WorksMathematics
Identification and curvature bias for shared inputs in multichannel dynamical models
Abstract
Pooling many channels with a shared unknown input can leave persistent bias when each channel retains a fixed number of observations and its own incidental parameters. For heterogeneous smooth channel manifolds under isotropic noise, the second-order bias is an information-weighted regression of mean curvature on normal input velocity. A feasible correction gives centred Gaussian inference as the noise level decreases with , compared with without correction. We extend the stochastic expansion and studentization to deterministic channel arrays with densely correlated Gaussian errors of bounded covariance operator norm. General marginal covariance adds a forcing term that can bias even flat channels, while cross-channel covariance changes uncertainty. For Gaussian errors with a separable frame-by-channel covariance whose channel factor is autoregressive in a known ordering, an independent zero-mean calibration field yields a covariance estimator accurate enough for corrected inference. Boundary accumulation and fixed working population models introduce distinct bias terms; rational-multiplier and zero-count arguments connect identification to frame-integrated compartment models. Simulations test both oracle and estimated-covariance inference, including independent PET replications. Moderate-noise undercoverage, optimization ambiguity and the need to validate reconstruction covariance limit practical inference. The real-data example supports sensitivity analysis, not clinically calibrated intervals.
Introduction
A collection of time series may be generated by different dynamical systems driven by the same unobserved input. In dynamic positron emission tomography (PET), the tracer concentration in each region or voxel is the response of a compartment system with its own rate constants to the arterial blood input. Total-body scanners now record such curves in – voxels [36], and the input is usually taken from an image of the aorta; recent likelihood-free and generative methods for voxelwise kinetic posteriors assume it is known [13, 43] (Gu et al. 2026). Voxelwise likelihoods that are nonlinear in a decay parameter also arise in positronium lifetime imaging with time-of-flight PET [25]. Estimating the input from the tissue curves themselves, by fitting all curves jointly, is the idea behind simultaneous estimation (SIME) and multichannel blind identification [16, 41, 38]. It is natural to expect that more channels give a better input. This paper asks what pooling channels can recover. The answer has three parts: identification, the input scale, and estimation.
Identification requires more than pooling. An alternative input can move poles and zeros between its transform and the rational channel transfer functions. A binding input pole budget and suitable separation of channel poles and zeros rule out these alternatives up to scale. For frame-integrated exponential-polynomial models, including the delayed Feng input, a zero-count argument transfers continuous-curve identification and local rank to frame means under explicit endpoint, onset-chart and channel-type conditions (§ 3). Structural identification does not guarantee good conditioning. An exact scale symmetry also remains: kinetic functionals that move with that symmetry retain external calibration error, however many channels are pooled.
The third part, and the principal statistical contribution, concerns estimation. Each channel brings its own kinetic parameters, so the number of parameters grows in proportion to the number of channels while the number of time frames stays fixed: the channel parameters are incidental in the sense of Neyman and Scott [33]. We show that the joint least-squares input estimator converges, as , to a pseudo-true input that differs from the truth by
where is the common noise level, is the sensitivity of a channel curve to the input, projects onto the normal space of that channel’s manifold of noise-free curves, and is the trace of the second fundamental form of the manifold, times its mean curvature vector. The deviation is an information-weighted regression of curvature on the normal input velocity. It is invariant under reparameterization of the channels, involves intrinsic but not parameter-effects curvature ([7]), and, under isotropic noise in the fitting metric, vanishes for flat channels such as unconstrained spectral basis models ([11]). Because the sampling standard deviation of the estimator is of order , its bias dominates as soon as is large: once , Wald intervals for every linear functional whose leading bias is nonzero cover with probability tending to zero. An analytic correction of the curvature term removes the bias to first order and leaves the variance of the inverse fixed-effects information whenever . A PET simulation with compact channel types and decreasing along an explicit sequence in examines these predictions (§ 7). Physical constraints in the joint fit change the order: channels on a boundary such as are projected onto a cone, and the bias becomes of order whenever their aggregate boundary forcing is nonzero.
A different approach is a population model, in which the channel parameters are random and a hierarchical posterior is computed for the input. With an unrestricted population and support identification this gives consistency (Proposition S8.1). With a working parametric population we show that the integrated pseudo-likelihood has its own second-order bias
with the velocity of the channel foot points; under Bernstein–von Mises conditions the same holds for the hierarchical posterior mean. The bias vanishes for the correct population, but it involves the population score in the directions of the weakly identified channels. In our design the second-order coefficient is 1.7–7.1 times the fixed-effects coefficient for Gaussian-mixture populations, although at realistic noise levels the realized bias is damped well below this first-order value. The analytic correction, in contrast, needs no population model.
Scope and contributions
The contribution lies in the combination of three features that template-estimation and incidental-parameter results do not cover jointly: the channels are heterogeneous manifolds that move with the shared input, the number of frames per channel is fixed so that the bias is an expansion in the noise level rather than in a per-unit sample size, and physical constraints make channels accumulate on boundary faces. Sections 3–4.2 connect frame-level identification to an invariant curvature-bias expansion and a feasible correction at fixed frame count, and Section 4.4 distinguishes boundary atoms from continuous accumulation near a face. Section 5 contrasts the population-score bias of integrated likelihood under a fixed working density. A covariance-aware extension (§ 4.3) permits dense Gaussian dependence, separates marginal bias from cross-channel uncertainty and gives an explicit sufficient covariance-calibration experiment. External calibration (§ 6) gives an additional limit. The numerical work tests the small-noise predictions in two prespecified studies of a compact-population design, the second with the computed estimator (domain, input chart and regular-set correction) fixed in advance; it separates bias from variance and from optimization error, records where the correction falls short at moderate noise, and illustrates the failure of two ad hoc averaging approximations in this heterogeneous design. The real-data example illustrates sensitivity rather than calibrated inference.
Related work
Table 1 summarizes the closest results. Inconsistency of maximum likelihood with incidental parameters goes back to Neyman and Scott [33]; Kiefer and Wolfowitz [28] showed that treating them as draws from an unknown distribution restores consistency, and Lancaster [30] reviews the econometric literature. Analytic and jackknife corrections for nonlinear panels [19, 5] and modified profile likelihoods [39] remove biases of order when each unit has observations; here is fixed and the expansion is in the noise level. Their large- guarantees do not directly cover this fixed- experiment. Splitting the frames can also remove the endpoint information required by our identification conditions; we do not establish that every possible jackknife construction must fail. Small-noise biases of orthogonal-distance estimators are known in several special settings: the essential bias of geometric circle fits [2], hyperaccurate corrections in geometric computer vision [27], nonlinear orthogonal regression [14] and principal curves [20]. Closest to Theorem 4.2 is the template-estimation literature: Miolane, Holmes and Pennec [31] show that the Fréchet mean in a quotient by an isometric group action has asymptotic bias , with the trace of the second fundamental form (unnormalized mean curvature vector) of the template orbit, and correct it by an iterative bootstrap; Devilliers et al. [12] and Bigot and Charlier [8] study inconsistency in related settings. There, all observations lie on one orbit of a group that acts by isometries, so every observation carries the same geometry. Our channels are different manifolds that move with the shared parameter, individual channels may be weakly informative, and the bias is the -weighted regression of their curvatures on their normal velocities. The scalar example below shows that the order of pooling and inversion matters even when every single-channel information matrix has condition number one. The PET comparisons in § 7 are illustrative approximations, not implementations of the bootstrap methods in the template literature. General-covariance geometric bias is also established [26, 37]. Our extension derives a uniform correction and studentization theorem for heterogeneous fixed-frame arrays under Gaussian dependence; it does not claim the first covariance-dependent bias calculation. Boundary accumulation, population-score expansion and frame-level identification provide further results. These distinctions specify the extension beyond that literature rather than a claim that curvature-induced bias itself is new. Kumon and Amari [29] study the differential geometry of estimating functions with many nuisance parameters in exponential families. Curvature in nonlinear regression and its effect on a single fit are classical [9, 7, 10]; Huang and He [23] review the intrinsic and parameter-effects curvature measures and their effect on Wald inference. For regularized low-rank matrix models, Huang et al. [24] use a curvature bound for the fixed-rank manifold, proportional to the inverse smallest nonzero singular value, to control estimation error. Here the curvature enters the leading incidental-parameter bias. The algebraic arguments relate to multichannel blind identification [32, 1] (Xu et al. 1995). For multi-region PET, Holler, Morina and Schramm [21] prove that noiseless point measurements from several regions identify the irreversible two-tissue parameters and the input up to a common scale, with a consistency result for Tikhonov-regularized estimation as the noise vanishes with a fixed number of regions, and Holler, Morina and Schramm (2025) extend this to reversible kinetics. Section 3 overlaps with these results; it differs in using a rational-multiplier argument that covers a delayed input with a repeated pole and a direct vascular term, in working with frame integrals, and in treating local rank. The statistical results, which concern channels with fixed frames, are not addressed there. Boundary parameters are well studied for a fixed number of parameters ([40, 3]); their accumulation in an incidental-parameter problem appears not to have been described. The empirical study of Ogden et al. [34] found a systematic bias of SIME for $1FDG and suggested that more regions with distinct kinetics might help; Theorem 4.2 identifies a mechanism, present even under correct specification, that more channels at fixed noise cannot remove, and § 7 shows that misspecification can produce larger shifts. The powers R\sigma^2R\sigma^6$ follow by comparing bias orders with the standard error; they are not asserted to be minimax phase boundaries.
| Work | Setting and result | This paper |
| Incidental-parameter panels ([19, 5, 39]) | observations per unit; bias , analytic and jackknife corrections | fixed; expansion in the noise level ; regime |
| Geometric fitting ([2, 26, 27]) | points on one curve; essential bias for circles | heterogeneous moving manifolds; covariance-aware correction with dense Gaussian dependence (Theorem 4.7) |
| Template estimation ([31]) | isometric group action, all observations on one orbit; bias ; bootstrap correction | non-isometric channel manifolds, individual channels may be weakly informative; bias is an information-weighted regression; analytic one-step correction with ; order- bias with nonzero boundary forcing; population-score expansion |
| Multi-region PET identification ([21, 22]) | noiseless point samples identify two-tissue kinetics and the input up to scale; Tikhonov consistency as noise vanishes with fixed regions | rational-multiplier proof with delay, repeated input pole and vascular term; frame integrals and local rank; inconsistency of joint estimation as at fixed noise |
| SIME ([16, 41, 34]) | joint estimation of input and kinetics; systematic bias observed empirically | a correctly specified bias mechanism that more channels cannot remove; misspecification shifts compared |
| Random effects ([28]) | unrestricted mixing distribution gives consistency | frame-level support condition (Proposition S8.1); second-order bias of parametric populations (Theorem 5.1) |
Table 1. Closest results and what this paper adds.
Multichannel experiments with a shared input
For channel , consider the causal linear compartment system
with a matrix whose eigenvalues have nonpositive real parts, transfer function and input . For frame boundaries , the observed frame means are
where are known whitening weights, is a channel noise scale and has identity covariance. Channel types are independent and identically distributed. The errors are independent and identically distributed across channels and independent of all channel types. § 4.3 replaces this assumption by a specified Gaussian covariance model, allowing dependence between channels. We consider with fixed and, in § 4.2, jointly with .
Two-tissue channels with a Feng input
Irreversible trapping permits a zero eigenvalue; asymptotic stability is therefore not required. Laplace transforms are taken in a right half-plane where they exist. For a reversible two-tissue system with fractional blood volume ,
here the same input drives tissue exchange and the vascular term. The irreversible model used for [F]fluorodeoxyglucose ([F]FDG) has , a pole at , and net influx rate .
Plasma and whole blood
Tissue exchange is driven by the arterial plasma concentration , whereas the vascular term contains whole blood , and an aortic image-derived input measures . If with constant , writing the model with plasma input gives vascular coefficient and tissue coefficient . To keep the normalized form (3) one must also set , provided ; merely replacing is incorrect. Alternatively, using whole blood as the input retains and replaces by . The rates are unchanged, whereas depends on this convention and on calibration. If , the plasma representation remains valid in free macro coefficients but not with a physical fractional blood volume. A time-varying ratio makes every channel a two-input system, outside the classes of § 3; we treat it, together with reversible kinetics, channel-specific delays and dispersion, as a stress test in § 7, where the analysis model is deliberately wrong. The Feng input [15] is
, , with and transform
For every the transformation
leaves every output unchanged. Thus the input scale is not identified, the fraction and all rate constants are scale free, and moves by the factor . We remove the scale by a gauge, a slice of input parameters transversal to the orbits of (4); for the Feng input we fix and write . For statistical estimation we use positive amplitudes and an ordered chart, , with all poles distinct; otherwise swapping and leaves the input unchanged. A sufficient physically positive subfamily has . Smooth local theory additionally restricts the delay to an interval strictly between two frame boundaries. Frame integration does not make the delayed onset six times differentiable across a boundary. These are restrictions of the statistical parameter space, not consequences of the algebraic theorem.
Identification
The admissible multiplier
Proposition 3.1 (Observational equivalence). Suppose for every channel, where all nonzero are rational and are nonzero transforms. Then there is a nonzero rational function with
Conversely, every for which and remain in the specified input and channel classes produces identical outputs. The transfer functions are identified up to scale exactly when every admissible multiplier is constant.
Supplement, Example S1.1, gives four positive-system counterexamples illustrating the role of the input pole budget and common channel poles and zeros.
Identification by pole separation
For a set , write for the inputs whose transform is with , monic of degree at most with all roots in , and of degree less than ; and for the rational channels with monic of degree at most and . Representations are in lowest terms.
Theorem 3.2 (Identification by pole separation). Let and nonzero , , satisfy
(i) and for every ;
(ii) for at least one channel ;
(iii) the polynomials have no common root in ;
(iv) the polynomials have no common root;
(v) for every .
If and satisfy for every , then has delay , and there is a constant with and .
With this is the usual pole-separation condition. Conditions (i)–(iv) cannot be dropped: Supplement Example S1.1(b) violates only the attainment of the input budget, (c) only (iii) and (d) only (iv); the supplement gives constructions for the channel half of (i) and for (ii). With , if the channel has a real pole and a distinct real root , the multiplier is admissible, so a single channel does not identify the input shape in these unrestricted rational classes. This construction need not preserve the narrower positive Feng family. The proof (Supplement S1) compares the root multiplicities of output denominators and numerators.
Corollary 3.3 (Reversible two-tissue channels). Let . Consider Feng inputs with and distinct and channels (3) with . Outside a Lebesgue-null set of parameters, any Feng input and channels of the form (3) that reproduce all outputs have the same delay and are related by (4).
Corollary 3.4 (Irreversible two-tissue channels). The conclusion of Corollary 3.3 holds for channels (3) with and , among alternative channels of the form (3) with ; in particular, no Feng input with reversible channels reproduces the outputs.
For irreversible channels all denominators share the root , so condition (iii) fails with ; it holds with , which contains every Feng pole. The exceptional sets are zero sets of products of resultants, nonzero at certified rational points (Supplement, Table S1). The next example shows that restricting is essential.
Example 3.5 (Trapping versus a non-decaying input). Let be any input and the channels irreversible with . For , the input and the reversible channels with
produce the same outputs; with an analogous map exists for small (Supplement S1.5). Here has a pole at .
Within the Feng class the exact equivalence is excluded. A slowly decaying Feng component can nevertheless resemble a persistent component over a finite scan, suggesting weak numerical separation between slow input dynamics and channel reversibility; no uniform approximation of the full transformed input by the Feng family is claimed. For the study analysed in Supplement S30, min exceeds the scan length.
Finite frames
Theorem 3.2 and its corollaries compare complete curves, whereas the data are frame means. For two-tissue channels and the Feng input both inputs and outputs are exponential polynomials, and a classical zero count transfers identification from curves to frames. A real exponential polynomial is with distinct real and real polynomials . Its order is the order of the smallest constant-coefficient differential equation it solves. A nonzero exponential polynomial of order has at most real zeros, counted with multiplicity [35 Part V, Problem 75]. Say that a model has exponential orders if every input vanishes before its onset and is for , with of order at most , real exponents and , and every channel has impulse response with of order at most and real exponents. Then the channel curve is , which we continue to all real as an exponential polynomial. The Feng input has . Two-tissue channels, reversible or irreversible, have real poles and , and so do the macro channels of § 7.
Theorem 3.6 (Finite frames). Consider a model with exponential orders , delays in , and frame endpoints that include at least points in .
(a) If two parameter values give equal frame means in channel , their curves agree on and for all real .
(b) Suppose further that one true curve is positive on and that the continuations have no common zero in . Then equal frame means in all channels imply and equal complete curves. In particular, the conclusions of Theorem 3.2 and Corollaries 3.3–3.4 hold for frame means, against alternatives in the same model of exponential orders.
(c) On an open real-analytic parameter chart with locally bounded real exponents and onset strictly inside a fixed frame, for any tangent direction , the derivative of the frame means along vanishes if and only if the derivative of every complete curve vanishes at all times other than its onset.
The proof (Supplement S4) applies the zero count to differences of cumulative curves, which frame means determine. Part (b) excludes an earlier alternative onset whose curves continue the true curves backwards. When alternative curves are nonnegative, it suffices, in place of the common-zero condition, that one true channel has and , because that curve then rises with positive slope at its onset. The bound is not claimed to be sharp. Part (c) reduces local rank for frames to an infinitesimal version of Theorem 3.2. In the next proposition and denote first-order variations of the transforms, and the delay variation enters through .
Proposition 3.7 (Infinitesimal identification). Assume conditions (i)–(v) of Theorem 3.2, and consider variations tangent to the classes: , , and . If for every , then , , , and , for a constant .*
Two degree conditions carry the delay step. The input numerator must keep its degree, and the channel of condition (ii) must attain its numerator degree. For the Feng input, exactly when , which holds for positive amplitudes when . For two-tissue and macro channels, the numerator degree is attained exactly when . If either condition fails, a delay change can be absorbed by the amplitudes.
Corollary 3.8 (Frame-level conditions for two-tissue channels). Let the input be Feng with fixed, positive amplitudes, , and delays in . Let the channels be reversible or irreversible two-tissue channels with and positive rates, or macro channels with and . Suppose at least 13 frame endpoints lie in , and the condition of Theorem 3.6(b) holds. For every set of channels containing two channels outside the exceptional null sets of Corollaries 3.3–3.4, defined for macro channels by the same products of resultants in macro coordinates:
frame means determine the delay, the input shape and every channel among Feng inputs in the chart and channels with at most two real poles (in the rational classes of Theorem 3.2), in particular the unconstrained macro extension;
the stacked frame Jacobian in has full column rank. Hence every channel Jacobian has full column rank, and the information matrix of § 4 is positive definite.
Consequently, for a channel-type law on a compact set of such types that charges every neighbourhood of two generic types, and whose support contains a type with continued curve negative on and a type with curve positive on , parts (c) and (d) of Assumption 4.1 in § 4 hold: each channel is identified given the input, , and vanishes only at .
The 35 frames of the study of Supplement S30 place 13 endpoints after every onset up to 10 min and 31 at or after the end of the frame containing the reference onset, so the delay chart is that frame, of width min. For condition (b) an analytic certificate replaces pointwise evaluation. If the input continuation satisfies on , then and is decreasing towards the onset there, and for an irreversible two-tissue channel with the continuation obeys
where is the continued tissue impulse response, on , and the convolution is continued by the oriented integral, so that the tissue part is nonnegative and at most . Hence on whenever , and one such channel type excludes a common zero. For positive Feng amplitudes with , for and for follow directly from the exponential terms (Supplement S4.4). The certificate holds for four of the eight organ parameter sets of the XCAT phantom used in § 7, for instance the lungs, where (Supplement S4.4). It concerns the simulation population, not the voxels. Under these conditions the identification and rank parts of Assumption 4.1 follow for the actual frames; the smoothness part still requires the delay to stay inside one frame.
Corollary 3.8 is a statement of structural identification: no two parameter values give the same frame means. It does not bound the conditioning, which governs estimation. On the compact set of channel types of the corollary, is a continuous positive function of the type, and a continuous positive function of the delay over a closed subinterval of the chart, so both are bounded below uniformly; but the bounds can be small. For the eight XCAT organ means and delays across the interior of the chart, ranges over and the condition number of over , while (Supplement S4.4). Weak numerical identification, a nearly flat criterion in a direction that is identified in principle, is therefore expected along the slow input components, and it is distinct from structural nonidentification, for which the criterion would be exactly flat.
Information for the input shape and scale
Consider the exact whitened tangent experiment , , with independent standard Gaussian noises, unrestricted channel nuisance , and common input perturbation . Let be the orthogonal projection onto and .
Proposition 3.9 (Profile information). The efficient information for is . If for every and , then is unidentifiable for every .
The scale symmetry (4) differentiates into a vector with for all .
Proposition 3.10 (Shape information grows, scale information does not). Suppose for every , for a unit vector , and is invertible. For the efficient least-squares estimator of Proposition 3.9, and
If are independent copies of with and is positive definite on , then for , with the Moore–Penrose inverse, and .
Its restriction to the gauge slice reappears below as the curvature of the profiled criterion.
Incidental channels and the curvature bias
The joint estimator and its pseudo-true value
The joint, or fixed-effects, least-squares estimator solves
It is the estimator of SIME and of joint blind identification. Profiling the channels, with
The channel manifold is the set of noise-free curves that channel can produce with input . Write , for a generic channel and for the expected criterion.
At the true curve define (), (), , and
In the coordinates , the second fundamental form of at is , and is its trace with respect to the induced metric : times the mean curvature vector. It depends on the manifold only. The tangential part measures parameter-effects curvature [7]. The columns of are the normal velocities of the channel manifold when the input moves. Put and .
ASSUMPTION 4.1. (a) , , and and are compact; the channel types have compact support, the true channel parameters lie in a compact subset of the interior of , and is interior to .
(b) , where is six times continuously differentiable in on a neighbourhood of , with derivatives bounded uniformly in and continuous in .
(c) (Channels given the input.) , and for every , , uniformly over channel types.
(d) (Input.) vanishes on only at , and is positive definite.
(e) (Noise.) is symmetric about zero, and for some .
Condition (d) is global identification in the finite-frame noise-free experiment: means that almost every true channel curve can be reproduced with input . Condition (c) asks that each channel be identified once the input is known. The continuous-curve theorem alone does not give either condition for frames. For Feng inputs in the ordered chart and two-tissue or macro channels, Corollary 3.8 derives both from the algebra under its stated conditions on the frames, the delay chart and the type law; the endpoint count alone does not suffice. For other models, (c) and (d) remain assumptions. The macro parameterization extends the physical channel manifold across and , so true channels with are interior. § 4.4 shows that imposing the physical boundaries in the joint fit changes the order of the bias. The data-generating law of the validation design in § 7 admits these conditions on suitable compact parameter sets; the computed fits are assessed separately from the exact constrained minimizers; the reference design, with unbounded log-normal channel draws, does not satisfy them literally, and the experiments report how often fits reach the solver box.
Theorem 4.2 (Curvature bias). Under Assumption 4.1, for every the function is continuous on and, almost surely, every limit point of lies in . There are and such that for :
(i) every satisfies
(ii) uniformly over a neighbourhood of ,
where is the point of nearest to the noise-free curve and is the curvature vector (10) of .
The criterion (12) explains the mechanism. Without noise, every input is penalized by the squared distance of the true curves from its channel manifolds, which is quadratic with Hessian at . Noise adds , the part of the noise normal to the manifold, minus a correction that is linear in the displacement and in the curvature: a manifold that bends towards the data captures more of the noise. The minimizer therefore moves in the direction in which the true curves come to lie on the concave side of their manifolds, by an amount set by the regression of the curvature on the normal input velocities. Symmetry of the noise removes the odd orders; with skewed noise the remainder in (11) is , with a leading term given by the third moments (Supplement, Remark S3.4).
Proposition 4.3 (Geometry of the bias). Under Assumption 4.1:
(a) (Invariance.) If for a diffeomorphism , the profiled criterion , hence and , are unchanged, and so are , and .
(b) (Flat channels.) Let be compact, with and of full column rank uniformly on , with continuous and bounded, and let have mean zero and identity covariance; assume also and ; parts (a), (c) and (e) of Assumption 4.1 are not needed. Then for every ; if vanishes only at , almost surely.
(c) (Minimal channels.) If almost surely, the term of (11) vanishes.
(d) (Channel bias.) For and , the channel estimate with the limiting input, , the minimizer near of , satisfies
where the expectation is over the event, of probability , on which the minimizer is defined.
The first term of (13) is Box’s coordinate-dependent bias of a single nonlinear regression (Box, [9]; Cook, Tsai and Wei, [10]). Suitable local coordinates can eliminate it at a chosen point; a single global reparameterization need not remove it everywhere, and nonlinear target functionals can still be biased. The second is the input bias propagated through the channel, which depends on the intrinsic curvature of all channels and cannot. For the two-tissue model in its macro form , the channel is linear in and only the direction of curves, so the curvature that matters is that of the curves as the washout rate changes; spectral models with fixed rates and unconstrained coefficients are flat and immune, whereas the nonnegativity constraints of Cunningham and Jones [11] and Gunn et al. [18] create boundaries, which change the order of the bias (Theorem 4.10).
Example 4.4 (Known special cases). (a) Circle fitting. Points with angles uniform on the circle have , , , and with , so and the radius of the geometric fit converges to , the essential bias of Al-Sharadqah and Chernov [2]. (b) Template orbits. If with in a compact Lie group acting orthogonally on , has trivial stabilizer, and the template is parameterized in a neighbourhood of in the normal slice of its orbit, then and , with the curvature vector of the orbit at , and (11) becomes , Theorem 2 of Miolane, Holmes and Pennec [31]. Both are verified in Supplement, Section S3.4.
The two cases use symmetry differently. For template orbits every channel sees the same geometry, and , so the bias is that of a single orbit. In circle fitting the channel quantities are not identical: has rank one and depends on the angle, so no single point identifies the circle. All points lie on one manifold, however, and the rotations about the centre map the channel quantities into one another, so with uniform angles and are averages over a single orbit of the group. With heterogeneous channels the leading bias is generally not the unweighted average of single-type biases. Individual may be singular or poorly conditioned. Global ambiguity in the unrestricted rational class does not imply singular local information in the narrower Feng family: a single channel can have full local rank while retaining a discrete global ambiguity. In the design of § 7, at the eight XCAT organ means, the single-organ matrices have condition numbers between and , so single-type biases are unstable or undefined, whereas the eight types together have condition number . The proof of Theorem 4.2 therefore cannot use the group structure or the quotient on which these arguments rest. It works instead with one profiled criterion that is uniform over channel types.
A well-conditioned reversal under pooling
Let , , and
Take and or with equal probabilities, inside the compact channel set and a compact input interval containing zero in its interior. At the two true types, , and , respectively. Thus the pooled coefficient is , whereas the average of individual coefficients is . Every scalar has condition number one. The first output identifies and identifies , so the example satisfies the identification conditions. The reversal is caused by information weighting, without singular inverses. It illustrates a distinction between two algebraic operations, not the performance of an existing competing estimator.
Inference and the curvature correction
The sampling fluctuation of is of order and its bias of order , so the ratio is . Let and ; for , .
Theorem 4.5 (Local asymptotics). Let Assumption 4.1 hold, and with . Then, for every ,
Consequently, if , then ; if , then in probability. For with , the interval with in probability has asymptotic coverage , , which is only for and tends to zero as .
For fixed with ( reduced if needed), every limit point of differs from by at least (Theorem 4.2), so every interval centred at with width has coverage tending to zero. Over all linear functionals, the largest standardized bias is with , which is when (Supplement S3.5). In the design of § 7, , so the worst functional is biased by two standard errors once at and at .
The bias in (11) is computable from the fit. Let , , and let and be the averages of and evaluated at ; when , may be replaced by the pooled estimate. The curvature-corrected estimator is
It needs the first and second channel derivatives and the first input derivatives of at the fit, using dense linear algebra at cost once derivatives are available; forward evaluation and differentiation costs are additional.
Theorem 4.6 (Curvature correction). Under the conditions of Theorem 4.5, if in addition , then
For Gaussian noise with , is the inverse of the efficient information for when are treated as unknown constants (Proposition 3.9), and the asymptotic variance of equals it to first order; no optimality is claimed among procedures that exploit the distribution of the channel types.
The correction extends valid inference from channels to ; the proof is in Supplement S3.6. At moderate noise higher-order terms matter: Supplement S12 gives a finite-design expansion of the next terms, and Supplement S7 a conditional antithetic bootstrap correction for them, whose conditions are not verified for the PET model.
The computed estimator. Theorems 4.5 and 4.6 concern the exact minimizer over the compact sets and of Assumption 4.1(a), and a computation must fix them. In § 7, is a solver box and an ordered input chart, both chosen before the second validation study. Near degenerate faces of a channel family ( and for the macro channel) is nearly singular and , which contains , is unbounded, so a few fitted channels can dominate while carrying little information about . We therefore use the regular-set correction
where contains the channels with interior to and for a fixed . If , every channel is regular on events of probability tending to one, so and Theorem 4.6 holds for (Supplement, Corollary S3.5). Unlike , the regular set depends on the channel coordinates. This is an asymptotic equivalence, not a finite-sample improvement: the two estimators differ only through fitted channels outside the simultaneous regular-projection event, and a large difference warns against relying on the local approximation. The assumptions concern the true channels and do not rule out such fitted channels at finite noise. The domain is part of the estimator, so enlarging defines a different estimator, which § 7 examines as such.
On a smooth branch, interval estimates use the sandwich variance of the profiled estimating equation, with the profile Hessian of the criterion, with independent channels as sampling units (Supplement S3.5). For half the criterion the channel score is , with the whitened residual. Replacing it by requires the nuisance normal equations; at an active box face the full-tangent projection generally changes the score (Supplement S23). Grouping frames within a voxel does not account for spatial dependence between voxels. Replacing by its Gauss–Newton form is asymptotically equivalent in the correctly specified small-noise regime, but the two differ at moderate noise (§ 7); neither is a fixed-noise misspecification-robust sandwich.
Covariance changes both bias and uncertainty
The curvature-only formula requires isotropic errors in the fitting metric. General covariance also couples tangential noise to the motion of the normal space. This matters when diagonal frame weights do not fully whiten a reconstruction. With isotropic marginal noise, dependence within independent clusters of fixed size changes only the score covariance, not the curvature bias (Supplement, Theorem S10.1); but independence of arbitrarily chosen image blocks is not guaranteed. We therefore give a deterministic-design extension allowing dense Gaussian dependence; it is not a model-free justification for reconstructed PET intervals.
Write , and . For a positive definite marginal covariance define
Here is an matrix. When , the second term vanishes and . Although its two terms use coordinates, their sum is invariant under smooth channel reparameterization (Supplement S13).
Theorem 4.7 (Gaussian dependence and covariance-aware correction). Let the channel types form deterministic triangular arrays satisfying the uniform compactness, smoothness and channel separation in Assumption 4.1(a)–(c). Replace (d) by uniform separation of away from , and assume . Let , where , and , for constants . Put , and assume . If and , the exact joint least-squares estimator satisfies
Evaluate hatted geometry at the joint fit and set
If also , then . The consistent covariance estimate is , where . The same conclusions hold for a regular-set correction with any fixed .
With and known scales bounded away from and , the forcing reduces to and the expansion and limits agree with those of Theorems 4.5 and 4.6 for deterministic channel arrays with Gaussian noise; the corrected estimators coincide when and the pooled noise scale is used. The proof combines a uniform profile expansion with a Gaussian Poincaré bound; the leading score is exactly Gaussian, so no artificial partition into independent spatial clusters is needed. For with and fixed , its operator norm is bounded. Conversely, unrestricted long-range dependence need not satisfy the theorem. Consistent estimation of is an additional problem: operator-norm consistency suffices for the variance, but replacing the in the bias correction also requires (Supplement, Corollary S13.2).
Example 4.8 (Bias without curvature). Let , , with , and . The channels are straight lines, but and the leading bias is . For unrestricted line projections the population minimizer is exactly . Spatial correlation changes the variance through while leaving this marginal expected criterion unchanged. Thus a variance adjustment alone does not correct the bias. Correct marginal generalized least squares also removes this population bias; this example does not establish superiority over a correctly specified fitting metric.
Proposition 4.9 (A sufficient covariance-calibration experiment). Under Theorem 4.7, suppose , where , , is fixed and . Observe independent calibration fields , independent of the fitted data, with arbitrary unknown . Define
Then satisfies . Replacing the covariance in Theorem 4.7 retains its corrected limit and studentization, even for .
The proof in Supplement S13.5 uses Gaussian quadratic-form variances and an operator-norm Lipschitz bound uniform in . Covariance estimation changes the standardized correction by . One calibration field contains correlated channel vectors; it is additional information, not the residuals of a single fitted image. The covariance structure and known zero mean are substantive assumptions.
Physical constraints and boundary channels
Kinetic parameters obey sign constraints such as and , and basis-function methods impose nonnegativity ([11, 18]). If these constraints are imposed in the joint fit, a channel whose true parameter lies on or near a boundary is projected onto a cone rather than a manifold. Consider one linear constraint per channel, with the other coordinates interior. Let , let be its part orthogonal to the other columns of , and put
Thus is the input velocity along the constrained direction, and is the linearized distance of the true curve from the face of its channel manifold, in the metric of the whitened data; is bounded below under Assumption 4.1(c). Write , , and for the standard normal density and distribution function, and for .
Theorem 4.10 (Boundary channels). Let Assumption 4.1 hold for the extended model with Gaussian noise, and let the true channels lie in , possibly on the face . Let be the joint estimator with in every channel, and let be its expected criterion. Then every limit point of lies in . For , every satisfies
where is the unique root of . Moreover , and as .
If the aggregate boundary forcing is nonzero, the constrained joint estimator is therefore biased by an amount of exact order , against without the constraint. A positive fraction of channels on the face is necessary for this but not sufficient, because boundary velocities of different channel types can cancel.
Example 4.11 (Cancelling boundary forcing). Let , , with , and let the channel types be equally likely, all with , and . Every channel lies on the face, , , and . Then , because , so . Indeed the reflection leaves the model and the noise law invariant, so the constrained estimator has no bias of order although every channel is at the boundary. With uniform on instead, and . Because the model is linear, exactly here. In 20 simulated data sets with channels at each of , and , the constrained estimate divided by averaged in the second case and was within of zero in the first.
When the forcing is nonzero, the order- bias is large relative to the nominal fixed-effects standard error, which is of order : the standardized bias grows like at every noise level, and the curvature correction does not remove it. Channels farther than from the face contribute only to the term. Newton's method from gives with ; this one-step value is accurate only when is small relative to , and we solve for the root in the experiments (proof in Supplement S5). Several active constraints per channel lead to a projection onto a polyhedral cone and a multivariate truncated-normal forcing term; we do not treat that case.
The forcing term has a direct reading. For the unconstrained fit the linearized slack estimate is , and . Thus is the average of the expected constraint violations of the unconstrained fits, weighted by the boundary velocities. More generally , so the unconstrained fits estimate without bias to first order under correct specification. Violations by unconstrained voxel fits, as in Supplement S30, therefore indicate the forcing term that a constrained input fit would create; on misspecified data the indication is only qualitative. These results suggest a two-stage rule. First estimate the shared input with the smooth extension, to which Theorems 4.2–4.6 apply. Then impose the physical constraints only when each channel is fitted with the estimated input. In the second stage the boundary is a fixed-dimensional problem for each channel (Self and Liang, [40]; Andrews [3]), and its error does not accumulate over channels. A parametric bootstrap of the constrained input fit is not expected to be a remedy, because the local parameters cannot be estimated consistently (Andrews, [4]); we have not studied it.
Continuous accumulation near a boundary differs from an atom on the face. If its slack density behaves as , , a nonzero leading boundary coefficient gives bias of order . Supplement, Proposition S11.1, states the precise regularity and noncancellation conditions and derives the coefficient. This dominates smooth curvature bias. A bounded slack density instead gives an displacement, whose coefficient can still differ from the unconstrained one.
Population models for the channels
An unrestricted mixing distribution can restore consistency under compactness and support identification (Supplement, Proposition S8.1), following Kiefer and Wolfowitz [28]. This does not assert consistency for a fixed parametric population. Here the channels are independent, and the noise is Gaussian. For a fixed working density on , let
and let maximize . Write and let denote columnwise divergence.
Theorem 5.1 (Population-score bias). Let Assumption 4.1 hold with Gaussian noise and , let be the closure of an open set with Lipschitz boundary (a box, for instance), and strengthen part (b) to on a neighbourhood, with uniformly bounded derivatives. Let be positive and on a neighbourhood of , with bounded derivatives and . Every limit point of maximizes , and there are and such that for every maximizer satisfies
The proof uses a uniform Laplace expansion and a volume identity that combines the profile curvature with the Laplace determinant (Supplement S9). For a smooth true population supported in the interior, integration by parts gives : the coefficient measures population-score misfit in the foot-point directions . The exact correctly specified likelihood also has the truth as an expected maximizer at every noise level, by the Kullback–Leibler inequality. This last fact does not follow by substituting into the theorem, whose positive-density assumptions differ.
Supplement S8 gives precise results for correct, flat and estimated populations. A jointly fitted working family needs additional uniformity and an identified limiting population; a hierarchical posterior conclusion further requires Bernstein–von Mises and moment conditions, unverified here. In the XCAT design, fixed Gaussian-mixture working populations leave second-order coefficients 1.7–7.1 times the fixed-effects coefficient in the information norm. These coefficients concern a small-noise expansion, not realized finite-noise bias. The analytic correction needs no population model. The attempted hierarchical sampler did not converge and remains a supplementary diagnostic, not evidence for posterior calibration.
Calibration limits
An exact scale symmetry leaves absolute kinetic targets dependent on external calibration, however many channels are pooled (Proposition 3.10). Supplement, Proposition S15.1, gives a van Trees bound along that symmetry orbit. With Gaussian log-scale calibration error of variance , its lower bound for aggregate squared log-influx error is on a regular orbit interval of half-width . This scale limitation is separate from the shape bias and from covariance-shape calibration in Proposition 4.9.
Numerical experiments
Common setting
All PET simulations use the 35 frames of the total-body study of Supplement S30 (12×5 s, 4×15 s, 6×1 min, 5×2 min, 4×3 min, 4×5 min) and, as true input, the Feng fit to its aortic image-derived input ( min, min; relative error of the fit 0.048). Channel parameters are built from the eight $1\mathrm{FDG} organ parameter sets of the XCAT phantom distributed with qurit/dynamicPET ((Ashrafinia et al., n.d.)), for myocardium, liver, lungs, kidneys, spleen, bone, white and grey matter, with k_4=0. An organ is chosen uniformly, its blood-volume value is floored at 0.005, and each log parameter is perturbed independently. The noise of frame ir\sigma\bar{f}_r u_i\bar{f}_ru_i^2 \propto \bar{c}_i/\Delta_i\bar{c}\sigma is a coefficient of variation; the data are whitened by these known shapes, so s_r \equiv 1(v,a_1,a_2,\log\beta); near the true curves it has the same channel manifold, so the same h_rN_rm_r (Proposition 4.3(a)), and it extends the manifold across the boundaries v=0k_3=0. For every candidate input each channel is fitted by linear least squares in (v,a_1,a_2)\beta$ followed by Levenberg–Marquardt refinement; the input is updated by Gauss–Newton steps on the profiled criterion, and the channel search is repeated at the final input until no channel improves. The procedure does not certify a global optimum, so every fit reports its standardized profile score (about zero at a stationary point, one at a point one standard error away); Supplement S16.2 states conditions under which the limit theory transfers to computed estimates, and Supplement S31 records seeds, run times and failures.
Computational provenance
An earlier implementation of the frame-integrated convolution was inaccurate when a channel washout rate nearly coincided with an input decay rate. The repaired kernel agrees with 50-digit quadrature to in 2,016 checks, and every result in this section was computed with it; earlier-kernel computations are historical analyses (Supplement S26), not used for any conclusion.
Designs
The validation design has a compact data-generating law compatible with Theorems 4.5 and 4.6. Each log parameter is perturbed by with standard normal truncated to , so the channel types have compact support, interior to the solver box ( min); the certificate ((6)) holds for every lung-type channel, so Corollary 3.8 gives Assumption 4.1(c)–(d). No estimator is started at the truth. Each data set has its own simulated image-derived input, the true input dispersed by an exponential kernel of mean 5 s with 5% multiplicative frame noise; its Feng fit starts the joint estimator, a median of 18 and at most 117 sandwich standard errors from the truth in the worst coordinate. The channel grid has 361 points, and each fit ends with a search along the two weakest directions of the information for a lower local minimum (Supplement S19). The reference design (unbounded log-normal perturbations, a 157-point grid, true start) is used below only for physical constraints; the stress design generates data from deliberately different models.
The predicted bias
In the coordinates the predicted deviation per unit in the validation design is , with ; has condition number , and in 5000 draws. The two weakest directions of are made of the amplitudes and decay rates of the two slow input components, which the trapping term can partly mimic (Example 3.5). We summarize an estimate by its standardized projection on the predicted direction , fixed in advance; the population minimizer has . This functional attains the largest standardized bias , so Theorem 4.5 predicts that nominal 95% intervals for it from the joint estimator cover with probability about , whereas Theorem 4.6 predicts 95% for the corrected estimator. Theorem 4.6 concerns distributional centring, not a finite-sample expectation, so empirical means assess higher-order terms, finite and computation together.
First validation study
The noise level decreases along for , 1024 and 4096, so that , 20 and 41 grows while . The study comprises all 120, 80 and 30 planned data sets, fixed before the runs and generated from disjoint, prespecified random-number streams; no computation failed (Table 2). The joint fit had , 1.16 and 1.13 (Monte Carlo standard errors 0.06, 0.04, 0.04) and its intervals covered in 36%, 8% and 0% of the data sets, against 46%, 17% and 1% predicted; the bias exceeded its first-order value by 13–16%. The smooth correction (4.8) had , 0.22 and 0.16, a residual bias of 0.5–0.7 standard errors, and its intervals covered in 88%, 88% and 97% (exact intervals 81–93%, 78–94%, 83–100%). The correction thus removed most of the bias but did not deliver nominal coverage for the projection at and 1024. Coordinatewise the shortfall was larger: at the corrected intervals for the six coordinates covered in 68–92% of the data sets, and the empirical spread exceeded the Gauss–Newton sandwich standard error by factors 1.20–1.97; at and 4096 the coverage was 86–96% and 97% (Supplement, Table S11 and Figure S6).
| estimator | bias/se | covered | 95% CI (%) | predicted (%) | sd/se | ||||
| 256 | 0.200 | 10.2 | joint | 1.14 (0.06) | 2.31 | 43/120 | [27, 45] | 46 | 1.26 [1.11, 1.39] |
| 256 | 0.200 | 10.2 | corrected | 0.26 (0.05) | 0.54 | 106/120 | [81, 93] | 95 | 1.21 [1.07, 1.35] |
| 1024 | 0.141 | 20.5 | joint | 1.16 (0.04) | 3.41 | 6/80 | [3, 16] | 17 | 1.03 [0.89, 1.15] |
| 1024 | 0.141 | 20.5 | corrected | 0.22 (0.04) | 0.64 | 70/80 | [78, 94] | 95 | 1.02 [0.88, 1.13] |
| 4096 | 0.100 | 41.0 | joint | 1.13 (0.04) | 4.68 | 0/30 | [0, 12] | 1 | 0.88 [0.68, 1.04] |
| 4096 | 0.100 | 41.0 | corrected | 0.16 (0.04) | 0.65 | 29/30 | [83, 100] | 95 | 0.88 [0.68, 1.04] |
Table 2. Validation design, regenerated with the repaired convolution kernel (all planned data sets, independent seeds fixed in advance): compact channel law, and a simulated image-derived start. Mean standardized projection on the predicted bias direction (Monte Carlo standard error), mean error divided by the mean sandwich standard error, coverage of nominal 95% intervals with exact 95% intervals, the first-order coverage predicted by Theorem 4.5 (joint) or Theorem 4.6 (corrected), and the ratio of the empirical standard deviation to the mean sandwich standard error with a paired bootstrap interval.
Bias and variance at smaller noise
A prespecified experiment with and 30 channel samples, each analysed at , 0.1 and 0.05 with antithetic noise and the same channels, noise array and start, separated residual bias from variance (Table 3). Half-sums of antithetic pairs contain the bias but no first-order noise: the joint estimator had , 1.09, 1.05 and the corrected one 0.32, 0.11, 0.05 (Monte Carlo standard errors 0.02, 0.02, 0.02), a corrected bias of 1.30, 0.23, 0.05 standard errors. At fixed the second-order bias of the fitted estimator is not the population coefficient, and a finite antithetic average retains random quadratic terms; Supplement S12 derives these finite-design coefficients. Their realized values average 1.039 (Monte Carlo s.e. 0.021) for the joint fit and 0.035 (0.020) for the correction; the remainders after subtracting them fall by factors of 3–5 per halving of , which explains most of the apparent discrepancy from 1 and 0 at small noise. Half-differences retain the first-order noise; their spread was 0.95, 1.01, 0.99 times the sandwich standard error along the projection and 0.70–1.27, 0.77–1.10, 0.76–1.09 times coordinatewise. At , then, the residual is higher-order bias that shrinks with , and the sandwich is accurate for the first-order noise term.
| pairs | joint | corrected | corr. bias/se | projection | coordinates | joint | corrected | corr. coord. | |
| 0.2 | 30 | 1.19 (0.02) | 0.32 (0.02) | 1.30 | 0.95 | 0.70–1.27 | 0 | 73 | 68–98 |
| 0.1 | 30 | 1.09 (0.02) | 0.11 (0.02) | 0.23 | 1.01 | 0.77–1.10 | 48 | 97 | 88–100 |
| 0.05 | 30 | 1.05 (0.02) | 0.05 (0.02) | 0.05 | 0.99 | 0.76–1.09 | 80 | 98 | 87–100 |
Table 3. Bias and variance at decreasing noise (, compact channel law, antithetic noise pairs with common channel samples, noise and simulated image-derived starts across ). Half-sums cancel odd noise terms and estimate the mean bias: mean standardized projection (Monte Carlo standard error over pairs) of the joint and corrected estimators and the corrected bias in units of the mean sandwich standard error. Half-differences retain odd terms, led by the first-order noise: their root mean square over the mean sandwich standard error, along the projection and over the six coordinates. Coverage (%) of nominal 95% sandwich intervals over both members of every pair, along the projection (joint, corrected) and the range over the six coordinates (corrected).
The solver domain
The solver domain. In 68 data sets a channel ended at the log box , near the faces and of the macro family, where the channel information degenerates. Refitted with the box , 16 joint estimates moved by more than half a standard error to other local minima, and the plug-in correction changed by up to 44 standard errors; the corrected coverage at fell from 106 to 88 of 120 (103 with a face-specific correction). The box restricts the numerical domain without by itself verifying Assumption 4.1, and the correction is sensitive to how close it lets channels come to these faces. Refits of the 31 data sets with the largest scores or two local minima, with tighter tolerance, an extended valley search and a second start, reduced the largest score from 0.51 to , moved the estimates by at most 2.26 standard errors and changed the corrected projection coverage at from 106 to 105 of 120, so optimization error of this size does not explain the residual bias or the coordinatewise undercoverage. The regular-set correction (16) with , applied post hoc to the stored fits, differs from the smooth correction by at most 0.29 standard errors at the original box, with projection coverage 106/120, 70/80 and 29/30. The second study recomputed the wider-box refits with saved fits; its selection rule, the lower criterion of the refit and the original fit, was fixed after the follow-up, so this is a recomputation of results already seen. With that rule the regular-set correction term changed by at most 0.57 standard errors (90th percentile 0.07), against up to 44 for the smooth term, and projection coverage at was 106 of 120 on both domains. The joint estimate still moved by more than half a standard error in 12 of the 68 data sets (at most 2.63), to other local minima along the slow-input valley (Supplement S20). The instability of the correction came from channels near degenerate faces, which the regular set excludes; the multiplicity of local minima remains.
Second validation study
Second validation study. After the first study we fixed the computed estimator: the box , the diagnostic input chart (delay within the onset frame less s at each end, , other coordinates within of the reference input), the regular-set correction with , and the original Gauss–Newton (primary) and profile-Hessian variance calculations. The threshold is below the empirical minimum 0.043 from 5000 draws; this is not a certified bound on the support or a verification that . It was applied to new prespecified data sets at , where the first study fell shortest, with , and and 400, 200 and 200 independent data sets, so that a 95% coverage has Monte Carlo standard error 1.1–1.5 points; the design file and code fingerprints were written before the first run (Supplement S20). All 800 data sets were computed and none failed. No fit left the chart. The solver checks this chart at the returned fit but does not enforce it along the search or certify minimization over it. Irregular channels occurred only at (in 290 of 400 data sets, 165 of them with a channel at the box), where the regular-set and smooth corrections differed by at most 0.49 standard errors.
At the observed corrected coverage was compatible with 95%, 94% for the projection (exact interval 89–96%) and 93–96% for the six coordinates, with spread ratios 0.99–1.08; the joint intervals covered the projection in 92%. At corrected coverage was 93% (88–96%) for the projection and 91–97% coordinatewise, with spread ratios 0.94–1.15. At the shortfall of the first study reappeared: projection 85% (81–88%), coordinates 72–95%, spread ratios 1.09–1.93, and a residual bias of 0.69 standard errors. A post hoc audit replaced the full-tangent projected scores by the envelope scores in the fixed-face profile-Hessian calculation (Supplement S23). This gave coverage 90% for the projection and 89–96% coordinatewise, with spread ratios 1.01–1.27; at the two smaller noise levels it changed little (Table 4). These results support a smaller variance discrepancy as decreases at fixed . They attribute part of the discrepancy to the variance approximation, without establishing finite-noise validity at active faces; the residual bias at is not removed by either sandwich. A second start changed the joint estimate by more than 0.1 standard errors in four of 120 data sets, by at most 0.50, and the standardized scores exceeded 0.1 in 56 data sets (maximum 0.61), all at .
| sets | cov. | bias/se | cov. [CI] | coord. | sd/se | cov. [CI] | coord. | sd/se | ||
| 0.2 | 400 | 35 | 0.35 (0.03) | 0.69 | 85 [81, 88] | 72–95 | 1.26 | 90 [87, 93] | 89–96 | 1.11 |
| 0.1 | 200 | 79 | 0.20 (0.07) | 0.21 | 93 [88, 96] | 91–97 | 1.06 | 93 [89, 96] | 94–98 | 1.03 |
| 0.05 | 200 | 92 | 0.06 (0.14) | 0.03 | 94 [89, 96] | 93–96 | 1.02 | 94 [89, 96] | 94–97 | 1.01 |
Table 4. Table 4. Second validation study, : the computational procedure fixed after the first study (box log , input chart checked at returned fits, regular-set correction (16) with ) on new data sets prespecified before the runs. Sets: number of data sets. All coverages are percentages of nominal 95% intervals. Projection on the predicted direction: coverage of joint-estimator intervals; mean of the corrected estimator (Monte Carlo standard error) and its bias over the mean standard error. Corrected intervals with the original Gauss–Newton calculation and the post hoc envelope-score profile-Hessian calculation: projection coverage [exact 95% interval], range of coordinatewise coverage, and the ratio of empirical standard deviation to mean standard error for the projection. The post hoc variant changes only the covariance, not the point estimates; the original secondary results are retained in the Supplement.
Heterogeneous channels
Heterogeneous channels. Two post hoc algebraic approximations illustrate why pooling and inversion cannot be interchanged: the average and the coefficient at the mean channel parameters. At the true input and compact-law channels their projections on the pooled direction are and (pooled value 1), at angles and in the metric. The first approximation is numerically fragile because individual information matrices have median condition number . Supplement S21 reports their plug-in interval summaries. Neither approximation implements a published template estimator; these comparisons do not establish superiority over the methods of Miolane, Holmes and Pennec [31]. The well-conditioned example in § 4 provides an independent algebraic demonstration of the weighting issue.
Numerical error
In the first study, refits from the true input agreed with the fits from the image-derived start within 0.1 standard errors in 26 cases and otherwise ended in other local minima at most 0.70 standard errors apart; a twice finer grid changed no estimate by more than standard errors (Supplement S19).
Physical constraints in the joint fit
A random quarter of the reference-design channels was given no trapping (, hence ). On matched data, a fit imposing was compared with the smooth macro fit, with and 8 data sets per noise level (Table 5). The constrained-minus-unconstrained difference divided by tracks the root of Theorem 4.10 at smaller noise. For at , the measured value is 1.71 (0.05) against 1.63 predicted, whereas the curvature contribution has size .
| coordinate | measured | predicted | measured | predicted | measured | predicted | curvature |
| 0.124 (0.003) | 0.122 | 0.121 (0.004) | 0.122 | 0.128 (0.008) | 0.122 | ||
| (0.003) | (0.005) | (0.005) | |||||
| 0.039 (0.003) | 0.034 | 0.039 (0.004) | 0.034 | 0.056 (0.004) | 0.034 | 1.42 | |
| 1.712 (0.053) | 1.626 | 1.604 (0.043) | 1.626 | 1.793 (0.101) | 1.626 |
Table 5. Physical constraints in the joint fit (Theorem 4.10): a quarter of the XCAT channels have , . Paired difference between the joint estimates with and without the constraint , divided by (mean over data sets, Monte Carlo standard error), and the prediction of Theorem 4.10. The last column is the unconstrained curvature bias per unit in this design, for comparison. The delay and rows are below 0.02 in all cells.
Population models and stress tests
Local quadrature diagnostics for working Gaussian mixtures agree with the population-score coefficient within 10.5% at (Supplement S22). Under misspecification (reversible kinetics, channel-specific delays or dispersion, a varying plasma-to-blood ratio) the noise-free limit of the joint estimator moved by more than the curvature bias at in every case but the last (Supplement, Table S28).
Correlated noise and covariance calibration
For the flat channels of Example 4.8, known covariance separates the roles of bias correction and spatial variance adjustment (Supplement S24). With and marginal and spatial correlations both 0.6, corrected 95% coverage at , 0.1 and 0.05 was 94.27, 94.88 and 95.06%, against 71.23, 72.29 and 72.53% when spatial correlation was ignored; correct marginal GLS gave similar coverage. Estimating the covariance from one independent calibration field, as in Proposition 4.9, gave 94.4, 95.1 and 95.4% (10,000 replicates per cell). The correction repairs the specified Euclidean estimator; no superiority over correct whitening is claimed.
A compartment-model experiment uses 50 independent datasets at each of and 0.05, with , temporal correlation 0.5, spatial correlation 0.4 and one independent calibration field per dataset; all datasets are retained. The estimated temporal matrix had a mean relative operator-norm error of about 41 and 40%, yet within each dataset the estimated covariance and oracle analyses agreed closely on the standard-error scale: ratios of their standard errors had median 0.99 and 0.99 and 5–95% range 0.94–1.08, and the corrected contrasts differed by at most 0.28 oracle standard errors. Coverage of the prespecified contrast was 86 and 98% without correction, 92 and 98% with oracle correction and 98 and 98% with estimated correction. The oracle and estimated intervals disagreed for 3 and 0 of 50 datasets at and 0.05, respectively, all borderline, so the study does not rank the two analyses; with 50 replicates per level its coverage estimates are imprecise (Supplement S24.4). These simulations assume known fitting weights and a correctly specified calibration field, neither of which is validated for a patient reconstruction.
Discussion
The theoretical distinction is between identification and estimation with incidental channels. A shared input can be identifiable from frame means yet estimated with a persistent curvature bias, whose leading term is an information-weighted regression of mean curvature on normal input velocity. The analytic correction removes this term under uniform smoothness and small-noise conditions; it is not a fixed-noise consistency result. Boundary constraints and population misspecification introduce different bias terms. With general marginal covariance, even flat moving channels can be biased. The Gaussian-dependence theorem distinguishes this marginal effect from the full score covariance needed for uncertainty.
In compact-population PET simulations with prespecified independent replicates, the joint intervals deteriorated as grew and the correction removed most of the bias. The two validation studies locate where the first-order theory is accurate. Along , the corrected intervals retained a residual bias of 0.5–0.7 standard errors and undercovered at and 1024. At with the estimator fixed in advance, they were close to nominal at and 0.1 but not at , where a residual bias remained and the Gauss–Newton sandwich understated the variance; the audited fixed-face profile-Hessian calculation reduced that understatement. Its score correction was post hoc, and it supplies no boundary-limit theorem. The finite-design expansion of the Supplement accounts for much of the gap between the fitted estimator and the population coefficient at fixed . We therefore claim calibrated inference only in the small-noise regime of Theorem 4.6; at in this design the corrected intervals are better than the uncorrected ones but not calibrated.
The computed estimator includes its domain. A box that lets fitted channels approach degenerate faces of the channel family makes the smooth plug-in correction unstable; the regular-set correction, equivalent to it under the assumptions, substantially reduced that instability of the correction term in the tested fits. Choosing the lower criterion among candidate fits keeps the reported optimum from worsening when the domain grows, although the joint fit can then move to another local minimum. Weak identification of the slow input components still makes the profiled criterion nearly flat along a valley, so starting values and optimization diagnostics matter at moderate noise. Tighter refits of the data sets with the largest profile scores moved the estimates by at most 2.26 standard errors and the projection coverage changed from 43 to 44 of 120 (joint, ) and from 106 to 105 of 120 (corrected, ). Finally, heterogeneity makes the order of pooling and inversion important: the two-type example proves this without ill-conditioning, and the PET diagnostics illustrate its numerical consequences.
Unrestricted non-Gaussian spatial dependence, estimation of reconstruction covariance with external information, multiple active constraints, a uniform bootstrap theory and reliable estimation of flexible populations remain outside the results; the hierarchical sampler we implemented did not converge. The historical single-subject PET example (Supplement S30), computed with the earlier kernel, has unresolved optimization, correlated reconstructed noise and uncertain calibration; it illustrates sensitivity and supports no calibrated-inference claim.
Acknowledgments
The author thanks the developers of PET-TRACER, vPET-ABC, dynamicpet and qurit/dynamicPET for making their code and data openly available.
Funding
H.-H. Huang acknowledges support from the National Science Foundation under grants DMS-1924792 and DMS-2318925 and from the National Institutes of Health (NIH), National Institute of Allergy and Infectious Diseases (NIAID) R01 award (1R01AI206603-01).
Supplementary Material
Supplement to “Identification and curvature bias for shared inputs in multichannel dynamical models”
Available from the author on request: proofs, additional theoretical results, simulation details and numerical diagnostics, the historical PET sensitivity analysis, and reproducibility code and data-access information.
References
- [1]Abed-Meraim, K., Qiu, W. and Hua, Y. (1997). Blind system identification. Proceedings of the IEEE 85 1310–1322.
- [2]Al-Sharadqah, A. and Chernov, N. (2009). Error analysis for circle fitting algorithms. Electronic Journal of Statistics 3 886–911. https://doi.org/10.1214/09-EJS419
- [3]Andrews, D. W. K. (1999). On a parameter is on a boundary. Econometrica 67 1341–1383.
- [4]Andrews, D. W. K. (2000). Inconsistency of the bootstrap when a parameter is on the boundary of the parameter space. Econometrica 68 399–405.
- [5]Arellano, M. and Hahn, J. (2007). Understanding Bias in Nonlinear Panel Models: Some Recent Developments. In Advances in Economics and Econometrics: Theory and Applications, Ninth World Congress (R. Blundell, W. Newey and T. Persson, eds.) 381–409. Cambridge University Press. https://doi.org/10.1017/CBO9780511607547.013
- [7]Bates, D. M. and Watts, D. G. (1980). Relative Curvature Measures of Nonlinearity. Journal of the Royal Statistical Society: Series B (Methodological) 42 1–25. https://doi.org/10.1111/j.2517-6161.1980.tb01094.x
- [8]Bigot, J. and Charlier, B. (2011). On the consequences of Fréchet means in deformable models and image analysis. Electronic Journal of Statistics 5 1054–1089. https://doi.org/10.1214/11-EJS633
- [9]Box, M. J. (1971). Bias in Nonlinear Estimation. Journal of the Royal Statistical Society: Series B (Methodological) 33 171–190. https://doi.org/10.1111/j.2517-6161.1971.tb00871.x
- [10]Cook, R. D., Tsai, C. L. and Wei, B. C. (1986). Bias in nonlinear regression. Biometrika 73 615–623. https://doi.org/10.1093/biomet/73.3.615
- [11]Cunningham, V. J. and Jones, T. (1993). Spectral analysis of dynamic PET studies. Journal of Cerebral Blood Flow and Metabolism 13 15–23.
- [12]Devilliers, L., Allassonnière, S., Trouvé, A. and Pennec, X. (2017). Template Estimation in Computational Anatomy: Fréchet Means Top and Quotient Spaces Are Not Consistent. SIAM Journal on Imaging Sciences 10 1139–1169. https://doi.org/10.1137/16M1083931
- [13]Fan, Y., Emvalomenos, G., Grazian, C. and Meikle, S. R. (2021). PET-ABC: fully Bayesian likelihood-free inference for kinetic models. Physics in Medicine and Biology 66 115002. https://doi.org/10.1088/1361-6560/afba37
- [14]Fazekas, I., Kukush, A. G. and Zwanzig, S. (2004). Correction of nonlinear orthogonal regression estimator. Ukrains’kyi Matematychnyi Zhurnal 56 1101–1118.DOI
- [15]Feng, D., Huang, S. C. and Wang, X. (1993). Models for computer simulation studies of input functions for tracer kinetic modeling with positron emission tomography. International Journal of Bio-Medical Computing 32 95–110.
- [16]Feng, D., Wong, K. P., Wu, C. M. and Siu, W. C. (1997). A technique for extracting physiological parameters and the required input function simultaneously from PET image measurements: theory and simulation study. IEEE Transactions on Information Technology in Biomedicine 1 243–254.
- [18]Gunn, R. N., Gunn, S. R., Turkheimer, F. E., Aston, J. A. D. and Cunningham, V. J. (2002). Positron emission tomography compartmental models: a basis pursuit strategy for kinetic modeling. Journal of Cerebral Blood Flow and Metabolism 22 1425–1439.
- [19]Hahn, J. and Newey, W. (2004). Jackknife and Analytical Bias Reduction for Nonlinear Panel Models. Econometrica 72 1295–1319. https://doi.org/10.1111/j.1468-0262.2004.00533.x
- [20]Hastie, T. and Stutzle, W. (1989). Principal Curves. Journal of the American Statistical Association 84 502–516.DOI
- [21]Holler, M., Morina, E. and Schramm, G. (2024). Exact parameter identification in PET pharmacokinetic modeling using the irreversible two tissue compartment model. Physics in Medicine and Biology 69 165008. https://doi.org/10.1088/1361-6560/ad539e
- [22]Holler, M., Morina, E. and Schramm, G. (2025). Exact parameter identification in PET pharmacokinetic modeling: extension to the reversible two tissue compartment model. Inverse Problems 41 095009. https://doi.org/10.1088/1361-6420/adfacc
- [23]Huang, H. H. and He, Q. (2023). Nonlinear Regression Analysis. In International Encyclopedia of Education fourth ed. (R. J. Tierney, F. Rizi and K. Erickan, eds.) 558–567. Oxford: Elsevier. https://doi.org/10.1016/B978-0-12-818630-5.10068-5
- [24]Huang, H. H., Yu, F., Fan, X. and Zhang, T. (2024). A framework of regularized low-rank matrix models for regression and classification. Statistics and Computing 34 10. https://doi.org/10.1007/s11222-023-10318-z
- [25]Huang, H. H., Zhu, Z., Boopasiri, S., Chen, Z., Pang, S. and Kao, C. M. (2025). A statistical reconstruction algorithm for positronium lifetime imaging using time-of-flight positron emission tomography. IEEE Transactions on Radiation and Plasma Medical Sciences 9 478–486. https://doi.org/10.1109/TRPMS.2025.3531225
- [26]Kanatani, K. (1994). Statistical Analysis of Geometric Computation. CVGIP: Image Understanding 59 286–306. https://doi.org/10.1006/ciun.1994.1020
- [27]Kanatani, K. and Sugaya, Y. (2013). Hyperaccurate Correction of Maximum Likelihood for Geometric Estimation. IPSJ Transactions on Computer Vision and Applications 5 19–29.
- [28]Kiefer, J. and Wolfowitz, J. (1956). Consistency of the Maximum Likelihood Estimator in the Presence of Infinitely Many Incidental Parameters. Annals of Mathematical Statistics 27 887–906. https://doi.org/10.1214/aoms/1177728066
- [29]Kumon, M. and Amari, S. I. (1984). Estimation of a structural parameter in the presence of a large number of nuisance parameters. Biometrika 71 445–459. https://doi.org/10.1093/biomet/71.3.445
- [30]Lancaster, T. (2000). The incidental parameter problem since 1948. Journal of Econometrics 95 391–413. https://doi.org/10.1016/S0304-4076(99)00044-5
- [31]Miolane, N., Holmes, S. and Pennec, X. (2017). Template Shape Estimation: Correcting an Asymptotic Bias. SIAM Journal on Imaging Sciences 10 808–844. https://doi.org/10.1137/16M1084493
- [32]Moulines, E., Duhamel, P., Cardoso, J. F. and Gassiat, Y. S. (1995). Subspace methods for the blind identification of multichannel FIR filters. IEEE Transactions on Signal Processing 43 516–525.
- [33]Neyman, J. and Scott, E. L. (1948). Consistent Estimates Based on Partially Consistent Observations. Econometrica 16 1–32. https://doi.org/10.2307/1914288
- [34]Ogden, R. T., Zanderigo, F., Foy, Y. S., Mann, J. J. and Parsey, R. V. (2010). Simultaneous Estimation of Input Functions: An Empirical Study. Journal of Cerebral Blood Flow and Metabolism 30 816–826. https://doi.org/10.1038/jcbfm.2009.245
- [35]Pólya, G. and Szegő, G. (1976). Problems and Theorems in Analysis II. Berlin: Springer. Translation of Aufgaben und Lehrsätze aus der Analysis, 1925.
- [36]Prenosil, G. A., Sari, H., Fürstner, M., Afshar-Oromieh, A., Shi, K., Rominger, A. and Hentschel, M. (2022). Performance Characteristics of the Biograph Vision Quadra PET/CT System with a Long Axial Field of View Using the NEMA NU 2-2018 Standard. Journal of Nuclear Medicine 63 476–484. https://doi.org/10.2967/jnumed.121.261972
- [38]Riabkov, D. Y. and Di Bella, E. V. R. (2002). Estimation of kinetic parameters without input functions: analysis of three methods for multichannel blind identification. IEEE Transactions on Biomedical Engineering 49 1318–1327.
- [39]Sartori, N. (2003). Modified profile likelihoods in models with stratum nuisance parameters. Biometrika 90 533–549. https://doi.org/10.1093/biomet/90.3.533
- [40]Self, S. G. and Liang, K. Y. (1987). Asymptotic properties of maximum likelihood estimators and likelihood ratio tests under nonstandard conditions. Journal of the American Statistical Association 82 605–610.
- [41]Wong, K. P., Feng, D., Meikle, S. R. and Fulham, M. J. (2001). Simultaneous estimation of physiological parameters and the input function—in vivo PET data. IEEE Transactions on Information Technology in Biomedicine 5 67–76.
- [43]Zhao, Y., Gu, Q., Angelis, G. I., Reader, A. J., Fan, Y. and Meikle, S. R. (2026). Generative Consistency Models for Estimation of Kinetic Parametric Image Posteriors in Total-Body PET. IEEE Transactions on Medical Imaging 45 4137–4149. arXiv:2509.13614. Code and data: https://github.com/yundunbledore/PET-TRACER. https://doi.org/10.1109/TMI.2026.3692917