{"id":"f657b280-e0e2-405b-a303-db2e78cf9da5","arxiv_id":"2505.17033","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Tensor-network approximations price binomial Asian and multi-asset American basket options with linear-in-size cost in tested regimes, beating Monte Carlo for high volatility and small time steps.","lead":"This preprint applies tensor network methods from quantum physics to price Asian and American basket options in binomial tree models. It reports large speedups over Monte Carlo in high-volatility settings, but contains an apparent error in the variational method's central formula.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Variational Asian lower bound rests on an exact MPS representation (Eq. 21) that fails a two-step contraction check; the displayed tensors mis-weight the strike term.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the claimed exact MPS representation of p(x)\\tilde v_A(x) in Eqs. (21)-(27). A hand contraction for N=2 shows the displayed tensors do not reproduce p(x)\\tilde v_A(x); the strike term is underweighted by one factor of path probability. This is not a stylistic or consensus-based objection; it is an internal correctness issue in the derivation of the variational lower-bound method. The numerical experiments for TTcross and the American basket method are plausible and may survive revision, but the headline claim that the variational method provides a rigorous lower bound depends directly on Eq. (21). Because the reader's CONDITIONAL verdict already accounts for this issue, no further adjustment is needed. The recommended test is a direct two-site contraction and an elementwise comparison, which would settle whether the representation is merely misprinted or whether the method needs a different exact construction.","tokens_in":18968,"tokens_out":3933,"duration_ms":39037,"concrete_test":"For N=2, compute the contraction B_1^{x_1} B_2^{x_2} using Eqs. (22)-(27) for all four strings x=(x_1,x_2), and compare elementwise with p(x)\\tilde v_A(x) from Eqs. (13) and (18) under the stated CRR/RB parameters. If any entry differs, Eq. (21) is false. Then rerun the variational lower-bound experiment with a corrected exact MPS and check whether the computed prices still bound the brute-force binomial price from below; if no correction is supplied, the lower-bound claim must be removed or weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript's central lower-bound claim for the variational Asian method rests on Eq. (21): p(x)\\tilde v_A(x) is claimed to be exactly representable as the MPS with tensors (22)-(27). A direct contraction for N=2 contradicts this. For x=(0,0), contracting the row vector B_1^{0} with the column vector B_2^{0} gives (1-p_u)\\big(S_0/N (1-p_u)(d+d^2)-K\\big), whereas p(x)\\tilde v_A(x)=(1-p_u)^2\\big(S_0/N(d+d^2)-K\\big) from Eqs. (13) and (18). The strike term is weighted by only the last N-1 path probabilities, not the full path probability. The same mismatch occurs for other binary strings. Thus Eq. (21) is not exact as written. Since the variational algorithm evaluates the cost K in Eq. (29) by contracting the filter MPS with this purported exact MPS, and the lower-bound statement in Eq. (30) plus the finite-bond-dimension argument relies on computing that cost correctly, the rigorous lower-bound property is unsupported. This is fixable if corrected tensors exist or if the exactness claim is downgraded to an approximation with controlled error, but as presented it is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper combines binomial-tree option pricing with matrix-product-state/tensor-train techniques. For arithmetic Asian options it proposes (i) a TTCross approximation of p(x)v_A(x) followed by a linear contraction, and (ii) a variational MPS method that maximizes an indicator-weighted sum over paths and claims to provide a rigorous lower bound. For multi-asset American basket options it combines decoupled trees with TTCross to approximate the early-exercise value at each backward step. Numerical experiments compare against Monte Carlo (up to 10^11 samples) and brute-force binomial prices for N up to 50 and baskets up to m=8, reporting faster convergence in low-N/high-volatility regimes. The central performance claims are linear scaling in N and m at fixed bond dimension and the 'stringent lower bound' property of the variational Asian method.","tokens_in":19219,"tokens_out":14317,"duration_ms":131325,"significance":"The TTCross results are a useful, if incremental, demonstration that low-rank tensor formats can reduce binomial exotic-pricing cost, and the numerical study is reasonably careful: prices are benchmarked against independent brute-force or large-sample Monte Carlo estimates, and each experiment is repeated multiple times. The paper is also honest about regime-dependent performance. If the variational lower-bound claim were established, it would be a distinctive selling point over Monte Carlo. As written, however, the claim rests on an unproven exact MPS representation that fails a two-step contraction check, so the main theoretical contribution is not yet supported. The paper would be publishable in revised form once the representation is corrected or the claim is appropriately weakened.","major_comments":[{"comment":"The claimed exact MPS representation of p(x)tilde_v_A(x) is not exact as written. For N=2 and x=(0,0), Eqs. (13) and (18) give p(x)tilde_v_A(x) = (1-p_u)^2 [ S_0/2 (d+d^2) - K ]. Contracting the boundary tensors in Eqs. (24) and (26), with no interior tensors, gives B_1^0 B_2^0 = [S_0/2 d (1-p_u), S_0/2 d (1-p_u) - K] dot [d (1-p_u), 1-p_u]^T = S_0/2 d (1-p_u)^2 + S_0/2 d^2 (1-p_u)^2 - K(1-p_u). The strike term carries only one factor of (1-p_u) rather than the full path probability (1-p_u)^2. The same mis-weighting occurs for other binary strings. Because the cost K in Eq. (29) is evaluated by contracting the filter MPS with this tensor, the lower-bound statement in Eq. (30) and the finite-bond-dimension argument are unsupported. The authors should either provide a corrected exact MPS construction, with boundary tensors and a proof, or downgrade the claim to an approximation with controlled error.","section":"Sec. IV.A.2, Eqs. (21)-(27)"},{"comment":"The lower-bound interpretation also requires that every MPS produced by the greedy sweep is a feasible binary tensor satisfying Eq. (28). Appendix A shows that products of matrices whose rows (for L) or columns (for R) have at most one nonzero entry are binary, but the greedy row-merge and row-drop procedure in Appendix B is not proved to preserve that structural invariant at every iteration, and the 'logical OR' merge of filter entries is not defined in terms of the tensor elements constrained by Eqs. (33)-(34). Without a feasibility invariant, an optimized psi could take non-binary values and K could exceed the true option price, which would void the lower-bound claim even if Eq. (21) were corrected. Please add a formal feasibility proof for the greedy update, or restrict the claim to the feasible subset of the ansatz.","section":"Sec. IV.A.2, Eqs. (31)-(37) and Appendix B"}],"minor_comments":[{"comment":"Equation (24) contains an unmatched square bracket in the second entry of the row vector; it should read [ S_0/N d(1-p_u), S_0/N d(1-p_u) - K ].","section":"Eq. (24)"},{"comment":"The symbol A is overloaded: it denotes both the site tensor in Eq. (31) and the derivative matrix A = partial K / partial A in Eq. (B2). Please use distinct symbols.","section":"Appendix B, Eq. (B2)"},{"comment":"The notation on line 9 is garbled: 'M[N],YN 1 M[N],Y1 N 2 ...' should be written consistently with Eq. (45), e.g. M_1^{[N],Y_N^1} M_2^{[N],Y_N^2} ... M_m^{[N],Y_N^m}.","section":"Algorithm 1, line 9"},{"comment":"The sentence 'Fig. 3 shows a higher convergence speed of the option price with the number of samples' lacks a comparison; it should say which method is being compared with which. Also, the y-axis is labeled 'Pricing Error' but the text says the standard deviation is used as a proxy for the error; the caption should state this explicitly.","section":"Fig. 3 and surrounding text"},{"comment":"The equation label (D5) is used twice, once for the outcome labeling and once for the transition-matrix definition; please renumber.","section":"Appendix D"},{"comment":"No code or data repository is provided, so the wall-clock comparisons in Figs. 3, 5, and 6 are not independently reproducible from the manuscript alone.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central technical issue is the exact MPS representation in Eqs. (21)-(27). A corrected construction is available, for example a bond-dimension-3 MPS tracking [W_i, A_i, P_i] with transition matrices [[p_i,0,0],[q_i,q_i,0],[0,0,p_i]] and boundary vectors [0,1,1] and [S_0/N,0,-K]^T, so the paper can be repaired in scope. If the authors cannot supply a corrected exact tensor or a controlled-error statement, I would expect the 'rigorous lower bound' language in the abstract and conclusions to be removed, and the TTCross material to carry the paper. I see no ethical concerns beyond the technical gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe TTCross treatment of binomial Asian options and the decoupled-trees American basket method look genuinely useful, and the numerical results are credible. But the advertised \"rigorous lower bound\" for the variational Asian method is not backed by the manuscript as written. I checked Eq. (21) against the definitions and the N=2 case fails: the displayed B tensors weight the strike K by only the last N-1 step probabilities, not the full path probability. For x=(0,0), the contraction gives (1-pu)(S0/N (d+d^2) - K), whereas p(x)tilde_v_A(x) = (1-pu)^2 (S0/N(d+d^2)-K). So the exactness claim is wrong.\n\nThe good news: this looks like a typo in the boundary tensors, not a dead end. Replacing the -K in B1^0 and B1^1 by -K(1-pu) and -K pu respectively would fix the weight, and the rest of the construction (the upper-triangular matrices with entries q r and q) is a standard way to accumulate partial sums. If that correction is right, the variational sweeps and the lower-bound argument may go through. But the authors need to supply the corrected tensors or an explicit derivation; as it stands, a reader cannot reproduce the method.\n\nOther soft spots: the linear-scaling claim is only shown for fixed bond dimension; we never learn how the bond dimension must grow with N or m to maintain accuracy, and the tested regimes are modest (N up to 50, m up to 8). For N=50 at sigma=0.5, Monte Carlo beats both tensor methods in their own plot. The paper is honest about the regime dependence, but the abstract's sweeping \"scale linearly with the parameters of interest\" is stronger than the evidence. There is also a small mismatch: the American basket payoff in the text is max-basket (Eq. 44), while Fig. 7's caption calls it a min-basket put.\n\nThe TTCross Asian section is the strongest part; the variational section needs repair. This deserves a serious referee, but the referee should be told to focus on the corrected tensors and a re-run of the lower-bound experiments. I would not cite the paper as it stands, but I would watch for the revision.\n\nRecommendation: send it to peer review with a request for major revision, focused on fixing the variational tensors and adding bond-dimension growth analysis.","headline":"TTCross parts are believable; the variational lower bound rests on an incorrect MPS tensor formula that is fixable but must be repaired before the claim is accepted.","tokens_in":19740,"tokens_out":13053,"would_cite":false,"duration_ms":107058,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91G20","91G60","15A69"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that binomial pricing of Asian and multi-asset American basket options can be compressed with tensor networks so the cost scales linearly in the number of time steps or assets instead of exponentially.","keywords":["option pricing","exotic options","binomial tree","tensor networks","matrix product states","tensor train cross approximation","Asian options","American basket options"],"falsifier":"Contract the tensors in Eqs. (21)–(27) for $N=2$ and compare the result with the exact product of path probability and (average price minus strike) for each of the four binary price paths; any mismatch disproves the exact-representation premise, and then the variational price could be compared with brute force on small $N$ to see whether it ever exceeds the exact binomial price.","tokens_in":18717,"feed_emoji":"📈","tokens_out":9061,"duration_ms":82266,"temperature":0.7,"pith_summary":"This paper tries to establish that two classes of exotic options that overload standard binomial trees—Asian options, whose payoff depends on the entire price path, and American basket options, whose payoff depends on $m$ correlated assets—can be priced by compressing the relevant payoff functions into matrix product states (MPS), reducing the cost from exponential to linear in $N$ or $m$. The central move is to approximate the product of path probability and payoff as an MPS, so the option price becomes a single tensor contraction. The paper presents two Asian methods (a tensor-train cross approximation and a variational MPS method argued to give a rigorous lower bound) and one multi-asset method that combines decoupled binomial trees with tensor-train cross. The contribution is a practical route around the curse of dimensionality for these exotics, with the largest numerical gains at high volatility.","feed_headline":"Tensor networks shrink exotic option pricing from exponential to linear","feed_subtitle":"Asian options and 8-asset American baskets run in linear time, beating Monte Carlo at high volatility.","key_machinery":"The machinery has four pieces. First, the matrix product state (MPS) representation stores an $N$-variable tensor as a chain of local tensors with bond dimension $D$, so memory and contraction cost scale linearly in $N$. Second, the tensor-train cross (TTCross) interpolation builds an MPS from a small number of evaluations of the target function. Third, the variational Asian method uses a binary MPS ansatz whose left and right tensors are constrained so that contracting them yields binary unit vectors, leaving only the central tensor free and keeping the entire state binary. Fourth, the decoupled-trees transform converts $m$ correlated assets into $m$ independent binomial trees via a Cholesky decomposition, which factorizes the pricing recursion over assets. Together these convert exponential path or asset sums into linear-time contractions.","core_discovery":"On the paper's own terms, the central discovery is that the exponential sums defining binomial prices for Asian and multi-asset American basket options admit low-rank tensor approximations that can be built and contracted in time linear in the problem size. For Asian options, the product $p(x)v_A(x)$ is approximated by an MPS via tensor-train cross, and the price follows from contracting that MPS with vectors of ones. The variational method instead maximizes a binary-filtered cost function over a binary MPS ansatz, and the paper argues that the optimum equals the exact binomial price while providing a lower bound. For American baskets, the terminal payoff is learned as an MPS, after which backward induction with early exercise is interleaved with repeated tensor-train cross recompressions so the value stays in MPS format at every step.","pith_inferences":["The constrained binary-MPS parametrization is a general device for restricting a tensor network to binary 0/1 values, so it could be transferred to discrete optimization problems with binary variables, such as portfolio selection or constraint satisfaction; the paper only demonstrates it on Asian payoffs.","The TTCross-plus-backward-induction loop is not specific to basket payoffs; it should extend to Bermudan swaptions or barrier options, with the open question being how the required bond dimension grows with the frequency of early-exercise decisions.","A direct check of the exact MPS representation for $N=2$ would settle whether the variational lower bound is rigorous as stated; if the check fails, the method might still converge from below numerically, but the proof would need to be amended."],"forward_implications":["For Asian options with smaller numbers of steps, the TTCross method reaches Monte Carlo accuracy at roughly 50–100 times less walltime in the reported tests.","At high volatility ($\\sigma=2$) and $N=50$, both TTCross and variational MPS reduce the pricing error by more than an order of magnitude compared with Monte Carlo.","The variational Asian method is constructed to return a price at or below the exact binomial value, a guaranteed one-sided error that Monte Carlo estimates do not provide.","American basket options on $m=4$ and $m=8$ correlated assets converge rapidly and monotonically with bond dimension on an $N=40$ tree, with the $m=4$ case validated against brute force.","The speed advantage is regime-dependent: at low volatility and larger $N$, Monte Carlo remains competitive, while tensor networks win at high volatility and smaller $N$."],"supporting_citations":[{"why":"supplies the standard binomial up/down scheme and parameters used to build the Asian option trees.","marker":"[3]"},{"why":"supplies the alternative binomial scheme used for the Asian examples and for the independent trees in the basket method.","marker":"[4]"},{"why":"provides the parallel cross-interpolation implementation used for the TTCross approximations.","marker":"[46]"},{"why":"defines tensor-train cross approximation, the algorithmic basis for compressing payoff functions into MPS form.","marker":"[47]"},{"why":"is the tensor-network package the MPS and TTCross implementation is built on.","marker":"[70]"},{"why":"introduces the decoupled-trees technique used for binomial multi-asset pricing that the basket method adapts.","marker":"[71]"},{"why":"supplies the decoupling transformation and its analysis for binomial pricing of multi-asset options.","marker":"[72]"},{"why":"supplies the tensor-by-tensor sweeping update on which the variational Asian optimizer is modelled.","marker":"[35]"}],"fun_headline_variants":["Tensor networks shrink exotic option pricing to linear","Exotic options priced in linear time via tensor trains","Asian and basket options: exponential cost becomes linear","MPS accelerates binomial exotic pricing to linear scaling","Tensor cross tames exotic option pricing explosion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The variational lower-bound claim rests on the assertion that the product of path probability and the average-minus-strike function has exactly the tensor form displayed in Eqs. (21)–(27); if that representation is not exact for every price path, the proof that the variational result is a rigorous lower bound does not go through.","fun_headline_variants_meta":{"raw":{"variants":["Tensor networks shrink exotic option pricing to linear","Exotic options priced in linear time via tensor trains","Asian and basket options: exponential cost becomes linear","MPS accelerates binomial exotic pricing to linear scaling","Tensor cross tames exotic option pricing explosion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000287,"raw_usage":{"total_tokens":1671,"prompt_tokens":914,"completion_tokens":757,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":687}},"tokens_in":530,"tokens_out":757,"duration_ms":7481,"temperature":1.0,"reasoning_tokens":687,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:24:52.108089+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Contract the tensors in Eqs. (21)–(27) for $N=2$ and compare the result with the exact product of path probability and (average price minus strike) for each of the four binary price paths; any mismatch disproves the exact-representation premise, and then the variational price could be compared with brute force on small $N$ to see whether it ever exceeds the exact binomial price.","supporting_citations":[{"cited_title":"Additionally, this form also ensures that the correlation matrix is positive definite as required by the decoupling trees approach","cited_arxiv_id":null,"evidence_quote":"supplies the standard binomial up/down scheme and parameters used to build the Asian option trees."},{"cited_title":"At any discrete time step k of the multi- dimensional tree, the original random variables Stk are given by Stk = exp(GYtk)","cited_arxiv_id":null,"evidence_quote":"supplies the alternative binomial scheme used for the Asian examples and for the independent trees in the basket method."},{"cited_title":"Bauernfeind, M","cited_arxiv_id":null,"evidence_quote":"provides the parallel cross-interpolation implementation used for the TTCross approximations."},{"cited_title":"Menczer, K","cited_arxiv_id":null,"evidence_quote":"defines tensor-train cross approximation, the algorithmic basis for compressing payoff functions into MPS form."},{"cited_title":"Jasra and P","cited_arxiv_id":null,"evidence_quote":"is the tensor-network package the MPS and TTCross implementation is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the decoupled-trees technique used for binomial multi-asset pricing that the basket method adapts."},{"cited_title":"Bouchard and X","cited_arxiv_id":null,"evidence_quote":"supplies the decoupling transformation and its analysis for binomial pricing of multi-asset options."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the tensor-by-tensor sweeping update on which the variational Asian optimizer is modelled."}],"review_version":1}