Mathematical Proofs
This chapter states and proves the model-level identities that SchroSIM implements for its CV Gaussian core and related runtime logic.
1) Notation and Assumptions
We use units with hbar = 1.
For n modes, define the quadrature vector
R = (q1, p1, q2, p2, ..., qn, pn)^T.
The canonical commutation relation is
[Ri, Rj] = i Omega_ij,
where Omega = direct_sum_k [[0, 1], [-1, 0]].
For a state rho, define:
- mean vector
d = <R>, - covariance matrix
V_ij = (1/2) <{Delta Ri, Delta Rj}>, withDelta R = R - d.
This chapter proves identities under the same modeling assumptions used by SchroSIM’s Gaussian path:
- affine symplectic gates for Gaussian unitaries,
- Gaussian channel updates for loss and thermal loss,
- linear-Gaussian conditioning formulas for homodyne and heterodyne,
- explicit compatibility predicates for backend routing.
2) Affine Symplectic Evolution
Theorem 2.1 (Mean and covariance update)
Assume a gate acts as:
R' = S R + u,
with real matrix S and real vector u. Then:
d' = S d + u,V' = S V S^T.
Proof
For the mean:
d' = <R'> = <S R + u> = S <R> + u = S d + u.
For covariance, note Delta R' = R' - d' = S(R - d) = S Delta R. Therefore
V' = (1/2)<{Delta R', Delta R'^T}>
= (1/2)<{S Delta R, (S Delta R)^T}>
= S ( (1/2)<{Delta R, Delta R^T}> ) S^T
= S V S^T.
QED.
Corollary 2.2 (Commutation preservation condition)
[R'i, R'j] = i (S Omega S^T)_ij.
Hence canonical commutation is preserved iff:
S Omega S^T = Omega.
This is exactly the symplectic condition checked by isSymplectic(...).
Corollary 2.3 (Physicality preservation under symplectic maps)
If a covariance matrix satisfies
V + i Omega/2 >= 0
and S Omega S^T = Omega, then V' = S V S^T also satisfies
V' + i Omega/2 >= 0.
Proof
V' + i Omega/2 = S V S^T + i Omega/2
= S V S^T + i S Omega S^T / 2
= S (V + i Omega/2) S^T.
For any vector x, let y = S^T x. Then
x^* (V' + i Omega/2) x = y^* (V + i Omega/2) y >= 0,
since V + i Omega/2 >= 0.
QED.
3) Gaussian Channel Correctness
3.1 Pure loss channel
Theorem 3.1
For one mode with transmissivity eta in [0, 1], model
R_out = sqrt(eta) R_in + sqrt(1-eta) R_env,
with vacuum environment (d_env = 0, V_env = I2/2). Then:
d_out = sqrt(eta) d_in,V_out = eta V_in + (1-eta) I2/2.
In n-mode embedding on mode m, this becomes:
d' = K d,
V' = K V K^T + N,
with K identity except the m-mode block scaled by sqrt(eta), and N zero except diagonal entries (q_m, p_m) equal to (1-eta)/2.
Proof
By linearity:
d_out = <R_out> = sqrt(eta) <R_in> + sqrt(1-eta) <R_env> = sqrt(eta) d_in.
For covariance, using independence of input and environment and zero cross-covariance:
V_out = eta V_in + (1-eta) V_env = eta V_in + (1-eta) I2/2.
Embedding into n modes yields the K/N form by block insertion.
QED.
3.2 Thermal loss channel
Theorem 3.2
With thermal environment mean photon number n_th >= 0, environment covariance is
V_env = ((2 n_th + 1)/2) I2.
Then:
d_out = sqrt(eta) d_in,
V_out = eta V_in + (1-eta) ((2 n_th + 1)/2) I2.
In n-mode embedding:
d' = K d,
V' = K V K^T + N,
where injected diagonal noise at target mode is
(1-eta)(2 n_th + 1)/2.
Proof
Same derivation as pure loss with thermal V_env.
QED.
Theorem 3.3 (Gaussian-channel physicality condition for loss/thermal-loss)
For affine Gaussian channels V' = K V K^T + N, complete positivity is guaranteed by:
N + i (Omega - K Omega K^T)/2 >= 0.
For one-mode loss/thermal-loss blocks in SchroSIM:
K = sqrt(eta) I2,
N = (1-eta) (2 n_th + 1) I2 / 2,
with n_th = 0 for pure loss.
Proof
Compute:
Omega - K Omega K^T = Omega - eta Omega = (1-eta) Omega.
So CP condition becomes:
N + i (1-eta) Omega / 2 >= 0
= (1-eta)/2 * ( (2 n_th + 1) I2 + i Omega ).
When eta in [0,1], prefactor (1-eta)/2 >= 0, so it is enough to show
(2 n_th + 1) I2 + i Omega >= 0.
The eigenvalues of i Omega are +1 and -1, so eigenvalues of the matrix above are:
(2 n_th + 1) +/- 1 = 2 n_th + 2 and 2 n_th.
Both are nonnegative for n_th >= 0. Therefore the channel is CP.
QED.
4) Measurement Conditioning Correctness
All formulas below are standard linear-Gaussian conditioning identities and match the direct implementation used in SchroSIM.
4.1 Homodyne update
Define scalar measurement model:
y = h^T R + nu,
with measurement-noise variance Var(nu) = v_meas (v_meas = 0 for ideal homodyne in the code path). Let prior be R ~ N(d, V).
Define:
mu = h^T d,
s2 = h^T V h + v_meas,
k = V h / s2.
Theorem 4.1
Conditioned on observed outcome y:
d' = d + k (y - mu),V' = V - (V h h^T V) / s2.
Equivalent code form:
d' = d + (Vh) (y-mu)/s2,
V' = V - (Vh)(Vh)^T/s2,
where Vh = V h.
Proof
The joint Gaussian of (R, y) has mean (d, mu) and covariance:
[[V, Vh], [h^T V, s2]].
Applying the Gaussian conditional formula gives exactly the above mean/covariance updates.
QED.
4.2 Heterodyne update
For one target mode, define:
y = H R + nu,
where H selects (q_m, p_m), and nu ~ N(0, Rm) with
Rm = (1/2) I2
for quantum-limited heterodyne added vacuum noise.
Define innovation covariance:
S = H V H^T + Rm.
Theorem 4.2
Conditioned on observed y:
K = V H^T S^{-1},d' = d + K (y - H d),V' = V - K H V.
Proof
Again from multivariate Gaussian conditioning for linear observation model.
QED.
5) Backend Routing Correctness Conditions
Routing is a semantic selection problem over finite predicates.
Let:
requiresFockPath(circuit)be true iff circuit containsinject_fockorinject_cat,fockPathIssues(circuit)enumerate compatibility violations (single-mode and supported-gate checks).
Theorem 5.1 (Routing safety)
Given requested backend in {gaussian, fock, auto, hybrid}:
- if request is
fock, execution is permitted ifffockPathIssuesis empty, - if request is
auto/hybridandrequiresFockPathis false, choose Gaussian path, - if request is
auto/hybridandrequiresFockPathis true, choose Fock path iff compatible; otherwise fail with explicit incompatibility.
Proof sketch
This follows directly from exhaustive case analysis in resolveExecutionBackend(...) and assertFockCompatible(...).
No silent unsafe downgrade exists for incompatible Fock-required circuits; failure is explicit.
QED (by case split over finite branch logic).
6) Fock Truncation and Exponential-Series Error
SchroSIM’s Fock displacement uses:
- finite cutoff
d, - finite Taylor terms
Tin matrix exponential.
Let truncated generator be G_d and truncated exponential approximation be
E_T(G_d) = sum_{k=0}^T G_d^k / k!.
For finite-dimensional operator norm:
||exp(G_d) - E_T(G_d)|| <= exp(||G_d||) * ||G_d||^(T+1) / (T+1)!.
Thus total modeling error has two parts:
- cutoff error from replacing infinite-dimensional dynamics with finite
d, - series truncation error bounded as above for fixed
d.
This chapter treats that as a controlled approximation regime, not a model-exact identity claim.
7) GKP Nearest-Lattice Decode/Correction Semantics
For syndrome value x, lattice spacing Delta > 0, define:
n = round(x / Delta),
x_lattice = n Delta,
correction = -x_lattice,
residual = x - x_lattice.
Theorem 7.1
residualis exactly the signed distance fromxto the nearest lattice point.- For nearest-integer rounding,
|residual| <= Delta/2(tie behavior at exactly half-spacing depends on rounding convention). - Post-correction displacement by
correctionshifts syndrome by-x_lattice.
Proof
By definition of nearest-integer rounding and lattice construction.
QED.
8) Implementation Correspondence Table
| Proven item | Implementation source |
|---|---|
| Affine symplectic update | core-swift/Sources/SchroSIM/core/evolution/symplectic_evolution.swift |
| Symplectic condition check | core-swift/Sources/SchroSIM/core/evolution/symplectic_evolution.swift |
| Loss and thermal-loss channel equations | core-swift/Sources/SchroSIM/core/evolution/gaussian_channels.swift |
| Homodyne/heterodyne conditioning | core-swift/Sources/SchroSIM/core/operators/measurement.swift |
| Fock displacement via matrix exponential | core-swift/Sources/SchroSIM/backends/fock/fock_ops.swift |
| Backend routing predicates | core-swift/Sources/SchroSIM/src/core/backend_routing.swift |
| GKP nearest-lattice decode semantics | core-swift/Sources/SchroSIM/src/core/gkp_decoder.swift |
9) Model-Exact Claim Boundary (Current)
Within this docs set, "model-exact" means exact with respect to the implemented Gaussian model equations and conditional-update rules, under the chosen parameterization and arithmetic behavior.
Approximation enters when:
- non-Gaussian dynamics require finite Fock cutoff,
- matrix exponentials are numerically truncated,
- optional effective models are used for workflow practicality.
This aligns with the taxonomy in 10-theory-and-exactness-scope.md: model-exact Gaussian core plus controlled approximation regimes outside that boundary.