{"id":"b16459b8-ae9d-4e77-b70a-8c9ba42b3c0c","arxiv_id":"1908.04779","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Random pulse computing circuits that preserve output entropy reduce cascaded calculation errors by one to two orders of magnitude, and the new RFF-based divider and subtractor designs achieve this.","lead":"The authors present five improved circuits for random pulse computing, a stochastic-computing style where numbers are represented by pulse probabilities. They introduce an entropy-budget analysis and show that preserving output randomness sharply reduces errors in cascaded calculations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The EBC-based impossibility and the precision claim both rest on requiring RPTs to be i.i.d. Bernoulli; the paper's own DIV3/SUB3 show deterministic circuits can compute the arithmetic correctly, so the causal conclusion is conditional on that definitional choice.","rationale":"The reader's weakest_assumption identifies exactly the dependence on the 'crucial assumption' that RPTs must be mutually independent Bernoulli variates, and notes that the paper's own DIV3 and SUB3 demonstrate that lower-entropy deterministic circuits can perform the arithmetic functions. My review finds the same load-bearing issue and adds specificity: the EBC is used to prove impossibility of deterministic division and half-sum, yet the paper immediately presents deterministic circuits that realize those operations with high precision. The resolution is that the EBC bound only forbids deterministic circuits that also produce full-entropy Bernoulli outputs; it does not forbid deterministic circuits that produce correlated pulse trains with the correct average rate. Section 5's central claim is similarly conditional on the downstream circuit's sensitivity to correlations: the squaring circuit of Fig. 23 fails for correlated inputs because its transfer function assumes independence, not because low entropy is intrinsically erroneous. The paper's conclusion that 'a non-maximal output entropy at the output of an RPC circuit causes error in subsequent calculations' is therefore too broad as stated; it should be scoped to circuits that assume Bernoulli inputs. This does not undermine the empirical contribution or the design guidance to prefer high-entropy outputs, but it requires a correction in the theoretical framing. Since the reader's CONDITIONAL verdict already flags this assumption as the weakest point, my assessment does not change the verdict.","tokens_in":16420,"tokens_out":5760,"duration_ms":60102,"concrete_test":"Re-derive the impossibility claim of Section 3.3 by replacing h(p_z) with the actual per-bit entropy H(x_z) of the output sequence (Eq. 3) on the left side of Eq. (8), using the measured output statistics of DIV3 (e.g., the p0=0.3, p1=0.5 case of Fig. 14a). If the inequality is then satisfied for DIV3, the EBC-based conclusion that deterministic division is impossible is an artifact of the i.i.d.-Bernoulli output requirement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 1 states the 'crucial assumption' that RPC circuits are fed by mutually independent random Binomial variates, and Section 2.2.1 defines entropy H(x) of an RPT via Eq. (3). The EBC argument (Eq. (8)) then treats the required output entropy as H(p_z)=h(p_z), i.e., assumes the output must be a maximal-entropy Bernoulli process with probability p_z. On this basis, Section 3.3 concludes 'a deterministic circuit for division is not possible' and Section 3.2 concludes the half-sum cannot be done deterministically. But the paper then presents DIV3 (Fig. 12) and SUB3 (Fig. 19), which are deterministic, have no internal entropy source, and are shown to compute p0/p1 and p1-p0 with better precision than the random circuits (Figs. 13, 20). The text reconciles this only by noting that these outputs have low relative entropy (Figs. 15, 21), i.e., that they are not Bernoulli processes. Thus 'cannot work correctly' in the EBC discussion really means 'cannot produce a Bernoulli-distributed RPT,' not 'cannot compute the arithmetic function.' The central precision claim in Section 5 is then also conditional: the squaring circuit in Fig. 23 computes y = z_i * z_{i-1}, whose expected value equals p_z^2 only when the input is memoryless Bernoulli. A low-entropy but rate-correct representation (e.g., DIV3's bursty output) breaks this particular downstream circuit, but the error is a property of the squaring circuit's independence assumption, not of entropy per se. The paper's own Fig. 14 and the discussion of DIV3 as ideal when division is the last circuit in a chain concede that such representations are acceptable in some contexts. The weakest load-bearing step is therefore the unstated identification of 'valid RPT' with 'i.i.d.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents five new or improved circuits for random pulse computing (RPC), of which DIV2 and DIV3 are division circuits, SUB2 and SUB3 are subtraction circuits, and a magnitude comparator is also introduced. The work uses random flip-flops (RFFs) driven by quantum-photon-detection randomness as entropy sources, and introduces an 'entropy budget criterion' (EBC), Eq. (8), as a necessary condition for a circuit to output a Bernoulli-distributed random pulse train (RPT) with a given probability. The paper claims, on the basis of EBC, that deterministic circuits for the half-sum, division, and subtraction are impossible, while also presenting deterministic circuits DIV3 and SUB3 that compute these functions with low output entropy. Experimental measurements with 8e7 clock intervals per point and matching C simulations compare the circuits' precision and output relative entropy, and a cascade with a squaring circuit (Fig. 23) is used to study how output entropy affects downstream precision.","tokens_in":16696,"tokens_out":5406,"duration_ms":50313,"significance":"The experimental work is extensive and careful: each transfer-function point uses 8e7 clock intervals, the stated statistical error is 0.00011, and C simulations agree with measurements. The entropy-budget inequality itself is a correct application of Shannon entropy bounds and is properly attributed to Cover and Thomas. If the central claims held, the paper would make a useful contribution by showing that output entropy, not just average pulse rate, matters for cascaded stochastic circuits. However, the main theoretical conclusion is conditional on the definitional choice that a valid RPT must be an i.i.d. Bernoulli process; the paper's own deterministic DIV3 and SUB3, which compute the arithmetic functions correctly but with low relative entropy, directly contradict the impossibility statements as written. The Section 5 error measurements also contain an apparent inconsistency that undermines the general claim that higher output entropy reduces downstream error.","major_comments":[{"comment":"The statement 'a deterministic circuit for division is not possible' is contradicted by the circuit DIV3 presented in the same section. DIV3 is deterministic, has no internal entropy source, and its transfer function and errors (Fig. 13) show that it computes the clipped division with better precision than DIV1 and DIV2. The text attempts to reconcile this by saying DIV3 'cannot work correctly' and then noting that its output has low relative entropy, but the circuit clearly does work correctly in the sense of computing the intended arithmetic function. The EBC in Eq. (8) is a criterion for whether a circuit can produce a Bernoulli-distributed output with the required probability, not for whether it can compute the function. The impossibility conclusion should be rephrased as 'a deterministic circuit cannot produce a maximum-entropy (Bernoulli) output for division,' and the paper should explicitly acknowledge that rate-correct deterministic representations exist.","section":"Section 3.3, Eq. (8), Fig. 12-15"},{"comment":"The claim that a deterministic half-sum circuit is impossible relies on the assertion that the operation is symmetric and therefore 'the only way to arrive to the half-sum symmetrically, is to select pulses from either input with equal probability,' and that this selection must be random, costing entropy of 1. This is an assertion, not a proof. Counter-based deterministic circuits such as DIV3 or SUB3 achieve symmetric operations without random selection, and a similar counter-based or stateful circuit could plausibly output a pulse train with average rate (p0+p1)/2. The argument needs to state precisely what class of circuits it excludes (e.g., memoryless Boolean gates) and prove the impossibility for that class, rather than relying on an unexplained 'only way' claim.","section":"Section 3.2, half-sum argument"},{"comment":"The text states that for the second type of error, |p_y - p_z^2|, 'circuits DIV1 and SUB1 yield 1-2 orders smaller error than the other two circuits, in any region.' If this is correct, it directly contradicts the paper's thesis that higher output entropy (DIV2 and SUB2 have higher relative entropy per Figs. 15 and 21) reduces the error made by the downstream squaring circuit. If the statement is a typo and the opposite is true, it must be corrected, because the current wording makes the Section 5 conclusion internally inconsistent. As written, the reader cannot tell whether the data support or undermine the central claim that non-maximal output entropy causes error in subsequent calculations.","section":"Section 5, Figs. 24c,d"}],"minor_comments":[{"comment":"The relative entropy H_rel(x) = H(x)/H1(x) is undefined when H1(x) = 0, i.e., when p = 0 or p = 1. The authors should state the limiting convention used for these boundary cases.","section":"Eq. (5)"},{"comment":"There are several typographical inconsistencies in terminology: 'RTP' is used in Section 2.2.1 where 'RPT' is meant, and 'PRC' appears in Section 2.1 where 'RPC' is intended. These should be corrected throughout.","section":"Section 2.1"},{"comment":"The approximate comparator transfer function in Eq. (17) is described as heuristic, but the paper does not report how the coefficient 9/8 was obtained, how many free parameters were fitted, or what the goodness of fit was. A brief statement of the fitting procedure and residuals would strengthen this part.","section":"Section 4, Eq. (17)"},{"comment":"The algorithm for generating a time-discrete RPT from the single-photon detector is described twice with slightly different wording ('if and only if one or more pulses have been received during the previous time segment' vs. 'during the previous time segments'). The duplication is confusing and should be merged or clarified.","section":"Section 2.1"},{"comment":"Reference [24] is missing page numbers and the journal volume; only the DOI is given. The authors should complete the bibliographic details.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper has solid experimental work and a correct entropy inequality, but the theoretical framing overreaches: the EBC-based impossibility results are conditional on a Bernoulli-output requirement that the paper itself violates with DIV3 and SUB3. The Section 5 error discussion appears internally inconsistent as written. These issues are fixable by carefully rephrasing the scope of the claims and correcting the Section 5 description, so I recommend major revision rather than rejection. I would also suggest that the authors engage with the broader stochastic-computing literature that already discusses correlation and burstiness (e.g., work on correlated stochastic bit streams) to position their contribution more accurately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real content here is the circuit work, and it is good. DIV2, DIV3, SUB2, SUB3, and the comparator are new, clearly described, and carefully measured. The long-run statistics (8e7 intervals, quoted error 0.00011) and the matching C simulations give the experimental claims real weight. The demonstration that the RFF-based DIV2/SUB2 lose less output entropy than their LFSR-based counterparts, and that this translates into one to two orders of magnitude smaller error when followed by the squaring circuit, is convincing and useful. Anyone building stochastic-computing chains should care about this. The entropy-budget analysis is where I part ways with the framing. Equation (8) is just the independence bound from Cover and Thomas, correctly cited and correctly applied. Fine. But the paper uses it to conclude that deterministic division and half-sum circuits are impossible. That conclusion only holds if a valid RPT must be an i.i.d. Bernoulli process with maximal entropy for its pulse probability. The paper states this as the 'crucial assumption' in Section 1, but then DIV3 and SUB3, which are deterministic and low-entropy, compute their arithmetic functions with better precision than the random circuits. The text reconciles this by saying those outputs have low relative entropy, i.e., they are not Bernoulli processes. So 'cannot work correctly' really means 'cannot produce a Bernoulli-distributed RPT,' not 'cannot compute the function.' The half-sum argument in Section 3.2 is even more heuristic: the claim that symmetric combination requires random equiprobable selection is asserted, not proven. The same conflation weakens the central precision claim. The squaring circuit computes y = z_i * z_{i-1}, whose expectation is p_z^2 only when the input is memoryless Bernoulli. A bursty but rate-correct representation breaks this specific downstream circuit. The error is a property of the squaring circuit's independence assumption, not of entropy per se. The paper's own admission that DIV3 is ideal when division is the last circuit in a chain concedes the point. Still, the measured facts stand. The circuits work, the entropy measurements are real, and the practical advice — prefer high-output-entropy circuits in a cascade — is reasonable even if the impossibility framing is overbroad. The paper deserves a serious referee, but the authors should be pushed to reframe the EBC as necessary but not sufficient, and to separate 'low entropy breaks downstream independence assumptions' from 'low entropy is inherently erroneous.' The lack of released code/data is a minor issue; the description is detailed enough for reimplementation. I would cite this for the circuits and measurements.","headline":"Solid experimental paper on random-pulse-computing circuits whose entropy-budget criterion is a textbook bound and whose impossibility claims are conditional on a definitional choice the paper itself later relaxes.","tokens_in":768,"tokens_out":900,"would_cite":true,"duration_ms":19390,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Entropy-preserving circuits cut random-pulse computing errors by 10 to 100 times.","keywords":["entropy budget criterion","random pulse computing","random flip-flop","relative entropy","stochastic computing","binomial random pulse train","division circuit","subtraction circuit"],"falsifier":"Build a downstream multiplication or squaring circuit and feed it pulse trains with identical pulse rate, one a true independent random stream and one deliberately patterned, such as an exactly alternating 0-1 sequence at probability 0.5. If the patterned train produces no larger error than the random train, then non-maximal output entropy is not itself the error source and the central claim fails; if the error grows as entropy drops while the rate stays fixed, the claim is supported.","tokens_in":16157,"feed_emoji":"🎲","tokens_out":5755,"duration_ms":59362,"temperature":0.7,"pith_summary":"This paper argues that a random pulse computer needs more than individually precise arithmetic circuits: it needs circuits whose output pulse trains stay genuinely random. The authors introduce an entropy budget criterion, stating that the entropy of a circuit's output cannot exceed the entropy supplied by its inputs plus any entropy generated inside the circuit. Because conventional deterministic division and subtraction circuits violate this budget over large parts of their input range, the paper concludes they cannot work correctly without an internal source of true randomness. It builds new versions around a quantum-random flip-flop and measures that, when followed by a squaring circuit, these entropy-preserving versions produce one to two orders of magnitude smaller total error than the older pseudorandom-register versions. The reader should care because this identifies output randomness itself as part of numerical precision, not a side effect.","feed_headline":"Entropy-preserving circuits cut random-pulse computing errors 10-100x","feed_subtitle":"An entropy budget criterion shows why deterministic dividers and subtractors corrupt downstream calculations, and quantum-random designs…","key_machinery":"The load-bearing object is the entropy budget criterion (EBC), applied to time-discrete random pulse trains in which each time slot independently contains a pulse with probability $p$, so the maximum entropy of the train is $H_1(p) = -p\\log_2 p - (1-p)\\log_2(1-p)$. The paper defines relative entropy $H_{\\mathrm{rel}} = H(\\boldsymbol{x})/H_1(\\boldsymbol{x})$ to quantify how close a circuit's output is to a true independent binomial stream. The second central component is the random flip-flop (RFF), a T-flip-flop whose clock acts with probability $1/2$ because it is driven by quantum-random photon detections; it supplies exactly the entropy that deterministic dividers and subtractors lack. The EBC identifies which operations demand internal randomness, the RFF supplies that randomness, and the squaring-circuit test exposes the downstream numerical cost when it is missing.","core_discovery":"The central claim is that the precision of a cascaded random-pulse computation is governed by the per-bit Shannon entropy of each intermediate pulse train. A circuit that outputs the correct average pulse rate but with low relative entropy, for example long bursts of pulses separated by long silences, hands its downstream neighbor signals that are no longer independent binomial variates, and calculations that assume independence accumulate error. The paper establishes an entropy budget criterion, $H(\\boldsymbol{x}_z) \\le H(\\boldsymbol{x}_{\\mathrm{circ}}) + \\sum_i H(\\boldsymbol{x}_i)$ for any circuit, and uses it to show that deterministic circuits for half-sum and division are impossible over significant regions of input space, while subtraction is not forbidden by the budget but remains unrealized deterministically. The measured demonstration is the cascade test: the RFF-based circuits DIV2 and SUB2 maintain high output relative entropy, and when followed by a squaring circuit their total error is one to two orders of magnitude smaller than that of the LFSR-based DIV1 and SUB1, whereas the cheap deterministic DIV3 and SUB3 are the most precise but only suitable as final stages in a calculation chain.","pith_inferences":["The entropy-budget screening should transfer directly to other stochastic-computing operations: any proposed gate whose output entropy exceeds its input entropy at some point of its domain will require an internal true-random source, so the criterion can serve as a fast pre-hardware filter.","A testable extension would sweep input relative entropy at fixed pulse probability into a fixed downstream circuit and measure error versus $H_{\\mathrm{rel}}$; the paper's mechanism predicts a monotone error increase, but no measurement in the paper isolates that curve.","The burstiness of DIV3 and SUB3 outputs makes their estimates converge faster than high-entropy exponential-distribution trains, so a hybrid architecture could deliberately use low-entropy fast circuits for early approximate answers and switch to entropy-preserving circuits for refinement; the paper notes the speed advantage but does not propose this hybrid.","Because the paper identifies counters, comparators, and spontaneous random neural firing as biologically available primitives, its entropy-budget view suggests that biological neurons may face the same precision constraint and may need intrinsic firing noise to compute reliably."],"forward_implications":["Circuit designers should screen stochastic-computing gates by output relative entropy, not only by transfer-function error, since low-entropy outputs become a source of error in every downstream stage.","The entropy budget criterion makes a structural prediction: half-sum and division cannot be built deterministically over their full input range, so a universal random-pulse computer needs internal random sources inside these units.","Interior nodes of a calculation network should favor the RFF-based DIV2 and SUB2, while the precise but low-entropy DIV3 and SUB3 are best reserved for final outputs or for restricted input regions such as divisors at or below 0.1.","The proposed comparator circuit supplies the missing flow-control primitive for programmable random-pulse computing, and its switching sharpness grows exponentially with counter bit-length, so decision logic can be made arbitrarily sharp at the cost of counter size."],"supporting_citations":[{"why":"Supplies the independence entropy bound theorem from which the entropy budget criterion is derived.","marker":"[14]"},{"why":"Defines the quantum random flip-flop used as the internal entropy source in DIV2 and SUB2.","marker":"[10]"},{"why":"Provides the earlier division, subtraction, and squaring circuits used as baselines and as the downstream test cascade.","marker":"[9]"},{"why":"Establishes the stochastic-computing circuit framework and the approximate division approach that DIV1 and SUB1 extend.","marker":"[6]"},{"why":"Documents the LFSR-based circuit setbacks including output correlation, seeding requirements, and resource cost that motivate replacing pseudorandom sources with RFFs.","marker":"[24]"},{"why":"Foundational random-pulse machine formulation defining random pulse trains and their pulse-rate representation.","marker":"[5]"}],"fun_headline_variants":["Entropy budget reveals why random-pulse dividers fail","Quantum-random circuits cut cascaded pulse error 100x","Shannon entropy governs random-pulse circuit precision","RFF-based blocks outperform LFSR in pulse cascades","Entropy criterion validates new random-pulse circuits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that a valid number in this computer must be a maximally random independent pulse train at the right average rate, so a lower-entropy train with the same rate does not count as the same number; if low-entropy representations are accepted, the entropy-budget impossibility results no longer apply.","fun_headline_variants_meta":{"raw":{"variants":["Entropy budget reveals why random-pulse dividers fail","Quantum-random circuits cut cascaded pulse error 100x","Shannon entropy governs random-pulse circuit precision","RFF-based blocks outperform LFSR in pulse cascades","Entropy criterion validates new random-pulse circuits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000538,"raw_usage":{"total_tokens":2551,"prompt_tokens":885,"completion_tokens":1666,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":1588}},"tokens_in":501,"tokens_out":1666,"duration_ms":12368,"temperature":1.0,"reasoning_tokens":1588,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:33:52.384939+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a downstream multiplication or squaring circuit and feed it pulse trains with identical pulse rate, one a true independent random stream and one deliberately patterned, such as an exactly alternating 0-1 sequence at probability 0.5. If the patterned train produces no larger error than the random train, then non-maximal output entropy is not itself the error source and the central claim fails; if the error grows as entropy drops while the rate stays fixed, the claim is supported.","supporting_citations":[{"cited_title":"Computing Polynomials Using Unipolar Stochastic Logic","cited_arxiv_id":null,"evidence_quote":"Provides the earlier division, subtraction, and squaring circuits used as baselines and as the downstream test cascade."},{"cited_title":"Energy - Efficient Stochastic Computing with Superparamagnetic Tunnel Junctions","cited_arxiv_id":null,"evidence_quote":"Documents the LFSR-based circuit setbacks including output correlation, seeding requirements, and resource cost that motivate replacing pseudorandom sources with RFFs."},{"cited_title":"Random-Pulse Machines","cited_arxiv_id":null,"evidence_quote":"Foundational random-pulse machine formulation defining random pulse trains and their pulse-rate representation."}],"review_version":1}