Simple Algorithm
Quantum Computing · Rydberg State · 30 min

Barren Plateaus

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)\operatorname{Var}[\partial_\theta C]\approx2^{-(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⁡[O U(θ)ρ U(θ)†]C(\boldsymbol\theta)=\Tr\big[O\,U(\boldsymbol\theta)\rho\,U(\boldsymbol\theta)^\dagger\big] on nn qubits, with parameters drawn from an initialization distribution.

DerivationWhy exponentially small variance kills training
  1. Pr⁡[∣∂kC∣≥δ]  ≤  Var⁡[∂kC]δ2\Pr\big[\abs{\partial_kC}\ge\delta\big]\;\le\;\frac{\operatorname{Var}[\partial_kC]}{\delta^2}

    Chebyshev's inequality turns a small variance into concentration:

  2. N  ≳  ∥O∥2Var⁡[∂kC]=Ω(b n)N\;\gtrsim\;\frac{\norm{O}^2}{\operatorname{Var}[\partial_kC]}=\Omega\big(b^{\,n}\big)

    So a random starting point has a gradient of size δ\delta only with probability O(b−n/δ2)\order{b^{-n}/\delta^2}. With the parameter-shift rule, each component is estimated with standard error ∼∥O∥/N\sim\norm{O}/\sqrt N; resolving a typical gradient, of size Var⁡∼b−n/2\sqrt{\operatorname{Var}}\sim b^{-n/2}, needs

  3. 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.

DerivationProof
  1. ∂θC=−i2Tr⁡(O~ [P,σθ])=Tr⁡(O~K),K:=i2[σθ,P]  (Hermitian, traceless)\partial_\theta C=-\tfrac i2\Tr\big(\tilde O\,[P,\sigma_\theta]\big)=\Tr\big(\tilde OK\big),\qquad K:=\tfrac i2[\sigma_\theta,P]\ \ (\text{Hermitian, traceless})

    Differentiate, using ∂θσθ=−i2[P,σθ]\partial_\theta\sigma_\theta=-\tfrac i2[P,\sigma_\theta], and write O~=A†OA\tilde O=A^\dagger OA:

  2. EA[(∂θC)2]=Tr⁡(K2)d2−1(Tr⁡(O2)−(Tr⁡O)2d)=d Tr⁡(K2)d2−1\mathbb E_A\big[(\partial_\theta C)^2\big]=\frac{\Tr(K^2)}{d^2-1}\Big(\Tr(O^2)-\frac{(\Tr O)^2}{d}\Big)=\frac{d\,\Tr(K^2)}{d^2-1}

    First moment: a 2-design is also a 1-design, so EA[A†OA]=Tr⁡OdI=0\mathbb E_A[A^\dagger OA]=\frac{\Tr O}{d}\id=0 and E[∂θC]=0\mathbb E[\partial_\theta C]=0. Second moment: the Haar (Weingarten) formula for E[Tr⁡(A†OAK)2]\mathbb E\big[\Tr(A^\dagger OAK)^2\big], with Tr⁡K=0\Tr K=0, gives

  3. Tr⁡(K2)=−14Tr⁡([σθ,P]2)=12(1−⟨P⟩σθ2)\Tr(K^2)=-\tfrac14\Tr\big([\sigma_\theta,P]^2\big)=\tfrac12\big(1-\expval{P}_{\sigma_\theta}^2\big)

    For pure σθ\sigma_\theta and P2=IP^2=\id, Tr⁡([σθ,P]2)=2Tr⁡(σθPσθP)−2Tr⁡(σθ2)=2(⟨P⟩2−1)\Tr\big([\sigma_\theta,P]^2\big)=2\Tr(\sigma_\theta P\sigma_\theta P)-2\Tr(\sigma_\theta^2)=2\big(\expval{P}^2-1\big), so

  4. Combine with d=2nd=2^n. Sanity check at n=1n=1: O~\tilde O is a uniformly random Pauli axis m^⋅σ⃗\hat m\cdot\vec\sigma, and (for ⟨P⟩=0\expval P=0) ∂θC\partial_\theta C is one Cartesian component of m^\hat m, whose variance is 13\tfrac13, matching 2⋅12⋅3\frac{2\cdot1}{2\cdot3}. □\square

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 2n2^n, not by depth as such: depth matters insofar as it makes the circuit 2-design-like.

Check your understanding

In the theorem, take O=Z1O=Z_1, a local observable, but keep AA a 2-design on all nn qubits. What happens to the variance?

  1. It becomes polynomial in nn, because OO is local
  2. It is unchanged, Θ(2−n)\Theta(2^{-n}): a global 2-design scrambles Z1Z_1 over all qubits
  3. It becomes exactly zero
  4. It grows exponentially with nn
Show the answer

B. It is unchanged, Θ(2−n)\Theta(2^{-n}): a global 2-design scrambles Z1Z_1 over all qubits

The formula only uses Tr⁡O=0\Tr O=0 and Tr⁡O2=2n\Tr O^2=2^n, which hold for Z1Z_1 as for any Pauli string. Locality helps only when the circuit is shallow, so that the backward light cone of Z1Z_1 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⟩\ket{\psi(\boldsymbol\theta)}=\bigotimes_{j=1}^nR_X(\theta_j)\ket0 with θj\theta_j uniform on [−π,π][-\pi,\pi], and the target ∣0⟩⊗n\ket0^{\otimes n}; qubit jj is found in ∣0⟩\ket0 with probability cos⁡2θj2\cos^2\frac{\theta_j}{2}.

DerivationGlobal versus local cost for a product ansatz
  1. CG=1−∏j=1ncos⁡2θj2,CL=1−1n∑j=1ncos⁡2θj2C_G=1-\prod_{j=1}^n\cos^2\tfrac{\theta_j}{2},\qquad C_L=1-\frac1n\sum_{j=1}^n\cos^2\tfrac{\theta_j}{2}

    Global cost (infidelity with the target) and local cost (average single-qubit infidelity):

  2. ∂1CG=12sin⁡θ1∏j≥2cos⁡2θj2,∂1CL=12nsin⁡θ1\partial_1C_G=\tfrac12\sin\theta_1\prod_{j\ge2}\cos^2\tfrac{\theta_j}{2},\qquad\partial_1C_L=\frac{1}{2n}\sin\theta_1

    Gradients, using ∂θcos⁡2θ2=−12sin⁡θ\partial_\theta\cos^2\frac\theta2=-\frac12\sin\theta:

  3. Var⁡[∂1CG]=18(38)n−1,Var⁡[∂1CL]=18n2\operatorname{Var}[\partial_1C_G]=\frac18\Big(\frac38\Big)^{n-1},\qquad\operatorname{Var}[\partial_1C_L]=\frac{1}{8n^2}

    Both have mean zero. With E[sin⁡2θ]=12\mathbb E[\sin^2\theta]=\frac12 and E[cos⁡4θ2]=E[(1+cos⁡θ2)2]=38\mathbb E[\cos^4\frac\theta2]=\mathbb E\big[(\frac{1+\cos\theta}{2})^2\big]=\frac38:

  4. CL  ≤  CG  ≤  n CL,C_L\;\le\;C_G\;\le\;n\,C_L,

    Same circuit, same minimum, exponentially different trainability. The local cost is still faithful: since ∏jpj≤pk\prod_jp_j\le p_k for every kk, and 1−∏jpj≤∑j(1−pj)1-\prod_jp_j\le\sum_j(1-p_j) by the union bound,

  5. so CLC_L and CGC_G vanish at exactly the same states.

Predict before you look

For n=20n=20 qubits, by how many orders of magnitude is Var⁡[∂1CG]\operatorname{Var}[\partial_1C_G] smaller than Var⁡[∂1CL]\operatorname{Var}[\partial_1C_L]? Enter log⁡10\log_{10} of the ratio.

Show the answer

5.5

Var⁡L/Var⁡G=(8/3)n−1/n2\operatorname{Var}_L/\operatorname{Var}_G=(8/3)^{n-1}/n^2. For n=20n=20: 19log⁡10(8/3)−log⁡10400≈8.09−2.60≈5.519\log_{10}(8/3)-\log_{10}400\approx8.09-2.60\approx5.5. Resolving the global gradient would take roughly 105.510^{5.5} 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+p I/d\mathcal D_p(\rho)=(1-p)\rho+p\,\id/d after each of LL layers. It commutes with unitaries, so the final state and every gradient are simply damped:

ρL=(1−p)L UρU†+(1−(1−p)L)Id⟹∂kCnoisy=(1−p)L ∂kCideal\rho_L=(1-p)^L\,U\rho U^\dagger+\big(1-(1-p)^L\big)\frac{\id}{d}\quad\Longrightarrow\quad\partial_kC_{\text{noisy}}=(1-p)^L\,\partial_kC_{\text{ideal}}
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\mathfrak g=\langle iH_1,\dots,iH_K\rangle_{\text{Lie}} be the dynamical Lie algebra generated by the ansatz's generators. Deep circuits explore the group ege^{\mathfrak g} nearly uniformly, and Ragone et al. and Fontana et al. (2024) showed that the cost variance is then set by dim⁡g\dim\mathfrak g and by how much of the input state and of the observable lies in g\mathfrak g:

Var⁡θ[C]=Pg(ρ) Pg(O)dim⁡g(g simple, O∈ig or ρ∈ig)\operatorname{Var}_{\boldsymbol\theta}[C]=\frac{\mathcal P_{\mathfrak g}(\rho)\,\mathcal P_{\mathfrak g}(O)}{\dim\mathfrak g}\qquad(\mathfrak g\ \text{simple},\ O\in i\mathfrak g\ \text{or}\ \rho\in i\mathfrak g)
Pg(X)\mathcal P_{\mathfrak g}(X) is the purity of the projection of XX onto g\mathfrak g. For g=su(2n)\mathfrak g=\mathfrak{su}(2^n), of dimension 4n−14^n-1, a Pauli OO and a pure ρ\rho, this gives 1/(2n+1)1/(2^n+1): a barren plateau. Polynomially large algebras (e.g. free-fermion circuits, with dim⁡g=O(n2)\dim\mathfrak g=O(n^2)) can avoid it.

In this language the known sources of plateaus are expressibility (large g\mathfrak g, or circuits close to 2-designs; Holmes et al., 2022), global observables (little weight of OO in small, local pieces), highly entangled or mixed inputs, and noise.

Key takeaways

  • Barren plateau: Var⁡[∂kC]∈O(b−n)\operatorname{Var}[\partial_kC]\in\order{b^{-n}}; by Chebyshev, resolving gradients then costs Ω(bn)\Omega(b^n) 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)\operatorname{Var}[\partial_\theta C]=2^n(1-\langle P\rangle^2)/(2(4^n-1))\approx2^{-(n+1)}.
  • Global costs plateau even at depth one (variance ∝(3/8)n\propto(3/8)^n in the product example); faithful local costs stay trainable in shallow circuits.
  • Noise damps gradients like qLq^L everywhere; the Lie-algebraic theory ties variance to dim⁡g\dim\mathfrak g, 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\operatorname{Var}[\partial_kC]\approx2^{-n}. Roughly how many shots does resolving a typical gradient component need at n=40n=40?

  1. about 40
  2. about 1600
  3. about 101210^{12}
  4. about 104010^{40}
Show the answer

C. about 101210^{12}

A typical gradient has size Var⁡≈2−20≈10−6\sqrt{\operatorname{Var}}\approx2^{-20}\approx10^{-6}; resolving it needs a standard error below that, i.e. N≳240≈1012N\gtrsim2^{40}\approx10^{12} shots, per component, per step.

Question 2

In the 2-design theorem, which ingredient produces the 2−n2^{-n} scaling?

  1. The number of parameters in the circuit
  2. The Pauli generator having eigenvalues ±1\pm1
  3. The Haar second moment, which spreads O~=A†OA\tilde O=A^\dagger OA evenly over a (4n−1)(4^n-1)-dimensional space of traceless operators
  4. Measuring in the computational basis
Show the answer

C. The Haar second moment, which spreads O~=A†OA\tilde O=A^\dagger OA evenly over a (4n−1)(4^n-1)-dimensional space of traceless operators

Averaging over AA makes O~\tilde O a uniformly random traceless operator of fixed norm. Its overlap with any fixed KK then has variance ∼Tr⁡(O2)Tr⁡(K2)/4n∼2−n\sim\Tr(O^2)\Tr(K^2)/4^n\sim2^{-n}.

Question 3

Why can the local cost CLC_L replace the global cost CGC_G when training toward ∣0⟩⊗n\ket0^{\otimes n}?

  1. Because CL=CGC_L=C_G for every state
  2. Because CL≤CG≤nCLC_L\le C_G\le nC_L, so both vanish at exactly the same states, while CLC_L keeps only polynomially small gradients in shallow circuits
  3. Because local costs have no minima
  4. Because CLC_L ignores entanglement, which never matters for the target
Show the answer

B. Because CL≤CG≤nCLC_L\le C_G\le nC_L, so both vanish at exactly the same states, while CLC_L keeps only polynomially small gradients in shallow circuits

The sandwich inequality makes CLC_L faithful: a small CLC_L certifies a small CGC_G up to a factor nn. And its gradients are Θ(1/n)\Theta(1/n) in the product example instead of (3/8)n/2(3/8)^{n/2}.

Question 4

Under global depolarizing noise of strength pp after each of LL layers, by what factor is every gradient scaled?

  1. 1−p1-p
  2. (1−p)L(1-p)^L
  3. exactly 1−pL1-pL
  4. pLp^L
Show the answer

B. (1−p)L(1-p)^L

The channel commutes with the unitary layers, so the state is (1−p)LUρU†+(1−(1−p)L)I/d(1-p)^LU\rho U^\dagger+\big(1-(1-p)^L\big)\id/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).