Piecewise-Ramp QAOA

• Piecewise-Ramp QAOA is a quantum optimization method that uses piecewise-linear schedules for objective and penalty terms to streamline parameter tuning.
• It replaces the traditional 2p parameter search with a low-dimensional schedule (4+3N_pen parameters) while maintaining high expressiveness for constrained problems.
• Empirical results show improved feasibility (up to 0.9) and optimality (α ≈ 0.85) on benchmarks like Maximum Independent Set and satellite mission planning compared to lr-QAOA.

Searching arXiv for 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 (Kremenetski et al., 2023). 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 $p$ (Ferrari et al., 23 Jun 2026). 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 (Montanez-Barrera et al., 2024, Ferrari et al., 23 Jun 2026).

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 (Ferrari et al., 23 Jun 2026).

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 $\gamma(s)$0 in the main experiments, and was presented as evidence for a universal set of QAOA parameters across nine combinatorial optimization problems (Montanez-Barrera et al., 2024).

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 (Ferrari et al., 23 Jun 2026).

Method	Schedule idea	Parameters
lr-QAOA	One linear ramp for mixer/objective	$\gamma(s)$1
$\gamma(s)$2-lr-QAOA	lr-QAOA plus one variational penalty weight per constraint	$\gamma(s)$3
Piecewise-ramp QAOA	Two-segment schedules for objective and each penalty	$\gamma(s)$4

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 $\gamma(s)$5 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

$\gamma(s)$6

with binary variables mapped by $\gamma(s)$7 and, for a weighted objective,

$\gamma(s)$8

Piecewise-ramp QAOA promotes the objective and penalty schedules to internal variational objects rather than fixing all $\gamma(s)$9 externally (Ferrari et al., 23 Jun 2026).

The core schedule is a two-segment piecewise-linear function

$\beta(s)$0

For the objective term, the parameters are $\beta(s)$1. For each constraint $\beta(s)$2, the corresponding penalty schedule has parameters $\beta(s)$3. The mixer retains a linear decay factor $\beta(s)$4 (Ferrari et al., 23 Jun 2026).

At Trotter depth $\beta(s)$5, with $\beta(s)$6, the unitary is

$\beta(s)$7

The total number of variational parameters is

$\beta(s)$8

which is independent of $\beta(s)$9 (Ferrari et al., 23 Jun 2026).

In the broader schedule-theoretic treatment, one may partition $p$0 into $p$1 pieces,

$p$2

and define $p$3 and $p$4 by linear interpolation on each interval. The discrete QAOA angles are then sampled as $p$5 and $p$6, with $p$7 or another sampling rule (Kremenetski et al., 2023). In that broader sense, the two-segment constrained ansatz is the $p$8 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 $p$9:

$2p$0

Infeasible strings contribute zero. Minimizing $2p$1 therefore drives amplitude toward high-quality feasible solutions (Ferrari et al., 23 Jun 2026).

To control the feasibility–optimality trade-off, the filtered variant introduces

$2p$2

As $2p$3, the loss focuses on the minimum-energy feasible string; as $2p$4, it recovers a more uniform expectation over feasibility. The paper states that $2p$5 gives a single knob to adjust the feasibility–optimality trade-off (Ferrari et al., 23 Jun 2026).

The optimization loop is correspondingly structured around feasible samples. One initializes

$2p$6

prepares $2p$7, applies $2p$8, samples $2p$9 bitstrings, identifies feasible samples, evaluates $$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$0, computes an empirical $$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$1 or $$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$2, and updates $$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$3 using gradient descent or gradient-free DE/L-BFGS-B. At the optimum, the reported diagnostics are the feasibility rate

$$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$4

the optimal feasible-solution probability

$$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$5

and the approximation ratio

$$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$6

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 (Ferrari et al., 23 Jun 2026).

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 (Montanez-Barrera et al., 2024). Piecewise-ramp schedules generalize this by allowing different linear slopes on different parts of the path (Kremenetski et al., 2023).

The theoretical analysis in the gradually changing unitaries framework introduces the pair unitary

$$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$7

and measures the maximum step size

$$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$8

Successive unitaries then satisfy

$$U(\gamma,\beta)=\prod_{\ell=1}^p e^{-i\beta_\ell H_{\mathrm{mix}}}e^{-i\gamma_\ell H_{\mathrm{QUBO}}},$$9

If the evolving target eigenphase remains separated by a non-vanishing spectral gap $H_{\mathrm{mix}}=\sum_i X_i.$0, the discrete adiabatic theorem yields an overlap bound of the form

$H_{\mathrm{mix}}=\sum_i X_i.$1

For a $H_{\mathrm{mix}}=\sum_i X_i.$2-piece ramp, applying the bound piecewise gives a total error scaling

$H_{\mathrm{mix}}=\sum_i X_i.$3

under uniform depth allocation and a uniform step-size bound (Kremenetski et al., 2023).

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 $H_{\mathrm{mix}}=\sum_i X_i.$4, an intermediate ridge where performance is best, and a large-$H_{\mathrm{mix}}=\sum_i X_i.$5 regime where unitary eigenvalues wrap around the circle and the mixer ground connects to a cost excited branch (Kremenetski et al., 2023). The same framework argues that piecewise segmentation can reduce error at fixed $H_{\mathrm{mix}}=\sum_i X_i.$6, and that slightly larger $H_{\mathrm{mix}}=\sum_i X_i.$7 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” $H_{\mathrm{mix}}=\sum_i X_i.$8 (Kremenetski et al., 2023).

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 $H_{\mathrm{mix}}=\sum_i X_i.$9 and $\{\beta_\ell,\gamma_\ell\}$0, standard lr-QAOA with fixed penalty $\{\beta_\ell,\gamma_\ell\}$1 often yields low feasibility $\{\beta_\ell,\gamma_\ell\}$2 at moderate depth, while $\{\beta_\ell,\gamma_\ell\}$3-lr-QAOA reaches $\{\beta_\ell,\gamma_\ell\}$4 for large $\{\beta_\ell,\gamma_\ell\}$5 but with modest growth in $\{\beta_\ell,\gamma_\ell\}$6. Piecewise-ramp QAOA is reported to attain both higher $\{\beta_\ell,\gamma_\ell\}$7 and much faster growth of $\{\beta_\ell,\gamma_\ell\}$8 versus $\{\beta_\ell,\gamma_\ell\}$9 (Ferrari et al., 23 Jun 2026).

At $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$0, the reported medians are approximately $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$1 and $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$2 for piecewise-ramp QAOA, compared with $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$3 and $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$4 for lr-QAOA. Under graph-size scaling at fixed $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$5 and $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$6 from $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$7, piecewise-ramp QAOA maintains median approximation ratio $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$8 and $$\gamma_i=\frac{i+1}{p}\,\Delta\gamma,\qquad \beta_i=\left(1-\frac{i}{p}\right)\Delta\beta,\qquad i=0,\ldots,p-1.$$9 close to $\Delta\beta=0.3$0, whereas both comparison methods degrade more rapidly as $\Delta\beta=0.3$1 increases (Ferrari et al., 23 Jun 2026).

On Earth-observation satellite mission planning formulated as a budget-constrained Maximum Weight Independent Set problem with $\Delta\beta=0.3$2 and $\Delta\beta=0.3$3, the comparison is especially explicit. A full 3D grid-search lr-QAOA yields $\Delta\beta=0.3$4, $\Delta\beta=0.3$5, and $\Delta\beta=0.3$6, while a single-run piecewise-ramp QAOA yields $\Delta\beta=0.3$7, $\Delta\beta=0.3$8, and $\Delta\beta=0.3$9 (Ferrari et al., 23 Jun 2026).

The same experiments also isolate the feasibility–optimality trade-off. By minimizing the filtered loss $\gamma(s)$00, one can increase $\gamma(s)$01 above $\gamma(s)$02 at the cost of reducing $\gamma(s)$03 to about $\gamma(s)$04 for large $\gamma(s)$05, or instead favor feasibility for small $\gamma(s)$06. The paper further reports that optimization with a finite number of shots, down to approximately $\gamma(s)$07–$\gamma(s)$08, remains stable because of the unbiased shot-based estimator of $\gamma(s)$09 and the robustness of differential evolution (Ferrari et al., 23 Jun 2026).

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 $\gamma(s)$10 qubits and $\gamma(s)$11, with an empirical success law

$\gamma(s)$12

for a problem-class-dependent constant $\gamma(s)$13 under stated conditions (Montanez-Barrera et al., 2024). 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 (Ferrari et al., 23 Jun 2026).

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 (Ferrari et al., 23 Jun 2026). In the gradually changing unitaries analysis, it denotes the more general family of QAOA schedules obtained by partitioning $\gamma(s)$14 into $\gamma(s)$15 pieces and interpolating linearly on each segment (Kremenetski et al., 2023). 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 $\gamma(s)$16-parameter search by a small number of schedule parameters plus a 2D grid search over endpoints (McDowall et al., 27 Apr 2026).

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

$\gamma(s)$17

with instantaneous gap $\gamma(s)$18 and $\gamma(s)$19, SGIR-QAOA constructs a monotonic reparameterization

$\gamma(s)$20

and sets

$\gamma(s)$21

This slows the schedule in regions where the spectral gap is small (McDowall et al., 27 Apr 2026).

On Grover’s problem and MIS, SGIR-QAOA improves constant-depth exponential fits relative to LR-QAOA. At $\gamma(s)$22, the reported fit exponents are $\gamma(s)$23 for Grover with LR-QAOA and $\gamma(s)$24 with SGIR-QAOA; for MIS on degree-3 graphs, $\gamma(s)$25 for LR-QAOA versus $\gamma(s)$26 with exact SGIR-QAOA and $\gamma(s)$27 with extrapolated SGIR-QAOA; and for MIS on Erdős–Rényi graphs with edge probability $\gamma(s)$28, $\gamma(s)$29 for LR-QAOA versus $\gamma(s)$30 exact and $\gamma(s)$31 extrapolated SGIR-QAOA (McDowall et al., 27 Apr 2026). For Grover’s search at threshold $\gamma(s)$32, the required depth is reported to be $\gamma(s)$33–$\gamma(s)$34 smaller for SGIR-QAOA than for LR-QAOA once $\gamma(s)$35, and under local depolarizing noise with strength $\gamma(s)$36 on every qubit after each two-qubit gate, SGIR-QAOA reaches its peak success probability at smaller $\gamma(s)$37 and at a higher peak probability on $\gamma(s)$38-node MIS instances (McDowall et al., 27 Apr 2026).

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 $\gamma(s)$39 qubits, Lanczos/ARPACK up to about $\gamma(s)$40, and an extrapolation protocol for larger MIS instances (McDowall et al., 27 Apr 2026). 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 (Montanez-Barrera et al., 2024). 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 (Montanez-Barrera et al., 2024, Ferrari et al., 23 Jun 2026, McDowall et al., 27 Apr 2026).