Gradient Troughs in ADAPT-VQE

• Gradient troughs are a non-monotonic regime in ADAPT-VQE where appending gradients drop sharply without reaching the true ground state.
• They arise from the position-dependent behavior of operator gradients, making the conventional appending rule prone to false convergence under noise.
• Mitigation protocols that vary operator insertion positions can enhance gradient signals and reduce shot cost, improving overall convergence.

Searching arXiv for the core paper on “gradient troughs” and a small set of related disambiguation papers. Gradient troughs are a failure mode of the adaptive derivative-assembled problem-tailored variational quantum eigensolver (ADAPT-VQE) in which the gradient norms “decrease rapidly even though the ground state has not yet been reached, only to increase again after some iterations.” In the paper that introduces the term, the phenomenon is not treated as ordinary convergence but as a transient, appending-specific low-gradient regime: the energy remains above the target, the operator-selection rule becomes unreliable under finite-shot noise and statistical uncertainty, and circuit-structure optimization can stagnate. The proposed explanation is that, in a noncommutative ansatz, the relevant gradients depend strongly on where a new operator is inserted, so apparently negligible appending gradients may coexist with much larger non-appending gradients (Stadelmann et al., 31 Dec 2025).

1. Definition and conceptual boundaries

The defining feature of a gradient trough is non-monotonic gradient behavior during adaptive ansatz growth. The standard ADAPT-VQE interpretation of small gradients is proximity to a variational minimum, because the algorithm ordinarily stops when some norm of the pool-gradient vector falls below a threshold. A gradient trough departs from that interpretation: the gradients become very small even though the minimum energy has not been reached, and later increase again. The paper therefore treats the trough as a misleading low-signal interval rather than evidence of successful convergence (Stadelmann et al., 31 Dec 2025).

This distinction matters because ADAPT-VQE uses gradient information twice: first to decide whether to terminate, and second to decide which operator to append next. In a trough, both decisions become vulnerable. The appending gradient norm may cross the stopping threshold and induce false convergence, or it may remain barely above threshold while finite-shot noise obscures the ranking of candidate operators, causing prolonged stagnation without appreciable energy reduction (Stadelmann et al., 31 Dec 2025).

The same paper explicitly distinguishes gradient troughs from barren plateaus. Barren plateaus are described there as exponentially vanishing gradients over large regions of parameter space, typically associated with random ansätze and poor initialization. By contrast, ADAPT-VQE uses chemically motivated pools, Hartree–Fock initialization, parameter recycling, and gradient-based adaptive growth. The claim is therefore not that ADAPT-VQE has rediscovered the barren-plateau problem under a new name, but that it exhibits a different pathology tied to its appending heuristic and to low-lying spectral structure (Stadelmann et al., 31 Dec 2025).

2. Why ADAPT-VQE is susceptible

The relevant variational objective is the Rayleigh–Ritz quotient,

$$E(\vec{\theta})=\frac{\bra{\psi (\vec\theta)} \hat{H} \ket{\psi (\vec\theta)} }{\bra{\psi (\vec\theta)} \ket{\psi (\vec\theta)}}.$$

ADAPT-VQE grows the ansatz iteratively by selecting, from an operator pool, the element with largest local energy-gradient magnitude. In the standard version, new operators are always added at the end of the ansatz. The pool gradients used for operator selection are

$$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$

The standard stopping rule is that some norm of the vector of all pool gradients falls below a threshold (Stadelmann et al., 31 Dec 2025).

The central claim of the paper is that a gradient trough is specifically a statement about the smallness of gradients in the appending position. This is why the pathology is not exhausted by the phrase “small gradients.” Standard ADAPT-VQE evaluates every candidate in the same insertion location, namely the end of the circuit, and therefore repeatedly samples a single position-dependent slice of the gradient landscape. The authors state that gradient troughs are more likely to arise when the same locations are used repeatedly for new operator insertions, which makes the conventional appending protocol a natural generator of the effect (Stadelmann et al., 31 Dec 2025).

A plausible implication is that the trough is not a property of the target Hamiltonian alone, nor of the current state alone, but of the pair formed by the adaptive selection rule and the ansatz-construction rule. In that reading, the trough is a geometric artifact of how ADAPT-VQE explores variational space.

3. Insertion-position dependence and noncommutative algebra

The algebraic basis of the phenomenon is the dependence of the local energy gradient on insertion position. For a new operator $e^{\theta_\mu A_\mu}$ inserted at position $p$, with $p=1$ corresponding to prepending and $p=n+1$ to appending, the generalized zero-initialization gradient is

$$\left.\frac{\partial E}{\partial\theta_\mu}\right|_{\theta_\mu=0} = \bra{\psi} \bigl[\hat{H},\; V(\theta) A_\mu V^\dagger(\theta) \bigr] \ket{\psi},$$

where

$$V(\theta) = \left(\prod_{i=p}^{n} e^{\theta_{p+n-i} A_{p+n-i}}\right).$$

In the appendix the same idea is derived by introducing a dressed Hamiltonian $\hat M$ and then rewriting the derivative as a commutator evaluated in the current ansatz state (Stadelmann et al., 31 Dec 2025).

This formulation makes explicit why insertion position matters. The candidate generator $A_\mu$ is conjugated by all subsequent ansatz layers. Because those layers generally do not commute, the effective operator seen by the gradient is different at different values of $$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$0. The same operator can therefore have a tiny gradient when appended and a much larger one when inserted earlier. The paper presents this not as a peripheral implementation issue, but as the core mechanism behind trough formation (Stadelmann et al., 31 Dec 2025).

A plausible implication is that the trough reflects a mismatch between the canonical appending coordinate system and the more informative directions available in the full ansatz manifold. What looks like vanishing first-order information in one coordinate chart can remain informative in another.

4. Empirical signatures and convergence diagnostics

The main diagnostic proposed in the paper is to measure the pool-gradient norm as a function of insertion position. During a trough, the observed signature is that the norm is much larger near the beginning of the ansatz and decreases steadily toward the appending position. At true convergence, by contrast, the gradient norms are “much more uniform across the positions,” without the systematic drop toward the end (Stadelmann et al., 31 Dec 2025).

The reported benchmark is a strongly correlated linear H$$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$1 chain at $$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$2 in STO-3G, mapped to 12 qubits, using the problem-specific qubit excitation pool. In that system, standard ADAPT-VQE exhibits a pronounced trough around iterations 20–29. The paper states that iteration 25, which lies inside the trough, shows strong position dependence with a clear downward trend from prepending to appending, whereas a truly converged iteration shows uniformly small gradients across positions. A “stair plot” over many iterations indicates that this position-dependent suppression appears during the trough interval and disappears before and after it (Stadelmann et al., 31 Dec 2025).

From these observations the paper suggests a revised conceptual stopping condition: not merely that the appending gradient norm be below threshold, but that all gradient norms over ansatz positions be small and of similar magnitude. The practical protocol proposed is to run standard ADAPT-VQE until the appending norm falls below the usual threshold, then probe one or a few other positions, preferably toward the beginning of the ansatz. If those non-appending positions exhibit significantly larger gradients, the algorithm is not converged but in a trough; if all tested positions are comparably small within a preset tolerance, convergence may be declared (Stadelmann et al., 31 Dec 2025).

5. Mitigation protocols and benchmark behavior

The mitigation strategy follows directly from the insertion-position diagnosis: if troughs are associated with suppressed appending gradients, then one should change where new operators are inserted. The paper develops four explicit protocols, designed to separate the effect of position choice from the effect of operator choice (Stadelmann et al., 31 Dec 2025).

Protocol	Operator choice	Position choice
OO/OP	optimized operator	optimized position
OO/RP	optimized operator	random position
RO/OP	random operator	optimized position
RO/RP	random operator	random position

In OO/OP, the algorithm first computes appending gradients and takes the 10 operators with largest appending gradients. For each of those 10 operators, it then computes gradients over all ansatz positions $$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$3,

$$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$4

with

$$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$5

For each candidate the position of maximal gradient magnitude is found, then the operator-position pair with largest overall $$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$6 is selected. The operator is inserted at that position, the new parameter is initialized to zero, and all variational parameters are reoptimized while recycling the previous ones. In OO/RP, a position is chosen randomly from $$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$7, and the operator with largest gradient magnitude at that fixed position is selected. The two remaining protocols, RO/OP and RO/RP, act as controls by weakening operator choice (Stadelmann et al., 31 Dec 2025).

On the H$$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$8 benchmark, standard ADAPT-VQE enters a pronounced trough around iterations 20–29. When the new protocols are activated after the trough begins, OO/OP and OO/RP successfully escape the trough and continue lowering the energy, whereas RO/OP and RO/RP perform worse than standard ADAPT-VQE. This comparison is central to the paper’s interpretation: varying insertion position matters, but it is not sufficient if the gradient-based operator-selection heuristic is abandoned. The best performance comes from protocols that preserve optimized operator choice while diversifying insertion position (Stadelmann et al., 31 Dec 2025).

The paper also studies temporary use of the enhanced protocols. If OO/OP or OO/RP is applied for 30 iterations after trough detection and the algorithm then returns to standard appending, the gradient norm rises rapidly, the trough is exited, and subsequent convergence is smoother and more monotonic. If the protocol is used only for 5 or 10 iterations, the algorithm often falls back into the trough after reverting to appending. The authors caution that the exact intervention length may be system-specific (Stadelmann et al., 31 Dec 2025).

6. Measurement cost, hardware relevance, and terminological scope

The practical importance of gradient troughs is amplified by shot noise. The paper estimates that distinguishing the leading gradient from zero requires roughly

$$\left.\frac{\partial E}{\partial\theta_{j}}\right|_{\theta_{j} = 0} = \bra{\psi_{n}} [\hat{H}, A_{j}] \ket{\psi_{n}}.$$9

shots, where $e^{\theta_\mu A_\mu}$0 is the largest gradient magnitude. For measuring all pool gradients in one ansatz position,

$e^{\theta_\mu A_\mu}$1

For the molecular Hamiltonians and pools considered, both $e^{\theta_\mu A_\mu}$2 and $e^{\theta_\mu A_\mu}$3 scale as $e^{\theta_\mu A_\mu}$4, giving

$e^{\theta_\mu A_\mu}$5

The crucial point is that appending gradients can become extremely small inside a trough, causing measurement cost to explode. The proposed protocols often increase the relevant gradient norm by two orders of magnitude, which the paper translates into an estimated $e^{\theta_\mu A_\mu}$6 reduction in shot cost for gradient evaluation during the trough, even in the 12-qubit H$e^{\theta_\mu A_\mu}$7 example. OO/OP and RO/OP do incur direct overhead from checking multiple positions, but the authors argue that this is usually subdominant relative to the gain from increased gradient magnitudes (Stadelmann et al., 31 Dec 2025).

The hardware motivation is explicit throughout. The paper frames the danger of troughs in terms of limited shots and statistical uncertainty: once the relevant gradients collapse, the algorithm may terminate incorrectly or add poor operators. It also notes a hardware subtlety: appending can use the commutator form and potentially commuting-group measurement strategies, whereas arbitrary insertion positions generally require parameter-shift rules. Even so, the asymptotic scaling is said to remain comparable, and the larger gradients can compensate in practice. At the same time, the numerical study is idealized and noise-free, detailed studies of decoherence and gate errors are left to future work, and the empirical evidence is confined to one benchmark system, linear H$e^{\theta_\mu A_\mu}$8 (Stadelmann et al., 31 Dec 2025).

Outside ADAPT-VQE, the word “trough” appears in unrelated technical senses. Other arXiv papers use the term for confined depressed regions in scroll-wave dynamics (Ke et al., 2015), for attractive potential channels guiding cubic–quintic solitons (Zeng et al., 1 Jan 2026), and for market-trough probabilities whose causal analysis involves an average partial derivative rather than a variational-optimization pathology (Rao et al., 7 Sep 2025). This suggests that “gradient troughs” is not a cross-disciplinary standard term: its precise technical meaning is set by the ADAPT-VQE literature, where it denotes a position-structured, non-monotonic, appending-specific low-gradient regime that can cause false convergence and stagnation unless insertion position is treated as an explicit control variable (Stadelmann et al., 31 Dec 2025).