betaFeedback

WorksMathematics

Counterexamples to RR-superlinear convergence of cyclic steepest descent

Abstract

The cyclic steepest descent method computes the exact steepest descent stepsize at the start of each cycle and reuses it for a fixed number mm of iterations. For strictly convex quadratics in nn variables, it has been suggested that the method converges RR-superlinearly whenever m≥(n+1)/2m\ge(n+1)/2. We show that this cannot hold for every initial point, already at the threshold cycle length. For n=m=2n=m=2 we exhibit an initial point from which the method follows the classical zigzag trajectory of steepest descent: the stepsize never changes and the error is halved at every iteration. The same mechanism yields linearly convergent trajectories for every Hessian with at least two distinct eigenvalues and every cycle length. For n=3n=3 and m=2m=2 we construct a second, genuinely three-dimensional counterexample: all components of the gradient remain nonzero, consecutive cycles use two different stepsizes, and every four iterations reduce the gradient exactly by the factor 1/491/49. Both trajectories are given in closed form. Consequently, any RR-superlinear convergence result for this method at the threshold must exclude a set of initial points, as recent almost-everywhere results do.

Introduction

Cyclic steepest descent (CSD) reuses an exact steepest-descent stepsize for a prescribed number of consecutive updates. It is a special delayed gradient method in the sense of Friedlander, Martínez, Molina, and Raydan [6]; Dai and Fletcher studied its asymptotic behavior on positive-definite quadratics [3]. With reference to that analysis, Dai’s ICM 2022 survey states that CSD is likely to converge RR-superlinearly when the cycle length satisfies m≥⌈(n+1)/2⌉m \ge\lceil(n+1)/2 \rceil [2]; the survey writes the condition as m≥(n+1)/2m \ge(n+1)/2, which for integer mm is the ceiling condition used here and is equivalent to n<2mn < 2m, the superlinear side of the threshold in the simplified model of Dai and Fletcher. The ICM Conjectures directory records the corresponding universal assertion as problem 2022-opt-18 [8]. Our counterexamples address this universal formulation, not an almost-everywhere interpretation of the survey’s conjectural statement.

Theorem 3.1 gives a counterexample at the smallest nontrivial dimension and cycle length (for n=1n = 1 the method terminates after one update), with error norm ratio 1/21/2 at every actual update. The statement describes the complete orbit, not a finite numerical trace. Proposition 3.2 extends this invariant family to arbitrary pairs of eigendirections and cycle lengths.

Theorem 3.3 gives a second obstruction at the three-dimensional threshold: A=diag⁡(1,2,3)A = \operatorname{diag}(1, 2, 3) and m=2m = 2. All three spectral components remain nonzero. The cycle-boundary weights alternate between two interior points of the simplex, and every four updates multiply the gradient and error by 1/491/49. The proof identifies the two Rayleigh quotients and verifies the two weight transitions exactly. Thus both counterexamples describe entire trajectories, but only the first uses a constant stepsize.

The underlying spectral mechanism is classical for ordinary exact-line-search steepest descent. Akaike’s probability-map analysis contains the equal-weight fixed distribution [1], and Yu recently stated the corresponding worst-case linear orbit and its higher-dimensional embedding explicitly [13 Proposition 3.1 of the arXiv version]. The decisive observation for CSD is that the exact stepsize is constant on this orbit. Consequently, cyclic reuse for any number of updates preserves the orbit exactly, turning the classical equal-weight trajectory into a threshold counterexample to the universal CSD assertion.

Periodic behavior under cyclic reuse has a direct precedent. Dai, Hager, Schittkowski, and Zhang [4 Section 3.3] identify the alternate-step method as CSD with m=2m = 2 and construct a five-dimensional cyclic example. Theorem 3.3 provides a closed-form full-support example at the three-dimensional threshold; periodic CSD behavior itself is already documented. The alternate minimization method of Dai and Yuan [5] uses different steps and has a two-dimensional QQ-superlinear result. Shen, Li, and Dai [11 Theorem 1] state RR-superlinear convergence of the cycle-boundary gradient subsequence for two-dimensional CSD; their argument parameterizes the initial gradient by the logarithm of its component ratio and uses that this logarithm is nonzero. The balanced initializations of Theorem 3.1 and Proposition 3.2 are exactly the fixed point of their recurrence at which this logarithm vanishes, where the cycle-boundary subsequence is geometric; this boundary case is treated here. Their revised preprint studies the triangle steepest-descent method instead [12].

The companion paper [9 Theorem 3.1] classifies CSD for Hessians with exactly two distinct eigenvalues and arbitrary eigenspace multiplicities. For m≥2m \ge2, the equal-norm spectral-block cone is precisely the nonterminating exceptional set; every other mixed initialization is RR-superlinear. In the fixed two-dimensional instance here this cone is the union ∣g0,1∣=∣g0,2∣>0|g_{0,1}| = |g_{0,2}| > 0 of two lines. The two papers therefore separate the construction of exact exceptions from the classification of all initializations for two-spectrum Hessians. A recent preprint of Gu [7 Theorem 2.1] states an almost-everywhere doubly exponential convergence theorem for general spectra when 2m2m exceeds the number of distinct eigenvalues. Both instances here satisfy that condition, so by that theorem the orbits constructed below must lie in its Lebesgue-null exceptional set; by Theorem 3.3 that set can contain full-support, nonresonant orbits, so it cannot be dropped. Exact exceptional orbits and almost-everywhere convergence are compatible. Neither positive result is needed in our proofs.

§ 2 fixes the method, the notation, and the rate terminology. § 3 proves the two counterexamples and the invariant family. § 4 summarizes the consequences.

Cyclic Steepest Descent

On Rn\mathbb{R}^n, let

q(x)=12xTAx−cTx,A=AT≻0,(1)q(x) = \frac{1}{2}x^{\mathsf{T}}Ax-c^{\mathsf{T}}x,\qquad A=A^{\mathsf{T}}\succ0, \tag*{(1)}

where c∈Rnc\in\mathbb{R}^n, and write gk=Axk−cg_k=Ax_k-c. All norms are Euclidean. The method terminates whenever gk=0g_k=0. Otherwise, the exact steepest-descent stepsize is

αkSD:=gkTgkgkTAgk.(2)\alpha_k^{\mathrm{SD}} := \frac{g_k^{\mathsf{T}}g_k}{g_k^{\mathsf{T}}Ag_k}. \tag*{(2)}

Let m≥1m\ge1 be an integer. At the boundary mℓm\ell of cycle ℓ\ell, CSD computes αmℓSD\alpha_{m\ell}^{\mathrm{SD}} from (2) and performs the mm updates

xmℓ+i+1=xmℓ+i−αmℓSDgmℓ+i,i=0,…,m−1.(3)x_{m\ell+i+1}=x_{m\ell+i}-\alpha_{m\ell}^{\mathrm{SD}}g_{m\ell+i},\qquad i=0,\ldots,m-1. \tag*{(3)}

Thus a newly computed exact stepsize is used for exactly mm consecutive updates. Equivalent index shifts do not change the method. The unique minimizer of (1) is x∗=A−1cx_\ast=A^{-1}c. Set

ek=xk−x∗,gk=Aek.(4)e_k=x_k-x_\ast,\qquad g_k=Ae_k. \tag*{(4)}

For any update with stepsize αk\alpha_k,

gk+1=(I−αkA)gk.(5)g_{k+1}=(I-\alpha_k A)g_k. \tag*{(5)}

A null sequence {ek}\{e_k\} converges RR-superlinearly if ∥ek∥21/k→0\lVert e_k\rVert_2^{1/k}\to0, that is, if its root-convergence factor lim sup⁡k∥ek∥21/k\limsup_k\lVert e_k\rVert_2^{1/k} is zero [10 Chapter 9]; equivalently, ∥ek∥2\lVert e_k\rVert_2 is bounded by a null sequence whose successive ratios tend to zero, which is the form used in [11]. It converges RR-linearly if this root factor lies in (0,1)(0,1). A positive geometric lower bound on the errors, even along a fixed-stride subsequence, forces a positive root factor and hence excludes RR-superlinear convergence. This elementary observation suffices for the exact orbits below.

Exact Counterexamples

The Balanced Zig-Zag Obstruction

Theorem 3.1 (Exact threshold counterexample) Let

n=2,m=2,A=(1003),c=(00),x0=(11/3).(6)n=2,\qquad m=2,\qquad A=\begin{pmatrix}1&0\\0&3\end{pmatrix},\qquad c=\begin{pmatrix}0\\0\end{pmatrix},\qquad x_0=\begin{pmatrix}1\\1/3\end{pmatrix}. \tag*{(6)}

Every stepsize recomputed by CSD is 1/21/2, and the complete orbit is

gk=2−k(1(−1)k),(7)g_k=2^{-k}\begin{pmatrix}1\\(-1)^k\end{pmatrix}, \tag*{(7)}
xk=ek=2−k(1(−1)k/3).(8)x_k=e_k=2^{-k}\begin{pmatrix}1\\(-1)^k/3\end{pmatrix}. \tag*{(8)}

It is nonterminating and satisfies

∥ek+1∥2∥ek∥2=12(k≥0).(9)\frac{\lVert e_{k+1}\rVert_2}{\lVert e_k\rVert_2}=\frac{1}{2}\qquad(k\geq0). \tag*{(9)}

Consequently, the assertion of universal RR-superlinear convergence for m≥⌈(n+1)/2⌉m\geq\lceil(n+1)/2\rceil does not hold.

Proof of Theorem 3.1. Here x∗=0x_\ast=0 and g0=(1,1)Tg_0=(1,1)^{\mathsf T}. Whenever a cycle starts with two gradient components of equal magnitude, its recomputed stepsize is

gTggTAg=12.\frac{g^{\mathsf T}g}{g^{\mathsf T}Ag}=\frac{1}{2}.

For this stepsize,

I−12A=(1/200−1/2).I-\frac{1}{2}A=\begin{pmatrix}1/2&0\\0&-1/2\end{pmatrix}.

The two component magnitudes therefore remain equal after each of the two updates in a cycle. Induction over the cycles gives (7); applying A−1A^{-1} gives (8), and (9) follows. In particular ∥ek∥2=(10/3) 2−k\lVert e_k\rVert_2=(\sqrt{10}/3)\,2^{-k}, so the error sequence has exactly RR-linear convergence with factor 1/21/2 and hence, by the observation of Section 2, is not RR-superlinearly convergent. Finally, m=2=⌈(2+1)/2⌉m=2=\lceil(2+1)/2\rceil is the smallest admissible integer cycle length, so the universal assertion fails at its threshold. The proof is complete. ☐

The same calculation extends to any pair of eigendirections and any cycle length.

Proposition 3.2 (Two-eigendirection invariant family) Let A≻0A\succ0 have orthonormal eigenvectors u,vu,v with distinct eigenvalues 0<a<b0<a<b, respectively. Fix any cycle length m≥1m\geq1, any t∈R∖{0}t\in\mathbb{R}\setminus\{0\}, and any σ∈{−1,1}\sigma\in\{-1,1\}. Initialize CSD so that

g0=t(u+σv).(10)g_0=t(u+\sigma v). \tag*{(10)}

Define

ρ:=b−ab+a∈(0,1).(11)\rho:= \frac{b-a}{b+a} \in(0,1). \tag*{(11)}

Then every stepsize recomputed at a cycle boundary is

αmℓSD=2a+b,ℓ=0,1,…,(12)\alpha^{\mathrm{SD}}_{m\ell} = \frac{2}{a+b}, \qquad\ell=0,1,\ldots, \tag*{(12)}

and the full orbit is

gk=tρk(u+(−1)kσv),(13)g_k=t\rho^k\left(u+(-1)^k\sigma v\right), \tag*{(13)}
ek=tρk(a−1u+(−1)kσb−1v).(14)e_k=t\rho^k\left(a^{-1}u+(-1)^k\sigma b^{-1}v\right). \tag*{(14)}

In particular,

∥gk+1∥2∥gk∥2=∥ek+1∥2∥ek∥2=ρ(k≥0).(15)\frac{\lVert g_{k+1}\rVert_2}{\lVert g_k\rVert_2} = \frac{\lVert e_{k+1}\rVert_2}{\lVert e_k\rVert_2} =\rho\qquad(k\ge0). \tag*{(15)}

None of the full gradient sequence, the full error sequence, and their cycle-boundary subsequences converges RR-superlinearly.

Proof of Proposition 3.2. For a vector of the form s(u+τv)s(u+\tau v) with s≠0s\ne0 and τ∈{−1,1}\tau\in\{-1,1\}, orthonormality and the eigenvalue relations give

gTggTAg=2s2(a+b)s2=2a+b.\frac{g^{T}g}{g^{T}Ag}=\frac{2s^2}{(a+b)s^2}=\frac{2}{a+b}.

For this stepsize,

(I−αA)u=ρu,(I−αA)v=−ρv.(I-\alpha A)u=\rho u,\qquad(I-\alpha A)v=-\rho v.

Equation (5) therefore changes the sign of the vv-component and multiplies both component magnitudes by ρ\rho. After any number of updates, and hence at every cycle boundary, the two components still have equal magnitude. Induction proves (12) and (13). Applying A−1A^{-1} gives (14).

The norms of the two parenthesized vectors in (13) and (14) are independent of kk, so (15) follows. The full sequences are geometric with factor ρ>0\rho>0, while their cycle-boundary subsequences are geometric with factor ρm>0\rho^m>0 in the cycle index; by the observation of § 2, none of them converges RR-superlinearly. The proof is complete. ⊓⊔

A Full-Support Non-Zig-Zag Obstruction

Theorem 3.3 (Full-support three-spectral period-two counterexample). Let

n=3,m=2,A=diag⁡(1,2,3),c=0,(16)n=3,\qquad m=2,\qquad A=\operatorname{diag}(1,2,3),\qquad c=0, \tag*{(16)}

and set

g0=((2+2)/41/2(2−2)/4),x0=A−1g0=((2+2)/41/4(2−2)/12).(17)g_0= \begin{pmatrix} (2+\sqrt{2})/4\\ 1/2\\ (2-\sqrt{2})/4 \end{pmatrix}, \qquad x_0=A^{-1}g_0= \begin{pmatrix} (2+\sqrt{2})/4\\ 1/4\\ (2-\sqrt{2})/12 \end{pmatrix}. \tag*{(17)}

Define

μ−:=2−22,μ+:=2+22,α±:=1μ±.\mu_- := 2-\frac{\sqrt{2}}{2},\qquad\mu_+ := 2+\frac{\sqrt{2}}{2},\qquad\alpha_{\pm} := \frac{1}{\mu_{\pm}}.

The CSD orbit is nonterminating, has full support (every component of gkg_k is nonzero for every kk), and is nonresonant: no cycle-boundary stepsize annihilates a spectral component, i.e., μ±≠λi\mu_{\pm} \ne\lambda_i for every ii. The orbit has projective period two at cycle boundaries: writing wk,i=gk,i2/∑j=13gk,j2w_{k,i}=g_{k,i}^2/\sum_{j=1}^3 g_{k,j}^2, the normalized squared component weights w2ℓw_{2\ell} alternate between two distinct interior points of the simplex. The stepsize pattern is α−,α−,α+,α+,…\alpha_-,\alpha_-,\alpha_+,\alpha_+,\ldots, and

gk+4=149gk,ek+4=149ek(k≥0).(18)g_{k+4}=\frac{1}{49}g_k,\qquad e_{k+4}=\frac{1}{49}e_k\qquad(k\ge0). \tag*{(18)}

Consequently the orbit is RR-linearly convergent with root factor 49−1/4=1/749^{-1/4}=1/\sqrt{7}, but not RR-superlinearly convergent, and it is not a two-dimensional balanced-zigzag orbit.

Proof of Theorem 3.3. For a nonzero gradient gg, write wi=gi2/∑j=13gj2w_i=g_i^2/\sum_{j=1}^3g_j^2 and μ=∑i=13λiwi\mu=\sum_{i=1}^3\lambda_iw_i, where (λ1,λ2,λ3)=(1,2,3)(\lambda_1,\lambda_2,\lambda_3)=(1,2,3). At a cycle boundary the exact stepsize is α=1/μ\alpha=1/\mu; after two updates the iith gradient component is multiplied by (1−λi/μ)2(1-\lambda_i/\mu)^2. Therefore the normalized squared component weights obey the map

Ti(w)=wi∣μ−λi∣4∑jwj∣μ−λj∣4.(19)T_i(w)=\frac{w_i\lvert\mu-\lambda_i\rvert^4}{\sum_j w_j\lvert\mu-\lambda_j\rvert^4}. \tag*{(19)}

Let w−=(g0,12,g0,22,g0,32)w^-=(g_{0,1}^2,g_{0,2}^2,g_{0,3}^2), whose entries sum to one, so that w−w^- is the initial weight vector; let w+w^+ be obtained by interchanging 2\sqrt{2} with −2-\sqrt{2} in the three entries. Explicitly,

w−=(3+228,14,3−228),w+=(3−228,14,3+228).w^-=\left(\frac{3+2\sqrt{2}}{8},\frac{1}{4},\frac{3-2\sqrt{2}}{8}\right),\qquad w^+=\left(\frac{3-2\sqrt{2}}{8},\frac{1}{4},\frac{3+2\sqrt{2}}{8}\right).

Both vectors are positive and sum to one, so they are interior points of the simplex. Their weighted means are μ−\mu_- and μ+\mu_+, respectively. Since

(1−μ−)(1−μ+)=12,(2−μ−)(2−μ+)=−12,(3−μ−)(3−μ+)=12,(1-\mu_-)(1-\mu_+)=\frac{1}{2},\qquad (2-\mu_-)(2-\mu_+)=-\frac{1}{2},\qquad (3-\mu_-)(3-\mu_+)=\frac{1}{2},

direct substitution gives

wi−∣μ−−λi∣4=14wi+,wi+∣μ+−λi∣4=14wi−.w_i^-\lvert\mu_- - \lambda_i\rvert^4=\frac{1}{4}w_i^+,\qquad w_i^+\lvert\mu_+ - \lambda_i\rvert^4=\frac{1}{4}w_i^-.

Hence T(w−)=w+T(w^-)=w^+ and T(w+)=w−T(w^+)=w^-, so w4ℓ=w−w_{4\ell}=w^- and w4ℓ+2=w+w_{4\ell+2}=w^+ for every ℓ≥0\ell\ge0. This proves the cycle-boundary projective period two and the displayed stepsize pattern. Since μ±∉{1,2,3}\mu_{\pm}\notin\{1,2,3\}, each update multiplies every gradient component by a nonzero factor 1−λiα±1-\lambda_i\alpha_{\pm}; hence all three components are nonzero at every iteration, so the orbit is nonterminating, has full support, and is nonresonant. Moreover μ−μ+=7/2\mu_{-}\mu_{+}=7/2, so over the four updates the iith gradient component is multiplied by

[(μ−−λi)(μ+−λi)μ−μ+]2=149\left[\frac{(\mu_{-}-\lambda_i)(\mu_{+}-\lambda_i)}{\mu_{-}\mu_{+}}\right]^2=\frac{1}{49}

for each ii. Any four consecutive updates contain two factors of each type, and these diagonal update matrices commute. Thus the identity holds for every starting index kk, proving (18); applying A−1A^{-1} gives the error relation. The subsequence {∥e4k∥2}\{\lVert e_{4k}\rVert_2\} is positive geometric with ratio 1/491/49, so by the observation of Section 2 the orbit is not RR-superlinearly convergent, while (18) gives the stated RR-linear root factor 49−1/449^{-1/4}. The proof is complete. ⊓⊔

Remark 3.4 (Higher-dimensional embedding). For every n≥2n \ge2 and m≥1m \ge1, augment diag⁡(a,b)\operatorname{diag}(a,b) by arbitrary positive diagonal entries and initialize the added gradient components at zero. They remain zero, so the active plane follows Proposition 3.2, and the obstruction occurs in every dimension and at every cycle length as a lower-dimensional support phenomenon. Theorem 3.3 is different: all three spectral components are nonzero, so its obstruction is full-support. Either theorem alone disproves the universal statement.

Conclusions

Cyclic reuse preserves the classical balanced orbit and also admits the full-support three-dimensional periodic orbit constructed above. Together these examples show why a universal convergence guarantee must account for exact exceptional initializations, even at the proposed cycle threshold. The three-dimensional construction also shows that full spectral support does not exclude periodic projective dynamics.

Sampling from a distribution absolutely continuous with respect to Lebesgue measure misses a fixed Lebesgue-null set with probability one, whereas prescribed or discretely sampled initializations may lie on it. Exceptional-orbit checks therefore matter for guarantees covering a specified initialization, alongside almost-everywhere convergence results.

Acknowledgements We thank Yu-Hong Dai for helpful comments on the exceptional-orbit calculation, its interpretation as a classical zigzagging trajectory, and the distinction between universal and almost-everywhere convergence. We acknowledge Gewu Intelligence Lab for providing the collaborative research environment in which this project was developed.

Declarations

Funding. No funding was received for this work.

Competing interests. The authors declare no competing interests.

Data availability. No datasets were generated or analysed. The results are established analytically in the article.

Disclosure of automated assistance

The main results of this paper were obtained through a generative-AI workflow using OpenAI GPT-5.6 Sol, Anthropic Claude Fable 5, OpenAI GPT-6 Astra, OpenAI GPT-6.1 Sol, Anthropic Claude Fable 5.1, and Anthropic Claude Opus 5.5. The workflow contributed to identifying and checking the exact counterexamples, reviewing the cited literature, and preparing and revising the LaTeX manuscript.

References

  1. [1]Hirotugu Akaike, On a successive transformation of probability distribution and its application to the analysis of the optimum gradient method, *Ann. Inst. Stat. Math.* 11 (1959), no. 1, 1–16. https://doi.org/10.1007/BF01831719.
  2. [2]Yu-Hong Dai, An overview of nonlinear optimization, in *Proceedings of the International Congress of Mathematicians—2022*, vol. 7, 5290–5313, EMS Press, Berlin, 2023. https://doi.org/10.4171/ICM2022/72.
  3. [3]Yu-Hong Dai and Roger Fletcher, On the asymptotic behaviour of some new gradient methods, *Math. Program.* 103 (2005), no. 3, 541–559. https://doi.org/10.1007/s10107-004-0516-9.
  4. [4]Yu-Hong Dai, William W. Hager, Klaus Schittkowski, and Hongchao Zhang, The cyclic Barzilai–Borwein method for unconstrained optimization, *IMA J. Numer. Anal.* 26 (2006), no. 3, 604–627. https://doi.org/10.1093/imanum/drl006.
  5. [5]Yu-Hong Dai and Ya-Xiang Yuan, Alternate minimization gradient method, *IMA J. Numer. Anal.* 23 (2003), no. 3, 377–393. https://doi.org/10.1093/imanum/23.3.377.
  6. [6]Ana Friedlander, José Mario Martínez, Brigida Molina, and Marcos Raydan, Gradient method with retards and generalizations, *SIAM J. Numer. Anal.* 36 (1998), no. 1, 275–289. https://doi.org/10.1137/S003614299427315X.
  7. [7]Ran Gu, *Doubly exponential convergence of the cyclic steepest descent method for strictly convex quadratics in arbitrary dimensions*, arXiv:2609.25800v1 [math.OC] (2026).arXiv
  8. [8]Conjectures, I. (2026). Entry 2022-opt-18: cyclic steepest descent, curated online directory. https://icmconjectures.com/2022-opt-18icmconjectures.com/2022-opt-18
  9. [9]Yu Li and Qihang Wang, *Convergence of cyclic steepest descent for quadratics with two distinct eigenvalues*, arXiv:2610.15335v2 [math.OC] (2026).arXiv
  10. [10]James M. Ortega and Werner C. Rheinboldt, *Iterative Solution of Nonlinear Equations in Several Variables*, Classics in Applied Mathematics, vol. 30, SIAM, Philadelphia, 2000. https://doi.org/10.1137/1.9780898719468.
  11. [11]Shen, Y., Li, Q.-N., & Dai, Y.-H. (2025). On the acceleration of gradient methods: the triangle steepest descent method. arXiv:2501.16731v1
  12. [12]Ya Shen, Qing-Na Li, and Yu-Hong Dai, *Triangle steepest descent: a geometry-based gradient algorithm with guaranteed R-linear convergence*, arXiv:2501.16731v2 [math.OC], revised October 17, 2025.arXiv
  13. [13]Thomas Yu, The essential best and average rate of convergence of the exact line search gradient descent method, *Numer. Algorithms* 103 (2026), no. 1, 683–710. https://doi.org/10.1007/s11075-025-02244-0. arXiv:2305.09140v4 [math.OC].

Paper details

vlnc:W_bcfs5jrv
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents