---
title: Shifted Gaussian Encoding
url: https://www.emergentmind.com/topics/shifted-gaussian-encoding
type: topic
---

# Shifted Gaussian Encoding

Searching arXiv for the specified paper and closely related uses of shifted Gaussian encoding.
Search results will help anchor citations and confirm related terminology across domains.
Searching arXiv: "Rethinking Positional Encoding" 2107.02561, shifted Gaussian encoding, GKP Gaussian breeding, surface-GKP designed bias.
Shifted Gaussian encoding denotes a family of constructions in which information is represented by Gaussian functions whose centers are deliberately translated across a domain. In the cited literature, this motif appears in several technically distinct forms: as a positional embedding for coordinate-based MLPs, with features $\phi(x) = [\exp(-\|x-\mu_k\|^2/(2\sigma^2))]_{k=1\ldots m}$ [2107.02561]; as approximate Gottesman–Kitaev–Preskill (GKP) codewords built from displaced squeezed Gaussians in propagating light [2212.05436]; as a squeezing-deformed GKP lattice used to bias logical noise in a concatenated surface–GKP code [2004.00541]; as a smooth integer representation based on localized Gaussian bumps with alternating coefficients [2505.02259]; and as shifted-center explicitly correlated Gaussians in pre-Born–Oppenheimer variational calculations [1807.11727]. Across these settings, the operative degrees of freedom are the Gaussian centers, widths, and, where relevant, amplitudes and projection operations.

## 1. Mathematical forms and shared structure

A basic shifted Gaussian positional embedding associates to a coordinate $x \in [0,C]$ an $m$-dimensional feature vector
\[
\phi(x) = \bigl[\exp(-\|x-\mu_k\|^2/(2\sigma^2))\bigr]_{k=1\ldots m},
\]
with equally spaced shifts $\mu_k = (k-1)s$ and $s=C/m$. In $d>1$, a separable construction samples $m_i$ points per axis with standard deviation $\sigma_i$ and concatenates the per-axis vectors, yielding total embedding dimension $\sum_i m_i$ [2107.02561].

In continuous-variable quantum information, approximate GKP codewords are superpositions of equally spaced, squeezed Gaussians in the position basis. With squeezing parameter $\Delta$ and envelope width $\kappa$, the unnormalized codewords are
\[
\ket{\tilde k_{\Delta,\kappa}}=
\sum_{s\in\mathbb Z}
\exp\!\Biggl[-\frac{1}{2\kappa^2}\bigl((2s+k)\sqrt\pi\bigr)^2\Biggr]
\,\hat D\!\bigl((2s+k)\sqrt\pi\bigr)\,\ket{S_{\Delta}},
\qquad k=0,1,
\]
where $\ket{S_\Delta}=\hat S(\Delta)\ket0$ and $\hat D(d)=e^{-i d \hat p}$ [2212.05436].

A smooth integer encoding uses a finite sum of shifted Gaussian bumps,
\[
f_N(t)=\sum_{n=1}^{N} a_n\,\exp\!\Bigl(-\frac{(t-\mu_n)^2}{2\sigma^2}\Bigr),
\]
with $\mu_n=n$, $\sigma=\delta>0$, and coefficients
\[
a_n=\frac{(1/2)^n+(-1)^n}{n}.
\]
Here the integer is not an explicit parameter of a symbolic representation; it is recovered from the integral balance of the smooth function [2505.02259].

In few-body quantum chemistry, a floating explicitly correlated Gaussian (FECG) is
\[
\phi(r;A,s)=\exp[-(r-s)^T A (r-s)],
\]
with real symmetric positive-definite block-diagonal $A\in\mathbb R^{3N_p\times 3N_p}$ and shift vector $s\in\mathbb R^{3N_p}$. The shifted center $s$ supplies the basis with localization flexibility unavailable to origin-centered ECGs [1807.11727].

These examples share a common translation mechanism but differ in what is being encoded: Euclidean coordinates, logical qubits, effective error bias, integers, or many-particle wavefunctions. This suggests that “shifted Gaussian encoding” is best understood as a structural motif rather than a single canonical formalism.

## 2. Shifted Gaussian positional encoding in coordinate MLPs

For coordinate-based MLPs, the shifted Gaussian embedder is analyzed through two metrics defined on the embedding matrix $\Phi\in\mathbb R^{m\times N}$ formed from sampled coordinates $\{x_j\}$. The stable rank
\[
\mathrm{srank}(\Phi)\coloneqq \|\Phi\|_F^2/\|\Phi\|_2^2
\]
measures the effective number of non-negligible singular values, and higher stable rank gives more capacity to memorize arbitrary $y$'s. The embedded-distance quantity
\[
d_{\mathrm{emb}}(x,x')\coloneqq \langle \phi(x),\phi(x')\rangle
\]
is desired to be a monotonic function of the original distance $\|x-x'\|$ so that nearby coordinates remain similar and far points remain dissimilar in feature space [2107.02561].

For the Gaussian embedder, the continuous-limit inner product satisfies
\[
D(x,x')=\int \exp(-(t-x)^2/(2\sigma^2))\,\exp(-(t-x')^2/(2\sigma^2))\,dt
=\sqrt{\pi}\sigma\,\exp(-\|x-x'\|^2/(4\sigma^2)).
\]
After zero-mean centering and normalizing, one may ignore the $\sqrt{\pi}\sigma$ prefactor and view
\[
d_{\mathrm{emb}}(x,x')\propto \exp(-\|x-x'\|^2/(4\sigma^2)).
\]
When $m$ and $N$ are large enough,
\[
\mathrm{srank}(\Phi)\approx \min\{N,(2\sqrt{\pi}\sigma)^{-1}\}.
\]

The bandwidth parameter $\sigma$ controls the central trade-off. Decreasing $\sigma$ raises the upper bound $1/(2\sqrt{\pi}\sigma)$ on stable rank and therefore increases memorization capacity for high-frequency content, but it also makes $d_{\mathrm{emb}}$ decay rapidly, so nearby points become nearly orthogonal. Increasing $\sigma$ reduces stable rank and can underfit high-frequency structure, while making interpolation overly smooth. The embedding dimension must satisfy $m \ge \mathrm{bandwidth}\ B \simeq 2C/\sigma$ to reach the stable-rank ceiling, and Nyquist sampling requires $s\lesssim \sigma$ to preserve the inner-product integral. A practical one-dimensional rule chooses $\sigma$ so that
\[
d_{\mathrm{emb}}(\Delta x)=\exp(-\Delta x^2/(4\sigma^2))
\]
stays above a small threshold $\epsilon=10^{-k}$ for nearest-neighbor spacing $\Delta x=C/m$, giving
\[
\sigma \approx \frac{C/m}{2\sqrt{k\ln 10}}.
\]
In $d$ dimensions, the same per-axis rule is applied under separable sampling, while keeping total dimension linear in $d$ through concatenation [2107.02561].

Empirically, the paper studies 1D and 2D image-signal reconstruction with coordinate-MLPs. Baselines include no encoding, fixed sinusoid (“basic”), Random Fourier Features (RFF), impulse, square wave, and random noise. In 1D experiments with a linear one-layer network, varying $m\in\{16,\ldots,10\,000\}$ over 10 seeds, the Gaussian embedder with $\sigma$ set by $\sigma\cdot m\approx 1/(2\sqrt{2\pi})$ yields test PSNR $\sim 34\,\mathrm{dB}$ with small error bars, whereas RFF is volatile at low $m$ and stabilizes only at large $m$. In 2D experiments with a 4-layer ReLU MLP, separable Gaussian sampling along $x,y$ and optionally additional rotated axes matches or exceeds RFF in test PSNR, for example $\sim 32\,\mathrm{dB}$ at $m\simeq 1\,000$, while requiring smaller $m$ for similar quality. Rank-versus-distance plots place Gaussian and RFF in a sweet spot of intermediate stable rank and reasonable distance preservation, whereas impulse and random noise over-rank and sine and square-wave embeddings under-rank. Training dynamics further show faster convergence, lower hidden-layer stable-rank growth, and higher final accuracy than unencoded or basic encodings [2107.02561].

## 3. Shifted Gaussian codewords and Gaussian breeding for GKP qubits

In propagating-light implementations of the GKP code, logical basis states are approximate codewords formed from equally spaced displaced squeezed vacua with a Gaussian envelope. In the square-lattice case one often takes $\kappa=\Delta$, and in the limit $\Delta,\kappa\to\infty$ the logical states approach ideal Dirac-comb codewords [2212.05436].

The paper’s central operation is the coherent bifurcation $\mathcal B_w$, which acts linearly on superpositions of displaced squeezed vacua according to
\[
\hat D(d)\ket{S_\Delta}\mapsto [\hat D(d-w)+\hat D(d+w)]\ket{S_\Delta}.
\]
Iterating it $N$ times with step size $w=\sqrt{\pi}$ yields
\[
\mathcal B_{\sqrt\pi}^{(N)}[\ket{S_\Delta}]
=
\sum_{l=0}^N \binom Nl\,\hat D\bigl((2l-N)\sqrt\pi\bigr)\ket{S_\Delta}
\approx
\ket{\tilde k_{\Delta,\sqrt{N\pi}}},
\qquad
k\equiv N\pmod 2.
\]
This constructs a comb of $N+1$ peaks [2212.05436].

A physical implementation starts from two single-mode squeezers and combines them either on a beam splitter of transmittance $T$ or through a QND gate $\exp[i g \hat p_1 \hat x_2]$, followed by photon-number-resolving detection on the second mode. Detecting $n$ photons heralds an approximate two-peak superposition
\[
\mathcal G_w[\ket{S_{\Delta_1}}]\simeq [\hat D(-w)+\hat D(w)]\ket{S_{\Delta_1}},
\qquad
w=\sqrt{2n}/\Delta_1,
\]
up to small overlap errors, and for even $n$ the wavefunction is symmetric. The QND variant $\tilde{\mathcal G}_w$ commutes with displacements, so linearity holds exactly and the coherent bifurcation becomes fully iterable [2212.05436].

Envelope control is implemented by heralding $n=0$ in the QND setup, which realizes an unsharp measurement in $p$ equivalent to multiplication by a Gaussian in momentum space, or operatorially
\[
e^{-t\hat p^2}.
\]
After a $90^\circ$ phase rotation one also obtains $e^{-t\hat x^2}$. These damping steps set the global envelope $\kappa$ without changing the internal peak spacing. To prepare an arbitrary superposition $\alpha\ket{0_L}+\beta\ket{1_L}$, one first constructs a seed
\[
\ket{\mathrm{seed}}_{\Delta,\alpha,\beta}
=
\mathcal B_w^{(2)}[\ket{S_\Delta}]
\approx
\alpha\ket{S_\Delta}+\beta\,\hat D(\sqrt\pi)\ket{S_\Delta},
\]
then applies $\mathcal B_{\sqrt\pi}^{(N)}$ to obtain
\[
\mathcal B_{\sqrt\pi}^{(N)}[\ket{\mathrm{seed}}]
\approx
\hat X^N\Bigl[\alpha\ket{\tilde 0_{\Delta,\sqrt{N\pi}}}
+\beta\ket{\tilde 1_{\Delta,\sqrt{N\pi}}}\Bigr].
\]
Via Bloch–Messiah reduction, all on-line squeezers and QND gates can be replaced by a fixed interferometer, off-line squeezing, and photon-number detection; in practice, using the same $(\Delta_1,n)$ at each stage reduces the beam-splitter count to $N$ [2212.05436].

The performance formulas quantify scalability. The per-round success probability is
\[
P(n)=\frac{\sqrt2\,(2n)!}{4^n(n!)^2}\;t^n\,(t+2)^{-n-\tfrac12},
\qquad
t=g^2\frac{\Delta_1^2}{\Delta_2^2},
\]
maximized at $t=4n$, giving
\[
P_{\max}(n)
=
\frac{(2n)!}{2^n(n!)^2}
\sqrt{\frac{1}{2n+1}\Bigl(\frac{n}{2n+1}\Bigr)^n}
\sim \mathcal O(n^{-1/2}).
\]
After $N$ bifurcations with $w=\sqrt\pi$, the comb envelope has variance $\kappa^2=N\pi$, and in the square-lattice case additional damping yields total squeezing parameter $\Delta^2=N\pi$. Each two-peak building block has overlap
\[
F_n\approx 1-0.03/n,
\]
so the overall infidelity remains small provided $n\gtrsim 10$. Threshold analyses cited in the summary typically require $\gtrsim 10\,\mathrm{dB}$ of squeezing, that is $\Delta^2\ge 10$, together with logical-state fidelity $>0.99$. With $n=16$, one achieves $\Delta^2\approx 10.6\,\mathrm{dB}$ and state fidelities $>0.998$ after $N=4$ rounds. For generation of $\ket{0_L}$ to $F>0.99$, the example resource count is $n=16$ photons per bifurcation, $N=3$–$4$ rounds, per-round success probability $P(n=16)\approx 2\times 10^{-2}$, and total success probability $\approx 10^{-6}\ldots 10^{-7}$, potentially improvable by multiplexing or quantum memory. For arbitrary magic states, the seed succeeds with probability $\approx 10^{-5}$ and the full process reaches $\approx 10^{-10}$ with $F>0.99$ [2212.05436].

## 4. Noise-biased surface–GKP encoding through squeezing deformation

A different use of shifted Gaussian structure appears in the concatenated surface–GKP code, where each bosonic GKP mode is first transformed by a single-mode squeezing unitary
\[
U(k)=\exp[-i(\ln k)(q p+p q)/2]
\]
associated with the symplectic matrix
\[
S_k=\mathrm{diag}(k,1/k).
\]
Under conjugation, the quadratures become $q'=kq$ and $p'=(1/k)p$, and displacement operators transform as
\[
D(\nu_q,\nu_p)\mapsto D(k\nu_q,\nu_p/k).
\]
The square-lattice GKP stabilizer lattice
\[
\mathcal L_\square=\{(2\sqrt\pi n_1,2\sqrt\pi n_2)\mid n_i\in\mathbb Z\}
\]
is deformed into the rectangular lattice
\[
\mathcal L_k=S_k\mathcal L_\square
=
\bigl\{(2\sqrt\pi k n_1,2\sqrt\pi n_2/k)\bigr\},
\]
with dual lattice
\[
\mathcal L_k^\perp
=
\bigl\{(\sqrt\pi k m_1,\sqrt\pi m_2/k)\bigr\}.
\]
In the rescaled coordinates, the logical Pauli operators are
\[
X_L=D(\sqrt\pi k,0)=e^{-i\sqrt\pi k p'},
\qquad
Z_L=D(0,\sqrt\pi/k)=e^{i\sqrt\pi q'/k}
\]
[2004.00541].

This deformation converts isotropic Gaussian displacement noise into anisotropic noise. An original channel
\[
\mathcal N_{\sigma^2}(\rho)
=
\frac1{2\pi\sigma^2}
\int e^{-\|\nu\|^2/(2\sigma^2)}D(\nu)\rho D(\nu)^\dagger\,d^2\nu
\]
becomes an anisotropic channel with covariance
\[
\Sigma_k=\mathrm{diag}(\tilde\sigma_Q^2,\tilde\sigma_P^2)
=
\mathrm{diag}(\sigma^2/k^2,k^2\sigma^2).
\]
Equivalently, the quadrature jitters satisfy
\[
\tilde\sigma_Q=\sigma/k,\qquad \tilde\sigma_P=k\sigma,
\]
and the bias parameter is
\[
B=\frac{\tilde\sigma_P}{\tilde\sigma_Q}=k^2.
\]
For $k\gg 1$, the $p'$-quadrature noise, which causes logical $X$ errors, is much larger than the $q'$-quadrature noise, which causes logical $Z$ errors [2004.00541].

Using nearest-lattice-point decoding at the GKP level, the residual logical-qubit error probabilities satisfy approximately
\[
q_X\approx 1-\Phi\!\Bigl(\tfrac14\sqrt{\tfrac{2\pi k^2}{\sigma^2}}\Bigr),
\qquad
q_Z\approx 1-\Phi\!\Bigl(\tfrac14\sqrt{\tfrac{2\pi}{k^2\sigma^2}}\Bigr),
\]
so $q_Z\ll q_X$ for large $k$. The encoding is then relabeled so that the small-noise quadrature is associated with the qubit’s $Y$ basis; in the paper the GKP “0/1” states are mapped to $\lvert \pm i\rangle$. The resulting single-qubit channel has
\[
p_X=q_Z(1-q_Y),\qquad p_Y=q_Y(1-q_Z),\qquad p_Z=q_Z q_Y,
\]
with $q_Y=q_X$, and in the large-$k$ limit one finds $p_Z\ll p_X\approx p_Y$. The purpose is to realize a Pauli-$Z$-biased qubit channel that the outer surface code can exploit [2004.00541].

For decoding, the study uses the Bravyi–Suchara–Vargo tensor-network decoder with bond dimension $\chi$ up to $100$, which approaches maximum-likelihood decoding as $\chi\to\infty$. Monte Carlo simulations show that, even without using GKP side information, an optimal choice of $k>1$ raises the threshold from $\sigma_c\simeq 0.54$ at $k=1$ to $\sigma_c\simeq 0.58$ for $k\approx \sqrt3$. Incorporating GKP analog-syndrome information into the prior further boosts the threshold to $\sigma_c\approx 0.61$; even for $k=2$ one observes $\sigma_c\approx 0.606$, and a similar effect persists on asymmetric hexagonal lattices with $\sigma_c\approx 0.605$ [2004.00541]. The paper frames this as a two-step map: single-mode squeezing reshapes isotropic displacements into an anisotropic Gaussian channel, and reinterpretation of the logical axes turns the small-variance direction into predominantly $Z$ errors.

## 5. Smooth integer encoding by shifted Gaussian integral balance

The integer-encoding construction of [2505.02259] represents $N\in\mathbb N$ by a smooth bump sum
\[
f_N(t)=\sum_{n=1}^{N} a_n\exp\!\Bigl(-\frac{(t-\mu_n)^2}{2\delta^2}\Bigr),
\qquad
\mu_n=n,
\]
with $K(N)=N$ bumps and coefficients
\[
a_n=\frac{(1/2)^n+(-1)^n}{n}
=
\frac{1}{n}(1/2)^n+\frac{(-1)^n}{n}.
\]
Because $\sum_{n=1}^\infty a_n=\ln 2-\ln 2=0$, the total integral tends to zero in the large-$N$ limit [2505.02259].

Exact Gaussian integration gives
\[
\int_{-\infty}^{\infty}
\exp\!\Bigl(-\frac{(t-n)^2}{2\delta^2}\Bigr)\,dt
=
\delta\sqrt{2\pi},
\]
hence
\[
I(N):=\int_{-\infty}^{\infty} f_N(t)\,dt
=
\delta\sqrt{2\pi}\sum_{n=1}^{N} a_n
=
\delta\sqrt{2\pi}\,S(N).
\]
Since $\lvert a_n\rvert\le C\rho^n$ for some $\rho<1$, the tail obeys $\sum_{n>N}a_n=O(\rho^{N+1})$, so $I(N)\to 0$ exponentially fast. Because the coefficients alternate in sign, $S(N)$ and $I(N)$ oscillate about zero with exponentially decaying amplitude. Recovery is defined by the first near-cancellation:
\[
|I(N)|<\varepsilon,
\qquad
I(N)\text{ locally minimal in magnitude},
\]
or operationally
\[
N=\min\{k\in\mathbb N:\ |I(k)|<\varepsilon\}.
\]
The summary states that no two integers share the same small-magnitude value once $\varepsilon$ lies below the preceding oscillation amplitude [2505.02259].

Several inversion procedures are given. Threshold-based inversion from a measured integral $I^*$ uses
\[
\hat N=\min\{k\ge 1:\ |I(k)-I^*|<\varepsilon\},
\]
and is stable if measurement noise satisfies $|\Delta I|<\varepsilon/2$. A tabulation-and-binary-search scheme precomputes $\{I(1),\dots,I(N_{\max})\}$ and locates the smallest compatible $k$ in $O(\log N_{\max})$ time. A spline interpolation $\mathcal I(x)\approx I(\lfloor x\rfloor)$ may be inverted numerically via
\[
x^*=\mathcal I^{-1}(I^*),\qquad \hat N=\mathrm{round}(x^*),
\]
and Newton’s method can be applied to $\mathcal I(x)-I^*=0$. The local error relation is
\[
\Delta N\approx \Delta I / (\delta\sqrt{2\pi}|a_{k+1}|).
\]
A piecewise analytical inversion is also supplied:
\[
\mathcal I(x)
=
\delta\sqrt{2\pi}\Bigl(\sum_{n=1}^{k} a_n +(x-k)a_{k+1}\Bigr),
\qquad
k\le x\le k+1,
\]
so that if $I\in[\mathcal I(k),\mathcal I(k+1)]$ then
\[
x=
k+\frac{I-\delta\sqrt{2\pi}\sum_{n=1}^{k} a_n}{\delta\sqrt{2\pi}\,a_{k+1}}.
\]

The construction extends to tuples $\vec N=(N_1,\dots,N_d)$ by summing $d$-dimensional Gaussian bumps centered at $(n_1,\dots,n_d)$ with coefficients such as
\[
a_{n_1,\dots,n_d}
=
\prod_{i=1}^{d}\bigl((1/2)^{n_i}+(-1)^{n_i}\bigr)/n_i,
\]
producing
\[
I(\vec N)
=
(2\pi)^{d/2}\delta^d
\sum_{n_1=1}^{N_1}\cdots\sum_{n_d=1}^{N_d}
a_{n_1,\dots,n_d}.
\]
The paper further notes that the map $N\mapsto I(N)$ is differentiable, or $C^\infty$ under a smooth blending replacement for the floor-based extension, enabling uses such as a differentiable layer “SmoothInteger,” a regularizer for integer constraints, and soft-argmax with guaranteed exact recovery [2505.02259].

## 6. Shifted-center explicitly correlated Gaussians in pre-Born–Oppenheimer calculations

In pre-Born–Oppenheimer quantum calculations, shifted Gaussian encoding appears as floating explicitly correlated Gaussian basis functions. For an $N_p$-particle system in three dimensions, a single basis element is
\[
\phi(r;A,s)=\exp[-(r-s)^T A (r-s)].
\]
With $x\equiv r$, $u\equiv As$, and $c\equiv -s^TAs$, one may equivalently write
\[
\phi(x;A,u)=\exp[-x^T A x + 2u^T x + c].
\]
In practice, $A=\bar A\otimes I_3$ with $\bar A\in\mathbb R^{N_p\times N_p}$, while the shifted center $s$ is a $3N_p$-vector of Gaussian centers in the laboratory frame. The role of the shift is to describe localized structures, including nuclei positions, more flexibly than origin-centered ECGs [1807.11727].

The overlap of two FECGs is analytic:
\[
\langle \phi_I\mid \phi_J\rangle
=
\Bigl(\frac{\pi^{3N_p}}{\det(A_I+A_J)}\Bigr)^{3/2}
\exp\!\bigl[
s_I^T A_I s_I+s_J^T A_J s_J-e^T(A_I+A_J)^{-1}e
\bigr],
\]
where $e=A_I s_I+A_J s_J$. For $I=J$,
\[
\langle \phi_I\mid \phi_I\rangle
=
\bigl(\pi^{3N_p}/\det(2A_I)\bigr)^{3/2},
\]
so the normalized basis function is
\[
\tilde\phi_I(x)=N_I\exp[-(x-s_I)^T A_I(x-s_I)],
\qquad
N_I=(\det(2A_I)/\pi^{3N_p})^{3/4}.
\]

Because a general FECG is not an eigenfunction of total angular momentum or parity, projection is required. Rotation-inversion projection uses
\[
P^L_{MM'}
=
\frac{2L+1}{8\pi^2}
\int_0^{2\pi}d\alpha
\int_0^\pi \sin\beta\,d\beta
\int_0^{2\pi}d\gamma\,
D^{L*}_{MM'}(\alpha,\beta,\gamma)\,R(\alpha,\beta,\gamma),
\]
with $R(\alpha,\beta,\gamma)=e^{-i\alpha L_z}e^{-i\beta L_y}e^{-i\gamma L_z}$. Parity projection is
\[
P^\pi=\tfrac12[1+\pi \hat I].
\]
The projected function
\[
\phi_M^{[L,\pi]}(r)=P^\pi P^L_{MM}\phi(r)
\]
then satisfies the requisite $L^2$, $L_z$, and parity eigenvalue equations [1807.11727].

The triple Euler-angle integral is evaluated numerically by a product Gauss–Legendre scheme. With $n$ nodes per angle,
\[
P^L_{MM}\phi(r)\simeq
\frac{2L+1}{8\pi^2}
\sum_{i,j,k=1}^{n}
w_i^{(\alpha)}w_j^{(\beta)}w_k^{(\gamma)}
D^{L*}_{MM}(\alpha_i,\beta_j,\gamma_k)\,
R(\alpha_i,\beta_j,\gamma_k)\phi(r).
\]
Typical values $n\approx 20$–$30$ suffice to converge $\langle L^2\rangle$ to a few $10^{-4}$, and the naive cost scaling $N_b^2 n^3$ can be reduced to $N_b^2 n$ by exploiting idempotency and Hermiticity of the projector. For on-the-fly optimization, a nested Gauss–Kronrod rule may be used [1807.11727].

Parameter optimization proceeds through competitive selection and Powell’s derivative-free refinement, often in a two-stage procedure: optimize non-projected FECGs by $\langle H\rangle$ minimization, then solve the linear variational problem with projected functions. Fully projected optimization is possible but more expensive. The reported benefit of shifted centers is that Gaussians can be localized at arbitrary interparticle distances, so fewer functions are needed to capture nuclear motion; the shift vector also explicitly encodes global translation, rotation, and internal equilibrium geometry [1807.11727].

The principal application reported is the five-particle $\mathrm{H}_3^+$ ion with target state $L=1$, $M=0$, parity $\pi=-1$. Basis sets range from a small $N_b=3$ test through $N_b=120$ to $N_b=1\,560$. For the largest projected basis,
\[
E=-1.322826\,E_h,\qquad \langle p\rangle=-0.999994,\qquad \langle L^2\rangle=2.000055,
\]
and extrapolation with $E(N_b)=a+b/N_b$ gives
\[
E(\infty)=-1.323005\,E_h.
\]
The previous best non-shifted ECG result cited is $E=-1.3143836\,E_h$, differing by approximately $+8.44\,\mathrm{m}E_h$, while a perturbative non-adiabatic model estimate is $E=-1.323146\,E_h$ [1807.11727].

## 7. Comparative interpretation and recurrent trade-offs

The cited literature does not present a single universal theory covering all uses of shifted Gaussian encoding. Instead, each field uses shifted centers to control a different structural property. In positional encoding, the decisive variables are stable rank and embedded-distance preservation, both governed by $\sigma$ and sampling density [2107.02561]. In GKP-state synthesis, the central issues are iterable superposition growth, envelope control, heralding probability, and fidelity under repeated coherent bifurcation [2212.05436]. In the surface–GKP setting, squeezing-induced anisotropy is used to transform isotropic Gaussian displacement noise into a biased qubit channel that is better matched to the surface code [2004.00541]. In smooth integer encoding, the essential mechanism is oscillatory near-cancellation of the total integral with exponentially decaying tails [2505.02259]. In shifted-center ECGs, the gain is variational flexibility at the cost of numerical projection onto symmetry sectors [1807.11727].

A common misconception would be to treat these constructions as interchangeable merely because they use translated Gaussians. The source material instead indicates domain-specific semantics for the same geometric operation. In one case the centers $\mu_k$ sample a coordinate domain; in another they mark lattice displacements of squeezed vacua; in another they are laboratory-frame centers of many-particle basis functions. This suggests that the unifying concept is not a shared application, but a shared representational device: localization through Gaussian basis elements whose positions are shifted to encode structure, constraints, or discrete alternatives.

A second recurring theme is that translation alone is insufficient; usefulness depends on accompanying control variables. Positional encoding requires an appropriate balance between rank and distance preservation. GKP breeding requires damping to shape the global envelope without changing the peak spacing. Surface–GKP biasing requires reinterpretation of logical axes after squeezing. Integer encoding requires carefully designed alternating coefficients and a recovery rule based on local minima of $|I(N)|$. Projected FECGs require explicit symmetry projection to restore good quantum numbers. The literature therefore presents shifted Gaussian encoding not as a generic recipe, but as a design pattern whose efficacy depends on how shifts, widths, amplitudes, and post-processing are coupled to the target problem.

Source: https://www.emergentmind.com/topics/shifted-gaussian-encoding