\to S_j) CSUB ( S B → S j ) , CSUM ( K j → K B ) \operatorname{CSUM}(K_j\to K_B) CSUM ( K j → K B ) becomes CSUM ( S j → K B ) \operatorname{CSUM}(S_j\to K_B) CSUM ( S j → K B ) , and the block CSUB ( S B → K B ) n \operatorname{CSUB}(S_B\to K_B)^n CSUB ( S B → K B ) n shortens to CSUB ( S B → K B ) n − ℓ \operatorname{CSUB}(S_B\to K_B)^{n-\ell} CSUB ( S B → K B ) n − ℓ . Recovery is then again exact, with every pair ( S j , K j ) (S_j,K_j) ( S j , K j ) , j ∈ L j\in L j ∈ L , returned to ∣ Φ ⟩ |\Phi\rangle ∣Φ ⟩ . We confirmed this numerically for d = 3 , 4 , 5 d=3,4,5 d = 3 , 4 , 5 with one and two replaced keys. Using the unmodified decoder on the substituted registers gives entanglement fidelity 1 / d 2 1/d^2 1/ d 2 in general.
Relation to other work. For d = 2 d=2 d = 2 the encoder coincides with that of Ref. [26 ] up to the global phase i i i . Two recent works also construct qudit encrypted cloning. Ref. [27 ] encodes with Chu sequences, which are quadratic-phase sequences closely related to P θ ∗ P_\theta^* P θ ∗ , and gives a decoder whose two-qudit gate count grows as d 3 d^3 d 3 . Ref. [28 ] works in the Weyl–Heisenberg picture, relates the construction to absolutely maximally entangled states, and gives the decoder as an operator sum. What the present construction adds is the explicit 3 n 3n 3 n -gate decoder, independent of d d d , exact transfer of entanglement with an arbitrary external system, and the statement that, for this encoder family, the flat angles are necessary as well as sufficient.
Coalition privacy and the all-shares coalition
Eq. (16) of Methods proves, in closed form and for all d d d and all n ≥ 2 n\ge2 n ≥ 2 , that every coalition C ⊊ { 1 , … , n } C\subsetneq\{1,\ldots,n\} C ⊊ { 1 , … , n } of signal qudits holds exactly I / d ∣ C ∣ \mathbb{I}/d^{|C|} I / d ∣ C ∣ , uncorrelated with the reference. This covers the entire threat model: in the deferred-recipient setting the selected site is chosen after distribution, so a coalition that already contains every site has nothing left to learn about. The bound is tight at ∣ C ∣ = n − 1 |C|=n-1 ∣ C ∣ = n − 1 and requires no numerical support. Direct evaluation over all subsets for d ≤ 7 d\le7 d ≤ 7 , n ≤ 4 n\le4 n ≤ 4 reproduces trace distance < 10 − 16 <10^{-16} < 1 0 − 16 in every case, as it must.
The one case the spectator argument cannot reach is C = { 1 , … , n } C=\{1,\ldots,n\} C = { 1 , … , n } — all n n n recipients colluding and pooling their shares before any selection is made. This lies outside the threat model, but it is worth recording that it is genuinely not private in general, rather than merely unproven. With no spectator pair to enforce orthogonality, cross terms ( p , q ) ≠ ( p ′ , q ′ ) (p,q)\ne(p',q') ( p , q ) = ( p ′ , q ′ ) survive the partial trace, and the residual correlation with B A B_A B A is governed by the overlap of the pooled signal registers across branches. The stabilizer count of Section S9 (case b = 0 b=0 b = 0 , s = n s=n s = n , k = 0 k=0 k = 0 ) gives a surviving correlation of
gcd ( n − η 2 , d ) = 1 , η = { 1 , d even , 2 − 1 m o d d , d odd , (29) \gcd(n-\eta^2,d)=1,\qquad
\eta=
\begin{cases}
1, & d\ \text{even},\\
2^{-1}\bmod d, & d\ \text{odd},
\end{cases} \tag*{(29)} g cd( n − η 2 , d ) = 1 , η = { 1 , 2 − 1 mod d , d even , d odd , (29)
so for example ( d , n ) = ( 2 , 3 ) (d,n)=(2,3) ( d , n ) = ( 2 , 3 ) fails this criterion and the pooled coalition recovers B A B_A B A -correlated information, with trace distance 1 2 \frac{1}{2} 2 1 between ρ B A C \rho_{B_AC} ρ B A C and ρ B A ⊗ I / d n \rho_{B_A}\otimes\mathbb{I}/d^n ρ B A ⊗ I / d n . We note two mitigating facts. First, the data qudit D A D_A D A remains perfectly protected even here: its reduced state conditioned on the full coalition is unchanged, trace distance 0 0 0 , because D A D_A D A is folded into B A B_A B A by a gate the encoder never touches. Second, any protocol run with gcd ( n − η 2 , d ) = 1 \gcd(n-\eta^2,d)=1 g cd( n − η 2 , d ) = 1 is private even against the full coalition. Choosing n n n to satisfy this congruence is free, since n n n is a design parameter; for n = 2 n=2 n = 2 it reduces to gcd ( 7 , d ) = 1 \gcd(7,d)=1 g cd( 7 , d ) = 1 for every even d d d , and holds for every odd d d d , the same condition as the ( 3 , 5 ) (3,5) ( 3 , 5 ) structure of Section S9 at θ ∗ \theta^* θ ∗ .
Access structure of the encrypted resource
Here we treat the encoded register as a secret-sharing scheme. The secret is the provider’s reference R R R — Alice’s half B A B_A B A , or anything B C B_C B C was entangled with — and the 2 n + 1 2n+1 2 n + 1 shareholders are B C , S 1 , … , S n , K 1 , … , K n B_C,S_1,\ldots,S_n,K_1,\ldots,K_n B C , S 1 , … , S n , K 1 , … , K n . For a set A A A of these registers we ask how much of the secret it holds, measured by the mutual information I ( R : A ) I(R:A) I ( R : A ) in units of its maximum 2 log d 2\log d 2 log d . The value 1 1 1 means that A A A can recover R R R ’s partner exactly, and 0 0 0 means that A A A is uncorrelated with R R R . Because the global state is pure, I ( R : A ) + I ( R : A ˉ ) = 2 log d I(R:A)+I(R:\bar A)=2\log d I ( R : A ) + I ( R : A ˉ ) = 2 log d for the complement A ˉ \bar A A ˉ , so a set is authorized exactly when its complement learns nothing. Classify each signal–key pair by how much of it lies in A A A . Let f f f be the number of pairs with both members in A A A , s s s and k k k the numbers with only S j S_j S j or only K j K_j K j in A A A , u = n − f − s − k u=n-f-s-k u = n − f − s − k the number of pairs entirely outside A A A , and b = 1 b=1 b = 1 if B C ∈ A B_C\in A B C ∈ A .
Sets that miss a whole pair. If u ≥ 1 u\ge1 u ≥ 1 , a pair outside A A A is a spectator exactly as in Eq. (16) . The cross terms of Eq. (17) vanish, and A A A holds an incoherent mixture over the labels ( p , q ) (p,q) ( p , q ) with weight ∣ μ p , q ∣ 2 / d 4 \lvert\mu_{p,q}\rvert^2/d^4 ∣ μ p , q ∣ 2 / d 4 . A single member of a pair is maximally mixed for every label, so half-pairs reveal nothing. A complete pair ( S j , K j ) (S_j,K_j) ( S j , K j ) reveals ( p , q ) (p,q) ( p , q ) exactly, through the orthonormal states ( W p , q ⊗ I ) ∣ Φ ⟩ (W_{p,q}\otimes\mathbb{I})|\Phi\rangle ( W p , q ⊗ I ) ∣Φ ⟩ . With B C B_C B C and one complete pair, the label can be read out and W p , q † W_{p,q}^{\dagger} W p , q † undone on B C B_C B C , restoring ∣ Φ ⟩ R B C |\Phi\rangle_{RB_C} ∣Φ ⟩ R B C , so I = 1 I=1 I = 1 . With B C B_C B C but no complete pair, B C B_C B C is Weyl-twirled, and at a flat angle the twirl is uniform, leaving R B C R B_C R B C in I / d 2 \mathbb{I}/d^2 I / d 2 , so I = 0 I=0 I = 0 . Without B C B_C B C , A A A never touches R R R ’s partner, so I = 0 I=0 I = 0 again. Hence, for u ≥ 1 u\ge1 u ≥ 1 ,
I ( R : A ) = 1 if b = 1 and f ≥ 1 , I ( R : A ) = 0 otherwise . (30) I(R:A)=1\ \text{if }b=1\text{ and }f\ge1,\qquad I(R:A)=0\ \text{otherwise}. \tag*{(30)} I ( R : A ) = 1 if b = 1 and f ≥ 1 , I ( R : A ) = 0 otherwise . (30)
Only the step with B C B_C B C and no complete pair uses the flat angle. Off the lattice the twirl is not uniform, and B C B_C B C keeps partial information.
Sets that touch every pair. If u = 0 u=0 u = 0 and f ≥ 1 f\ge1 f ≥ 1 , the complement contains no complete pair and misses pair f f f ’s partner entirely, so by Eq. (30) it learns nothing, and A A A is authorized. The remaining sets hold exactly one member of each pair, with or without B C B_C B C . Here no spectator exists, and the answer depends on arithmetic. At the flat angle θ = π r / d \theta=\pi r/d θ = π r / d it is
I ( R : A ) = { 1 − log g 1 2 log d , b = 1 , log g 0 2 log d , b = 0 , (31) I(R:A)=
\begin{cases}
1-\frac{\log g_1}{2\log d}, & b=1,\\
\frac{\log g_0}{2\log d}, & b=0,
\end{cases} \tag*{(31)} I ( R : A ) = { 1 − 2 l o g d l o g g 1 , 2 l o g d l o g g 0 , b = 1 , b = 0 , (31)
with g 1 = gcd ( r 2 k ( s + 1 ) − 1 , d ) g_1=\gcd(r^2k(s+1)-1,d) g 1 = g cd( r 2 k ( s + 1 ) − 1 , d ) and g 0 = gcd ( r 2 s ( k + 1 ) − 1 , d ) g_0=\gcd(r^2s(k+1)-1,d) g 0 = g cd( r 2 s ( k + 1 ) − 1 , d ) , for s + k = n s+k=n s + k = n . The two cases are complements of each other and sum to 1, as purity requires. The case b = 0 b=0 b = 0 , s = n s=n s = n , k = 0 k=0 k = 0 is the pooled coalition of all n n n shares of Section S8, and g 0 = gcd ( r 2 n − 1 , d ) g_0=\gcd(r^2n-1,d) g 0 = g cd( r 2 n − 1 , d ) reproduces its condition gcd ( n − η 2 , d ) = 1 \gcd(n-\eta^2,d)=1 g cd( n − η 2 , d ) = 1 (for odd d d d , r = 2 r=2 r = 2 and 4 n − 1 = 4 ( n − η 2 ) 4n-1=4(n-\eta^2) 4 n − 1 = 4 ( n − η 2 ) with 4 invertible; for even d d d , r = 1 r=1 r = 1 and η = 1 \eta=1 η = 1 ).
Relation to earlier classifications. For qubits and sets without B C B_C B C , g 0 g_0 g 0 reproduces the parity rule of Ref. [37 ] : a set with s s s signal and k k k key qubits leaks exactly when s s s and n n n are both odd. Sets containing the source were classified for qubits in Ref. [38 ] . For even d d d (r = 1 r=1 r = 1 ), g 0 = gcd ( s ( k + 1 ) − 1 , d ) g_0=\gcd(s(k+1)-1,d) g 0 = g cd( s ( k + 1 ) − 1 , d ) is the leakage criterion of Ref. [39 ] , derived there for sets without the source; for odd d d d that work uses a different quadratic phase, and the factor r 2 r^2 r 2 reflects the encoder of Eq. (2) . New here are the case b = 1 b=1 b = 1 for d > 2 d>2 d > 2 and the derivation of both cases, together with Eq. (30) , from a single stabilizer count.
Derivation. Both Eq. (30) and Eq. (31) follow from a single counting argument, because at a flat angle the encoder is a Clifford unitary and the encoded register is a stabilizer state. Let T = ∑ c t c T=\sum_c t_c T = ∑ c t c be the sum over the N = n + 1 N=n+1 N = n + 1 carriers (B C , S 1 , … , S n B_C,S_1,\ldots,S_n B C , S 1 , … , S n ); because r d rd r d is even, e i θ T 2 e^{i\theta T^2} e i θ T 2 depends only on T T T modulo d d d , and e i θ [ ( T + 1 ) 2 − T 2 ] = e i θ ω r T e^{i\theta[(T+1)^2-T^2]}=e^{i\theta\omega rT} e i θ [( T + 1 ) 2 − T 2 ] = e i θ ω r T . Hence U Z U_Z U Z maps each carrier shift X c X_c X c to X c Z a l l r X_cZ_{\mathrm{all}}^r X c Z all r , up to phase, and fixes every Z c Z_c Z c ; conjugating by F F F gives the action of U X U_X U X . On a carrier Weyl string labeled by ( x , z ) ∈ Z d 2 N (x,z)\in\mathbb{Z}_d^{2N} ( x , z ) ∈ Z d 2 N , U e n c U_{\mathrm{enc}} U enc therefore acts as the linear map
x ↦ x − ( r 2 N σ x + r σ z ) 1 , z ↦ z + r σ x 1 , (32) x\mapsto x-(r^2N\sigma_x+r\sigma_z)\mathbf{1},\qquad z\mapsto z+r\sigma_x\mathbf{1}, \tag*{(32)} x ↦ x − ( r 2 N σ x + r σ z ) 1 , z ↦ z + r σ x 1 , (32)
with σ x = ∑ c x c \sigma_x=\sum_c x_c σ x = ∑ c x c , σ z = ∑ c z c \sigma_z=\sum_c z_c σ z = ∑ c z c and 1 \mathbf{1} 1 the all-ones vector. Before encoding, the 2 N 2N 2 N qudits R R R , B C B_C B C , { S j , K j } \{S_j,K_j\} { S j , K j } are in Bell pairs, stabilized by X ⊗ X X\otimes X X ⊗ X and Z ⊗ Z − 1 Z\otimes Z^{-1} Z ⊗ Z − 1 on each pair. A general stabilizer element is labeled by ( a , β , a j , b j ) ∈ Z d 2 N (a,\beta,a_j,b_j)\in\mathbb{Z}_d^{2N} ( a , β , a j , b j ) ∈ Z d 2 N , and after Eq. (32) it acts as
R : ( a , β ) , B C : ( a − γ , − β + r α ) , K j : ( a j , − b j ) , S j : ( a j − γ , b j + r α ) , (33) \begin{aligned}
R &: (a,\beta), & B_C &: (a-\gamma,-\beta+r\alpha),\\
K_j &: (a_j,-b_j), & S_j &: (a_j-\gamma,b_j+r\alpha), \tag*{(33)}
\end{aligned} R K j : ( a , β ) , : ( a j , − b j ) , B C S j : ( a − γ , − β + r α ) , : ( a j − γ , b j + r α ) , (33)
where a pair ( x , z ) (x,z) ( x , z ) denotes X x Z z X^xZ^z X x Z z up to phase, α = a + ∑ j a j \alpha=a+\sum_j a_j α = a + ∑ j a j and γ = r 2 N α + r ( − β + ∑ j b j ) \gamma=r^2N\alpha+r(-\beta+\sum_j b_j) γ = r 2 N α + r ( − β + ∑ j b j ) . For a stabilizer state on m m m qudits, the reduced state on any subset A A A is ρ A = d − ∣ A ∣ ∑ g ∈ S A g \rho_A=d^{-|A|}\sum_{g\in S_A}g ρ A = d − ∣ A ∣ ∑ g ∈ S A g , where S A S_A S A is the subgroup of stabilizers supported on A A A , so S ( A ) = ∣ A ∣ log d − log ∣ S A ∣ S(A)=|A|\log d-\log|S_A| S ( A ) = ∣ A ∣ log d − log ∣ S A ∣ for every d d d , prime or not [48 ] . Since R R R alone is maximally mixed, this gives
I ( R : A ) = log ∣ S R A ∣ ∣ S A ∣ = log ∣ π R ( S R A ) ∣ , (34) I(R:A)=\log\frac{|S_{RA}|}{|S_A|}=\log\left|\pi_R(S_{RA})\right|, \tag*{(34)} I ( R : A ) = log ∣ S A ∣ ∣ S R A ∣ = log ∣ π R ( S R A ) ∣ , (34)
the number of values ( a , β ) ∈ Z d 2 (a,\beta)\in\mathbb{Z}_d^2 ( a , β ) ∈ Z d 2 that occur on R R R among stabilizers vanishing outside R A R A R A . Computing I I I is thus a matter of solving linear congruences.
For a set with one member of each pair, every half-pair outside A A A imposes vanishing conditions: if only S j ∈ A S_j\in A S j ∈ A , then a j = b j = 0 a_j=b_j=0 a j = b j = 0 ; if only K j ∈ A K_j\in A K j ∈ A , then a j = γ a_j=\gamma a j = γ and b j = − r α b_j=-r\alpha b j = − r α . Substituting, α = a + k γ \alpha=a+k\gamma α = a + k γ and γ = r 2 ( s + 1 ) α − r β \gamma=r^2(s+1)\alpha-r\beta γ = r 2 ( s + 1 ) α − r β . If B C ∈ A B_C\in A B C ∈ A , eliminating α \alpha α leaves the single condition ( r 2 k ( s + 1 ) − 1 ) γ = r [ β − r ( s + 1 ) a ] (r^2k(s+1)-1)\gamma=r[\beta-r(s+1)a] ( r 2 k ( s + 1 ) − 1 ) γ = r [ β − r ( s + 1 ) a ] . Since r r r is invertible modulo d d d , it is solvable for exactly d 2 / g 1 d^2/g_1 d 2 / g 1 pairs ( a , β ) (a,\beta) ( a , β ) , which gives the first line of Eq. (31) . If B C ∉ A B_C\notin A B C ∈ / A , its part must vanish as well, a = γ a=\gamma a = γ and β = r α \beta=r\alpha β = r α , which forces β = r ( k + 1 ) a \beta=r(k+1)a β = r ( k + 1 ) a and ( r 2 s ( k + 1 ) − 1 ) a ≡ 0 (r^2s(k+1)-1)a\equiv0 ( r 2 s ( k + 1 ) − 1 ) a ≡ 0 . That has g 0 g_0 g 0 solutions, giving the second line. If instead some pair lies entirely outside A A A , both of its members must vanish, which forces α = γ = 0 \alpha=\gamma=0 α = γ = 0 . Then ( a , β ) (a,\beta) ( a , β ) ranges over all of Z d 2 \mathbb{Z}_d^2 Z d 2 when B C ∈ A B_C\in A B C ∈ A and A A A contains a complete pair, whose free ( a j , b j ) (a_j,b_j) ( a j , b j ) absorb the other constraints, and only over ( 0 , 0 ) (0,0) ( 0 , 0 ) otherwise. This is Eq. (30) . The flat angle enters only through Eq. (32) : off the lattice U e n c U_{\mathrm{enc}} U enc is not Clifford and the counting does not apply, consistent with the partial information retained by B C B_C B C there.
Consequences. (i) The key register alone (k = n k=n k = n , b = 0 b=0 b = 0 , s = 0 s=0 s = 0 ) has g 0 = gcd ( − 1 , d ) = 1 g_0=\gcd(-1,d)=1 g 0 = g cd( − 1 , d ) = 1 and carries no information for any d d d and n n n . Intercepting the keys alone, or discarding them, reveals nothing. (ii) The recipient’s set ( S B , K 1 , … , K n ) (S_B,K_1,\ldots,K_n) ( S B , K 1 , … , K n ) has s = 1 s=1 s = 1 and u = 0 u=0 u = 0 and is authorized without B C B_C B C , which is why the provider may discard B C B_C B C after encoding. (iii) For n = 2 n=2 n = 2 there are five shareholders. By Eq. (30) , every pair of registers carries nothing except possibly the half-pair sets ( S 1 S 2 , S 1 K 2 , K 1 S 2 , K 1 K 2 ) (S_1S_2,S_1K_2,K_1S_2,K_1K_2) ( S 1 S 2 , S 1 K 2 , K 1 S 2 , K 1 K 2 ) , and every triple is authorized except possibly B C B_C B C plus two half-pairs. By Eq. (31) , these exceptions have ( s , k ) ∈ { ( 2 , 0 ) , ( 1 , 1 ) , ( 0 , 2 ) } (s,k)\in\{(2,0),(1,1),(0,2)\} ( s , k ) ∈ {( 2 , 0 ) , ( 1 , 1 ) , ( 0 , 2 )} , and every one of them reduces to either gcd ( − 1 , d ) = 1 \gcd(-1,d)=1 g cd( − 1 , d ) = 1 or gcd ( 2 r 2 − 1 , d ) \gcd(2r^2-1,d) g cd( 2 r 2 − 1 , d ) , so all of them vanish exactly when gcd ( 2 r 2 − 1 , d ) = 1 \gcd(2r^2-1,d)=1 g cd( 2 r 2 − 1 , d ) = 1 . The scheme is then a ( 3 , 5 ) (3,5) ( 3 , 5 ) threshold scheme. That encrypted cloning with two pairs is equivalent to a ( 3 , 5 ) (3,5) ( 3 , 5 ) scheme in every dimension was shown by Lin and Lo [28 ] (their Theorem 3); what the present count adds is the exact condition for the encoder of Eq. (2) , which fails at d = 7 d=7 d = 7 , θ ∗ \theta^* θ ∗ , and the structure for n ≥ 3 n\ge3 n ≥ 3 below. At θ ∗ \theta^* θ ∗ (r = 2 r=2 r = 2 for odd d d d , r = 1 r=1 r = 1 for even d d d ) the condition reads gcd ( 7 , d ) = 1 \gcd(7,d)=1 g cd( 7 , d ) = 1 for odd d d d and holds automatically for even d d d . For d = 7 d=7 d = 7 another flat angle, for example r = 4 r=4 r = 4 , satisfies it, and we find such an angle for every d < 3000 d<3000 d < 3000 . (iv) For n ≥ 3 n\ge3 n ≥ 3 the structure is no longer a threshold one. Sets of one member per pair can carry exactly half of the information, for example the three shares at ( d , n ) = ( 2 , 3 ) (d,n)=(2,3) ( d , n ) = ( 2 , 3 ) , as already noted in Section S8.
Verification. As an independent check on the derivation, the verification notebook (Code availability) computes I ( R : A ) I(R:A) I ( R : A ) from the exact state vector for every proper subset A A A and compares it with Eqs. (30) and (31) . It covers d = 2 , … , 9 d=2,\ldots,9 d = 2 , … , 9 at n = 2 n=2 n = 2 , d = 2 , … , 5 d=2,\ldots,5 d = 2 , … , 5 at n = 3 n=3 n = 3 , ( d , n ) = ( 2 , 4 ) , ( 3 , 4 ) , ( 5 , 4 ) (d,n)=(2,4),(3,4),(5,4) ( d , n ) = ( 2 , 4 ) , ( 3 , 4 ) , ( 5 , 4 ) at θ ∗ \theta^* θ ∗ , and six further flat angles, 4182 subsets in all. The largest deviation is ≲ 10 − 14 \lesssim10^{-14} ≲ 1 0 − 14 .
Illustration of privacy, recovery and access structure
The statements illustrated in Figure S2 are proved in closed form: privacy of every proper coalition by the Weyl twirl, Eq. (16) ; recovery on the flat angles by the chirp identity, Eq. (14) , and the explicit decoder; the access structure by Eqs. (30) and (31) ; and the spent-key fidelities by the post-decoding product form, Eq. (7) . The figure makes the separation between privacy and recovery visible. A single unselected share holds no information about the reference at any phase angle, whereas the recipient’s set recovers the reference completely only on the flat lattice and partially elsewhere (panel a). The tuning of θ \theta θ therefore buys recovery, not privacy. Panel b shows every subset of the five registers for n = 2 n=2 n = 2 . At ( d , θ ) = ( 3 , θ ∗ ) (d,\theta)=(3,\theta^*) ( d , θ ) = ( 3 , θ ∗ ) the information jumps from 0 to 1 between two and three registers, which is the ( ( 3 , 5 ) ) ((3,5)) (( 3 , 5 )) threshold structure. At d = 7 d=7 d = 7 , where gcd ( 2 r 2 − 1 , d ) = 7 \gcd(2r^2-1,d)=7 g cd( 2 r 2 − 1 , d ) = 7 at θ ∗ \theta^* θ ∗ , three two-register sets and three-register sets carry exactly half the information, and the flat angle 4 π / 7 4\pi/7 4 π /7 restores the threshold.
Restriction II: deterministic non-local CSUMk ^k k
Removing the cloning layer and fixing the recipient in advance leaves the teleportation stage of Figure S1 , in which the resource pair ( B A , B B ) (B_A,B_B) ( B A , B B ) is supplied directly. B B B_B B B plays the role of the recovered share S B S_B S B of the full protocol. This is the qudit version of the qubit non-local CNOT [4 , 5 ] , and it extends to every power of CSUM.
Let Alice apply CSUM k 1 ( D A → B A ) \operatorname{CSUM}^{k_1}(D_A\to B_A) CSUM k 1 ( D A → B A ) and Bob apply CSUM k 2 ( B B → D B ) \operatorname{CSUM}^{k_2}(B_B\to D_B) CSUM k 2 ( B B → D B ) , for any k 1 , k 2 ∈ Z d k_1,k_2\in\mathbb{Z}_d k 1 , k 2 ∈ Z d . Alice measures B A B_A B A in the computational basis ( m A ) (m_A) ( m A ) . Bob applies F F F to B B B_B B B and measures it in the computational basis ( m B ) (m_B) ( m B ) . They exchange the two outcomes, and Bob applies X d k 2 m A X_d^{k_2m_A} X d k 2 m A to D B D_B D B while Alice applies Z d k 1 m B Z_d^{k_1m_B} Z d k 1 m B to D A D_A D A . Following a basis input ∣ c , t ⟩ |c,t\rangle ∣ c , t ⟩ as in Methods, each branch ( m A , m B ) (m_A,m_B) ( m A , m B ) occurs with probability 1 / d 2 1/d^2 1/ d 2 and leaves
∣ c ⟩ D A ∣ t + k c ⟩ D B , k ≡ k 1 k 2 ( m o d d ) , (35) |c\rangle_{D_A}|t+kc\rangle_{D_B},\qquad k\equiv k_1k_2\pmod d, \tag*{(35)} ∣ c ⟩ D A ∣ t + k c ⟩ D B , k ≡ k 1 k 2 ( mod d ) , (35)
with the overall phase ω m A m B \omega^{m_A m_B} ω m A m B . That phase does not depend on c c c or t t t . After correction, each branch therefore applies a fixed unitary, CSUM k \operatorname{CSUM}^k CSUM k , up to a global phase. By linearity the same is true for superposed inputs and for inputs entangled with external systems, with no averaging over branches and nothing discarded.
Several choices of ( k 1 , k 2 ) (k_1,k_2) ( k 1 , k 2 ) are worth singling out. Panel c of Figure S1 is k 1 = k 2 = 1 k_1=k_2=1 k 1 = k 2 = 1 , which gives CSUM with corrections X d m A X_d^{m_A} X d m A and Z d m B Z_d^{m_B} Z d m B , the teleportation stage of the main protocol. The naive transcription of the qubit circuit, in which Bob’s gate adds ( k 1 = 1 , k 2 = − 1 ) (k_1=1,k_2=-1) ( k 1 = 1 , k 2 = − 1 ) , is not wrong so much as mislabeled: it teleports CSUB exactly, provided Bob applies X d − m A X_d^{-m_A} X d − m A . It fails only when combined with the CSUM correction X d m A X_d^{m_A} X d m A , and then it succeeds on just a fraction ν ( d ) / d \nu(d)/d ν ( d ) / d of (input, branch) pairs (Section S12). Finally, Bob’s operations depend on k 2 k_2 k 2 alone. If Bob always uses k 2 = 1 k_2=1 k 2 = 1 , Alice selects k = k 1 k=k_1 k = k 1 , and hence which power of CSUM is applied. Bob’s operations and the public outcomes ( m A , m B ) (m_A,m_B) ( m A , m B ) , which are uniformly distributed for every k k k , are then independent of k k k , so Bob learns about k k k only what the action of CSUM k \operatorname{CSUM}^k CSUM k on his own qudit D B D_B D B reveals, as it would for an ideal local implementation of the gate (Methods, “Hiding the power of CSUM”). All of this carries over unchanged to the full deferred-recipient protocol, whose final stage is exactly this circuit acting on the recovered pair ( B A , S B ) (B_A, S_B) ( B A , S B ) .
The resources are one maximally entangled qudit pair and two classical dits, one in each direction, as in the qubit case [5 ] . For gcd ( k , d ) = 1 \gcd(k,d)=1 g cd( k , d ) = 1 the entanglement cost is optimal. In that case C S U M k = ( M k − 1 ⊗ I ) C S U M ( M k ⊗ I ) \mathrm{CSUM}^{k}=(M_k^{-1}\otimes I)\,\mathrm{CSUM}\,(M_k\otimes I) CSUM k = ( M k − 1 ⊗ I ) CSUM ( M k ⊗ I ) with M k ∣ c ⟩ = ∣ k c ⟩ M_k|c\rangle=|kc\rangle M k ∣ c ⟩ = ∣ k c ⟩ , so the gate turns the product state F ∣ 0 ⟩ ⊗ ∣ 0 ⟩ F|0\rangle\otimes|0\rangle F ∣0 ⟩ ⊗ ∣0 ⟩ into a maximally entangled state of Schmidt rank d d d . Local operations and classical communication cannot create that entanglement, so at least one maximally entangled pair must be consumed. For general k k k , the operator Schmidt rank of C S U M k = ∑ c ∣ c ⟩ ⟨ c ∣ ⊗ X d k c \mathrm{CSUM}^{k}=\sum_c |c\rangle\langle c|\otimes X_d^{kc} CSUM k = ∑ c ∣ c ⟩ ⟨ c ∣ ⊗ X d k c is d / gcd ( k , d ) d/\gcd(k,d) d / g cd( k , d ) , and k = 0 k=0 k = 0 is the identity. Because the gate carries information in both directions (D A → D B D_A\to D_B D A → D B in the computational basis and D B → D A D_B\to D_A D B → D A in the Fourier basis), classical communication is needed both ways. We have not shown that one dit in each direction is the minimum for every d d d and k k k .
The literal qubit transcription
By Eq. (8) the teleported gate is fixed by the product k 1 k 2 k_1k_2 k 1 k 2 , so CSUM requires the two entangling gates to have opposite signs; whether Alice or Bob carries the inverse is immaterial, provided the corrections are chosen accordingly. Here we quantify what happens when the qubit circuit is transcribed literally: both parties apply CSUM (k 1 = 1 k_1=1 k 1 = 1 , k 2 = − 1 k_2=-1 k 2 = − 1 ), and the qubit corrections X d m A X_d^{m_A} X d m A , Z d m B Z_d^{m_B} Z d m B are kept. This circuit teleports CSUB, and after the correction X d m A X_d^{m_A} X d m A the target D B D_B D B is left displaced by 2 ( m A − c ) 2(m_A-c) 2 ( m A − c ) , where c c c is the (generally superposed) value of D A D_A D A . No classical function of ( m A , m B ) (m_A,m_B) ( m A , m B ) applied by Bob can supply this c c c -dependent shift without measuring D A D_A D A ; the shift is removed only by changing Alice’s side as well (her gate to CSUB and her correction to Z d − m B Z_d^{-m_B} Z d − m B ), i.e. by restoring k 1 k 2 ≡ 1 k_1k_2\equiv1 k 1 k 2 ≡ 1 . At d = 2 d=2 d = 2 the displacement 2 ( m A − c ) ≡ 0 2(m_A-c)\equiv0 2 ( m A − c ) ≡ 0 for every c c c , which is why the sign is invisible for qubits.
The displacement vanishes exactly on the solutions of 2 ( m A − c ) ≡ 0 ( m o d d ) 2(m_A-c)\equiv0\pmod d 2 ( m A − c ) ≡ 0 ( mod d ) . For odd d d d , 2 2 2 is invertible and each c c c is matched by the single branch m A = c m_A=c m A = c . For even d d d the congruence is m A ≡ c ( m o d d / 2 ) m_A\equiv c\pmod{d/2} m A ≡ c ( mod d /2 ) , solved by two values of m A m_A m A for each c c c . Either way the transcribed circuit is exact on a fraction
# { ( c , m A ) : 2 ( m A − c ) ≡ 0 } d 2 = ν ( d ) d , ν ( d ) = gcd ( 2 , d ) , (36) \frac{\#\{(c,m_A):2(m_A-c)\equiv0\}}{d^2}=\frac{\nu(d)}{d},\qquad\nu(d)=\gcd(2,d), \tag*{(36)} d 2 # {( c , m A ) : 2 ( m A − c ) ≡ 0 } = d ν ( d ) , ν ( d ) = g cd( 2 , d ) , (36)
of the d 2 d^2 d 2 (input, branch) pairs — all of them at d = 2 d=2 d = 2 , 1 / 3 1/3 1/3 at d = 3 d=3 d = 3 (Figure 3 c), 1 / 2 1/2 1/2 at d = 4 d=4 d = 4 , 1 / 5 1/5 1/5 at d = 5 d=5 d = 5 , and → 0 \to0 → 0 as d d d grows — with infidelity 1.000 1.000 1.000 on every remaining pair, to be contrasted with < 10 − 15 <10^{-15} < 1 0 − 15 on all d 2 d^2 d 2 pairs when k 1 k 2 ≡ 1 k_1k_2\equiv1 k 1 k 2 ≡ 1 . The surviving successes carry no operational content: whether a given branch is one of them is fixed by c c c , which is precisely the unknown the protocol must not measure, so the parties can neither predict nor post-select them. The failure is a statement about Z d \mathbb{Z}_d Z d arithmetic, not a numerical finding: CSUM is its own inverse on Z 2 \mathbb{Z}_2 Z 2 and on no other Z d \mathbb{Z}_d Z d . It is a pitfall in carrying qubit circuits over to qudits rather than an obstruction, and the protocol avoids it by construction. Supplementary Section S11 shows the transcribed circuit is nonetheless an exact protocol for CSUB, with the correction X d − m A X_d^{-m_A} X d − m A .
Comparison with simpler deferral schemes, and threat model
The introduction notes that the obvious ways of deferring the recipient each fail one of the requirements met by our construction. Table S2 summarizes the comparison; here we spell out the two simplest alternatives, comment on operation sharing and on the cost of selection, and then state the scope of our guarantees.
Table S2. Ways to defer or share a non-local gate. Columns give the resource distributed in advance, the quantum communication needed once the recipient is chosen, who can block the gate, and who can complete it alone without the selected node. QOS: quantum operation sharing [20 , 21 ] ; CT: threshold-controlled teleportation [22 ] .
Entanglement swapping. The provider shares a Bell pair with each candidate, keeps B C B_C B C , and once the recipient is known swaps B C B_C B C onto that candidate’s pair with one Bell measurement and two classical dits. Unselected candidates learn nothing, and the decision costs two dits rather than n n n key qudits. But whoever holds B C B_C B C holds half of a Bell pair with Alice and can complete the gate directly: the right to choose the recipient cannot be separated from the ability to use the gate, and the resource is lost with B C B_C B C . In our construction the key register carries no information about Alice’s qudit, nor does any single share or B C B_C B C , and B C B_C B C may be discarded after encoding.
GHZ-based information splitting. The provider prepares a ( n + 1 ) (n+1) ( n + 1 ) -qudit GHZ state d − 1 / 2 ∑ j ∣ j ⟩ B A ∣ j ⟩ ⊗ n d^{-1/2}\sum_j|j\rangle_{B_A}|j\rangle^{\otimes n} d − 1/2 ∑ j ∣ j ⟩ B A ∣ j ⟩ ⊗ n over B A B_A B A and the candidates [13 , 15 ] . To leave a Bell pair between Alice and a chosen site, each of the n − 1 n-1 n − 1 other sites must measure in the Fourier basis and announce its outcome for a phase correction. Every unselected site must therefore participate, any one of them can block or corrupt the gate by withholding or misreporting its outcome, and all of them learn that a selection has been made. Moreover each GHZ share is classically correlated with B A B_A B A in the computational basis: measuring one share reveals the outcome Alice would obtain. In our construction unselected sites take no action, each share is exactly maximally mixed and uncorrelated with B A B_A B A , and they cannot tell whether a selection has occurred.
Operation sharing and controlled teleportation. In quantum operation sharing [20 , 21 ] and threshold-controlled teleportation [22 ] a dealer enables an operation on, or a teleportation to, a recipient fixed in advance, and the sharers are controllers who must each measure and announce. These schemes therefore share the participation, veto and visibility properties of GHZ splitting. Here the recipient is one of the sharers, chosen after distribution, the unselected sharers take no action, and the shared object is an entangling gate acting on data that never enter the scheme.
Cost of selection. Swapping is faster at selection: it moves two dits, whereas our construction moves n n n key
qudits to the recipient. This cost can be moved to the distribution stage by pre-sharing n n n Bell pairs between the key custodian and each candidate (n 2 n^2 n 2 pairs) and teleporting the keys at selection with 2 n 2n 2 n dits, after which selection is classical, as for swapping. The advantage of our construction is therefore not raw selection latency but the distribution of trust: the provider holds nothing usable, authority can be split among custodians, and unselected nodes neither act nor learn of the selection.
Scope and threat model. Apart from the noise analysis, our results assume perfect state preparation, gates and measurements and lossless classical communication, and n ≥ 2 n \ge2 n ≥ 2 candidate sites with 2 n + 4 2n+4 2 n + 4 qudits in total (for n = 1 n=1 n = 1 recovery remains exact, but the single share is private only when gcd ( 1 − η 2 , d ) = 1 \gcd(1-\eta^2,d)=1 g cd( 1 − η 2 , d ) = 1 ; Section S8). The privacy guarantee covers honest-but-non-participating parties—sites that observe without deviating from the protocol—and not active corruption of a site’s operations. The provider is trusted to prepare the specified states and to release the keys to one site; it and Alice know the recipient’s identity, and anonymity is claimed only toward the unselected sites, at the level of their registers and the public transcript. The mere presence of a classical exchange between Alice and some site is not hidden by the protocol. If that matters, the dits can be sent privately, or uniformly random dits can be published in rounds without a selection, which is indistinguishable because the genuine ( m A , m B ) (m_A,m_B) ( m A , m B ) are uniform. The two operational requirements stated in the Protocol section mark the boundary explicitly. (i) Key delivery: the key register alone carries nothing, but an unselected site that intercepts all of it holds an authorized set together with its own share and could redeem the gate in Bob’s place, so the keys must travel over an authenticated quantum channel [35 ] . (ii) Classical messages: m A m_A m A and m B m_B m B may be public but must be authenticated. An adversary that alters them by ( δ A , δ B ) (\delta_A,\delta_B) ( δ A , δ B ) learns nothing and can only append the Weyl error Z d k 1 δ B ⊗ X d k 2 δ A Z_d^{k_1\delta_B} \otimes X_d^{k_2\delta_A} Z d k 1 δ B ⊗ X d k 2 δ A of Eq. (24) to the gate. When the keys are held by independent custodians, as discussed in the main text, each custodian’s key is by itself uncorrelated with Alice’s qudit (Section S9).