{"id":"be70d9f6-a7c5-4502-8447-ebb355d699ed","arxiv_id":"2502.06014","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For recurrent resetting via finite-time harmonic trap stiffness changes, the authors derive the mean and variance of the work and an optimal finite-time protocol whose quasistatic cost falls below the equilibrium free-energy difference when the initial state is out of equilibrium.","lead":"This paper calculates the work and fluctuations needed to repeatedly erase a Brownian particle's memory by changing an optical trap's stiffness in finite time. It finds that starting from an out-of-equilibrium state can make erasure cost less than the equilibrium Landauer value, and derives an optimal protocol with a closed-form mean cost for all durations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified.","rationale":"The central assertion is Eq. (24): in the long-time limit the optimal mean erasure work lies below the equilibrium free-energy difference ΔF_{U→V} whenever ζ²<1. I re-derived this expression independently. Using the optimal variance trajectory V_t = c_1(1+c_2t)^2 with c_1 = Dt_Uζ², the boundary-term-plus-action work formula of Appendix D gives, as τ→∞, c_2τ = sqrt(t_V/(t_Uζ²))−1, V_τ → Dt_V, and W_2 = ΔF + (k_BT/2)(ln ζ² + 1−ζ²). The same result follows from the physical decomposition: jump to λ_U/ζ² (work (k_BT/2)(1−ζ²)), then quasi-static equilibrium change from variance Dt_Uζ² to Dt_V (cost ΔF + (k_BT/2)ln ζ²). The non-Gaussianity of P_{ta,U} affects only heat dissipation, not the mean work, because the mean work in a harmonic trap depends only on the variance. The full-cycle total work W = W1+W2 remains non-negative: for t_V/t_U ≤ ζ² ≤ 1, W/(k_BT/2) = ln(t_U/t_V)+t_V/t_U−ζ²+ln ζ² ≥ 0, with equality only in the trivial zero-exploration limit. Thus the below-ΔF result is thermodynamically consistent. The factor-2 typo in Eq. (7)/A12 is real but localized: it changes the printed general-protocol mean formula, whereas Eq. (22) and Eq. (24) come from Appendix D and pass the standard ζ²=1 and δ-function limits. The reader's chosen weakest assumption, exact Boltzmann reset after holding in V, is an idealization whose error is exponentially small in the holding time and does not affect the sign or magnitude of the central shortfall. Therefore no load-bearing objection to the central claim is identified, though the manuscript's typos still justify the conditional verdict.","tokens_in":27084,"tokens_out":48253,"duration_ms":494618,"concrete_test":"Simulate the full recurrent cycle with the optimal protocol λ_opt(t) from Eq. (D11) for a large finite duration (e.g., τ=100 t_U), using a Gamma(shape=2) exploration-time distribution with t_U=4, t_V=0.5, D=k_BT=1, and measure the mean W2 over at least 10^6 cycles. Compare to Eq. (22) and to the τ→∞ prediction Eq. (24); agreement within statistical error (e.g., <1% of k_BT) would confirm the central claim and rule out a hidden factor error propagating from Eq. (7).","verdict_should_be":"UNCHANGED","load_bearing_attack":"No load-bearing objection to Eq. (24) was found. An independent check confirms the structure: the long-τ optimal protocol is an instantaneous jump λ_U → λ_U/ζ² (work (k_BT/2)(1−ζ²)) followed by a quasi-static leg from an equilibrium Gaussian of variance Dt_U ζ² to V, costing ΔF + (k_BT/2)ln ζ²; the sum is exactly Eq. (24). The full cycle still obeys W_tot = W1+W2 ≥ 0, so there is no second-law violation. The factor-2 discrepancy in printed Eq. (7)/A12 affects the general-protocol mean-work formula but not the Appendix D optimal-protocol derivation; Eq. (22)/(24) reproduce the known ζ²=1 limit and the f(t)=δ(t) limit. The reader's equilibrium-reset assumption is a standard, exponentially-correct idealization and does not threaten the long-time bound.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the thermodynamic cost of a recurrent resetting protocol in which a Brownian particle in a harmonic trap is switched instantaneously from a stiff trap V to a shallow trap U, explores U for a random time drawn from f(t), and is then returned to V by a finite-time stiffness protocol λ(t). The authors derive a moment generating function for the work per cycle, compute mean and variance for a tanh-type protocol, and derive an optimal protocol that minimizes the mean work for any duration τ. The central result is Eq. (24): in the long-time limit the optimal erasure cost is ΔF_{U→V}^{eq} + (k_B T/2)(ln ζ² + 1 − ζ²), which lies below the equilibrium free-energy difference whenever the recurrently prepared initial state is not equilibrated in U. The authors explain this deficit via an information-geometric projection of the non-Gaussian initial state onto Gaussian trap-compatible states, and connect the finite-time corrections to Wasserstein distances.","tokens_in":27208,"tokens_out":12276,"duration_ms":110142,"significance":"If the main results are correct, the paper makes a useful contribution to finite-time information erasure and stochastic resetting. The explicit moment generating function and the closed-form optimal work for all τ are nontrivial, and the result that recurrent erasure from an out-of-equilibrium state can cost less than the equilibrium free-energy difference is physically interesting and consistent with the generalized Landauer bound of Esposito and Van den Broeck. The paper contains no fitted parameters: Eq. (24) is derived from the work functional, and the generalized Landauer bound is rederived in Appendix C from the non-negativity of entropy production. The interpretation of the shortfall as the KL divergence between the Gaussian projection of the time-averaged initial state and the Boltzmann state in U is plausible and testable. The central long-time formula and the ζ²=1 and δ(t) limits check out; the reported numerical simulations are a further strength, provided they are run with the corrected version of the mean-work formula (see major comments).","major_comments":[{"comment":"The coefficient of the noise integral in the mean work is printed as 4, but the correct coefficient is 2. In Eq. (A5) the authors themselves write ⟨x²⟩ = x0² G(t,0) + 2D ∫_0^t ds G(t,s); substituting this into W2 = (1/2)∫ dt λ̇ ⟨x²⟩ gives W2 = (D/2)∫ dt λ̇ [t_U ζ² G(t,0) + 2∫_0^t ds G(t,s)]. The factor 4 therefore contradicts Eq. (A5) and also contradicts Appendix B, Eq. (B5)/(B9), where the factor 2 is used. Because Eq. (7) is the central general-protocol moment formula and is applied to the tanh protocol, this inconsistency must be fixed and the numerical comparisons in Fig. 2 must be checked against the corrected expression.","section":"Appendix D, Eq. (D9)"},{"comment":"The action term in the expression for W2 is misprinted. From the Euler–Lagrange solution Vt = c1(1+c2t)² with c1 = D t_U ζ², the integral (γ/4)∫ dt (V̇²/V) equals k_B T t_U ζ² c2² τ, not the printed term involving c2 linearly. As printed, the last term is dimensionally inconsistent (it has dimensions of energy times time), and it cannot lead to the quadratic equation whose solution is the c2 given in Eq. (D10). The intended term is k_B T t_U ζ² c2² τ, which corresponds to the factor 2(c2τ)²ζ² t_U/τ appearing inside the brackets of Eq. (22).","section":"Eq. (30) and Eq. (D13)"},{"comment":"The final τ-dependent correction term is printed without a square on the bracket, but consistency with Eq. (22) and with the large-τ expansion in Eq. (31) requires that the bracket be squared. With the printed un-squared bracket, the large-τ coefficient of the correction is t_U(1 − √(t_V/t_U)) rather than the water-square term (1/2)(√(2t_V) − ζ√(2t_U))² of Eq. (31); the two disagree numerically. Since Eq. (30) is advertised as the closed-form optimal work for all protocol durations, this typo should be corrected in both the main text and the appendix.","section":"Eq. (30) and Eq. (D13)"}],"minor_comments":[{"comment":"In the displayed expression for the O(α¹) term, the combination r²t_V²/(r²λ_U²) is dimensionally inconsistent and should read t_V²/t_U² (as in Eq. (B45)).","section":"Appendix B, Eq. (B44)"},{"comment":"The model assumes that after a finite holding time in V the system is exactly at the Boltzmann distribution P_eq,V before the next instantaneous jump. This is a standard idealization, but since all subsequent formulas depend on P_ta,U through Eq. (13), a brief quantitative statement about the exponential smallness of the correction for finite holding times would be helpful.","section":"Sec. II A, Fig. 1"},{"comment":"The notation for the optimal work switches between W_2^opt, W_2^{opt}, and W₂^* in Eqs. (22), (24), and the surrounding text; please standardize it.","section":"Sec. II D 2, Eq. (22)"}],"recommendation":"major_revision","confidential_remarks":"The central physical claim, Eq. (24), appears sound and is not affected by the factor-4 typo in Eq. (7), because the optimal-protocol derivation in Appendix D bypasses that formula. However, the factor-4 error appears in the paper's main displayed mean-work equation and is used for the tanh-protocol analysis; the authors should confirm that the numerical simulations in Fig. 2 were performed with the factor-2 expression and state this explicitly. The typographical errors in Eq. (D9) and in the missing square in Eq. (30) are also easily fixable but are important because Eq. (30) is a headline result. The paper is within scope for a statistical mechanics journal; once these display issues are corrected, a revised version would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is worth a serious look. The core idea—recurrent resetting with a time-averaged, non-equilibrium initial state—is new, and the closed-form optimal work for all durations, Eq. (30), plus the long-time bound Eq. (24) below ΔF, are real contributions. I checked the logic of Eq. (24) independently: an instantaneous jump to a matched variance followed by a quasi-static leg gives exactly that expression. The projection interpretation (only the KL divergence between the projected Gaussian and the equilibrium state is harvestable as work) is clean and plausible.\n\nThe main problem is a factor of 2 in the central formula. Eq. (7) and Eq. (A12) have a 4 multiplying ∫ds G(t,s), but the variance equation for an overdamped harmonic trap gives 2D ∫G, and Appendix B's Eq. (B5), which yields the correct quasi-static limit, uses the 2. So the printed general-protocol mean-work expression is off by a factor of 2 in the noise integral. The figures and the optimal-protocol derivation in Appendix D appear to use the correct version, so this is a typo-class error, not a fatal flaw, but it has to be fixed in both the main text and the appendix. Related to this, the paper claims a 'moment generating function' but only delivers the first two moments for harmonic potentials; the formal MGF in Eq. (5) is never evaluated. That overstatement should be toned down.\n\nOther soft spots: the equilibrium-reset assumption (hold in V for a few relaxation times) is a standard idealization and fine for the long-time bound, though the paper could be clearer that it is an assumption. The small-τ expansion of Eq. (30) could be shown explicitly; Section II.D.4 mentions it but Appendix D only summarizes. These are minor.\n\nWho is this for? People working on stochastic resetting, finite-time Landauer erasure, and optimal control in stochastic thermodynamics. It's a competent paper with a genuinely interesting result, but it needs to be handled carefully by the referee because of the factor-of-two inconsistency. I would send it to peer review, expecting major revisions on the presentation, and I'd cite the corrected result. Bring it to reading group.","headline":"Solid result on finite-time erasure below Landauer cost, but the printed mean-work formula has a factor-2 slip that needs correcting; worth refereeing after the fix.","tokens_in":27753,"tokens_out":4126,"would_cite":true,"duration_ms":35251,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82C31","82C41"],"pacs":["05.40.-a","05.70.Ln"],"model":"deepseek-v4-flash","headline":"Repeatedly erasing a resetting particle's memory can cost less than the equilibrium free-energy difference.","keywords":["stochastic resetting","thermodynamic cost of resetting","information erasure","Landauer principle","optimal finite-time protocols","harmonic traps","work fluctuations","non-equilibrium steady state"],"falsifier":"Measure the per-cycle mean work in an optical-trap resetting experiment with exponentially distributed exploration times and a slow, optimal stiffness protocol in a regime with $\\zeta^2 < 1$, e.g., $t_U = 4$, $t_V = 0.5$ in units of $r^{-1}$. Equation (24) predicts that the long-time work falls below $\\Delta F_{U\\to V}^{\\mathrm{eq}}$ by $(k_B T/2)(\\ln \\zeta^2 + 1 - \\zeta^2)$; observing $W_2 \\ge \\Delta F_{U\\to V}^{\\mathrm{eq}}$ in that limit would falsify the central claim.","tokens_in":26876,"feed_emoji":"⚛️","tokens_out":6533,"duration_ms":59996,"temperature":0.7,"pith_summary":"This paper analyzes the thermodynamic cost of repeatedly “erasing” a Brownian particle by switching a harmonic trap from a stiff potential $V$ to a shallow potential $U$ for a random exploration time, then back to $V$ over a finite duration $\\tau$, so that the particle is restored to the Boltzmann distribution in $V$. The authors derive the moment generating function of the work per erasure cycle and use it to compute the mean and variance for a hyperbolic-tangent protocol and for the protocol that minimizes mean work. Their central result is that in the long-time limit the optimal mean erasure cost equals $\\Delta F_{U\\to V}^{\\mathrm{eq}} + (k_B T/2)(\\ln \\zeta^2 + 1 - \\zeta^2)$, which is strictly below the equilibrium free-energy difference whenever the time-averaged state at the start of the protocol has not equilibrated in $U$ ($\\zeta^2 < 1$). A sympathetic reader would care because this quantifies how much of the information stored in a non-equilibrium state can be converted into negative work, and it extends Landauer-style erasure bounds to the recurrent, steady-state setting used in resetting experiments.","feed_headline":"Recurrent erasure can beat Landauer's cost bound","feed_subtitle":"The saving comes from information lost when the pre-erasure state is squeezed into Gaussian trap states.","key_machinery":"The central object is the per-cycle work moment generating function $C(k,\\tau) = \\langle e^{k w}\\rangle$, assembled by averaging the instantaneous-switch work $U - V$ over the equilibrium distribution in the stiff trap, the exploration-phase propagator, the exploration-time density $f(t)$, and the work moment generating function of the finite-time stiffness protocol. For harmonic potentials, the mean and variance reduce to quadratures involving $G(t,s) = e^{-2\\beta D \\int_s^t \\lambda(t') dt'}$, organized by the dimensionless length $\\zeta^2 = 1 - (1 - t_V/t_U)\\tilde{f}(2/t_U)$, which measures how far the time-averaged initial state $P_{\\mathrm{ta},U}$ is from equilibrium in $U$. The paper minimizes the work functional $W_2[\\lambda(t)]$ by the Euler–Lagrange method on the variance dynamics, giving the variance trajectory $V_t = c_1(1 + c_2 t)^2$, and from it a closed-form optimal work for all $\\tau$. The generalized Landauer bound of Eq. (21) and the Gaussian $m$-projection of Eq. (27) supply the information-geometric interpretation that identifies the work shortfall with accessible information.","core_discovery":"Starting each cycle from the equilibrium distribution in the stiff trap $V$, the particle jumps instantaneously to $U$, evolves for a time drawn from $f(t)$, and then the stiffness is switched back to $V$ by a protocol $\\lambda(t)$ of duration $\\tau$. Because the cycle repeats, the state at the start of the finite-time erasure is the time-averaged, generally non-Gaussian distribution $P_{\\mathrm{ta},U}(x)$, and the paper's aim is to compute the true work of that erasure rather than the ideal Landauer cost. For harmonic traps the paper obtains explicit formulas for the mean and variance of the per-cycle work for any protocol, and for the optimal protocol it obtains a closed-form expression valid for all $\\tau$. The long-time limit gives Eq. (24): $\\lim_{\\tau\\to\\infty} W_2^{\\mathrm{opt}} = \\Delta F_{U\\to V}^{\\mathrm{eq}} + (k_B T/2)(\\ln \\zeta^2 + 1 - \\zeta^2)$. Since $\\ln \\zeta^2 + 1 - \\zeta^2$ is non-positive for $\\zeta^2 \\le 1$, the cost falls below the equilibrium free-energy difference whenever $\\zeta^2 < 1$, i.e., whenever the shallow-trap exploration is too short for full equilibration. The paper interprets the shortfall through the Gaussian $m$-projection: only the KL divergence between the projected state and the $U$-equilibrium, $k_B T\\, D_{\\mathrm{KL}}(\\pi[P_{\\mathrm{ta},U}]\\,\\|\\,P_{\\mathrm{eq},U})$, can be harvested as work; the rest of the non-equilibrium information is inaccessible under harmonic-trap control.","pith_inferences":["A natural extension the paper leaves implicit is whether the “replacement by a Gaussian with matching variance” works for other parametric trap families; one could test this with quartic traps, where the projection argument predicts the accessible-information formula would need modification.","The moment generating function formalism invites optimizing higher cumulants or Pareto fronts of mean work versus work variance for recurrent erasure; the paper notes the possibility but does not carry it out.","If the post-protocol hold in the stiff trap is shortened from “a few relaxation times” to a finite time, residual non-equilibrium correlations will enter the steady-state initial condition, and one would expect extra corrections to Eq. (24) that could be probed experimentally.","The Gamma-distribution analysis suggests a practical design rule: allowing stochastic exploration durations with more sub-mean mass lowers the erasure cost, because such fluctuations push $\\zeta^2$ down toward $t_V/t_U$."],"forward_implications":["If Eq. (24) is correct, a resetting trap operated cyclically from a non-equilibrium initial state can erase information in the quasistatic limit at strictly less than the equilibrium free-energy cost, with the saving set by $\\zeta^2$.","The optimal protocol's mean work is monotonic in duration, so longer protocols always cost less, unlike the hyperbolic-tangent protocol, which can be non-monotonic in some parameter regions.","The generalized Landauer bound (21) is valid but not tight; Eq. (29) locates the gap precisely as the information lost when the true initial state is projected onto the Gaussian family compatible with harmonic traps.","In the vanishing-exploration-duration limit, the two instantaneous-switch contributions cancel to leading order, so the total cycle work can vanish, meaning erasure can be nearly free in that limit.","At large but finite $\\tau$, the excess cost takes the form $(k_B T/2\\tau)(\\sqrt{2t_V} - \\zeta\\sqrt{2t_U})^2$, a squared $L^2$-Wasserstein distance between the Boltzmann state in $V$ and the Gaussian projection of the initial state."],"supporting_citations":[{"why":"provides the experimental resetting-trap setup whose finite-time stiffness switching motivates the protocol and model.","marker":"[16]"},{"why":"supplies the Euler–Lagrange optimal-protocol method for harmonic traps used to derive $W_2^{\\mathrm{opt}}$.","marker":"[41]"},{"why":"supplies the work distribution results for the driven harmonic oscillator with equilibrium initial conditions that the paper extends to out-of-equilibrium initial states.","marker":"[59]"},{"why":"gives the slow-driving expansion technique adapted here to compute the long-time correction for the tanh protocol.","marker":"[65]"},{"why":"provides a closed-form work moment generating function for a driven harmonic oscillator with arbitrary initial condition, the basis for the non-equilibrium MGF construction.","marker":"[66]"},{"why":"reports experimental erasure from a non-equilibrium state with cost below the Landauer bound, the phenomenon the paper analyzes in a recurrent steady state.","marker":"[70]"},{"why":"derives the generalized Landauer work bound for transformations between non-equilibrium states, quoted as Eq. (21).","marker":"[71]"}],"fun_headline_variants":["Finite-time erasure beats Landauer bound via Gaussian projection","Recurrent erasure costs less than Landauer's limit","Erasure cost dips below Landauer via Gaussian projection","Non-equilibrium erasure exploits Gaussian KL divergence","Harmonic-trap resetting undercuts Landauer bound"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cycle model assumes that after each resetting protocol the particle is held in the stiff potential $V$ long enough to reach exactly the Boltzmann distribution before the next instantaneous jump to $U$; if the hold is finite or the relaxation incomplete, the steady-state initial distribution and all the work formulas change.","fun_headline_variants_meta":{"raw":{"variants":["Finite-time erasure beats Landauer bound via Gaussian projection","Recurrent erasure costs less than Landauer's limit","Erasure cost dips below Landauer via Gaussian projection","Non-equilibrium erasure exploits Gaussian KL divergence","Harmonic-trap resetting undercuts Landauer bound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000431,"raw_usage":{"total_tokens":2262,"prompt_tokens":1066,"completion_tokens":1196,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":682,"completion_tokens_details":{"reasoning_tokens":1116}},"tokens_in":682,"tokens_out":1196,"duration_ms":9720,"temperature":1.0,"reasoning_tokens":1116,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:02:57.989116+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the per-cycle mean work in an optical-trap resetting experiment with exponentially distributed exploration times and a slow, optimal stiffness protocol in a regime with $\\zeta^2 < 1$, e.g., $t_U = 4$, $t_V = 0.5$ in units of $r^{-1}$. Equation (24) predicts that the long-time work falls below $\\Delta F_{U\\to V}^{\\mathrm{eq}}$ by $(k_B T/2)(\\ln \\zeta^2 + 1 - \\zeta^2)$; observing $W_2 \\ge \\Delta F_{U\\to V}^{\\mathrm{eq}}$ in that limit would falsify the central claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Euler–Lagrange optimal-protocol method for harmonic traps used to derive $W_2^{\\mathrm{opt}}$."},{"cited_title":"Zhong and M","cited_arxiv_id":null,"evidence_quote":"supplies the work distribution results for the driven harmonic oscillator with equilibrium initial conditions that the paper extends to out-of-equilibrium initial states."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides a closed-form work moment generating function for a driven harmonic oscillator with arbitrary initial condition, the basis for the non-equilibrium MGF construction."},{"cited_title":"Rosales-Cabara, G","cited_arxiv_id":null,"evidence_quote":"reports experimental erasure from a non-equilibrium state with cost below the Landauer bound, the phenomenon the paper analyzes in a recurrent steady state."},{"cited_title":"Speck and U","cited_arxiv_id":null,"evidence_quote":"derives the generalized Landauer work bound for transformations between non-equilibrium states, quoted as Eq. (21)."}],"review_version":1}