{"id":"e08bc570-49ea-407e-9ccb-327aeb76bd34","arxiv_id":"2412.13754","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"This paper claims the semi-supervised exact-recovery threshold for CSBM is I(aτ,bτ,cτ)>1 and that PCA, ridge regression on a linear GCN, and a trained GCN achieve it; the GCN proof is invalidated by an inconsistent estimator of the optimal self-loop weight.","lead":"This paper derives the information-theoretic limit for perfectly labeling all unlabeled nodes in the contextual stochastic block model and proposes spectral and graph-convolution methods that, in theory, reach it. The main threshold is plausible, but the GCN training procedure contains a scaling error that breaks its central guarantee.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma D.5's claimed estimator of the optimal self-loop is off by a factor Θ(√K); under Lemma D.3 the quantity 2 y_L^T X_L W^(1)a/(n^2 q_m) is o(1), not 2cτ, so Theorem 3.9's GCN threshold is unsupported.","rationale":"The reader's weakest assumption identifies precisely the point on which the paper's central claim hinges. The paper's Theorem 3.9 is the advertised conclusion that a gradient-trained GCN achieves the information-theoretic threshold, and its proof depends entirely on s^(1) converging to the oracle self-loop weight. Lemma D.5 asserts this convergence, but a direct scaling calculation using the paper's own Lemma D.3 shows the relevant quantity is o(1), not 2cτ. This is not a disagreement with a prevailing consensus; it is an internal inconsistency between two stated results in the manuscript. The spectral estimator and ridge-regression analyses may survive independently, but the strongest advertised contribution, the GCN achievability claim, is unsupported by the proof. Therefore the reader's REJECT verdict is appropriate, and no relaxation to CONDITIONAL is warranted without a corrected estimator and proof for the GCN self-loop learning step.","tokens_in":40603,"tokens_out":6158,"duration_ms":55604,"concrete_test":"Independently re-derive the left-hand side of Lemma D.5 using Lemma D.3: substitute W^(1)a = √K ζ μ + o(√K) and compute (2/(n^2 q_m)) y_L^T X_L W^(1)a = 2ζ√K(θn + O_P(√n))/(n^2 q_m). If this scale is o(1) rather than 2cτ, Lemma D.5 is refuted. A numerical check: simulate CSBM with N = 10^4, τ = 0.25, q_m = log m, choose aτ = 4, bτ = 1, cτ = 1 so that I(aτ,bτ,cτ) = 1 but I(aτ,bτ,0) = 0.5; run Algorithm D.1 with one GD step for W, compute s^(1), and compare it with s_opt = 2cτ/log(aτ/bτ) ≈ 2.885. If s^(1) is O(1/√N) rather than 2.885, the mechanism of Theorem 3.9 fails.","verdict_should_be":"REJECT","load_bearing_attack":"The load-bearing failure is in the GCN achievability proof (Theorem 3.9). The proof reduces the trained GCN to the ridge-regression setting by showing that s^(1) in (D.2) is close to s_opt = 2cτ/log(aτ/bτ) (Lemma D.7), via Lemma D.5's claim that (2/(n^2 q_m)) y_L^T X_L W^(1)a → 2cτ. This claim contradicts the paper's own Lemma D.3, which gives W^(1)a = √K·ζμ + O(η1/K) with ζ = Θ(1), √K ≍ √N, and θ ≍ √q_m. Since y_L^T X_L μ = θn + y_L^T Z_L μ = Θ(n√q_m), the unscaled numerator is Θ(n^{3/2}√q_m). Therefore (2/(n^2 q_m)) y_L^T X_L W^(1)a = Θ(1/(√n√q_m)) = o(1), not the constant 2cτ. Even replacing W^(1)a by its normalized version W^(1)a/√K gives Θ(θ/(n q_m)) = o(1). Consequently s^(1) is o(1), not s_opt, and the subsequent application of Proposition D.1 at s = s_opt does not apply. The actual estimator is effectively running with no self-loop, whose exact-recovery threshold is I(aτ,bτ,0) by Lemma C.2, not the advertised information-theoretic threshold I(aτ,bτ,cτ). Thus the abstract's headline claim that a gradient-trained GCN achieves the information threshold is unsupported by the proof as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies exact recovery of all unlabeled nodes in a semi-supervised Contextual Stochastic Block Model (CSBM). It derives an information-theoretic threshold I(aτ,bτ,cτ)=[(√aτ−√bτ)^2+cτ]/2, proves an impossibility theorem below that threshold, gives a lower bound on the mismatch exponent, and constructs a spectral estimator claimed to achieve the threshold. It also analyzes linear ridge regression on a graph-convolutional feature map with self-loops and claims that a one-step-gradient-trained two-layer linear GCN with a learned self-loop weight attains the same information-theoretic threshold. The main advertised novelties are the exact-recovery threshold for semi-supervised CSBM and the GCN achievability claim.","tokens_in":40959,"tokens_out":34262,"duration_ms":290476,"significance":"If correct, the information-theoretic threshold and the matching spectral estimator would provide a useful benchmark for semi-supervised node classification on CSBM, and the ridge-regression analysis would clarify the role of self-loops in graph-convolutional methods. The paper contains a substantial amount of technical apparatus: large-deviation rate functions, leave-one-out arguments, and numerical simulations. However, the GCN achievability theorem, which is a headline contribution of the abstract, rests on several scaling errors in Appendix D; as written, the theorem is not established. The spectral and information-theoretic parts are plausible and could be valuable after repair, but the current manuscript cannot be accepted with the GCN claim in its present form.","major_comments":[{"comment":"Lemma D.5 claims that (2/(n^2 q_m)) y_L^T X_L W^(1)a converges to 2cτ, but this is inconsistent with Lemma D.3. Lemma D.3 gives W^(1)a/√K = ζμ + O(η1/K) with ζ = Θ(1) and K ≍ N, while Assumption 3.1 together with d = o(q_m^2) gives θ^2 ≍ q_m. Hence y_L^T X_L W^(1)a has dominant term θ n (√K) ζ = Θ(√q_m · N · √N) = Θ(N^{3/2}√q_m), and dividing by n^2 q_m yields Θ(1/(√N√q_m)) = o(1), not the constant 2cτ. Replacing W^(1)a by its normalized version W^(1)a/√K gives Θ(θ/(n q_m)) = o(1) as well. Therefore s^(1) in Eq. (D.2) is o(1) under the paper's own scalings, and the subsequent application of Proposition D.1 at s = s_opt in the proof of Theorem 3.9 is unsupported.","section":"Appendix D.3, Lemma D.5 and Eq. (D.2)"},{"comment":"Lemma D.6 asserts that (1/n) y_L^T A_L y_L = ((aτ−bτ)/2) q_m + o(q_m). This is off by a factor of τ. With balanced training labels, E[y_L^T A_L y_L | y_L] = α·(# same ordered pairs) − β·(# different ordered pairs) = (n^2/2)(α−β) + O(n). Since α−β = (a−b) q_m/m and m = (1−τ)N, this expectation is τ^2 N(a−b)q_m/(2(1−τ)) + O(n), so dividing by n = τN gives (τ/2)(aτ−bτ) q_m + o(q_m). Consequently the logarithm in Eq. (D.2) does not converge to log(aτ/bτ) = log(a/b) for fixed τ ∈ (0,1); the argument of the logarithm converges to [(aτ+bτ)+τ^2(aτ−bτ)]/[(aτ+bτ)−τ^2(aτ−bτ)], whose logarithm is not log(a/b). Thus even if the numerator scaling in Lemma D.5 were repaired, s^(1) would not be a consistent estimator of the optimal self-loop weight.","section":"Appendix D.3, Lemma D.6 and Eq. (D.2)"},{"comment":"Lemma C.5 contains a scaling mismatch that affects the ridge-regression proof. The signal part of h(X) = (N q_m)^{-1/2} D_ρ^{-1} A_ρ X is θ/(ar d√(N q_m)) (EA_ρ y) μ^T, which equals R θ/(√(N q_m)) y μ^T with R = (aτ−bτ+2s)/(aτ+bτ+2s) for ρ = s q_m. This requires κ_m = R·θ/√q_m in the definition H = κ_m/√N y μ^T. The displayed κ_m := ((α−β+2ρ)/(α+β+2ρ))·θ√q_m is instead O(q_m) and its prefactor tends to 1 rather than R. With the displayed κ_m, the later identification κ_m τ/(κ_m^2 τ + λ) = ζ(1+o(1)) in the proof of Theorem 3.6 is false for θ^2 ≍ q_m. The theorem statement itself defines κ as an O(1) constant, so this appears to be a fixable typo, but the proof as written does not support the theorem.","section":"Appendix C, Lemma C.5 and proof of Theorem 3.6"}],"minor_comments":[{"comment":"Lemma A.5 and Lemma A.6 contain apparent typographical errors: the ratio in Lemma A.5 should presumably be p_X(X|yi,y−i)/p_X(X|−yi,y−i), and Lemma A.6 writes p_X in the denominator of what should be a graph likelihood ratio.","section":"Appendix A, Lemmas A.5 and A.6"},{"comment":"The displayed formula for s^(1) in Section 3.4 is typeset ambiguously; the quotient structure is clear only after consulting Eq. (D.2) in Appendix D.","section":"Section 3.4, Eq. (3.13) and Eq. (D.2)"},{"comment":"The heading 'Comparation with unsupervised regime' should read 'Comparison with the unsupervised regime'.","section":"Section 3.2.4"},{"comment":"The estimator is denoted byLLR in Theorem 3.6 but byLRR in the surrounding text and in Eq. (3.9); the notation should be unified.","section":"Theorem 3.6"},{"comment":"The simulation plots report the number of independent trials but do not include error bars or standard deviations; given the sharp-threshold claims, some measure of variability would be helpful.","section":"Section 4"}],"recommendation":"reject","confidential_remarks":"The information-theoretic and spectral parts of the paper may be salvageable, and the ridge-regression result is plausible once the κ_m scaling issue in Lemma C.5 is corrected. The GCN section, however, is not a local fix: the proposed self-loop estimator is inconsistent under the paper's own scalings, and the claimed GCN achievability of the information threshold is therefore not established. I would be willing to look at a substantially revised version that either supplies a correctly scaled GCN analysis or removes the GCN achievability claim from the abstract and main contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this before you cite it. The paper claims the first information-theoretic threshold for exact recovery in semi-supervised CSBM, and gives three achievability results. The threshold, I(aτ,bτ,cτ)>1, looks genuinely new and useful, and the spectral estimator is a plausible adaptation of Abbe–Fan–Wang. The ridge regression with oracle self-loop is coherent, and the self-loop optimality formula is a nice practical insight.\n\nThe soft spot is the GCN theorem. Theorem 3.9 leans on Lemma D.5, which claims that (2/(n^2 q_m)) y_L^T X_L W^(1)a concentrates at 2cτ. But the paper's own Lemma D.3 gives W^(1)a = √K ζ μ + o(√K), with √K ≍ √N, and y_L^T X_L μ = Θ(n θ) = Θ(n√q_m). So the numerator is Θ(n^{3/2}√q_m), and the normalized quantity is Θ(1/(√n√q_m)) = o(1), not a constant. Hence s(1) is o(1), so the trained GCN is effectively running with no self-loop. The LDP in Proposition D.1 only reaches the advertised threshold at s = 2cτ/log(a/b); at s=0 the threshold is I(aτ,bτ,0), which is strictly smaller when cτ>0. So the abstract's headline claim that a gradient-trained GCN hits the information threshold is unsupported.\n\nThe spectral and ridge-regression parts may survive a correction, but the GCN result is load-bearing and broken. I'd still send this to referees, because the threshold and spectral method are substantive and the error is concrete and checkable. For a reading group I'd say maybe — it's a useful case study in why you do a scaling check on every lemma that claims a constant limit. I'd cite the threshold result, but not the GCN theorem.","headline":"Genuinely new semi-supervised CSBM threshold and a clean spectral estimator, but the GCN achievability is sunk by a scaling error in Lemma D.5.","tokens_in":41533,"tokens_out":4936,"would_cite":true,"duration_ms":39805,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C80","60F10","62H30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper identifies the exact information-theoretic boundary at which all test nodes in a contextual stochastic block model become jointly recoverable, and shows that a PCA spectral estimator, graph ridge regression with an optimal…","keywords":["semi-supervised learning","contextual stochastic block model","exact recovery","information-theoretic threshold","spectral clustering","graph neural networks","large deviations","graph ridge regression"],"falsifier":"Compute $(2/(n^2 q_m))\\, y_L^\\top X_L W^{(1)}a$ under the paper's scaling $n\\asymp K\\to\\infty$, $\\theta^2\\asymp q_m$, $d=o(q_m^2)$, with $W^{(1)}a$ aligned to $\\zeta\\mu$: if the result is $o(1)$ rather than $\\Theta(1)$, then the estimator $s^{(1)}$ in (D.2) is inconsistent and Theorem 3.9's GCN achievability proof would not hold. Experimentally, run the GCN just above $I=1$ at increasing $m$; if exact-recovery frequency does not tend to 1, the GCN achievability claim is falsified.","tokens_in":40304,"feed_emoji":"🎯","tokens_out":10768,"duration_ms":84642,"temperature":0.7,"pith_summary":"This paper studies semi-supervised node classification on the contextual stochastic block model (CSBM): a two-community random graph in which each node also carries a Gaussian feature vector whose mean depends on its label, with only a fraction $\\tau$ of labels revealed. It claims to pin down, for the first time, the information-theoretic threshold for exact recovery of all unrevealed test labels: recovery is possible if and only if $I(a_\\tau,b_\\tau,c_\\tau)>1$, where $I=[(\\sqrt{a_\\tau}-\\sqrt{b_\\tau})^2+c_\\tau]/2$ and $a_\\tau,b_\\tau,c_\\tau$ are rescaled edge probabilities and feature signal-to-noise parameters. It further constructs a spectral PCA-type estimator that provably achieves exact recovery down to this threshold, and shows that graph ridge regression and a gradient-trained graph convolutional network achieve the same threshold when the graph is augmented with the optimal weighted self-loop. If correct, the result gives a benchmark for when any semi-supervised algorithm can perfectly label all test nodes, and shows that simple spectral methods and GCNs can be information-theoretically optimal in this model.","feed_headline":"Exact test-node recovery pinned to a sharp threshold in CSBM","feed_subtitle":"Below the boundary I=1 no algorithm can label every test node; above it, PCA and graph nets succeed.","key_machinery":"The rate function $I(a_\\tau,b_\\tau,c_\\tau)=[(\\sqrt{a_\\tau}-\\sqrt{b_\\tau})^2+c_\\tau]/2$, with $a_\\tau=a/(1-\\tau)$, $b_\\tau=b/(1-\\tau)$, and $c_\\tau=\\theta^4/[q_m(\\theta^2+(1-\\tau)d/m)]$ in the sparse regime $\\alpha=a q_m/m$, $\\beta=b q_m/m$, is the object that carries the argument: it is the large-deviation exponent governing whether a single flipped test node or a small cluster can mimic the truth, so the condition $I>1$ is the boundary of exact recoverability. The constructive half of the paper rests on two spectral estimators added together: $\\hat{y}_{\\mathrm{SBM}}$, which uses the log-ratio of the two largest eigenvalues of the unlabeled adjacency matrix together with $A_{UL}y_L/\\sqrt{m}$, and $\\hat{y}_{\\mathrm{GMM}}$, which uses the leading eigenvector of the Gram matrix together with $G_{UL}y_L/\\sqrt{m}$. For the graph-convolution methods, the key mechanism is the weighted self-loop $\\rho I_N$ added to the adjacency matrix; choosing $\\rho=2c_\\tau/\\log(a_\\tau/b_\\tau)\\cdot q_m$ makes the graph-convolution operator exactly balance the SBM and GMM information so that the large-deviation rate of the ridge or GCN predictor equals $I(a_\\tau,b_\\tau,c_\\tau)$.","core_discovery":"On the paper's own terms, the central discovery is a sharp phase transition: in the semi-supervised CSBM with training ratio $\\tau$ and test size $m$ growing, all $m$ test-node labels can be recovered exactly, with probability tending to 1, exactly when $I(a_\\tau,b_\\tau,c_\\tau)>1$, with $I$ defined by $I(a_\\tau,b_\\tau,c_\\tau)=[(\\sqrt{a_\\tau}-\\sqrt{b_\\tau})^2+c_\\tau]/2$. Necessity is proved by an MLE large-deviation argument: when $I<1$, the maximum-likelihood configuration differs from the truth on some flipped pair with probability tending to 1, so every algorithm misclassifies at least two test nodes; the paper also gives an exponential lower bound $\\mathbb{E}\\psi_m \\ge e^{-I q_m}$ on the best possible mismatch ratio. Sufficiency is constructive: the spectral estimator $\\hat{y}_{\\mathrm{PCA}}$, formed by adding an SBM-based term using the adjacency matrix and labeled nodes and a GMM-based term using the Gram matrix and labeled nodes, recovers all test nodes with probability at least $1-m^{1-I}$ when $I>1$, and matches the optimal error exponent when $I\\le 1$. The paper then claims that linear ridge regression on a graph convolution, and a two-layer GCN trained by one large gradient step plus a spectrally estimated self-loop weight, cross the same threshold when the self-loop weight is set to $\\rho=2c_\\tau/\\log(a_\\tau/b_\\tau)\\cdot q_m$.","pith_inferences":["A scaling check of the proposed self-loop estimator $s^{(1)}$ suggests that $(2/(n^2q_m))\\,y_L^\\top X_L W^{(1)}a$ may be $o(1)$ rather than a positive constant; if so, Theorem 3.9's consistency claim for $s^{(1)}$ would not hold, while the PCA and ridge-regression results would stand.","The predicted exponent $1-I$ for the failure probability suggests a finite-size test: near the boundary, plot $\\log P(\\text{failure})$ against $\\log m$ and check that the slope approaches $1-I$.","The same MLE-switching argument should transfer to other attributed random graph models, such as XOR-SBM or random geometric graphs with Gaussian features, yielding analogous thresholds with modified $a_\\tau$ and $c_\\tau$.","The optimal self-loop formula $\\rho=2c_\\tau/\\log(a_\\tau/b_\\tau)\\cdot q_m$ gives a concrete design rule: estimate $a,b,\\tau,\\theta,d$ from degrees, Gram spectrum, and labeled features, then set the self-loop weight accordingly."],"forward_implications":["When $I(a_\\tau,b_\\tau,c_\\tau)>1$, the spectral PCA estimator $\\hat{y}_{\\mathrm{PCA}}$ achieves exact recovery with probability at least $1-m^{1-I}$, so the information-theoretic threshold is attained by a polynomial-time spectral algorithm.","When $I(a_\\tau,b_\\tau,c_\\tau)<1$, every algorithm misclassifies at least two test nodes with probability tending to 1, and the best achievable expected mismatch ratio decays like $e^{-I q_m}$.","Graph ridge regression without self-loops recovers all test nodes only under the weaker threshold $I(a_\\tau,b_\\tau,0)>1$; adding the optimal self-loop lifts the threshold to $I(a_\\tau,b_\\tau,c_\\tau)>1$.","For heterophilic graphs with $a<b$, the optimal self-loop strength is negative, so the analysis predicts that downweighting self-loops is required for optimal graph-convolution performance.","A gradient-trained GCN with the estimated self-loop weight reaches exact recovery for $I(a_\\tau,b_\\tau,c_\\tau)>1$; the paper leaves open whether its error rate below the threshold matches the optimal exponent."],"supporting_citations":[{"why":"Introduces the CSBM model and the cavity-method detection threshold that this paper extends to semi-supervised exact recovery.","marker":"[18]"},{"why":"Supplies the unsupervised exact-recovery threshold and the entrywise eigenvector and large-deviation machinery used in Theorems 3.4 and 3.5.","marker":"[4]"},{"why":"Proves the sharp weak-recovery threshold in CSBM, the unsupervised comparison point for the semi-supervised boundary.","marker":"[33]"},{"why":"Introduces the GCN architecture and self-loop convention whose semi-supervised behavior the paper analyzes.","marker":"[30]"},{"why":"Analyzes graph convolution for semi-supervised classification on CSBM and provides the transductive setting and linear convolution tools.","marker":"[9]"},{"why":"Provides the one-step gradient feature-learning analysis used to align W(1) with the feature direction in the GCN proof.","marker":"[8]"},{"why":"Supplies the representation-learning-by-gradient-descent result used in the GCN training step.","marker":"[17]"},{"why":"Provides the concentration and Bernstein inequalities used throughout the spectral and ridge-regression proofs.","marker":"[44]"}],"fun_headline_variants":["Sharp threshold found for exact recovery in semi-supervised learning","Optimal spectral estimator hits information limit in CSBM","PCA and GCNs match exact recovery threshold with tuned self-loops","Exact recovery in CSBM iff I>1, even for GCN","Semi-supervised exact recovery possible above information threshold"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The GCN achievability claim rests on the assumption that the two-stage estimator $s^{(1)}$ from equation (D.2) is consistent for the oracle self-loop weight $2c_\\tau/\\log(a_\\tau/b_\\tau)$; if the scaled quantity $y_L^\\top X_L W^{(1)}a/(n^2 q_m)$ does not converge to a positive constant, that proof step would need to be replaced.","fun_headline_variants_meta":{"raw":{"variants":["Sharp threshold found for exact recovery in semi-supervised learning","Optimal spectral estimator hits information limit in CSBM","PCA and GCNs match exact recovery threshold with tuned self-loops","Exact recovery in CSBM iff I>1, even for GCN","Semi-supervised exact recovery possible above information threshold"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3368,"prompt_tokens":1108,"completion_tokens":2260,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":724,"completion_tokens_details":{"reasoning_tokens":2188}},"tokens_in":724,"tokens_out":2260,"duration_ms":14867,"temperature":1.0,"reasoning_tokens":2188,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:50:58.072447+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $(2/(n^2 q_m))\\, y_L^\\top X_L W^{(1)}a$ under the paper's scaling $n\\asymp K\\to\\infty$, $\\theta^2\\asymp q_m$, $d=o(q_m^2)$, with $W^{(1)}a$ aligned to $\\zeta\\mu$: if the result is $o(1)$ rather than $\\Theta(1)$, then the estimator $s^{(1)}$ in (D.2) is inconsistent and Theorem 3.9's GCN achievability proof would not hold. Experimentally, run the GCN just above $I=1$ at increasing $m$; if exact-recovery frequency does not tend to 1, the GCN achievability claim is falsified.","supporting_citations":[{"cited_title":"Deshpande, S","cited_arxiv_id":null,"evidence_quote":"Introduces the CSBM model and the cavity-method detection threshold that this paper extends to semi-supervised exact recovery."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the unsupervised exact-recovery threshold and the entrywise eigenvector and large-deviation machinery used in Theorems 3.4 and 3.5."},{"cited_title":"Lu and S","cited_arxiv_id":null,"evidence_quote":"Proves the sharp weak-recovery threshold in CSBM, the unsupervised comparison point for the semi-supervised boundary."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the GCN architecture and self-loop convention whose semi-supervised behavior the paper analyzes."},{"cited_title":"Baranwal, K","cited_arxiv_id":null,"evidence_quote":"Analyzes graph convolution for semi-supervised classification on CSBM and provides the transductive setting and linear convolution tools."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the one-step gradient feature-learning analysis used to align W(1) with the feature direction in the GCN proof."},{"cited_title":"Vershynin","cited_arxiv_id":null,"evidence_quote":"Provides the concentration and Bernstein inequalities used throughout the spectral and ridge-regression proofs."}],"review_version":1}