---
title: Quantum Shifted Phase Estimation
url: https://www.emergentmind.com/topics/quantum-shifted-phase-estimation
type: topic
---

# Quantum Shifted Phase Estimation

Quantum Shifted Phase Estimation denotes a family of phase-estimation procedures in which the phase variable is recentered, offset, or left-shifted before inference. In adaptive interferometric metrology, the shift is a controllable offset $\theta_k$ that steers Ramsey-type likelihoods into high-sensitivity regions and resolves phase wrapping; in algorithmic quantum phase estimation, the shift may mean estimating the fractional part of $2^s\phi$ or canceling already-estimated suffix bits; in quantum signal processing and quantum phase processing, the shift is the transformed variable $\theta=\phi-\phi_0$ or the shifted unitary $U_{\mathrm{shift}}=e^{-i\phi_0}U$ [2303.01517][2507.06711][2604.01205]. The common motivation is that phase is periodic, local sensitivity is nonuniform, and explicit shifts can convert a globally ambiguous problem into a locally informative one, with consequences for error scaling, circuit depth, and noise robustness [2209.14278][2211.11973].

## 1. Terminological scope and conceptual variants

A central feature of the literature is that “quantum shifted phase estimation” is not a single standardized protocol. The term is used across at least three partially overlapping settings: interferometric quantum metrology, circuit-level eigenphase estimation, and signal-engineered phase learning.

| Context | Shift mechanism | Representative source |
|---|---|---|
| Adaptive metrology | controllable Ramsey offset $\theta_k$ | [2303.01517] |
| Circuit QPE | left shift of binary phase via $2^s\phi$ | [2507.06711] |
| Kitaev-style sampling | suffix-canceling phase shifts | [1902.11168] |
| QSP/QPP | shifted variable $\theta=\phi-\phi_0$ | [2604.01205] |
| Interferometric blind-spot avoidance | reference phase $\phi_{\mathrm{ref}}$ | [1611.01498] |
| Time-series estimation | demodulation by test frequency $\theta$ | [2211.11973] |

The metrological and algorithmic usages are conceptually distinct. Interferometric phase-shift metrology studies estimation of a physical phase encoded by a generator and analyzes sensitivity through FI, QFI, CRB, Bayesian bounds, and circular estimators. By contrast, quantum computing’s QPE algorithm estimates an eigenphase of a unitary via controlled operations and inverse QFT; “shifted” variants there refer to circuit transformations, bit targeting, or reduced-resource inference rather than interferometric operating-point control [1804.10048].

This terminological multiplicity matters because statements about optimality, scaling, and robustness are not interchangeable across these settings. A claim about $O(1/N_{\mathrm{tot}}^2)$ MSE in an adaptive Ramsey protocol, for example, is not the same type of claim as a query-complexity bound of $O((1/\delta)\log(1/\epsilon)\log(1/\delta))$ for a QPP-based eigenphase search [2303.01517][2209.14278].

## 2. Adaptive Bayesian shifted estimation in interferometric metrology

In the adaptive Bayesian formulation, the problem is to estimate an unknown $2\pi$-periodic phase $\phi\in[0,2\pi)$ encoded by a unitary such as $U(\phi)=e^{-i\phi H}$ or $R_z(\phi)=e^{-i\phi Z/2}$, using coherence rather than inter-probe entanglement [2303.01517]. The protocol assumes an effective two-level subspace with eigenstates $|\psi_0\rangle,|\psi_1\rangle$ such that
$$
U(\phi)|\psi_0\rangle=|\psi_0\rangle,\qquad U(\phi)|\psi_1\rangle=e^{i\phi}|\psi_1\rangle,
$$
and prepares the QFI-optimal superposition
$$
|\Psi\rangle=\frac{|\psi_0\rangle+|\psi_1\rangle}{\sqrt{2}}.
$$

At step $k$, a Ramsey-like experiment uses interrogation length $m_k$, applies a known phase shift $\theta_k$, and returns a binary outcome $x_k\in\{0,1\}$. The single-shot likelihood is
$$
P(x_k=0\mid \phi,m_k,\theta_k)=\frac{1+V_k\cos(m_k\phi+\theta_k)}{2},
$$
$$
P(x_k=1\mid \phi,m_k,\theta_k)=\frac{1-V_k\cos(m_k\phi+\theta_k)}{2},
$$
with visibility $V_k\in[0,1]$ capturing SPAM errors, decoherence, and amplitude damping. In the instantiated noise model,
$$
V_k=\alpha \beta^{m_k},\qquad p_0(\phi;m_k,\theta_k)=\frac{1+\alpha\beta^{m_k}\cos(m_k\phi+\theta_k)}{2}.
$$
If $\nu_k$ repetitions are performed at fixed $(m_k,\theta_k)$, the resource budget is $N_{\mathrm{tot}}=\sum_k m_k\nu_k$ and the data are modeled binomially [2303.01517].

Because $\phi$ is circular, inference is performed on $[0,2\pi)$ with circular statistics. The posterior update is
$$
p_k(\phi)\propto p(x_k\mid \phi,m_k,\theta_k,\nu_k)\,p_{k-1}(\phi),
$$
with estimators including the posterior MAP and the circular mean
$$
\hat\phi_{\mathrm{CM}}=\arg \int_0^{2\pi} e^{i\phi}p_k(\phi)\,d\phi.
$$
Uncertainty is quantified by shortest circular credible arcs. This is not incidental: multiple peaks arise from $\cos(m_k\phi+\theta_k)$, so phase identification is not automatic. The shifted design explicitly constructs arcs $\Theta_k$ on which inversion becomes single-valued [2303.01517].

The adaptive step is the defining shifted-phase mechanism. The offset is chosen as
$$
\theta_k=\frac{\pi}{2}-m_k\theta_c,
$$
where $\theta_c$ is the center of the current credible arc, so that $m_k\phi+\theta_k\approx \pi/2$ and the likelihood derivative with respect to $\phi$ is maximal. The next arc is
$$
\Theta_k=\left[\theta_c-\frac{\pi}{2m_{k+1}},\ \theta_c+\frac{\pi}{2m_{k+1}}\right],
$$
which ensures that $\cos^{-1}(\cdot)$ is single-valued there. The interrogation length doubles according to $m_k\approx 2^{k-1}$ up to the smaller of the hardware limit and the noise-optimal depth. The paper states that this “phase shifting” strategy is the central mechanism that turns ambiguous multi-peak posteriors into a single dominant peak while using minimal resources, and that it is exactly what shifted phase estimation calls for [2303.01517].

In the noiseless regime, with visibility approximately one up to the largest used depth and with resource allocation concentrated at the largest interrogation length, the adaptive protocol achieves
$$
\mathrm{MSE}=O(1/N_{\mathrm{tot}}^2),\qquad \mathrm{MAE}=O(1/N_{\mathrm{tot}}).
$$
The paper further states that this family of coherence-based, non-entangling algorithms had previously not exhibited optimal quadratic scaling in MAE and MSE, and that the adaptive Bayesian design closes that gap [2303.01517].

## 3. Shifted binary-phase extraction in circuit quantum phase estimation

In algorithmic QPE, the shift typically refers to a transformation of the binary expansion of the eigenphase rather than to interferometric operating-point control. If
$$
U|\psi\rangle=e^{2\pi i\phi}|\psi\rangle,\qquad \phi=0.b_1b_2b_3\ldots,
$$
then the shifted method introduces an integer $s\ge 0$ and estimates the fractional part of $2^s\phi$:
$$
2^s\phi=\lfloor 2^s\phi\rfloor+\{2^s\phi\},
$$
with
$$
\{2^s\phi\}=0.b_{s+1}b_{s+2}\ldots.
$$
QSPE then returns an $n$-bit estimate of $0.b_{s+1}\ldots b_{s+n}$, directly targeting lower-order phase bits [2507.06711].

Two constructions are given. In the exponentiation-based version, standard controlled-$U^{2^{k-1}}$ gates are replaced by controlled-$U^{2^{s+k-1}}$, producing a post-kickback state peaked near $x\approx 2^n\{2^s\phi\}$. In the ancilla-removal version, the first $s$ ancilla wires and associated gates of a standard QPE circuit are deleted, which left-shifts the effective phase register by discarding the most significant part [2507.06711]. The exponentiation-based variant satisfies
$$
|\{2^s\phi\}-\widehat{2^s\phi}|\le 2^{-n},
$$
and mapping back gives
$$
|\phi-\hat\phi_{\mathrm{shift}}|\le 2^{-(n+s)}.
$$

This circuit-level use of shifting was developed partly for error mitigation in HHL-type workflows. In the reported HHL application, repeated QPE runs identify a binary matrix of eigenvalue estimates, from which a distinguishing column set is extracted. QSPE removes non-distinguishing leading qubits, and Quantum Punctured Phase Estimation removes known constant bits after the leading distinguishing qubit. On IBM ibm_kingston, the Hybrid25 implementation reduced total qubits from $9$ to $6$, reduced `cz` gates from $189$ to $54`, and reduced `sx` gates from $405$ to $108$; the hardware output probabilities tracked theory more closely than the earlier Hybrid19 circuit [2507.06711].

A different shifted mechanism appears in practical sampling schemes for Kitaev’s algorithm. There, previously inferred suffix bits are used to define a phase shift
$$
\theta_j=-\frac{\tilde\varphi_{j+1}}{2},
$$
so that the residual angle is centered near $0$ if the next bit is $0$ and near $1/2$ if the next bit is $1$. This converts two-dimensional sine/cosine estimation into a sign-of-cosine test. The paper proves that, after a constant-size bootstrap, one can estimate $\varphi$ to accuracy $2^{-(m+2)}$ with success probability at least $1-\epsilon$ using $N_\epsilon+m$ measurements, where $N_\epsilon$ depends only on $\epsilon$ and the chosen sampling algorithm [1902.11168].

An intermediate approach between Kitaev and QFT-based QPE uses arbitrary constant-precision controlled phase shift operators. If only $R_2,\ldots,R_m$ are available, the already-determined bits are used to coarsely cancel the ancilla phase, leaving a residual obeying $|\theta_j|<1/2^m$. The next bit is then revealed by a Hadamard measurement with success probability
$$
p(m)=\cos^2(\pi/2^m),
$$
and majority voting yields the desired per-bit confidence. For $m=3$, the paper gives $p=\cos^2(\pi/8)\approx 0.8536$ and a per-bit repetition count on the order of $4\ln(n/\epsilon_{\mathrm{total}})$, substantially reducing the trial count relative to Kitaev’s original scheme while avoiding exponentially small phase rotations [1012.4727].

## 4. Programmable signal shaping and shifted QSP/QPP methods

Quantum phase processing and quantum signal processing reinterpret phase estimation as the synthesis of a bounded trigonometric response tailored to a target phase interval. In the QPP framework, one starts from a unitary
$$
U=\sum_j e^{i\phi_j}|\psi_j\rangle\langle\psi_j|,
$$
and constructs a single-ancilla circuit with alternating controlled-$U$ and controlled-$U^\dagger$ gates interleaved with ancilla rotations. On each eigenspace, the processor reduces to trigonometric QSP evaluated at the eigenphase, and for suitable angle sequences it realizes a trigonometric polynomial transform $F(\phi)$ in the ancilla block or in the ancilla-$Z$ expectation value [2209.14278].

This machinery yields a no-QFT phase-estimation algorithm based on phase classification and interval search. The Phase Interval Search procedure repeatedly shifts the unitary by the current midpoint,
$$
U_{\mathrm{shift}}=e^{-im}U,
$$
uses a bounded trigonometric approximation to a square wave to classify which half-interval contains the phase, and then refines the interval. Once the interval is sufficiently small, powers of the shifted unitary amplify the local phase for further binary search. The paper states a total query complexity of
$$
O\!\left(\frac{1}{\delta}\log\frac{1}{\epsilon}\log\frac{1}{\delta}\right)
$$
controlled uses of $U$ and $U^\dagger$, with one ancilla qubit and additive error $\delta$ at success probability at least $1-\epsilon$ [2209.14278]. In this setting, “shifted phase estimation” is operationalized by replacing $U$ with $e^{-i\phi_0}U$ so that the relevant variable is $\phi-\phi_0$.

Programmable signal design via QSP makes this centering explicit. At iteration $k$, one keeps an uncertainty interval $I_k=[\phi_0-\Delta,\phi_0+\Delta]$, defines the shifted variable $\theta=\phi-\phi_0$, and synthesizes a QSP signal whose sensitivity is uniformly large over $\theta\in[-\Delta,+\Delta]$ [2604.01205]. For “0→0” measurements, the response takes the form
$$
p(\theta)=g^2(\theta),\qquad s(\theta)=2g^2(\theta)-1,
$$
with $g$ a bounded trigonometric polynomial. For “0→+” measurements, useful near $\theta\approx 0$,
$$
p(\theta)=\frac{1}{2}+\sin\theta\,w(\cos\theta)\,u(\cos\theta).
$$
The design objective is a max–min optimization over admissible signals, maximizing either the minimum derivative $|f'(\theta)|$ or the minimum Fisher information on the current interval [2604.01205].

The 2026 programmable-signal framework introduces the sensitivity efficiency
$$
\kappa(f,\mathcal I)=\frac{\min_{\theta\in\mathcal I}|f'(\theta)|}{d},
$$
and an information-efficiency score based on $\min_{\theta\in\mathcal I}I(\theta)/d$. The paper reports that fixed-signal robust phase estimation has baseline $\kappa_{\mathrm{RPE}}=1/\sqrt{2}$, whereas optimized QSP signals achieve $\kappa\approx 0.9$–$0.95$ on narrow intervals, with mean-squared-error improvement ratio $\mathrm{IR}\approx 2.46$ relative to RPE in experiments at depths $d=8,16,32,64$ and shots $2{,}000$–$50{,}000$ [2604.01205]. This suggests that, within the admissible QSP family, shifting the design interval and programming the signal itself can improve prefactors without changing the Heisenberg-limited asymptotic scaling.

## 5. Statistical regimes, periodicity, and noise

Because shifted phase estimation is fundamentally a periodic-parameter problem, statistical interpretation is inseparable from circular geometry. For interferometric phase shifts, frequentist and Bayesian analyses answer different questions. Frequentist precision bounds such as the CRB and QCRB constrain the sampling spread of estimators at fixed $\phi_0$, while Bayesian posterior variances and Bayes MSE quantify degree of belief averaged over a prior. The literature explicitly warns that CRB/QCRB do not apply to Bayesian strategies and vice versa, and that bounds for fluctuating parameters do not constrain fixed-parameter estimation [1804.10048].

This distinction is especially sharp when the phase is shifted or wrapped. For periodic parameters one must use circular costs, circular means, and circular confidence or credible sets. In Bayesian shifted metrology, posteriors live on the circle and may become multi-peaked because of $\cos(m\phi+\theta)$ likelihoods; the adaptive offset is what restores local identifiability [2303.01517]. In frequentist interferometric settings, confidence intervals quantify repeated-sampling coverage at fixed phase, while Bayesian credible intervals quantify posterior plausibility for the specific data and prior [1804.10048].

Noise changes the meaning of optimal shifting. In the adaptive Bayesian protocol with visibility decay
$$
V(m)=\alpha \beta^m,
$$
the Fisher information per repetition behaves approximately as $V_k^2m^2$, so deep interrogations eventually cease to be beneficial. The noise-optimal depth is
$$
m_{\mathrm{opt}}=-\frac{1}{2\ln\beta},
$$
and for dephasing with interrogation time $t_k=m_k\tau$ and coherence time $T_2$, identifying $\beta=e^{-\tau/T_2}$ gives
$$
m_{\mathrm{opt}}=\frac{T_2}{2\tau}.
$$
At that optimum, the minimal achievable variance scales as
$$
\mathrm{Var}_{\min}(\hat\phi)\approx -\frac{2e\ln\beta}{\alpha^2N_{\mathrm{tot}}},
$$
which is SQL scaling with an improved constant compared to classical separable strategies [2303.01517].

Interferometric blind spots provide another noise-sensitive motivation for shifting. For a two-photon Mach–Zehnder interferometer under depolarization, the Fisher information
$$
F(\phi;p)=\frac{8(1-p)^4\sin^2(2\phi)}{1-(1-p)^4\cos(4\phi)-(p-2)p[(p-2)p+2]}
$$
vanishes at $\phi\in\{0,\pi/2,\pi\}$. A traditional remedy is to insert a known reference phase $\phi_r$, shifting the effective operating point to $\phi-\phi_r$ and attaining
$$
F_{\phi_r}(p)=4(1-p)^4
$$
at $\phi-\phi_r=\pi/4$; the paper states that this remains super-sensitive for $p<0.1591$ [1611.01498]. The same work proposes an ancilla-based adaptive alternative that avoids a reference phase and achieves global super-sensitivity up to $p<0.072$ by optimizing polarization ancilla parameters [1611.01498].

Under genuine phase diffusion, phase and diffusion strength can themselves be jointly estimated, but a trade-off appears because the SLD-optimal measurements are incompatible. For the effective qubit models studied in joint phase-and-diffusion metrology, separable measurements obey
$$
\frac{F_{11}}{H_{11}}+\frac{F_{22}}{H_{22}}\le 1,
$$
which formalizes the incompatibility between optimal phase and optimal diffusion readout [1410.5353]. For optical Gaussian probes under phase diffusion, the QFI obeys the empirical scaling law
$$
H(N,\Delta)=\frac{N}{\Delta}\gamma(\xi),\qquad \xi=N\Delta,
$$
and the optimal squeezing fraction moves from $\beta_{\mathrm{opt}}=1$ at very small noise toward $\beta_{\mathrm{opt}}\to 0$ at large noise, so strong diffusion favors coherent over strongly squeezed probes [1012.1123]. In shifted-phase practice, these results imply that the usefulness of deeper, sharper, or more strongly nonclassical phase shifts depends sensitively on the relevant visibility or diffusion model.

## 6. Applications, eigenvalue problems, and short-depth variants

Beyond metrology and textbook QPE, shifted phase estimation appears in several application-driven eigenvalue workflows. A statistical and variational IPEA-like method uses a prior guess $\phi_0$ and applies a compensation $e^{-im\phi_0}$ on the control, so that the interferometric signal depends on the residual $\Delta=\phi_\lambda-\phi_0$ rather than on the full eigenphase. For an eigenstate input, the single-ancilla outcome probabilities are
$$
P_X(+\mid \Delta,m)=\frac{1}{2}[1+\cos(m\Delta)],\qquad
P_Y(+\mid \Delta,m)=\frac{1}{2}[1+\sin(m\Delta)],
$$
and using both $X$ and $Y$ measurements gives classical Fisher information approximately $\sum_i m_i^2$. The same paper uses the control-$|0\rangle$ probability as an eigenstate–eigenphase proximity metric and reports mean absolute phase error on the order of $10^{-2}$ radians in simulated decompositions, with mean per-eigenphase absolute error decreasing from approximately $2.8\times 10^{-2}$ rad at $d_c=2$ to approximately $1.2\times 10^{-2}$ rad at $d_c=8$ in a $16$-dimensional water-molecule example [2104.10285].

Time-series phase estimation yields another shifted interpretation. The QCELS algorithm measures
$$
Z_n\approx \langle\psi|e^{-in\tau H}|\psi\rangle,
$$
fits a single complex exponential by minimizing
$$
L(r,\theta)=\frac{1}{N}\sum_{n=0}^{N-1}|Z_n-r e^{-i\theta n\tau}|^2,
$$
and equivalently maximizes
$$
f(\theta)=\left|\sum_{n=0}^{N-1} Z_n e^{i\theta n\tau}\right|^2.
$$
For each trial $\theta$, the data are rephased by $e^{i\theta t_n}$, which is a digital frequency shift. The paper states that the maximal coherent evolution time can be made
$$
T_{\max}=\delta/\epsilon,
$$
with $\delta=\Theta(\sqrt{1-p_0})$ as the initial overlap $p_0$ approaches one, while the total coherent evolution time remains Heisenberg-limited up to polylogarithms. Numerical experiments on transverse-field Ising and Hubbard-model instances reported around two orders of magnitude reduction in circuit depth relative to standard QPE under several settings [2211.11973].

Shifted QPP/QSP methods also connect directly to Hamiltonian simulation, amplitude estimation, entanglement spectroscopy, and entropy estimation. The QPP-based phase-search framework proposes a phase-estimation routine without quantum Fourier transform that uses one ancilla qubit, while the broader QPP toolkit is applied to Hamiltonian simulation and quantum entropies estimation [2209.14278]. Programmable QSP signal design extends to Hamiltonian eigenvalue estimation in higher dimensions via qubitization, preserving the same interval-centered signal-shaping logic [2604.01205].

The practical significance of these variants is clearest when shift operations reduce hardware cost or improve conditioning. In the HHL example, left-shifting and puncturing of phase bits lowered qubit and gate counts on superconducting hardware [2507.06711]. In the adaptive Bayesian metrological setting, controllable phase offsets removed posterior ambiguity without requiring entanglement or fault tolerance [2303.01517]. In QCELS, time-domain demodulation and optional Fourier filtering turned large initial overlap into a smaller depth constant [2211.11973]. Taken together, these developments indicate that “shifted” phase estimation is less a single algorithm than a recurrent design principle: reparameterize the phase so that the experiment probes the informative part of a periodic signal rather than the signal in its raw, globally wrapped form.

Source: https://www.emergentmind.com/topics/quantum-shifted-phase-estimation