Why random parameterized circuits are untrainable.
Initialize a deep parameterized circuit at random and its gradients are not just small: they are exponentially small in the number of qubits, for almost every parameter setting. This lesson proves it, shows how cost functions and noise create or avoid such barren plateaus, and confronts an uncomfortable recent question: does provable trainability imply classical simulability?
What you’ll learn
Define barren plateaus and convert gradient variance into a lower bound on shots.
Prove Var[∂θC]≈2−(n+1) when the circuit after a parameter forms a unitary 2-design.
Compute gradient variances exactly for global and local costs with a product ansatz, and state the cost-function-dependent results.
Show how noise damps gradients exponentially in depth.
Explain the Lie-algebraic theory of barren plateaus and the debate linking their absence to classical simulability.
Flat landscapes
Consider a cost C(θ)=Tr[OU(θ)ρU(θ)†] on n qubits, with parameters drawn from an initialization distribution.
DerivationWhy exponentially small variance kills training
Pr[∣∂kC∣≥δ]≤δ2Var[∂kC]
Chebyshev's inequality turns a small variance into concentration:
N≳Var[∂kC]∥O∥2=Ω(bn)
So a random starting point has a gradient of size δ only with probability O(b−n/δ2). With the parameter-shift rule, each component is estimated with standard error ∼∥O∥/N; resolving a typical gradient, of size Var∼b−n/2, needs
shots per component: exponential resources, however clever the optimizer. Gradient-free methods do not escape, because cost differences concentrate in the same way (Arrasmith et al., 2021); in fact barren plateaus are equivalent to exponential concentration of the cost itself (Arrasmith et al., 2022).
The 2-design theorem
McClean et al. (2018) showed that if the circuit on either side of a parameter is random enough (a unitary 2-design), gradients vanish exponentially. Here is a clean, exact version of the calculation.
First moment: a 2-design is also a 1-design, so EA[A†OA]=dTrOI=0 and E[∂θC]=0. Second moment: the Haar (Weingarten) formula for E[Tr(A†OAK)2], with TrK=0, gives
Tr(K2)=−41Tr([σθ,P]2)=21(1−⟨P⟩σθ2)
For pure σθ and P2=I, Tr([σθ,P]2)=2Tr(σθPσθP)−2Tr(σθ2)=2(⟨P⟩2−1), so
Combine with d=2n. Sanity check at n=1: O~ is a uniformly random Pauli axis m^⋅σ, and (for ⟨P⟩=0) ∂θC is one Cartesian component of m^, whose variance is 31, matching 2⋅32⋅1. □
Two remarks. The theorem needs randomness only after the parameter, between the gate and the measurement; McClean et al. also treat randomness before it, with the same exponential conclusion. And the decay is set by the dimension 2n, not by depth as such: depth matters insofar as it makes the circuit 2-design-like.
Check your understanding
In the theorem, take O=Z1, a local observable, but keep A a 2-design on all n qubits. What happens to the variance?
It becomes polynomial in n, because O is local
It is unchanged, Θ(2−n): a global 2-design scrambles Z1 over all qubits
It becomes exactly zero
It grows exponentially with n
Show the answer
B. It is unchanged, Θ(2−n): a global 2-design scrambles Z1 over all qubits
The formula only uses TrO=0 and TrO2=2n, which hold for Z1 as for any Pauli string. Locality helps only when the circuit is shallow, so that the backward light cone of Z1 stays small: the content of the cost-function-dependent results below.
Cost functions and noise
Cerezo, Sone, Volkoff, Cincio & Coles (2021) showed that what you measure decides trainability for shallow circuits. The cleanest illustration needs no 2-designs at all. Take the depth-1 product ansatz ∣ψ(θ)⟩=⨂j=1nRX(θj)∣0⟩ with θj uniform on [−π,π], and the target ∣0⟩⊗n; qubit j is found in ∣0⟩ with probability cos22θj.
DerivationGlobal versus local cost for a product ansatz
CG=1−j=1∏ncos22θj,CL=1−n1j=1∑ncos22θj
Global cost (infidelity with the target) and local cost (average single-qubit infidelity):
∂1CG=21sinθ1j≥2∏cos22θj,∂1CL=2n1sinθ1
Gradients, using ∂θcos22θ=−21sinθ:
Var[∂1CG]=81(83)n−1,Var[∂1CL]=8n21
Both have mean zero. With E[sin2θ]=21 and E[cos42θ]=E[(21+cosθ)2]=83:
CL≤CG≤nCL,
Same circuit, same minimum, exponentially different trainability. The local cost is still faithful: since ∏jpj≤pk for every k, and 1−∏jpj≤∑j(1−pj) by the union bound,
so CL and CG vanish at exactly the same states.
Predict before you look
For n=20 qubits, by how many orders of magnitude is Var[∂1CG] smaller than Var[∂1CL]? Enter log10 of the ratio.
Show the answer
5.5
VarL/VarG=(8/3)n−1/n2. For n=20: 19log10(8/3)−log10400≈8.09−2.60≈5.5. Resolving the global gradient would take roughly 105.5 times more shots, at depth one.
Noise flattens landscapes from the other direction. The cleanest case is global depolarizing noise Dp(ρ)=(1−p)ρ+pI/d after each of L layers. It commutes with unitaries, so the final state and every gradient are simply damped:
Exponential in depth, at every point of the landscape, whatever the initialization.
Where plateaus come from, and the simulability debate
A unifying picture arrived in 2024 from Lie theory. Let g=⟨iH1,…,iHK⟩Lie be the dynamical Lie algebra generated by the ansatz's generators. Deep circuits explore the group eg nearly uniformly, and Ragone et al. and Fontana et al. (2024) showed that the cost variance is then set by dimg and by how much of the input state and of the observable lies in g:
Varθ[C]=dimgPg(ρ)Pg(O)(gsimple,O∈igorρ∈ig)
Pg(X) is the purity of the projection of X onto g. For g=su(2n), of dimension 4n−1, a Pauli O and a pure ρ, this gives 1/(2n+1): a barren plateau. Polynomially large algebras (e.g. free-fermion circuits, with dimg=O(n2)) can avoid it.
In this language the known sources of plateaus are expressibility (large g, or circuits close to 2-designs; Holmes et al., 2022), global observables (little weight of O in small, local pieces), highly entangled or mixed inputs, and noise.
Key takeaways
Barren plateau: Var[∂kC]∈O(b−n); by Chebyshev, resolving gradients then costs Ω(bn) shots, and gradient-free methods fare no better.
If the circuit after a parameter is a 2-design, Var[∂θC]=2n(1−⟨P⟩2)/(2(4n−1))≈2−(n+1).
Global costs plateau even at depth one (variance ∝(3/8)n in the product example); faithful local costs stay trainable in shallow circuits.
Noise damps gradients like qL everywhere; the Lie-algebraic theory ties variance to dimg, and provable absence of plateaus often signals classical simulability.
Check yourself
In the interactive lesson these are scored with your confidence, and what you miss comes back for review.
Question 1
Suppose Var[∂kC]≈2−n. Roughly how many shots does resolving a typical gradient component need at n=40?
about 40
about 1600
about 1012
about 1040
Show the answer
C. about 1012
A typical gradient has size Var≈2−20≈10−6; resolving it needs a standard error below that, i.e. N≳240≈1012 shots, per component, per step.
Question 2
In the 2-design theorem, which ingredient produces the 2−n scaling?
The number of parameters in the circuit
The Pauli generator having eigenvalues ±1
The Haar second moment, which spreads O~=A†OA evenly over a (4n−1)-dimensional space of traceless operators
Measuring in the computational basis
Show the answer
C. The Haar second moment, which spreads O~=A†OA evenly over a (4n−1)-dimensional space of traceless operators
Averaging over A makes O~ a uniformly random traceless operator of fixed norm. Its overlap with any fixed K then has variance ∼Tr(O2)Tr(K2)/4n∼2−n.
Question 3
Why can the local cost CL replace the global cost CG when training toward ∣0⟩⊗n?
Because CL=CG for every state
Because CL≤CG≤nCL, so both vanish at exactly the same states, while CL keeps only polynomially small gradients in shallow circuits
Because local costs have no minima
Because CL ignores entanglement, which never matters for the target
Show the answer
B. Because CL≤CG≤nCL, so both vanish at exactly the same states, while CL keeps only polynomially small gradients in shallow circuits
The sandwich inequality makes CL faithful: a small CL certifies a small CG up to a factor n. And its gradients are Θ(1/n) in the product example instead of (3/8)n/2.
Question 4
Under global depolarizing noise of strength p after each of L layers, by what factor is every gradient scaled?
1−p
(1−p)L
exactly 1−pL
pL
Show the answer
B.(1−p)L
The channel commutes with the unitary layers, so the state is (1−p)LUρU†+(1−(1−p)L)I/d, and the constant part has zero gradient.
Further reading
J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush & H. Neven, "Barren plateaus in quantum neural network training landscapes", Nature Communications 9, 4812 (2018).
M. Cerezo, A. Sone, T. Volkoff, L. Cincio & P. J. Coles, "Cost function dependent barren plateaus in shallow parametrized quantum circuits", Nature Communications 12, 1791 (2021).
S. Wang, E. Fontana, M. Cerezo, K. Sharma, A. Sone, L. Cincio & P. J. Coles, "Noise-induced barren plateaus in variational quantum algorithms", Nature Communications 12, 6961 (2021).
M. Ragone et al., "A Lie algebraic theory of barren plateaus for deep parameterized quantum circuits", Nature Communications (2024).
M. Cerezo et al., "Does provable absence of barren plateaus imply classical simulability? Or, why we need to rethink variational quantum computing" (2023).