Barren Plateaus in Variational Landscapes

The barren plateau is the reason a quantum computer can be perfectly capable of representing the answer to a problem and still be unable to find it. Variational quantum algorithms — the whole near-term programme of chemistry solvers, combinatorial optimizers, and quantum machine-learning models — rest on a single act of faith: that a classical optimizer, handed the measured gradient of a cost function, can walk a parameterized circuit downhill to a good solution. The barren plateau result, proved by McClean and co-workers in 2018, shows that faith is misplaced for the most natural circuit designs. Make the ansatz expressive enough to reach anywhere in state space, and its cost landscape becomes flat almost everywhere — flat not by a little, but to a precision that shrinks exponentially with the number of qubits. This deep-tier synthesis takes the cost-function and gradient machinery from aq-qci-cost-function-measurement-estimation and aq-qci-parameter-shift-rules-gradients and asks what happens when the circuit is random enough to average like the Haar measure. Three ideas organize everything. First, the mechanism is not bad luck but concentration of measure: on the exponentially large unitary group, almost every circuit sits exponentially close to a trivial average, so both the mean gradient and its variance are computable exactly by Weingarten calculus and both come out crushed. Second, the obstruction is information-theoretic, not algorithmic — a landscape flat to precision $2^{-n}$ cannot be navigated by any optimizer, gradient-based or not, without resolving cost differences that cost $2^{n}$ measurement shots to see. Third, the escape is structural: locality, shallow depth, and problem-informed ansätze keep the circuit from ever behaving as a 2-design, trading some expressibility for the trainability the plateau destroys.

Key takeaways

A barren plateau is a region of a variational cost landscape where the gradient has zero mean and a variance that vanishes exponentially in the qubit number $n$ : $\mathrm{Var}[\partial_\mu C]\sim 2^{-n}$ , so the whole landscape is nearly flat.
The cause is concentration of measure: once a parameterized circuit is expressive enough to form an approximate unitary 2-design, Haar averaging by Weingarten calculus forces both the first and second moments of the gradient to their trivial, exponentially suppressed values.
On a plateau the cost itself concentrates on the maximally-mixed value $\mathrm{Tr}(H)/2^{n}$ with fluctuations only $\mathcal{O}(2^{-n/2})$ — an optimizer sees a near-constant function buried in shot noise.
Resolving such a gradient above statistical noise needs $N_{\mathrm{shots}}\sim 2^{n}$ measurements, so even one optimization step becomes intractable at scale — the barrier is information-theoretic, defeating gradient-free and second-order methods equally.
Plateaus come in flavours with one shared exponential signature: expressibility (deep random circuits), global-cost-function, entanglement-induced, and noise-induced, where the variance decays as $q^{L}$ with circuit depth $L$ .
Every mitigation — local cost functions, shallow or structured ansätze, identity-block initialization, warm starts — works by lowering expressibility so the circuit stops being a 2-design.
Key idea: if a parameterized circuit ensemble is expressive enough to be an approximate unitary 2-design, then averaging the cost-function gradient over circuit parameters gives zero mean and a variance that decays exponentially in the qubit number — the landscape is a barren plateau, untrainable without exponentially many measurements.

Where this lives today

The barren plateau is the organizing obstacle of the entire variational, NISQ-era programme, and diagnosing it has become a routine first step in any serious ansatz design. Before committing to a hardware-efficient or chemistry-inspired circuit, practitioners estimate the gradient variance versus qubit number on small instances and extrapolate; an exponential decay is a red flag that scaling up will fail regardless of the optimizer. That diagnostic now shapes practice across the field. In VQE, global observables are rewritten or decomposed into local terms and paired with shallow circuits so the lightcone argument keeps the variance polynomial; adaptive constructions such as ADAPT-VQE grow the ansatz gate by gate to stay in the trainable regime. In QAOA and quantum machine learning, the noise-induced plateau sets a hard depth budget on real devices, and the entanglement-induced variant (Marrero et al., Patti et al.) warns against ansätze that scramble a bipartition too aggressively. The active research fronts are all responses to the theorem: identity-block and small-angle initializations that start the circuit off the Haar-typical region; classically informed warm starts; symmetry-preserving and equivariant ansätze that break the circuit's randomness; and layerwise or block-local training. A sharper and more sobering thread runs alongside them — recent "provable absence of barren plateaus" results show that many circuits engineered to be trainable are, for the same structural reasons, also classically simulable, raising the live and unresolved question of whether escaping the plateau and retaining a quantum advantage are compatible at all.

The physical picture

Imagine hunting for a single valley on a continent-sized plain that is flat to within the width of an atom everywhere except one tiny crater. Random parameter initialization parachutes you somewhere onto the plain; the slope beneath your feet is so faint it is drowned out by the trembling of your only instrument — the shot noise of a finite number of measurements. Add a qubit and the plain doubles in size while the residual slope halves in a mean-square sense: the featureless expanse grows and flattens at the same exponential rate. This is not a pathological landscape chosen to defeat you. It is the typical landscape, because on the exponentially large space of circuits almost every instance looks like the average, and the average has no slope at all. The first figure makes the concentration visible: as $n$ grows the cost curve collapses onto its mean and its wiggles shrink like $2^{-n/2}$ .

Figure

As the qubit number n grows the cost landscape C(θ) collapses onto the maximally-mixed value Tr(H)/2ⁿ, its peak-to-peak features shrinking like 2·2⁻ⁿ⁄² — an n = 2 curve still swings widely, an n = 6 curve is muted, and an n = 12 curve hugs the mean line: the barren plateau

The deep reason is symmetry. A highly expressive circuit reaches essentially any unitary with roughly equal probability, so averaging a fixed observable's gradient over the ensemble is very nearly averaging over the entire unitary group with the Haar measure — and positive and negative contributions cancel almost perfectly. Concentration of measure then guarantees that not just the average but almost every individual instance lies exponentially close to that trivial value: you cannot dodge the plateau by re-rolling the initial parameters, because nearly all of them land you on it. The one genuine escape is locality. If the cost reads out only a few qubits and the circuit is shallow, the observable's causal "lightcone" touches only a polylogarithmic patch of gates; the 2-design never forms over the whole register, and the gradient variance decays only polynomially. The second figure is the quantity a practitioner actually measures — the gradient variance versus qubit number — and it shows exactly where trainability dies.

Figure

Measured gradient variance Var(∂C) ≈ 2⁻ⁿ collapses exponentially with qubit number: from 6.3×10⁻² at n = 4 to 9.5×10⁻⁷ at n = 20 to 8.9×10⁻¹⁶ at n = 50, crossing the shot-noise floor 1/N = 10⁻⁶ of a million-shot estimate near n = 20 — beyond that the gradient is buried in statistical noise, the trainability cliff

The core equations

Everything begins with the cost function of a variational algorithm: the expectation value of a problem observable in the state a parameterized circuit prepares. It is the quantity the optimizer tries to lower, and the object whose gradient will turn out to vanish.

C(\boldsymbol\theta)=\langle 0|\,U^{\dagger}(\boldsymbol\theta)\,H\,U(\boldsymbol\theta)\,|0\rangle \qquad (1)

where: $U(\boldsymbol\theta)$ — parameterized unitary built from $L$ gate layers; $H$ — problem Hamiltonian or cost observable; $|0\rangle$ — fixed reference input $|0\rangle^{\otimes n}$ ; $\boldsymbol\theta$ — vector of variational parameters. A larger $H$ term count or a deeper $U$ both push the circuit toward the Haar-random behaviour that creates the plateau. To differentiate with respect to a single angle, split the circuit at the gate that carries it. This is the standard device that localizes the derivative onto one Hermitian generator.

U(\boldsymbol\theta)=U_{-}\,e^{-i\theta_\mu V_\mu}\,U_{+} \qquad (2)

where: $V_\mu$ — Hermitian generator of the $\mu$ -th gate, taken with $V_\mu^2=\mathbb{1}$ as every Pauli rotation has; $U_{+},U_{-}$ — sub-circuits after and before the chosen gate; $\theta_\mu$ — the single parameter being differentiated. Only the middle factor depends on $\theta_\mu$ , so the product rule will strike just once. Carrying the derivative through and grouping the pieces, the gradient becomes the expectation of one commutator — a compact form that is ready to be averaged over the circuit ensemble.

\partial_\mu C = i\,\mathrm{Tr}\!\big(\rho\,[V_\mu,\,\tilde H]\big),\qquad \rho=U_{+}|0\rangle\langle 0|U_{+}^{\dagger},\;\;\tilde H=U_{-}^{\dagger}HU_{-} \qquad (3)

where: $\rho$ — state entering the differentiated gate; $\tilde H$ — the observable propagated back through $U_{-}$ ; $[\,\cdot,\cdot\,]$ — commutator. Read it as: the slope in direction $\mu$ is how much the generator $V_\mu$ fails to commute with the back-propagated observable, sampled in the state $\rho$ . Averaging this commutator over a Haar-random (or 1-design) sub-circuit sends the observable to a multiple of the identity, and the identity commutes with everything — so the mean gradient is exactly zero.

\mathbb{E}_{\mathrm{Haar}}[\partial_\mu C]=0 \qquad (4)

where: $\mathbb{E}_{\mathrm{Haar}}$ — average over Haar-random $U_{+}$ or $U_{-}$ . The plateau has no preferred direction: the expected slope is flat everywhere, so the optimizer starts with no bias toward the solution. The mean vanishing is only half the story — a fluctuating gradient with zero mean could still be usable. The second moment settles it: computed with the 2-design (Weingarten) integral, the variance carries an explicit factor of $1/(2^{2n}-1)$ .

\mathrm{Var}[\partial_\mu C]\;\approx\;\frac{\mathrm{Tr}(H^2)\,\mathrm{Tr}(V_\mu^2\rho)}{2^{2n}-1}\;\sim\;2^{-n} \qquad (5)

where: $\mathrm{Tr}(H^2)$ — Hilbert–Schmidt norm of the observable, at most $\|H\|^2\,2^{n}$ ; $2^{2n}=d^2$ — square of the Hilbert-space dimension $d=2^{n}$ . Because $\mathrm{Tr}(H^2)\le\|H\|^2 d$ while the denominator is $d^2$ , the ratio scales as $1/d=2^{-n}$ : the barren plateau. An exponentially small gradient can be detected only against exponentially small statistical noise, so the number of measurement shots per gradient estimate is set by the reciprocal of the variance.

N_{\mathrm{shots}}\;=\;\mathcal{O}\!\left(\frac{1}{\mathrm{Var}[\partial_\mu C]}\right)\;\sim\;\mathcal{O}(2^{n}) \qquad (6)

where: $N_{\mathrm{shots}}$ — measurement repetitions to resolve the gradient above shot noise. This is the operational face of the plateau: not merely a small number but an exponentially expensive one to measure. Finally, the cost value itself concentrates. Under a 2-design the trial energy collapses onto the value of the maximally mixed state, with fluctuations only $\mathcal{O}(2^{-n/2})$ .

C(\boldsymbol\theta)\;\xrightarrow[\ \text{2-design}\ ]{}\;\frac{\mathrm{Tr}(H)}{2^{n}}\;\pm\;\mathcal{O}\!\big(2^{-n/2}\big) \qquad (7)

where: $\mathrm{Tr}(H)/2^{n}$ — trivial mean, the energy of the maximally mixed state $\mathbb{1}/2^{n}$ . For a traceless observable this is simply $0\pm\mathcal{O}(2^{-n/2})$ : the optimizer sees a nearly constant function, and the plateau in the gradient is mirrored by concentration in the cost.

Derivation AVanishing mean and exponentially small variance of the gradient

The goal is the flagship result of the whole subject: to show that when even one sub-circuit is drawn from an approximate unitary 2-design, the parameter-shift gradient of the cost has exactly zero mean and a variance that decays like $2^{-n}$ , so that the landscape is provably a barren plateau rather than merely a difficult one. We localize the gradient onto a single commutator, average with the first Weingarten moment to kill the mean, then square and average with the second moment to expose the $1/(d^2-1)$ suppression, and finally bound the surviving trace. Take $C=\langle 0|U^{\dagger}HU|0\rangle$ with $U=U_-\,e^{-i\theta_\mu V_\mu}\,U_+$ and $V_\mu^2=\mathbb{1}$ .

Step 1. Differentiate and group: only the middle factor carries $\theta_\mu$ , so the product rule gives $\partial_\mu C = i\,\langle 0|U_+^{\dagger}[V_\mu,\tilde H]U_+|0\rangle$ with $\tilde H=U_-^{\dagger}HU_-$ . Why: the derivative of a long circuit collapses to a single commutator sandwiched by the front sub-circuit, which is what makes the Haar average tractable.

Step 2. Average the observable over $U_-$ using the first Weingarten moment $\int\!\mathrm{d}U\,U^{\dagger}HU=\frac{\mathrm{Tr}(H)}{d}\mathbb{1}$ , so $\mathbb{E}_{U_-}[\tilde H]=\frac{\mathrm{Tr}(H)}{d}\mathbb{1}$ . Why: the Haar average of any conjugated operator is proportional to the identity — the ensemble forgets everything about $H$ except its trace.

Step 3. Take the mean of the gradient: $\mathbb{E}[\partial_\mu C]\propto \mathrm{Tr}\!\big(\rho\,[V_\mu,\mathbb{1}]\big)=0$ . Why: a commutator with the identity vanishes identically, so the expected slope is exactly flat — a unitary 1-design already suffices for a zero mean.

Step 4. Now the second moment. Squaring brings two factors of $U$ and two of $U^{\dagger}$ , which require the 2-design (second Weingarten) moment; the surviving contraction yields $\mathbb{E}[(\partial_\mu C)^2]\approx \dfrac{\mathrm{Tr}(H^2)\,\mathrm{Tr}(V_\mu^2\rho)}{d^2-1}$ . Why: the first-moment average sent the mean to zero, so the entire variance lives in this second-moment integral, whose prefactor is $1/(d^2-1)$ .

Step 5. Bound the trace and read off the scaling: for bounded $\|H\|$ , $\mathrm{Tr}(H^2)\le\|H\|^2 d$ , so with $d=2^{n}$ , $\dfrac{d}{d^2-1}\sim\dfrac1d=2^{-n}$ and hence $\mathrm{Var}[\partial_\mu C]\sim 2^{-n}$ . Why: the observable norm can grow at most like the dimension, but the denominator grows like the dimension squared — the mismatch is precisely the exponential suppression that defines the plateau. *Two quick checks confirm the result: $\mathrm{Tr}[V_\mu,\mathbb{1}]=0$ (a commutator with the identity vanishes identically), and $d/(d^2-1)\sim 2^{-n}$ follows immediately by inspection — these are the two key steps that seal the exponential suppression.*

Go deeper · Derivation B · 1 minFrom gradient variance to exponentially many shots

Derivation BFrom gradient variance to exponentially many shots

The aim is to prove that a barren plateau is not just an inconveniently small gradient but a genuine information-theoretic wall: a landscape flat to precision $2^{-n}$ demands of order $2^{n}$ measurement shots merely to take one reliable optimization step, and no cleverness in the optimizer can lower that count, because the bottleneck is the statistics of reading a quantum expectation value, not the search strategy. We model the finite-sample energy estimate, propagate its shot noise into the parameter-shift gradient, impose a signal-to-noise condition, and solve for the shot budget. Assume a bounded observable so each cost estimate has $\mathcal{O}(1)$ single-shot variance.

Step 1. Write the exact gradient as a parameter shift, $\partial_\mu C=\tfrac12\big[C(\theta_\mu+\tfrac\pi2)-C(\theta_\mu-\tfrac\pi2)\big]$ , each cost measured from $N$ shots. Why: the gradient is obtained from two ordinary cost evaluations, so it simply inherits their shot noise rather than adding any new source of error.

Step 2. Propagate the noise: with single-shot variance $\sigma^2=\mathcal{O}(1)$ , the estimator obeys $\mathrm{Var}[\widehat{\partial_\mu C}]\approx \sigma^2/(2N)$ . Why: two independent shot-noise variances add, and dividing by the $N$ shots per evaluation gives the statistical error bar on the measured slope.

Step 3. Impose reliable descent — the deterministic gradient must exceed its own fluctuation, $(\text{typical }|\partial_\mu C|)^2\gtrsim \mathrm{Var}[\widehat{\partial_\mu C}]$ . Why: if the signal is smaller than the noise, the measured "gradient" points in a random direction and the optimizer performs a random walk, making no systematic progress.

Step 4. Solve for the shot count using $(\text{typical }|\partial_\mu C|)^2\sim\mathrm{Var}[\partial_\mu C]\sim 2^{-n}$ : $N\gtrsim \dfrac{\sigma^2}{2\,\mathrm{Var}[\partial_\mu C]}\sim \dfrac{\sigma^2}{2}\,2^{n}$ . Why: because the resolvable signal is set by the landscape variance and that variance is exponentially small, the shots needed to see it are exponentially large — a single step is intractable at scale. *The $N\sim 1/\mathrm{Var}$ scaling follows directly from the standard $1/\sqrt{N}$ shot-noise law. Numerically: $n=20\Rightarrow N\sim 2^{20}\approx 1.05\times10^{6}$ ; $n=50\Rightarrow N\sim 2^{50}\approx1.13\times10^{15}$ .*

Go deeper · Derivation C · 1 minLocality and shallow depth restore trainability

Derivation CLocality and shallow depth restore trainability

The purpose here is to show why the plateau is escapable rather than a death sentence: if the cost is a sum of geometrically local terms and the circuit is shallow, only a polylogarithmic patch of gates lies in each term's causal lightcone, so the 2-design never forms over the full register, the effective Hilbert-space dimension that enters the Weingarten estimate is polynomial rather than $2^{n}$ , and the gradient variance decays only polynomially — but the same argument locates a depth threshold beyond which the plateau returns. Let $H=\sum_j O_j$ be a sum of local operators and $U$ a brickwork circuit of depth $L$ .

Step 1. Bound the causal cone: a local term $O_j$ is affected only by gates in its backward lightcone, of width $\mathcal{O}(2^{L})$ qubits for a brickwork circuit of depth $L$ . Why: information in a local-depth circuit spreads at a finite "speed," so a shallow circuit couples each observable to only a small, contiguous block of qubits.

Step 2. Identify the effective dimension: only $|\mathrm{lightcone}|$ qubits are scrambled, so the relevant Weingarten dimension is $d_{\mathrm{eff}}=2^{|\mathrm{lightcone}|}$ , which is $\mathrm{poly}(n)$ when $L=\mathcal{O}(\log n)$ . Why: the exponential suppression in the variance is governed by however many qubits actually average like Haar, not by the total register size.

Step 3. Replace $d$ by $d_{\mathrm{eff}}$ in the variance estimate: $\mathrm{Var}[\partial_\mu C]\sim 1/d_{\mathrm{eff}}\sim 1/\mathrm{poly}(n)$ . Why: with a polynomial effective dimension the gradient shrinks only polynomially, so a polynomial number of shots resolves it and the landscape is trainable.

Step 4. Cross the threshold: once $L\gtrsim\mathcal{O}(n)$ the lightcone spans the whole register, $d_{\mathrm{eff}}\to 2^{n}$ , and $\mathrm{Var}[\partial_\mu C]\sim 2^{-n}$ again. Why: enough depth scrambles the entire system into a 2-design, so even a local cost eventually plateaus — trainability is a statement about depth as much as about locality.

Limiting cases

How the gradient variance scales is decided by three levers — cost locality, circuit depth, and hardware noise:

Case	Result	Meaning
Global cost, deep ( $\geq$ 2-design) ansatz	$\mathrm{Var}[\partial_\mu C]\sim 2^{-n}$	Full barren plateau: once a 2-design forms the variance decays exponentially regardless of further depth; trainability is lost.
Local cost, shallow ( $\mathcal{O}(\log n)$ depth)	$\mathrm{Var}[\partial_\mu C]\sim \mathrm{poly}(1/n)$	Trainable: the lightcone confines the randomness, so the variance decays only polynomially (Cerezo et al.).
Local cost, linear depth	$\mathrm{Var}[\partial_\mu C]\sim 2^{-n}$	Even a local cost plateaus once depth grows enough to scramble globally — there is a depth threshold at $L\sim\mathcal{O}(n)$ .
Noisy circuit, depth $L$	$\mathrm{Var}[\partial_\mu C]\sim q^{L},\ q<1$	Noise-induced plateau: the state contracts toward maximally mixed, so the variance decays exponentially in depth (Wang et al.).

Common trapA barren plateau is a bad local minimum a smarter optimizer could escape

It is neither a local minimum nor a symptom of a weak optimizer. The landscape is genuinely flat — flat to a precision that shrinks like $2^{-n}$ — over an exponentially large region, so there is simply no slope for any method to follow. Switching from vanilla gradient descent to a gradient-free method (Nelder–Mead, SPSA), a second-order method (Newton), or the quantum natural gradient does not help: all of them still have to resolve cost differences of size $2^{-n}$ , and all of them drown in the same shot noise. Nor can you re-roll your way out by trying many random initializations, because concentration of measure puts almost all of them on the plateau. A related trap: the finite-sample "gradient" you measure on a plateau is dominated by shot noise, not landscape structure, and mistaking its random fluctuations for genuine descent is one of the most common ways a plateaued run is misread as progress.

Worked examples

1 · The trainability cliff, fully worked

A hardware-efficient ansatz on $n$ qubits reaches an approximate unitary 2-design with a bounded observable of order-one norm, so $\mathrm{Var}[\partial_\mu C]\approx 2^{-n}$ . Tabulate the gradient variance, the typical gradient magnitude, and the shots needed to resolve it for $n=4,10,20,50$ , and locate the cliff for a million-shot budget.

Step 1. The typical magnitude of the gradient is the square root of its variance, since the mean is zero: $|\partial_\mu C|_{\mathrm{typ}}\sim\sqrt{\mathrm{Var}}=\sqrt{2^{-n}}=2^{-n/2}$ . The shots needed follow from equation (6): $N_{\mathrm{shots}}\sim 1/\mathrm{Var}\sim 2^{n}$ . Both are pure functions of $n$ .

Step 2. Evaluate at $n=4$ : $\mathrm{Var}=2^{-4}=\mathbf{6.25\times10^{-2}}$ , $|\partial_\mu C|_{\mathrm{typ}}=2^{-2}=0.25$ , $N\sim 2^{4}=\mathbf{16}$ . A four-qubit toy is comfortably trainable — a handful of shots sees the slope.

Step 3. Evaluate at $n=10$ and $n=20$ : at $n=10$ , $\mathrm{Var}=2^{-10}=9.77\times10^{-4}$ , $|\partial_\mu C|_{\mathrm{typ}}=2^{-5}\approx0.031$ , $N\sim1024$ . At $n=20$ , $\mathrm{Var}=2^{-20}=\mathbf{9.54\times10^{-7}}$ , $|\partial_\mu C|_{\mathrm{typ}}=2^{-10}\approx9.8\times10^{-4}$ , $N\sim2^{20}\approx\mathbf{1.05\times10^{6}}$ . This is the cliff: a standard $N=10^{6}$ -shot estimate has a shot-noise floor $1/N=10^{-6}$ , and the variance $9.5\times10^{-7}$ has just fallen through it — beyond about 20 qubits the gradient is buried in statistical noise.

Step 4. Evaluate at $n=50$ : $\mathrm{Var}=2^{-50}=\mathbf{8.88\times10^{-16}}$ , $|\partial_\mu C|_{\mathrm{typ}}=2^{-25}\approx3.0\times10^{-8}$ , $N\sim 2^{50}\approx\mathbf{1.13\times10^{15}}$ . A single optimization step would need of order a quadrillion measurements. The plateau turns a "50-qubit trainable model" into a physical impossibility — not because the answer is unrepresentable, but because its gradient is unmeasurable.

2 · A noise-induced plateau and the depth budget

A depth- $L$ circuit on a device with weak per-layer noise contracts the state toward maximally mixed, so the gradient variance decays as $\mathrm{Var}[\partial_\mu C]\sim q^{L}$ with an effective per-layer factor $q=0.99$ (a one-percent contraction per layer). Find how the gradient shrinks with depth and the deepest circuit that keeps the variance above a usable floor of $10^{-4}$ .

Step 1. The gradient magnitude scales as $\sqrt{\mathrm{Var}}\sim q^{L/2}$ . At $L=100$ , $q^{L}=0.99^{100}=\mathbf{0.366}$ , so the gradient is reduced to $\sqrt{0.366}\approx0.61$ of its noiseless value — a mild but real erosion after a hundred layers.

Step 2. Push deeper. At $L=500$ , $0.99^{500}=6.57\times10^{-3}$ , and at $L=1000$ , $0.99^{1000}=\mathbf{4.32\times10^{-5}}$ : the variance has fallen by more than four orders of magnitude purely from noise, with no 2-design assumed. This is a distinct mechanism from the expressibility plateau — it is driven by depth and error rate, not by Haar randomness.

Step 3. Solve for the depth budget. Requiring $q^{L}>10^{-4}$ gives $L<\dfrac{\ln 10^{-4}}{\ln 0.99}=\dfrac{-9.21}{-0.01005}=\mathbf{916\ \text{layers}}$ . A one-percent-per-layer device can run at most about 900 layers before noise alone flattens the landscape — a hard ceiling on usable depth that tightens as the error rate rises, and the quantitative basis for depth budgets in QAOA and quantum machine learning.

Physics of the portrait

The barren plateau is where a piece of pure high-dimensional geometry — concentration of measure, a nineteenth- and twentieth-century idea about spheres — reappeared in 2018 as the central obstacle of near-term quantum computing.

Origins · 1 minbarren plateau and how this was discovered

barren plateau

the exponential flattening of variational landscapes, McClean et al. 2018 · trainability of variational quantum algorithms

The barren plateau is not a person but a phenomenon, and its lineage runs through both mathematics and physics. Its mathematical root is concentration of measure: Paul Lévy observed early in the twentieth century, and Vitali Milman sharpened in the 1970s, that a smooth function on a very high-dimensional sphere is almost constant — nearly all of its mass sits exponentially close to the mean. The unitary group is such a high-dimensional space, and Haar integrals over it were made computable by the Weingarten calculus of Don Weingarten and, in the modern combinatorial form, Benoît Collins and Piotr Śniady in 2006. The quantum-computing chapter opened in 2018, when Jarrod McClean, Sergio Boixo, Vadim Smelyanskiy, Ryan Babbush, and Hartmut Neven at Google proved in "Barren plateaus in quantum neural network training landscapes" that a random, sufficiently deep parameterized circuit inherits exactly this concentration: its cost gradient has zero mean and variance $\sim 2^{-n}$ . The result reframed the field. Marco Cerezo, Patrick Coles, and collaborators showed in 2021 that the severity depends on cost-function locality, rescuing shallow local-cost circuits; the same year, Samson Wang and co-workers established that hardware noise induces plateaus of its own; and Zoë Holmes with Coles tied the whole effect to a single scalar, the ansatz's expressibility. Together they turned a discouraging theorem into a design principle — trainability demands structured, low-expressibility ansätze, not generic universal ones.

History

1971

— Milman sharpens concentration of measure on high-dimensional spheres

2006

— Collins and Śniady formalize the Weingarten calculus for Haar unitary integrals

2018

— McClean, Boixo, Smelyanskiy, Babbush, and Neven prove barren plateaus in random parameterized circuits

2021

— Cerezo and co-workers link plateau severity to cost-function locality and circuit depth

2021

— Wang and co-workers identify noise-induced barren plateaus

Going further

Texts and reviews. McClean, Boixo, Smelyanskiy, Babbush, Neven, "Barren plateaus in quantum neural network training landscapes," Nature Communications 9, 4812 (2018) · Cerezo, Sone, Volkoff, Cincio, Coles, "Cost function dependent barren plateaus in shallow parametrized quantum circuits," Nature Communications 12, 1791 (2021) · Wang, Fontana, Cerezo, Sharma, Sone, Cincio, Coles, "Noise-induced barren plateaus in variational quantum algorithms," Nature Communications 12, 6961 (2021) · Holmes, Sharma, Cerezo, Coles, "Connecting ansatz expressibility to gradient magnitudes and barren plateaus," PRX Quantum 3, 010313 (2022) · Cerezo et al., "Variational Quantum Algorithms," Nature Reviews Physics 3, 625 (2021) · Preskill, "Quantum Computing in the NISQ era and beyond," Quantum 2, 79 (2018) · Nielsen and Chuang, Quantum Computation and Quantum Information (gates and channels) · Watrous, The Theory of Quantum Information (Haar measure, Weingarten calculus). The key analytic checks are: $\mathrm{Tr}[V_\mu,\mathbb{1}]=0$ (zero mean gradient), $d/(d^2-1)\sim 2^{-n}$ (exponential variance suppression from equation (5)), and $N\sim 1/\mathrm{Var}$ (shot cost from standard shot-noise estimation). The worked-example figures are exact: $2^{-4}=6.25\times10^{-2}$ , $2^{-20}=9.54\times10^{-7}$ , $2^{-50}=8.88\times10^{-16}$ ; the noise example uses $0.99^{L}$ with depth budget $\ln(10^{-4})/\ln(0.99)\approx916$ .

Prerequisites. Variational Quantum Eigensolver · Parameterized Quantum Circuits and Ansatz Design · Parameter-Shift Rules and Quantum Gradients · Cost-Function and Measurement Estimation in Variational Algorithms. Leads to. Strategies to Mitigate Barren Plateaus · Expressibility, Entangling Capability, and Trainability Trade-offs · Classical Optimizers for Variational Quantum Algorithms · Quantum Machine-Learning Models.

Ask the tutor: Ask me to walk the Weingarten derivation of the $2^{-n}$ variance line by line, to work out the shot budget or noise depth-limit for a circuit of your size, or to compare the expressibility, locality, entanglement, and noise routes to a barren plateau.

Is a barren plateau the same thing as a local minimum?

No, and conflating them is the most common misconception. A local minimum is a specific point where the gradient is zero but the surrounding landscape still has structure — you are stuck in a basin, and a lucky perturbation or a different optimizer might escape it. A barren plateau is an entire exponentially large region where the gradient is essentially zero everywhere, to a precision that shrinks like $2^{-n}$ . There is no basin to climb out of because there is no structure at all: the cost is nearly constant. That is why the two demand opposite responses — you fight local minima with better search heuristics, but you can only fight barren plateaus by changing the circuit or the cost so the landscape stops being flat in the first place.

What is a unitary 2-design, and why does it matter here?

A unitary $t$ -design is an ensemble of circuits whose first $t$ statistical moments exactly match those of the fully random Haar distribution. A 2-design matches the first two moments — averages of products of up to two circuit copies and two conjugates. That is precisely what the gradient's mean (one copy of $U$ ) and variance (two copies) depend on, so once your ansatz reaches an approximate 2-design, those two quantities take their Haar values, which the Weingarten calculus shows are exponentially suppressed. Crucially, you do not need genuine randomness: generic expressive hardware-efficient circuits approach a 2-design with only polynomial depth, which is why the plateau shows up in ordinary, non-exotic ansätze rather than only in contrived ones.

If almost every initialization is on the plateau, why do small experiments still train?

Because the plateau is an asymptotic, large- $n$ statement, and small circuits are far from that limit. At $n=4$ the gradient variance is $2^{-4}\approx0.06$ — a perfectly healthy slope that a handful of shots resolves, as the worked example shows. The exponential suppression only becomes fatal once $n$ is large enough that $2^{-n}$ falls below your shot-noise floor, which for a million-shot budget happens near $n=20$ . This is exactly what makes barren plateaus so dangerous in practice: a method validated on a few qubits looks like it works, and the failure only appears when you scale up to the sizes where a quantum computer would actually be useful.

Does switching to a gradient-free optimizer avoid the problem?

No. It is tempting to think that because the plateau is defined through the gradient, a method that never computes a gradient — Nelder–Mead, particle swarm, SPSA, Bayesian optimization — might sidestep it. But every one of these methods ultimately decides where to move by comparing cost values at nearby points, and on a plateau those cost values differ by only $\mathcal{O}(2^{-n/2})$ , which is buried in shot noise for exactly the same reason the gradient is. The obstruction is information-theoretic: the information needed to descend is exponentially expensive to measure, and no reformulation of the search — gradient-based, gradient-free, or second-order — changes the number of shots that measurement requires.

How is a noise-induced plateau different from the standard one?

They share the exponential signature but arise from different physics. The standard, expressibility-induced plateau is a property of an ideal noiseless circuit that is random enough to be a 2-design; it depends on how expressive the ansatz is. A noise-induced plateau needs no randomness at all — it comes from hardware imperfections that, layer by layer, contract the quantum state toward the maximally mixed state, so the gradient variance decays as $q^{L}$ in the circuit depth $L$ rather than $2^{-n}$ in the qubit count. The practical distinction matters: you fight the expressibility plateau by making the ansatz less expressive (locality, structure), but you fight the noise plateau by making the circuit shallower or the hardware cleaner, since it is depth and error rate, not expressibility, that drive it.

Can you just rescale the cost or use a huge learning rate to climb the plateau?

No — the flatness is in the measured landscape, not in a units convention you can redefine away. Rescaling $H$ by a constant multiplies both the gradient and its shot noise by the same factor, leaving the signal-to-noise ratio, and therefore the shot cost, unchanged. A large learning rate does not help either: you cannot take a confident step in a direction you cannot reliably measure, and on a plateau the measured direction is dominated by statistical noise, so a big step is just a big random jump. The only genuine levers are the ones that change the landscape itself — a local cost, a shallow or structured ansatz, an initialization away from the Haar-typical region — because they attack the variance $\mathrm{Var}[\partial_\mu C]$ rather than its cosmetic scale.

Does escaping the barren plateau guarantee a quantum advantage?

Not necessarily, and this is one of the sharpest open questions in the field. The structural features that make a circuit trainable — locality, shallow depth, limited entanglement, restricted expressibility — are, in several recent results, the very features that also make the circuit's outputs efficiently classically simulable. In other words, some ansätze proven to have no barren plateau turn out to compute nothing a laptop could not. This does not mean the two goals are always incompatible, but it does mean that "we made it trainable" and "it beats classical methods" are separate claims that must each be established. Reconciling trainability with genuine quantum advantage — finding the regime that is both learnable and classically hard — is where much of the current effort is aimed.