---
title: Piecewise-Ramp QAOA
url: https://www.emergentmind.com/topics/piecewise-ramp-qaoa
type: topic
---

# Piecewise-Ramp QAOA

Searching arXiv for recent papers on piecewise-ramp and linear-ramp QAOA.
Piecewise-Ramp QAOA denotes a class of QAOA parameter schedules in which the layer angles are sampled from piecewise-linear functions of a normalized time variable rather than optimized independently at every depth. In a broad usage, it includes any QAOA schedule obtained by partitioning the interval $[0,1]$ into pieces and linearly interpolating $\gamma(s)$ and $\beta(s)$ on each piece [2305.04455]. In a narrower and more recent usage, it refers specifically to a constrained-optimization ansatz that replaces the single linear ramps of linear-ramp QAOA (lr-QAOA) by two-segment piecewise schedules for the objective term and for each penalty term, while keeping the parameter count independent of the circuit depth $p$ [2606.25117]. The construction inherits the de-variationalizing motivation of lr-QAOA—namely, replacing a $2p$-dimensional search by a low-dimensional schedule—but increases expressiveness in settings with multiple heterogeneous constraints [2405.09169, 2606.25117].

## 1. Position within the QAOA scheduling literature

Standard QAOA alternates mixer and problem unitaries with depth-dependent angles. For a QUBO or Ising encoding, one writes
$$
U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},
$$
with transverse-field mixer
$$
H_{\mathrm{mix}}=\sum_i X_i.
$$
The central scheduling question is how to choose $\{\beta_\ell,\gamma_\ell\}$ without an expensive instance-specific outer loop [2606.25117].

Montañez-Barrera et al. proposed linear-ramp QAOA as a fixed-schedule alternative. In the notation used for large-scale simulations and hardware runs,
$$
\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad
\beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.
$$
This uses only two global ramp parameters, typically $\Delta\beta=0.3$ and $\Delta\gamma=0.6$ in the main experiments, and was presented as evidence for a universal set of QAOA parameters across nine combinatorial optimization problems [2405.09169].

The two-segment piecewise-ramp construction generalizes that lr-QAOA baseline. Instead of one straight-line interpolation for the objective and all penalties, it introduces a single breakpoint per scheduled Hamiltonian term. The resulting ansatz remains low-dimensional but is more expressive than a single linear ramp, particularly for constrained problems with several penalty terms [2606.25117].

| Method | Schedule idea | Parameters |
|---|---|---|
| lr-QAOA | One linear ramp for mixer/objective | $2$ |
| $\Lambda$-lr-QAOA | lr-QAOA plus one variational penalty weight per constraint | $2+N_{\mathrm{pen}}$ |
| Piecewise-ramp QAOA | Two-segment schedules for objective and each penalty | $4+3N_{\mathrm{pen}}$ |

This parameterization makes explicit that the relevant design variable is not layerwise freedom but schedule shape. A plausible implication is that piecewise-ramp QAOA should be viewed less as a new optimizer over $2p$ angles than as a structured scheduling ansatz positioned between fixed linear ramps and fully variational QAOA.

## 2. Formal definition of the two-segment piecewise-ramp ansatz

For constrained binary optimization, the Hamiltonian is written as
$$
H_{\mathrm{QUBO}}=H_{\mathrm{obj}}+\sum_k \lambda_k H_{\mathrm{pen}}^{(k)},\qquad \lambda_k>0,
$$
with binary variables mapped by $x_i\mapsto (1-Z_i)/2$ and, for a weighted objective,
$$
H_{\mathrm{obj}}=\sum_i w_i Z_i.
$$
Piecewise-ramp QAOA promotes the objective and penalty schedules to internal variational objects rather than fixing all $\lambda_k$ externally [2606.25117].

The core schedule is a two-segment piecewise-linear function
$$
f_{\mathrm{pw}}(t;t_a,a,\tilde a)=
\begin{cases}
a\, (t/t_a), & 0\le t\le t_a,\\[4pt]
a+(\tilde a-a)\,[(t-t_a)/(1-t_a)], & t_a<t\le 1.
\end{cases}
$$
For the objective term, the parameters are $(t_\gamma,\gamma,\tilde\gamma)$. For each constraint $k$, the corresponding penalty schedule has parameters $(t_{\Lambda_k},\Lambda_k,\tilde\Lambda_k)$. The mixer retains a linear decay factor $\beta(1-t)$ [2606.25117].

At Trotter depth $p$, with $t_\ell=\ell/p$, the unitary is
$$
U_{\mathrm{pw}}=\prod_{\ell=1}^p
\left[
e^{-i\beta(1-t_\ell)H_{\mathrm{mix}}}
\cdot
e^{-i f_{\mathrm{pw}}(t_\ell;t_\gamma,\gamma,\tilde\gamma)H_{\mathrm{obj}}}
\cdot
\prod_k e^{-i f_{\mathrm{pw}}(t_\ell;t_{\Lambda_k},\Lambda_k,\tilde\Lambda_k)H_{\mathrm{pen}}^{(k)}}
\right].
$$
The total number of variational parameters is
$$
1+3+3N_{\mathrm{pen}}=4+3N_{\mathrm{pen}},
$$
which is independent of $p$ [2606.25117].

In the broader schedule-theoretic treatment, one may partition $[0,1]$ into $P$ pieces,
$$
0=f_0<f_1<\cdots<f_P=1,
$$
and define $\gamma(s)$ and $\beta(s)$ by linear interpolation on each interval. The discrete QAOA angles are then sampled as $\gamma_j=\gamma(s_j)$ and $\beta_j=\beta(s_j)$, with $s_j=j/(p+1)$ or another sampling rule [2305.04455]. In that broader sense, the two-segment constrained ansatz is the $P=2$ specialization applied separately to the objective and each penalty schedule.

## 3. Feasibility-driven training and constraint handling

The distinctive feature of the constrained piecewise-ramp construction is not only the schedule shape but also the training criterion. Instead of optimizing a standard expectation value of the full penalized Hamiltonian, the method uses a feasibility-driven loss over the feasible set $F$:
$$
L_F=\sum_{\sigma\in F} E_{\mathrm{obj}}(\sigma)\,|\langle \sigma|\psi\rangle|^2,
\qquad
E_{\mathrm{obj}}(\sigma)=\langle \sigma|H_{\mathrm{obj}}|\sigma\rangle \quad (<0 \text{ for feasible }\sigma).
$$
Infeasible strings contribute zero. Minimizing $L_F$ therefore drives amplitude toward high-quality feasible solutions [2606.25117].

To control the feasibility–optimality trade-off, the filtered variant introduces
$$
L_F^{(G)}(\mu)= -\frac{1}{\mu}\log\!\left[\sum_{\sigma\in F} e^{-\mu E_{\mathrm{obj}}(\sigma)}|\langle \sigma|\psi\rangle|^2\right].
$$
As $\mu\to\infty$, the loss focuses on the minimum-energy feasible string; as $\mu\to 0$, it recovers a more uniform expectation over feasibility. The paper states that $\mu$ gives a single knob to adjust the feasibility–optimality trade-off [2606.25117].

The optimization loop is correspondingly structured around feasible samples. One initializes
$$
\theta=\{\beta;\, t_\gamma,\gamma,\tilde\gamma;\, t_{\Lambda_k},\Lambda_k,\tilde\Lambda_k\},
$$
prepares $|\psi_{\mathrm{mix}}\rangle=|+\rangle^{\otimes n}$, applies $U_{\mathrm{pw}}(\theta)$, samples $M$ bitstrings, identifies feasible samples, evaluates $E_{\mathrm{obj}}$, computes an empirical $L_F$ or $L_F^{(G)}(\mu)$, and updates $\theta$ using gradient descent or gradient-free DE/L-BFGS-B. At the optimum, the reported diagnostics are the feasibility rate
$$
p_{\mathrm{feas}}=\Pr[\sigma\in F],
$$
the optimal feasible-solution probability
$$
p_{\mathrm{opt}}=\Pr[\sigma \text{ is an optimum feasible solution}],
$$
and the approximation ratio
$$
\alpha=L_F(\theta^\*)/E_{\mathrm{opt}}.
$$
This suggests that piecewise-ramp QAOA is best understood as a joint schedule-and-loss design for constrained optimization, rather than only a change in angle parameterization [2606.25117].

## 4. Adiabatic interpretation and piecewise-ramp theory

The schedule-based view of QAOA is closely tied to an adiabatic interpretation. For lr-QAOA, QAOA can be seen as a Trotterized quantum-annealing evolution in which the problem Hamiltonian is turned on gradually while the transverse mixer is turned off. In Montañez-Barrera et al., the linear ramp is described as the simplest first-order discretization of a linear anneal [2405.09169]. Piecewise-ramp schedules generalize this by allowing different linear slopes on different parts of the path [2305.04455].

The theoretical analysis in the gradually changing unitaries framework introduces the pair unitary
$$
U_j=e^{-i\beta_j H_M}e^{-i\gamma_j H_C}
$$
and measures the maximum step size
$$
\Delta=\max_j\{|\gamma_{j+1}-\gamma_j|,\; |\beta_{j+1}-\beta_j|\}.
$$
Successive unitaries then satisfy
$$
\|U_{j+1}-U_j\|=O\bigl(\Delta(\|H_M\|+\|H_C\|)\bigr).
$$
If the evolving target eigenphase remains separated by a non-vanishing spectral gap $g(s)$, the discrete adiabatic theorem yields an overlap bound of the form
$$
\bigl\|\,|\psi_p\rangle-|v_0(1)\rangle\bigr\|
\le
C\,\frac{\Delta\,p}{g_{\min}},
\qquad
g_{\min}=\min_{s\in[0,1]} g(s).
$$
For a $P$-piece ramp, applying the bound piecewise gives a total error scaling
$$
\epsilon(p,\Delta)=O\!\left(\frac{p\Delta}{P}\right)
$$
under uniform depth allocation and a uniform step-size bound [2305.04455].

This analysis was used to explain the common structure of QAOA performance diagrams. The cited work identifies three regimes: an initial-state region with small $\Delta p$, an intermediate ridge where performance is best, and a large-$\Delta$ regime where unitary eigenvalues wrap around the circle and the mixer ground connects to a cost excited branch [2305.04455]. The same framework argues that piecewise segmentation can reduce error at fixed $p\Delta$, and that slightly larger $\Delta$ can sometimes reduce the required depth by exploiting diabatic jumps. The two-node MaxCut example in that work serves as an explicit case where a 2-piece ramp outperforms a single-segment ramp at the same total “time” $\Delta$ [2305.04455].

## 5. Empirical behavior on constrained optimization benchmarks

The strongest direct numerical evidence for the two-segment constrained ansatz comes from random Maximum Independent Set and satellite mission-planning Maximum Weight Independent Set benchmarks. On random MIS instances generated from Erdős–Rényi graphs with $|V|=18$ and $p\in\{5,\ldots,50\}$, standard lr-QAOA with fixed penalty $\lambda=1.1$ often yields low feasibility $p_{\mathrm{feas}}\approx 0.5$ at moderate depth, while $\Lambda$-lr-QAOA reaches $p_{\mathrm{feas}}\to 1$ for large $p$ but with modest growth in $p_{\mathrm{opt}}$. Piecewise-ramp QAOA is reported to attain both higher $p_{\mathrm{feas}}$ and much faster growth of $p_{\mathrm{opt}}$ versus $p$ [2606.25117].

At $p=20$, the reported medians are approximately $p_{\mathrm{feas}}\approx 0.9$ and $p_{\mathrm{opt}}\approx 0.3$ for piecewise-ramp QAOA, compared with $p_{\mathrm{feas}}\approx 0.4$ and $p_{\mathrm{opt}}\approx 0.1$ for lr-QAOA. Under graph-size scaling at fixed $p=20$ and $|V|$ from $6\to 20$, piecewise-ramp QAOA maintains median approximation ratio $\alpha\gtrsim 0.6$ and $p_{\mathrm{feas}}$ close to $1$, whereas both comparison methods degrade more rapidly as $|V|$ increases [2606.25117].

On Earth-observation satellite mission planning formulated as a budget-constrained Maximum Weight Independent Set problem with $|V|=24$ and $B=20$, the comparison is especially explicit. A full 3D grid-search lr-QAOA yields $p_{\mathrm{opt}}\approx 0.20$, $\alpha\approx 0.44$, and $p_{\mathrm{feas}}\approx 0.49$, while a single-run piecewise-ramp QAOA yields $p_{\mathrm{opt}}\approx 0.33$, $\alpha\approx 0.85$, and $p_{\mathrm{feas}}\approx 0.94$ [2606.25117].

The same experiments also isolate the feasibility–optimality trade-off. By minimizing the filtered loss $L_F^{(G)}(\mu)$, one can increase $p_{\mathrm{opt}}$ above $50\%$ at the cost of reducing $p_{\mathrm{feas}}$ to about $70\%$ for large $\mu$, or instead favor feasibility for small $\mu$. The paper further reports that optimization with a finite number of shots, down to approximately $200$–$500$, remains stable because of the unbiased shot-based estimator of $L_F$ and the robustness of differential evolution [2606.25117].

These results should be read against the lr-QAOA baseline established by Montañez-Barrera et al., where fixed linear ramps were studied on random instances of nine combinatorial optimization problems up to $N_q=42$ qubits and $p=400$, with an empirical success law
$$
\mathrm{probability}(x^\*)\approx 2^{-(\eta N_q/p)}
$$
for a problem-class-dependent constant $\eta$ under stated conditions [2405.09169]. Piecewise-ramp QAOA does not present an analogous universal scaling conjecture in the cited data; instead, its empirical claim is better feasibility and better feasible-optimum concentration for constrained instances [2606.25117].

## 6. Variants, terminology, and open technical issues

The phrase “piecewise-ramp QAOA” is not fully standardized across the schedule-design literature. In the constrained-optimization paper, it denotes the two-segment ansatz with separate schedules for the objective and for each penalty term [2606.25117]. In the gradually changing unitaries analysis, it denotes the more general family of QAOA schedules obtained by partitioning $[0,1]$ into $P$ pieces and interpolating linearly on each segment [2305.04455]. In the spectral-gap-informed scheduling work, “piecewise-ramp QAOA” is used in an umbrella sense for low-dimensional schedule families such as LR-QAOA and SGIR-QAOA that replace the $2p$-parameter search by a small number of schedule parameters plus a 2D grid search over endpoints [2604.24580].

The SGIR-QAOA variant is a particularly relevant extension because it changes not only the number of pieces but the distribution of evolution time along the path. Defining the adiabatic interpolation
$$
H(s)=(1-s)H_X+sH_C,
$$
with instantaneous gap $g(s)=E_1(s)-E_0(s)$ and $g_{\min}=\min_{s\in[0,1]} g(s)$, SGIR-QAOA constructs a monotonic reparameterization
$$
f(s)=
\frac{\int_0^s [g(s')-g_{\min}]^\kappa ds'}{\int_0^1 [g(s')-g_{\min}]^\kappa ds'},
\qquad \kappa=2,
$$
and sets
$$
\gamma(t)=\Delta_\gamma f(t),\qquad \beta(t)=\Delta_\beta [1-f(t)].
$$
This slows the schedule in regions where the spectral gap is small [2604.24580].

On Grover’s problem and MIS, SGIR-QAOA improves constant-depth exponential fits relative to LR-QAOA. At $p=10$, the reported fit exponents are $0.56\pm 0.02$ for Grover with LR-QAOA and $0.42\pm 0.02$ with SGIR-QAOA; for MIS on degree-3 graphs, $0.56\pm 0.02$ for LR-QAOA versus $0.41\pm 0.02$ with exact SGIR-QAOA and $0.46\pm 0.02$ with extrapolated SGIR-QAOA; and for MIS on Erdős–Rényi graphs with edge probability $0.4$, $0.62\pm 0.03$ for LR-QAOA versus $0.49\pm 0.03$ exact and $0.53\pm 0.03$ extrapolated SGIR-QAOA [2604.24580]. For Grover’s search at threshold $P_s^{\mathrm{th}}=0.1$, the required depth is reported to be $30$–$50\%$ smaller for SGIR-QAOA than for LR-QAOA once $n\gtrsim 7$, and under local depolarizing noise with strength $p_{\mathrm{noise}}=10^{-3}$ on every qubit after each two-qubit gate, SGIR-QAOA reaches its peak success probability at smaller $p$ and at a higher peak probability on $10$-node MIS instances [2604.24580].

The main technical limitation of these more informed schedule families is classical side information. Exact spectral-gap evaluation may be as hard as solving the optimization problem, so SGIR-QAOA uses dense diagonalization only up to about $14$ qubits, Lanczos/ARPACK up to about $20$, and an extrapolation protocol for larger MIS instances [2604.24580]. By contrast, lr-QAOA was motivated precisely by the absence of an instance-specific outer optimizer and by a single schedule reused across many problems [2405.09169]. A common misconception is therefore to treat all ramped QAOA variants as equally “parameter free.” The cited literature suggests a more precise distinction: lr-QAOA fixes a universal-style linear schedule; piecewise-ramp QAOA in the constrained setting adds a small number of schedule parameters to improve feasibility handling; and spectral-gap-informed schedules trade additional classical preprocessing for better depth-efficiency [2405.09169, 2606.25117, 2604.24580].

Source: https://www.emergentmind.com/topics/piecewise-ramp-qaoa