Mathematical Pharmacology
Introduction
Pharmacokinetics is usually presented as a small, self-contained piece of linear algebra. A drug is administered into a compartment, a well-stirred pool in which every molecule of drug is equally likely to leave by any available route. A single concentration and a handful of rate constants then describe it completely. Drug moves between compartments by first-order exchange and leaves the body by first-order elimination. The resulting differential equation is linear. The literature writes it in matrix notation whenever it writes it formally: a vector of compartment concentrations, a matrix of rate constants and a solution built from a single exponential.
Pharmacodynamics is taught as something else. Its standard object is a curve to fit: the Hill equation or the Emax model, four parameters estimated by nonlinear regression against a scatter of measured effect against dose. The dynamical system that produced the curve is left implicit or omitted, and the literature rarely mentions the linear machinery next door.
Quantitative systems pharmacology (QSP) looks like a third and much larger subject. A QSP model may have tens to hundreds of state variables (free drug, bound drug, receptor, ligand-receptor complex, and a cascade of downstream signalling intermediates) coupled by nonlinear mass-action and Michaelis-Menten terms. Each new pathway is diagrammed and coded as its own construction and little of the last model's structure carries over.
These three literatures rarely cite one another and share less mathematics. They read as three subjects because they are taught, reviewed and published as three. Each is an instance of one object: a state that evolves under a control input and is read out through a measurement ,
Pharmacokinetics is the case in which and are linear. Pharmacodynamics is the case in which the response curve is the output map evaluated at a fixed point of a compartmental flow, linear or Michaelis-Menten regardless of its compartment count; Section II.6 works the simplest, one-compartment instance. QSP is the case in which and are left fully nonlinear on the pharmacokinetic and pharmacodynamic sides at once, and the state dimension is allowed to grow large. Nothing about the object in [eq:state-eq-preview] changes across these three cases. Only the fragment of its theory that each field imports differs, as Section II.1 explains.
Two anchors run through everything that follows.
Control theory before physics. Every tool built in this Part (controllability, observability, BIBO stability, a positivity structure specific to compartmental sign patterns, and a structural test for observability that needs no numbers at all) was developed by control engineers and mathematical economists between the 1940s and the 1970s to answer questions about feedback systems, well before anyone asked them of a compartmental drug model. We import the finished theory and build no pharmacology-specific analogues. The theory ignores what measures and so do we until Part II.
Observability at scale. The question is what happens to controllability and observability once the state dimension grows from a handful of compartments to the range QSP models occupy. Does every state coordinate remain identifiable from the few outputs a laboratory can measure? Once forming a matrix by hand stops being the bottleneck, can the answer even be computed? Part IV answers both.
The construction proceeds from the ground up. Part I (this Part) builds the skeleton: the state-space object itself, its closed-form solution, controllability, observability, BIBO stability, the positivity structure that compartmental sign patterns force on the flow, and a structural observability test that needs only the wiring diagram of the system. Every statement is made for an arbitrary state dimension . Part II specialises the skeleton to the reversible/dissipative split of compartmental kinetics, recovers the pharmacodynamic dose-response curve as a fixed point, and introduces a single worked pharmacokinetic example that the rest of the article grows compartment by compartment. Part III lifts the skeleton from a matrix to a tensor, the setting of a QSP reaction network. Part IV puts every tool to work at the scale QSP demands and closes with a numeric verdict on when an added compartment adds dynamical structure and when it adds only a direction no measurement can see.
The reader needs linear algebra: vector spaces, matrices, rank, eigenvalues and the ordinary differential equations built from them. No pharmacology is assumed. Every pharmacological term (starting with compartment above) is defined in prose where it first does mathematical work.
Part I. Control theory before physics
I.1. State-space object
The three literatures the introduction named are one object once its state dimension is left free. This section fixes that object, first in the nonlinear generality that covers PK, PD and QSP, then in the linear form the rest of Part I uses.
A state-space system with state dimension , input dimension , and output dimension consists of a state , a control input , and a measured output , related by
for maps and , not assumed linear. The integers , and are arbitrary and fixed for a given system.
The linear state-space system specialises [eq:state-eq-general] to
with system matrices , , , and . Written entrywise, with the entry of in row and column for ,
In the compartmental reading that occupies the rest of the article, is the concentration of drug in compartment , the absolute amount (mass or moles) divided by a constant volume of distribution . Section I.6 derives this from conservation of mass. is the administered input (an infusion rate, a bolus, a dosing schedule), and is whatever a laboratory can measure, almost always a small number of compartments against a state dimension that may be large. The off-diagonal entry for is the rate at which drug leaves compartment for compartment ; the diagonal entry collects everything that leaves compartment by any route, including out of the system entirely (elimination, whose rate constant times the volume of distribution is the clearance, the volume of a reference fluid fully cleared of drug per unit time). Every off-diagonal entry is nonnegative and the diagonal absorbs the total outflow. Section I.6 shows that this sign structure is common to every compartmental system at every and derives its consequence for the sign of the whole flow.
plays no role in most compartmental models, since there is rarely a direct algebraic path from dose rate to a measured concentration that bypasses the state. We keep it throughout because Section I.4 and Section I.5 need to say precisely what it does and does not affect. Sections I.2 through I.7 develop the linear specialisation [eq:state-eq-linear] in full; Part III returns to the fully nonlinear [eq:state-eq-general] for QSP networks.

I.2. Matrix exponential and superposition
Every structural property of [eq:state-eq-linear] is a statement about its trajectory. The linear system has a closed-form solution, and every section from Section I.3 onward is a statement about that formula.
For and , the matrix exponential is the convergent series
an element of for every , satisfying and .
The linear state-space system [eq:state-eq-linear] with initial state has the unique solution
for .
Proof
Differentiate the proposed using Leibniz's rule and :
The proposed satisfies and (the integral term vanishes at ), and since is globally Lipschitz in , the Picard-Lindelöf theorem gives uniqueness.
The solution in [eq:variation-of-parameters-eq] splits the trajectory into a free response , the state relaxing (or not, Section I.5) on its own, and a forced response, the convolution of the input against the same exponential kernel. Superposition (the forced response to a sum of inputs is the sum of the forced responses to each) follows from linearity of the integral. It lets pharmacokinetics reduce every dosing regimen (bolus, infusion, repeated dosing) to a sum or convolution of exponentials at any compartment count . Evaluated numerically, the same formula produces the concentration-time curves of Section II.4.
I.3. Controllability
The free response of Section I.2 describes what the state does with no input at all. Controllability asks which states the input can reach and reduces that question to a finite rank computation.[1]
The pair is controllable if for every pair of states and every horizon there exists an input such that the solution of with satisfies . Equivalently, by linearity, taking and arbitrary: every state is reachable from the origin in finite time.
The Kalman controllability matrix is
The pair is controllable if and only if . Equivalently, with the controllability Gramian over a horizon defined as
is controllable if and only if for some, equivalently every, . When every eigenvalue of has negative real part, the limit exists and is the unique solution of the Lyapunov equation
and is controllable if and only if this .
Proof
Fix . First, , so : if , i.e. for , then by the Cayley-Hamilton theorem every higher power is a linear combination of , so for every and hence
so . Conversely, if , the same identity forces on ; differentiating times at gives for , i.e. .
Second, implies controllability, by explicit construction: given any , the input steers to exactly, since substituting into [variation-of-parameters] gives
Third, controllability implies , by contrapositive: if is singular there is a nonzero with on (by the first paragraph), so every reachable state satisfies for every input : the reachable set from is confined to a proper affine hyperplane of and the system is uncontrollable.
The three paragraphs close the equivalence for the time-invariant systems of this article. Kalman's argument covers the time-varying case by the same construction with a time-dependent Gramian.[1]
Controllability, as characterised by [eq:kalman-c-matrix] or [eq:lyapunov-c], is a property of the pair alone. It makes no reference to , or any output. It asks only whether the input can steer the state anywhere in . Section II.5 computes the Gramian explicitly with numbers.
I.4. Observability
Controllability asks what the input can do to the state; observability asks what the output reveals about the state. The construction is dual and Section I.3's theorem transposes without change.
The pair is observable if, for every horizon , the zero input together with the resulting output trajectory for determines the initial state uniquely; equivalently, no two distinct initial states produce the same zero-input output trajectory.
The observability matrix is
The pair is observable if and only if , if and only if the observability Gramian
is positive definite for some, equivalently every, . When every eigenvalue of has negative real part, the infinite-horizon limit solves
and is observable if and only if this . Every statement in this section is Section I.3's, applied to the pair in place of : is observable if and only if is controllable. , like , is computed explicitly in Section II.5.
Proof
By definition is observable exactly when no nonzero initial state is indistinguishable from the origin under zero input, i.e. when for all forces ; transposing, for all forcing is exactly [kalman-rank-c]'s hypothesis (with the roles of state and steering direction exchanged) applied to the pair in place of . So is observable if and only if is controllable, and every clause of the theorem (the rank of , the positivity of , the Lyapunov equation [eq:lyapunov-o]) is [kalman-rank-c]'s corresponding clause for , transposed back: for the transposed pair, and solves [eq:lyapunov-o] exactly as solves [eq:lyapunov-c] for . No new argument is needed beyond [kalman-rank-c]'s proof, read through this transpose.
Controllability and observability, as characterised above, do not depend on . enters only the algebraic term added to the output at each instant , with no memory and no dynamics: given a measured output and a known input , the quantity is available before is reconstructed from . can always be subtracted first. It plays no role in Section I.3 or in this section and none in internal stability, as Section I.5 makes precise.
I.5. BIBO stability
Every dosing regimen a clinician administers is bounded: a finite infusion rate and a finite number of finite doses. BIBO stability asks whether boundedness alone guarantees a bounded measured response. This is weaker than asking whether the state itself decays.
The linear system [eq:state-eq-linear], with , is BIBO stable (bounded-input, bounded-output stable) if every input bounded in the sense produces an output bounded in the sense .
The transfer function is the Laplace-domain input-output map. Since , every signal here is at rest before , so the relevant transform is the unilateral (one-sided) Laplace transform,
with a complex frequency variable, the integral understood for large enough that it converges. Taking Laplace transforms of [eq:state-eq-linear] at zero initial state,
is the instantaneous, memoryless part of the input-output map and the only term of that remains as .
Let denote the restriction of to the subspace that is both controllable (Section I.3) and observable (Section I.4), the minimal-order realisation obtained from by cancelling every uncontrollable or unobservable mode; equivalently, the poles of in [eq:transfer-function] are exactly the eigenvalues of . The system is BIBO stable if and only if the spectral abscissa of ,
satisfies : every pole of the transfer function has strictly negative real part.
Proof
Write the minimal realisation's impulse response as for , so that whenever .
If , every entry of is, by the Jordan form of , a finite sum of terms with , each absolutely integrable on , so . For at every ,
for every , so the system is BIBO stable.
Conversely, if some eigenvalue of has , minimality of means the mode at is both controllable and observable, so some bounded input tuned to that mode (a constant if , a sinusoid at frequency if ) drives the corresponding component of unbounded by resonance. If the mode grows without bound under any input that excites it. In both cases a bounded input produces an unbounded output and the system is BIBO unstable.
This is weaker than internal (asymptotic) stability of , which requires every eigenvalue of the full to have negative real part. A mode that is uncontrollable, unobservable, or both is invisible to . It never appears in the measured output and its eigenvalue cancels out of [eq:transfer-function] regardless of its sign. A system can therefore be BIBO stable and contain an internally unstable mode that no input reaches and no output sees. Section I.7 and Part IV return to this gap between what decays and what is seen to decay. Section II.3 shows that this criterion collapses to a single positivity check once the reversible/dissipative split of Section II.2 is in place.
I.6. Metzler matrices and positivity
Compartmental kinetics is linear with a sign pattern, since mass or concentration can neither go negative nor be created from nothing. This section isolates that sign pattern as a property of alone, derives it from conservation of mass, and shows that it forces positivity on the whole flow for every .
A matrix is Metzler if every off-diagonal entry is nonnegative,
Let be the absolute amount (mass or moles) of drug in compartment , related to the concentration by for a constant volume of distribution . Transport between compartments neither creates nor destroys drug, so the amounts obey their own linear system for a matrix whose column sums by bookkeeping alone are exactly the negative of whatever compartment eliminates,
for the elimination rate constant of compartment (zero exactly when compartment only exchanges and never itself eliminates). Writing , gives : is Metzler exactly when is, and [eq:amount-mass-balance] becomes the volume-weighted mass balance
which collapses to the unweighted
that Section II.4 onward states and checks, exactly when every compartment shares one volume , the convention of the rest of the article. and are similar matrices, hence share every eigenvalue: nothing from Section I.5 onward depends on which of the two is used.
Proof
Write entrywise, ; since , and share a sign for every , so is Metzler if and only if is. Multiplying [eq:amount-mass-balance] through by and substituting termwise gives exactly [eq:weighted-mass-balance], which reduces to [eq:mass-balance] on setting every equal. Similarity of and is immediate from with invertible.
Column of collects every rate at which compartment loses concentration: to other compartments (the off-diagonal entries, each nonnegative by [eq:metzler-def]) and out of the system entirely (elimination, the excess by which [eq:mass-balance] falls short of equality). Equality in [eq:mass-balance] for a given column means compartment only exchanges with other compartments and is never itself a route out of the system. Section II.2 isolates this reversible, non-eliminating structure algebraically.
If is Metzler, then is entrywise nonnegative for every . Consequently, if is entrywise nonnegative, the input is entrywise nonnegative for all , and the initial state is entrywise nonnegative, then the trajectory of [eq:variation-of-parameters-eq] is entrywise nonnegative for every . A nonnegative dose administered to a compartmental system from a nonnegative state never produces a negative amount anywhere in the system at any compartment count . This is a consequence of the sign pattern in [eq:metzler-def] alone and is independent of the numerical values of the rate constants.
Proof
Choose , so that every diagonal entry of satisfies ; combined with [eq:metzler-def], is entrywise nonnegative. Its matrix exponential is a nonnegative combination of nonnegative matrices,
hence entrywise nonnegative for . Since commutes with , , and is a positive scalar, so is entrywise nonnegative for every . Nonnegativity of then follows termwise from [eq:variation-of-parameters-eq]: is a nonnegative matrix applied to a nonnegative vector, and the integrand is a product of entrywise nonnegative factors at every , so the integral is entrywise nonnegative.
I.7. Structural (generic) observability
The rank test of Section I.4 is decisive, but it needs numbers: a specific and to form and check its rank. At the compartment counts of Part IV forming that matrix is undesirable and its rank is fragile: a coincidence among the rate constants can make a system look observable (or not) in a way that a small parameter change would reverse. This section replaces the numerical test with a combinatorial one that depends only on which entries of and may be nonzero.
Fix only the sparsity pattern of and : which entries and (for ) are free parameters and which are structurally zero. Such a pair is a structured pair , and a numerical realisation is any choice of real values for its free entries. Associate to a structured pair the directed graph with a vertex for each state coordinate and each output coordinate , an edge whenever is a free parameter, and an edge whenever is a free parameter. is exactly the compartment diagram: an arrow for every rate constant the modeller has allowed to be nonzero, plus an arrow to every sensor.
A structured pair is structurally observable, meaning observable (Section I.4) for every choice of its free parameters outside a proper algebraic subset (a set of Lebesgue measure zero: the exceptional non-observable parameter choices are a knife-edge coincidence), if and only if both of the following hold in :
- Output connectivity. Every state vertex has a directed path in to some output vertex .
- Saturating matching. The bipartite graph with left vertices and right vertices , with an edge from left to right whenever is free and to right whenever is free, admits a matching of size (one that saturates every left vertex).
This is the dual of Lin's structural controllability theorem under the transpose of Section I.4. Lin proved the single-input case and Shields and Pearson (1976) the multi-input case. It holds for every state dimension and every output dimension and uses no numerical rate constant.[2]
Proof
(Necessity.) If condition 1 fails, some state vertex has no directed path to any output vertex in . The entry is a sum over walks of length from to , so column of every vanishes identically in the free parameters and for every numerical choice. If condition 2 fails, the bipartite graph admits no matching saturating every left vertex. Its maximum matching then has size strictly less than , and the generic rank of a matrix with a given zero pattern equals its maximum bipartite matching size, so the stacked matrix has rank less than for every numerical choice. A nonzero with and lies in the kernel of every , so again .
(Sufficiency, sketch.) Treat every free entry of and as an independent indeterminate. Every minor of is then a polynomial in the free parameters. If one numerical realisation is observable, some minor is nonzero there and hence is not the zero polynomial, and a polynomial that is not identically zero vanishes only on a proper algebraic subset of parameter space and so on a set of Lebesgue measure zero. It remains to exhibit one observable realisation when both conditions hold. By the Hautus test, observability fails exactly when drops rank at some : condition 2 excludes a generic drop at , and the two conditions together exclude one at . Lin's argument covers the graph by disjoint cycles and output-rooted paths and builds a realisation with no drop at any .[2]
Condition 1 excludes a state coordinate that no measurement can reach however the free parameters are chosen. Condition 2 excludes two or more state coordinates competing for the same limited routes to the outputs, a structural degeneracy that no numerical coincidence fixes generically. Both conditions are decidable in time polynomial in (the Hopcroft-Karp algorithm finds a saturating matching in a bipartite graph). Section IV.2 and Part IV run this test at a compartment count where forming numerically is the last resort. The same combinatorial test extends unchanged to the tensor networks of Section III.1, where the wiring diagram is a reaction network.
II.1. Fragmentation of the fields
The introduction opened on three literatures that read as unrelated. Sections I.1 through I.7 have built one object at one level of generality that contains all three. We can now say which fragment of it each field uses.
Pharmacokinetics uses Section I.1's linear specialisation and Section I.2's exponential solution almost always at a small fixed compartment count. It relies on Section I.6's positivity constantly, since a negative concentration is meaningless. It rarely names Metzler matrices or states the constraint as a theorem. The sign pattern is enforced by writing rate constants as manifestly nonnegative symbols and no argument shows that concentrations stay nonnegative for every input. Controllability (Section I.3) is rarely asked at all, since the clinical protocol fixes the dosing route. Observability (Section I.4) is asked implicitly as the question of which compartment can be sampled, but rarely by name and never structurally (Section I.7).
The introduction noted that pharmacodynamics is typically posed with no dynamical system. Section II.6 shows that its central object, the Hill or Emax dose-response curve, is the output map evaluated at a fixed point of a one-compartment nonlinear flow: the missing differential equation is present all along, solved at equilibrium before pharmacodynamics ever states it.
QSP uses the fully nonlinear generalisation of [eq:state-eq-general] at the large state dimension the other two fields avoid and needs every tool built in this Part: Metzler positivity generalised to nonlinear mass-action kinetics, and observability generalised from a rank condition on a matrix to a rank condition on Lie derivatives (Section III.3). Historically it built this machinery field by field and model by model. Part III and Part IV import it from control theory as a settled body of results.
Each field solved the problem in front of it. Historical accident left the common object underneath unnamed. Part II names it and starts from the sign structure that Section I.6 exposed.
Part II. Pharmacokinetics as a reversible/dissipative split
II.2. Reversible/irreversible split
Section I.6 isolated a sign pattern in , off-diagonal entries never negative, and showed it alone forces the whole flow nonnegative. That sign pattern does not yet distinguish two processes pharmacokinetics treats as different in kind: drug moving between compartments and drug leaving the body outright. Non-equilibrium thermodynamics has an algebra for this distinction and it applies to unconditionally, before any numbers are chosen.
Every decomposes uniquely into an antisymmetric part and a symmetric part,
entrywise and for all . This holds for every square real matrix and needs no hypothesis on . admits a GENERIC split when, in addition, the symmetric part is positive semi-definite, ; call the reversible (or conservative) part and the irreversible (or dissipative) part. The name is borrowed from Grmela and Öttinger's GENERIC formalism (General Equation for Non-Equilibrium Reversible-Irreversible Coupling) for non-equilibrium thermodynamics[3]. Section II.3 shows what the borrowing gives.
The extra content in [eq:generic-split-def] is entirely the sign condition . It is not automatic for a compartmental and must be checked case by case. Section II.4 checks it for the first worked example below. The two parts differ. alone generates a flow that leaves the Euclidean norm exactly unchanged, a pure circulation with no preferred direction of relaxation. alone generates a flow that never increases , pure relaxation towards the origin. Section II.3 proves both claims at once.
The reversible part has a name in another language too. Antisymmetric matrices are exactly the generators of rotations: under the standard identification of antisymmetric matrices with bivectors (grade-2 elements of a geometric algebra, which form a space of dimension ), corresponds to a bivector , and is exactly the rotor sandwich for , in the sense this site's geometric-algebra treatment of classical mechanics develops in full for rigid-body rotation. Section II.4's is the simplest possible instance: a single bivector generating planar rotation at rate , with exactly the ordinary rotation matrix at that rate.
Every eigenvalue of is purely imaginary. Its resolvent therefore has every pole on the imaginary axis of the Laplace variable ([eq:laplace-def]): driven by the reversible part alone, the system oscillates forever at the rotor's own rate with no decay, as the absence of dissipation in above shows.
Proof
is orthogonal for every , since using ; and since an antisymmetric matrix is traceless. This excludes the reflection branch of the orthogonal group. For the spectrum, let for a nonzero (possibly complex) eigenvector and let denote its conjugate transpose:
the second line using for real . Since , the two expressions for force , so is purely imaginary.
The GENERIC term "reversible" differs from the everyday pharmacological sense of a route that runs both ways. A compartmental system whose micro-rate-constants happen to be symmetric, for every pair , has symmetric, so and the entire flow is irreversible in the GENERIC sense even though every route in it runs both ways. Symmetric exchange is the linear analogue of diffusion, the paradigm case of a process that produces entropy. isolates the asymmetry between the forward and backward rates of an exchanging pair of compartments (one pair suffices, as Section II.4's shows) and is the only part of a linear compartmental system that can fail to dissipate.
II.3. BIBO stability as a corollary
Section I.5's criterion asks for the spectral abscissa of a restricted matrix . This computation is redone whenever changes. The GENERIC split of Section II.2 gives a sufficient condition for the same conclusion that mentions no eigenvalue and is phrased in terms of dissipation.
Suppose admits a GENERIC split ([generic-split]) with , and suppose further that the only subspace of invariant under and contained in is (connectivity: no direction of state space escapes dissipation entirely and indefinitely). Then every eigenvalue of has strictly negative real part, and the system is BIBO stable by [bibo-criterion].
Proof
Let , a free-energy-like function, quadratic and bounded below. Along a trajectory of ,
is a scalar, hence equal to its own transpose, and since is antisymmetric; a quantity equal to its own negation is zero, so for every . [eq:lyapunov-generic] collapses to , since : never increases along any trajectory and the system is Lyapunov stable. Equality at a state means , which for symmetric positive semi-definite forces , i.e. . By the connectivity hypothesis, the largest subset of invariant under the flow is ; LaSalle's invariance principle upgrades Lyapunov stability to asymptotic stability. Every trajectory converges to the origin, so every eigenvalue of has negative real part. [bibo-criterion] concludes BIBO stability immediately, since asymptotic stability of the full is strictly stronger than its hypothesis on the controllable-observable restriction alone.
For a compartmental system the connectivity hypothesis has a reading from Section I.6: it fails only when some subset of compartments exchanges among itself forever without reaching a compartment whose column sum in [eq:mass-balance] is a strict inequality, an isolated conservative pocket with no route to elimination. Every compartment diagram in which every compartment has a route to elimination, direct or indirect, satisfies the hypothesis. Section II.4 checks the hypotheses of this proposition on numbers and finds that is strictly positive definite. Connectivity then needs no separate check.
II.4. Worked example for k=2
Every tool built in Part I and specialised in Sections II.2 and II.3 has so far been stated for an arbitrary , , . This section fixes the smallest case pharmacokinetics uses, a central compartment (blood or plasma, the compartment a dose enters and a laboratory samples) exchanging with a single peripheral compartment (a tissue pool with no direct access to either the dose or the assay), and computes every quantity defined so far on numbers.
Let be the central-compartment concentration and the peripheral-compartment concentration, with micro-rate-constants (central to peripheral transfer), (peripheral to central return), and (elimination from the central compartment only). The mass-balance equations are
an infusion or bolus entering the central compartment and a single assay reading the central-compartment concentration directly. In the notation of [eq:state-eq-linear],
is Metzler ([metzler]): its only off-diagonal entries, and , are both nonnegative. Its columns sum to and . The first is strictly negative and is the elimination rate leaving through the central compartment. The second is exactly zero, the equality case of Section I.6 for a compartment that only exchanges and never eliminates.
The eigenvalues of follow from its characteristic polynomial, . Here and . This gives with discriminant , so
Both eigenvalues are real and strictly negative: alone is internally asymptotically stable.
Section II.2's split, applied to [eq:pk-k2-matrices], gives
's characteristic polynomial is (, ), with discriminant , so
Both are strictly positive, so is positive definite. This is stronger than the that [generic-split] requires and that [bibo-generic] needs. Connectivity is automatic since . [bibo-generic] reconfirms the conclusion of [eq:pk-k2-eigenvalues] by a different route through dissipation.

II.5. Controllability and observability Gramians for the k=2 example
Section I.3 and Section I.4 stated the Kalman rank tests and the Gramians for an arbitrary and . This section forms both for [eq:pk-k2-matrices] explicitly and reads what the numbers say about the peripheral compartment that nothing measures directly.
The Kalman controllability matrix ([eq:kalman-c-matrix]) is ; since is the first column of ,
so and is controllable ([kalman-rank-c]). The observability matrix ([eq:kalman-o-matrix]) is ; since is the first row of ,
so and is observable ([kalman-rank-o]).
Equivalently, the Gramians solve the Lyapunov equations [eq:lyapunov-c] and [eq:lyapunov-o] directly. Writing , expands entrywise to three equations, , , , with unique solution , , :
Writing , expands to , , , with unique solution , , :
Both are positive definite (, , both with positive trace). The Gramian criterion of [kalman-rank-c] and [kalman-rank-o] agrees with the rank computation above.
The peripheral compartment is never sampled and reads only , yet says is uniquely recoverable from alone over any positive time window. The mechanism is visible in [eq:pk-k2-kalman-o] itself: the second entry of is , nonzero because feeds into and leaves an imprint on the slope of the measured trajectory that a single snapshot of cannot contain. Full rank is the outcome that the structural test of Section I.7 predicts for any nonzero . It is the generic case and holds for every choice of the free rate constants outside the single knife-edge , which severs 's only route back to a sampled compartment.

II.6. PD as a gradient flow
The introduction promised that the pharmacodynamic dose-response curve is a fixed point of a flow. This section shows it for the simplest receptor-binding mechanism: one drug, one receptor and one bound complex.
Let be the concentration of drug-receptor complex, the total receptor abundance, and , the binding and unbinding rate constants of the mass-action reaction , where is the free-drug concentration of Section II.4's central compartment, now an input into this second state:
For fixed [eq:receptor-binding-ode] has the unique fixed point
and the relaxation towards it is the gradient flow of the quadratic potential . Identifying and makes [eq:receptor-fixed-point] the standard Hill/Emax law at Hill coefficient ,
Proof
For fixed [eq:receptor-binding-ode] is a scalar linear-affine equation in , . Setting and solving gives the unique fixed point [eq:receptor-fixed-point]. Writing the equation in terms of the fixed point,
since by [eq:receptor-fixed-point], exhibits the flow as for , whose unique minimum sits exactly at [eq:receptor-fixed-point] and whose curvature is exactly the relaxation rate. Substituting , into [eq:receptor-fixed-point] gives [eq:hill-emax-law].
[eq:receptor-binding-ode] is itself a one-dimensional instance of Section II.2's split: a matrix has no antisymmetric part, so and sets the whole dynamics. Receptor-binding relaxation is the fully dissipative endpoint of the decomposition that gave the two-compartment flow of Section II.4 both a rotational and a contracting part. A single bound complex has no second coordinate for a rotation.
With , , (so ), [eq:hill-emax-law] reads , . At the half-maximal concentration as the name demands. At above the value lies strictly between and as [eq:hill-emax-law] requires.
[eq:hill-emax-law] at is a rectangular hyperbola in : saturating, monotonically increasing and concave throughout. The S-shaped curve reported for Hill coefficients needs cooperative multi-site binding that this single-site mechanism does not model. This is the fixed point that Section II.1 promised. The four-parameter curve fitted in the pharmacodynamics literature is this equilibrium, reached without writing down [eq:receptor-binding-ode].


II.7. First look at scale
Section II.6's receptor-binding state was kept separate from Section II.4's two-compartment flow: entered [eq:receptor-binding-ode] only as an input and never as a coupled state. Once receptor engagement is tracked as dynamics the separation ends: is one three-coordinate state-space system, , with coupled to exactly as is. This is the move that the second anchor of the introduction anticipated: each mechanism added to a model is one more state coordinate coupled to those already built, and repeating this dozens of times yields a QSP network.
Section III.5 carries out this construction with its own rate constants and its own Gramians, in the same style as Sections II.4 and II.5, and nothing here anticipates its numbers. What can already be said without them is qualitative. stays the only sampled coordinate. , like , leaves an imprint on and remains observable by the structural argument Section II.5 made for . A binding rate constant can differ in order of magnitude from an inter-compartmental exchange rate, and a rank computation does not show the mismatch: stays regardless of how weak 's footprint on becomes, while the observability Gramian's smallest eigenvalue shrinks continuously as that footprint weakens. Whether a coordinate is observable and how well it is observable are different questions, and only the first has a rank test. Part IV answers the second at a compartment count where the answer is no longer visible by inspection.
Part III. QSP networks as tensors
III.1. From matrix to tensor
Every reaction network built so far has been linear: each rate constant governs an exchange between exactly two coordinates, or one coordinate and the outside world, and the whole flow is generated by acting once with a fixed matrix on the state. A QSP network is rarely built entirely from such reactions. A drug binding a receptor to form a complex, the paradigm reaction for the rest of this Part, has a rate proportional to the product of two concentrations. No matrix acting on can express a rate law of this shape because the rate is quadratic in the state.
Reaction network theory has the bookkeeping this needs. A network of species undergoing elementary reactions is fixed by its stoichiometric matrix , whose th column records the net change in each species per firing of reaction , together with a rate vector giving each reaction's instantaneous rate as a function of the current state. The dynamics of every mass-action network, whatever the order of its reactions, take the single form . The order of the reactions changes only the shape of .
A reaction has order (or molecularity) when mass action assigns it a rate proportional to a product of reactant concentrations, counted with multiplicity: for the unimolecular exchanges and eliminations of Part I and Part II, for a bimolecular reaction such as ligand binding receptor, for the rarer termolecular case. Collecting the net rate of production of species contributed by every order- reaction defines the order- rate tensor , symmetric under any permutation of its reactant indices , and the dynamics decompose additively by order,
with the highest molecularity occurring anywhere in the network, finite and almost always . At , is entrywise exactly Section I.1's matrix, for the matrix of first-order rate constants; every tool built for arbitrary in Part I and specialised in Part II is a statement about [eq:tensor-dynamics].
The sign structure that Section I.6 isolated extends beyond . Mass action assigns every rate constant a nonnegative value, and every reaction removes mass from exactly the species that supply it and adds it to exactly the species it produces. Entrywise, this forces every entry with to be nonnegative. This is the tensor generalisation of the off-diagonal sign condition of [metzler]. The series argument that proved [perron-frobenius-positivity] at needs a matrix and does not transfer. A boundary argument replaces it: on the face of every term of [eq:tensor-dynamics] with among its reactant indices vanishes and every other term is nonnegative, so there and no trajectory crosses the face. The flow of [eq:tensor-dynamics] never leaves at any order.
The structural test of Section I.7 reads the sparsity pattern of a linear pair as a directed graph on compartments and sensors. The same reading applies unchanged to a tensor network, with each nonzero entry of [eq:tensor-dynamics] read as an edge whatever its number of reactant indices. The sparsity figure below previews this at the scale that Part IV needs, before any of that Part's numbers exist.

III.2. GENERIC bracket formalism
Section II.2's split was built from a single fixed matrix and does not extend to a flow generated by an order- tensor: there is no longer one object to take an antisymmetric part and a symmetric part of. Non-equilibrium thermodynamics generalises this split to nonlinear state-dependent flows under the name that Section II.2 borrowed.
Let be two smooth functionals of the state, energy and entropy, and let and be matrices depending smoothly on , with (antisymmetric, and such that the bracket below satisfies the Jacobi identity) and (symmetric positive semi-definite). For smooth functionals of the state, define the Poisson bracket and the dissipative bracket . Every functional evolves along the flow by
so that taking recovers the state equation , subject to the degeneracy conditions
The two degeneracy conditions in [eq:generic-degeneracy] are exactly what makes [eq:generic-bracket-evolution] thermodynamically consistent, and the argument is the one-line collapse that [bibo-generic] used for a single antisymmetric matrix. Taking , is a scalar equal to its own negation ( antisymmetric), hence zero, and by the second degeneracy condition. Hence and energy is conserved exactly along every trajectory for any choice of . Taking , by the first degeneracy condition, using 's antisymmetry to move it across the inner product, and since . Hence and entropy never decreases. Neither law of thermodynamics is assumed in [generic-bracket]. Both follow from antisymmetry, positive semi-definiteness and the two degeneracy conditions alone.
and generalise and exactly as Section III.1's tensor generalises Section I.1's matrix: an antisymmetric generator of conservative circulation and a symmetric positive semi-definite generator of monotone relaxation, now allowed to depend on the state and to act through two independent functionals. The purely linear setting of Sections II.2 and II.3 is the furthest reduction of this structure: a single quadratic functional stands in for two independent functionals and , and constant matrices stand in for state-dependent . With one functional in play neither degeneracy condition in [eq:generic-degeneracy] has content, and the direct Lyapunov argument of [bibo-generic] does the work in place of the bracket algebra above. That argument is the , single-functional case of the same fact.
III.3. Nonlinear controllability and observability
Sections I.3 and I.4 built controllability and observability by iterating a single linear map, , and stacking the results into a rank condition. Once the flow is nonlinear, as the case of Section III.1 forces, no linear map is left to iterate. "Apply again" is replaced by an operation that manufactures new directions from nonlinear vector fields, and control theory supplies it as the Lie bracket.
For smooth vector fields , their Lie bracket is ( the Jacobian matrices of ), the infinitesimal difference between flowing along then and flowing along then : a direction unavailable to either field alone whenever it is nonzero. For a control-affine system (drift , control directions ), the accessibility distribution at is the span, evaluated at , of every vector field obtainable from by repeated Lie bracketing.
If the accessibility distribution of has dimension at , the system is locally accessible at : the states reachable from in arbitrarily small time contain a nonempty open set. When and each is the constant field (the th column of ), every iterated Lie bracket collapses to iterated application of , and so on, and at , where itself vanishes, the accessibility distribution is spanned exactly by : the rank condition above is [kalman-rank-c]'s Kalman condition on , recovered as its case.
Dually, for , , define the Lie derivative of along by and , the rate of change of the output and of its own rate of change along the flow. If
the system is locally (weakly) observable at . When , , identically, so for every , and the differentials in [eq:nonlinear-obs-rank] stack into exactly [eq:kalman-o-matrix]'s : [kalman-rank-o] is this theorem's case.
Proof
The reductions stated above are direct computation: when , constant, since (constant field) and ; iterating, every bracket of against a constant field reduces to one more application of , so the accessibility distribution at the origin, where drops out of the span, is exactly by Cayley-Hamilton (higher powers add nothing new), which has dimension exactly when . Dually, collapses by induction to since and each Lie derivative differentiates once more along , so identically and [eq:nonlinear-obs-rank] is [eq:kalman-o-matrix]'s stacked by rows.
The general nonlinear statements, that a -dimensional accessibility distribution gives local accessibility and that the rank condition [eq:nonlinear-obs-rank] gives local weak observability, are theorems of nonlinear geometric control theory in their own right, resting on the Frobenius theorem for the first and on an inverse-function argument for the second; neither reduces to linear algebra once , and both are proved in full generality by Hermann and Krener[4], who build on the accessibility theory of Sussmann and Jurdjevic.[5]
[eq:nonlinear-obs-rank] is, entrywise, the tool that structural identifiability analysis of a real QSP model runs: a state coordinate or parameter direction is structurally unidentifiable exactly when it lies in the common kernel of every , however many derivatives are taken. Software auditing a nonlinear model before it is fit to data checks this condition by differential-algebraic elimination. The tensor structure of Section III.1 keeps computable for beyond the first two or three. Without it the symbolic expressions in a network of any size explode long before a rank can be read off. Section IV.3 therefore replaces this symbolic computation with a numerical surrogate once grows large enough.
III.4. One geometric remark
Section III.1's stoichiometric matrix constrains independently of the nonlinearity of . The constraint has a name in reaction network theory. It is stated once before the article returns to ordinary vector-space language.
Whatever is, always lies in the column space of , the stoichiometric subspace , so for every . Equivalently, a trajectory starting at is confined for all time to the coset , an invariant affine submanifold of that Feinberg calls a stoichiometric compatibility class[6], cut out by the conservation laws for every in the left null space of : a total that moves between species and is conserved, such as free plus bound receptor, which Section III.5 fixes at .
This is the only point in the article at which the word submanifold or any of its usual apparatus (a chart, a tangent bundle, coordinate-free notation), appears. It appears because it names an object implicit in every column-sum computation since [metzler]: a stoichiometric compatibility class is the affine set to which the mass-balance column sums confine the flow, made explicit once nonlinearity rules out writing that confinement as a single fixed matrix identity. Nothing introduced in this remark is used again. Every section before and after it works within and its ordinary vector-space structure.
III.5. Worked example extended to k=3
Section II.7 promised a third coordinate coupled to the flow and called it . This section builds it, gives it its own rate constants and calls it . The mechanism is target-mediated drug disposition (TMDD)[7], the standard QSP extension of the two-compartment model of Section II.4 for a drug whose target is abundant enough or whose affinity is high enough that binding measurably depletes the free-drug pool. Section II.6 read the binding off at equilibrium. The bound complex is a full third state coordinate coupled to by mass action. It is exactly the order-2 term that the tensor formalism of Section III.1 was built to represent.
Extend [eq:pk-k2-odes] by a bound-receptor-complex compartment , with total receptor abundance and binding, unbinding and internalisation rate constants :
with , , unchanged from [eq:pk-k2-matrices], and , , , the new TMDD parameters. Linearising [eq:qsp-k3-x1]-[eq:qsp-k3-x3] about the drug-free steady state (a fixed point of all three equations at once, since every term on the right is a product involving at least one of ) gives
a bolus dose still entering only the central compartment and the assay still sampling it alone.
A constant is a special case of the Mager–Jusko model[7]. There free receptor has its own turnover, synthesised at rate and degraded at rate constant . Its equation is . Adding [eq:qsp-k3-x3] gives the total the equation . When and the receptor starts at its baseline , the total stays at for all time and recovers [eq:qsp-k3-x1] and [eq:qsp-k3-x3]. The example therefore fixes and . With the free receptor is a fourth state and the linearisation about the drug-free state is unchanged in its block: every term coupling into the other three equations contains a factor , which vanishes there.
The linearisation is a direct Jacobian computation. Differentiating [eq:qsp-k3-x1] and evaluating at , reduces at to ; everywhere; reduces at to . [eq:qsp-k3-x2] is already linear and contributes the row unchanged by linearisation. Differentiating [eq:qsp-k3-x3], reduces at to , and reduces at to ; every entry of [eq:qsp-k3-matrices] is accounted for. In the language of Section III.1 the term appears with coefficient in [eq:qsp-k3-x1] and in [eq:qsp-k3-x3]. It is an order-2 rate tensor contribution that couples two reactant indices. A linearisation discards it by construction: has only the order-1 part of the true dynamics and is valid near only.
's off-diagonal entries, (and two structural zeros), are all nonnegative: is Metzler ([metzler]). Its column sums are , , and , exactly . This follows from mass balance: every exchange between and contributes once with each sign to the first two columns, and every binding/unbinding step between and contributes once with each sign to the first and third. Both cancel in the column sums, column 1 collapsing to and column 3 to . Only the two true eliminations, drug elimination and complex internalisation , remove mass from the system. Mass conservation forces the column sums to equal independently of the values of , and the computed matrix confirms this.
's characteristic polynomial has , sum of principal minors , and . Hence
All three roots are real and strictly negative: the linearised network is asymptotically stable at the drug-free state. This extends the conclusion of Section II.4 to .
Section II.2's split, applied to the linearisation [eq:qsp-k3-matrices] exactly as it was to [eq:pk-k2-matrices] in Section II.4, gives
's characteristic polynomial is (, the sum of its principal minors , ), with roots
All three are strictly positive, so is positive definite. This extends the conclusion of Section II.4 to and reconfirms the stability verdict of [eq:qsp-k3-eigenvalues] by the dissipative route of [bibo-generic], independent of the root-finding above.

The Kalman controllability matrix is built from , the first column of , and . Hence
The determinant expands along the first column. Hence and is controllable ([kalman-rank-c]). Dually, is built from , the first row of , and , the first row of . Hence
The determinant expands along the first row. Hence and is observable ([kalman-rank-o]). Both determinants are nonzero, so both Gramians satisfy by the equivalence of [kalman-rank-c] and [kalman-rank-o]. The promise of Section II.7 of a third compartment with its own controllability and observability certificate is discharged without solving either Lyapunov equation by hand. Section IV.2 confirms the observability half of this conclusion by a different route that reads 's sparsity pattern as a graph ([structural-observability]) before any of these numbers are fixed.


Part IV. Observability at scale
IV.1. Compartment overload
Section III.5 added a third compartment. The bound complex had its own controllability and observability certificate and both Gramians came back positive definite. This discharged the promise of Section II.7 of a coordinate that adds structure. A real QSP model reaches dozens of coordinates because each mechanism it includes becomes one more state (the second anchor of the introduction). This Part asks what happens to the certificate as that count grows. Each new compartment either adds dynamical structure, a direction that the single sampled trajectory can resolve, or adds only a direction that the data cannot separate from those already present. We call the second outcome compartment overload. The rest of Part IV builds the machinery that tells the two apart at a scale where the determinant of [kalman-rank-o] cannot be read off by hand and then runs it on a family of networks grown from that of Section III.5 until overload appears.
IV.2. Applied structural observability
Before any Gramian is formed, the combinatorial test of Section I.7 runs on the wiring diagram alone. It is the first screen because it needs no numerics: a network that fails it is unobservable for every choice of rate constants and no Lyapunov equation needs solving. We run it on the network of Section III.5, whose graph is read off the sparsity pattern of [eq:qsp-k3-matrices].
The nonzero off-diagonal entries of in [eq:qsp-k3-matrices] are . They give the directed edges , (from , each free entry reading as the edge ), and . The single sensor adds the edge . This is exactly the bidirectional network drawn in Section III.5's figure, now read as the graph of a structured pair ([structural-observability]).
Output connectivity. is direct. and route both peripheral coordinates back to the sensor through the central compartment they each feed. Every state vertex reaches the output and condition 1 holds.
Saturating matching. With every diagonal entry nonzero, each left vertex already has a self-edge to the right vertex , so the identity is a matching of size . A matching that uses the sensor, , saturates all three left vertices as well. Condition 2 holds.
Both conditions hold, so the pair is structurally observable: observable for every choice of its rate constants outside a measure-zero exceptional set, and the full rank of [eq:qsp-k3-kalman-o] is the generic case. The network passes the screen before any Gramian is formed, as a well-posed model should, and proceeds to the finer value-dependent questions that the screen cannot answer. The screen sees only whether a route to the sensor exists. It is blind to whether two coordinates take near-identical routes, the failure that Section IV.6 engineers and only a Gramian detects.
IV.3. Empirical Gramians
Section III.3's Lie-derivative rank test is exact and symbolic, and it stops being computable well before a real QSP network's compartment count: the differentials in [eq:nonlinear-obs-rank] grow in size with each order taken. The differential-algebraic elimination that reads a rank off them fails on the resulting expressions once is large and is nonlinear. The tool that reaches that scale replaces symbolic differentiation with simulation and reduces to the analytic Gramians of Sections I.3 and I.4 when the dynamics are linear.
For the nonlinear system , operated about a steady state , perturb each input channel in turn and record the state response, and perturb each initial-state coordinate in turn and record the output response. The empirical controllability Gramian is the time-integrated covariance of the state trajectories produced by a set of input perturbations,
with the response to the th input perturbation of size ; the empirical observability Gramian is the time-integrated cross-covariance of the outputs produced by perturbing each initial-state direction, entrywise
with the output from the initial state .
When , , each response is a matrix exponential and the integrals in [eq:empirical-Wc]-[eq:empirical-Wo] collapse to the analytic solutions of the Lyapunov equations [eq:lyapunov-c] and [eq:lyapunov-o]: and exactly.
This construction, due to Lall, Marsden and Glavaski[8], is what an observability analysis of a large nonlinear QSP model runs when the symbolic test of Section III.3 is out of reach. It requires only that the model can be simulated. It returns a pair of matrices in the coordinates of the analytic Gramians, so every question the next two sections ask of can be asked of unchanged. The family of Section IV.6 is linearised, so its coincide with the analytic Gramians and the numbers reported there are computed from the Lyapunov equations.
IV.4. Balanced truncation and Hankel singular values
A positive-definite certifies that every initial state is recoverable, and a positive-definite that every state is reachable, but neither verdict says how well. Section II.7 isolated the gap: stays full while the smallest eigenvalue of shrinks continuously as a coordinate's imprint on the sensor weakens. The quantity that grades a coordinate on this continuous scale, and does so in a way no change of state coordinates can distort, combines both Gramians at once.
The Hankel singular values of a stable system are
the square roots of the eigenvalues of the product of the two Gramians, ordered decreasingly. Though and each depend on the choice of state coordinates (under , and ), the eigenvalues of the product are invariant, since is a similarity transform: the are properties of the input-output map alone, independent of how the state is written.
Each grades one input-output direction by how strongly it is jointly reachable from the dose and visible at the sensor: a direction with large is excited hard by the input and imprints hard on the output, and a direction with near zero is one the dose barely excites, or that barely reaches the assay, or both. Balanced truncation uses this grading.
Every stable minimal (controllable and observable) admits a balanced realisation, a choice of state coordinates in which : each coordinate is exactly as reachable as it is observable, and both equal to its Hankel singular value.[9] Truncating this realisation to its first coordinates (those with the largest ) yields a reduced model of order whose input-output map differs from the original in the norm by at most twice the discarded tail,
Proof
(Existence of a balanced realisation.) Minimality gives and , so every and below exists. Write (Cholesky). The matrix is symmetric positive semi-definite, hence diagonalises as for orthogonal and diagonal , and its eigenvalues equal those of (a matrix product and its factors reversed share eigenvalues), so is exactly [eq:hankel-sv]'s Hankel singular values. Take the coordinate change ; using the transformation rule stated after [eq:hankel-sv],
and, since ,
as well. Both Gramians equal in the new coordinates and the realisation is balanced.
The truncation bound [eq:truncation-bound] is a harder estimate on the transfer function of the truncated system and is due to Glover, who builds on Moore's existence result above.[10]
[eq:truncation-bound] is what turns the spectrum [eq:hankel-sv] into a verdict. When the Hankel singular values fall off a cliff, a sharp drop of one or more orders of magnitude between and , the tail beyond the cliff is negligible and the model has an effective order well below its nominal : the extra coordinates have a joint reachability-observability weight that the data cannot resolve. Reading it off the spectrum is the quantitative form of the overload question that Section IV.1 posed. The location of the cliff is the number of directions the single sampled trajectory can resolve. Everything past it is compartment overload. The picture resembles the sloppy models of Gutenkunst et al.[11], whose Fisher eigenvalues for rate-constant parameters spread over many decades; Section V.3 makes the relation precise.
IV.5. Computability
Forming and the way Section II.5 did, by solving the Lyapunov equations [eq:lyapunov-c] and [eq:lyapunov-o] densely takes arithmetic and storage, as much as one dense eigendecomposition. At the of this article the time is negligible. At the compartment counts of a systems-scale QSP model it separates a computation that runs from one that does not, and a scalable solver exploits the sparsity that Section I.6, Section I.7 and Section III.1 established.
When has few columns, as it does here with a single dosing route, the controllability Gramian typically has rapidly decaying eigenvalues. A low-rank alternating-direction-implicit or rational-Krylov iteration then builds a tall thin factor with such that . Its only operations on are sparse shifted solves. The column space of approximates the dominant reachable subspace, with an error set by the eigenvalues of that the factor drops. A second factor from the observability equation completes the pair: the singular values of approximate the Hankel singular values of [hankel-singular-values], and the reduced model of [balanced-truncation] is built from the two factors by the square-root method. Neither dense Gramian is ever formed.
The sparsity that the positivity of Section I.6 and the graph test of Section I.7 read as structure is the sparsity that a low-rank Lyapunov solver uses in its shifted sparse solves. The low-rank factors it returns are the input from which [balanced-truncation] builds its reduced model.
IV.6. Scaling family
We grow the network of Section III.5 in two steps and watch the Hankel spectrum. The first step reaches by adding two near-duplicate peripheral compartments and one slow deep compartment alongside the retained TMDD complex. The second reaches by adding a third near-duplicate peripheral, a second deep compartment and a two-step cascade that ends in an internalised product. The rates keep both matrices Metzler and mass-conserving. The only variable is the Hankel spectrum as the count climbs.
The network keeps Section III.5's central compartment and TMDD complex (now , with , , ), adds two near-duplicate peripherals (rates and ) and one slow deep compartment (). The system matrices are
The network adds a third near-duplicate peripheral (), a second deep compartment (), and turns the complex into a cascade: complex internalises with rate into a product that only degrades, with rate . The matrix is
with and as before: one dose into , one assay on .
Check. Both matrices are Metzler (every off-diagonal entry nonnegative) and mass-conserving: 's column sums are and 's are . Each column of an internal exchange cancels to zero and only the true eliminations (drug elimination and complex or product turnover) remove mass, as Section III.5 predicted for any such network. At the network stays full rank, controllable and observable at , with Hankel singular values
a spectrum that descends by factors of roughly and then falls off a cliff of about into the fifth value: five resolvable directions, the last only barely so. At the rank verdict changes. The pair loses full rank and is controllable and observable at only . The Hankel spectrum collapses,
the last two values dropping to and to zero: one direction the dose and assay cannot resolve at all, and one more they resolve only at the level of numerical noise.

IV.7. Verdict
The spectrum [eq:scaling-hsv8] lost a direction outright. The balancing coordinates of [balanced-truncation] identify it. We form for [eq:scaling-A8], take the eigenvector of belonging to its smallest (zero) eigenvalue, and read off the state coordinates that dominate it.
The zero Hankel singular value of [eq:scaling-hsv8] aligns to numerical precision with the standard basis vector , the eighth and last coordinate of the state vector. The unobservable direction is the internalised product alone, which occupies that position, and the smallest eigenvalue of has eigenvector with exactly. The reason is structural: is a terminal sink with no off-diagonal entry in the last column of [eq:scaling-A8]. It has no directed path back to the sampled and fails [structural-observability]'s output-connectivity condition. The structural screen of Section IV.2 catches this direction without any Gramian: the two-step cascade added a coordinate downstream of every route to the assay, and no rate choice could make it observable. The smallest nonzero Hankel value, the of [eq:scaling-hsv8], is a distinct failure and lives in the near-duplicate peripheral subspace : those three compartments share the coupling ratio and are excited near-identically by the single central dose. also has an exact null direction, a combination of , and : the coincidence leaves the mode with eigenvalue and eigenvector uncontrollable from the dose. That mode is the unobservable one above and adds no second zero to [eq:scaling-hsv8]. The weakest of the remaining directions of , with eigenvalue near , is a difference between and . The input barely separates it and the assay barely sees it. This second failure passes the structural screen (each peripheral routes back to ) and is visible only in the Gramian spectrum. It is compartment overload: real coordinates that the data cannot tell apart.
The two failures are the two answers that the question of Section IV.1 admits. The internalised product is unobservable exactly: no assay on can recover it, and a graph test that needs no arithmetic catches it. The near-duplicate peripherals are unobservable only in practice: the Hankel weight of the direction has fallen more than four orders of magnitude below the leading one, and only the continuous scale of [hankel-singular-values] grades it. A modeller who added either coordinate to gain new structure added a direction that the single sampled trajectory cannot resolve. The machinery of this Part returns that verdict before a single parameter is fit.
Part V. Sensitivity and identifiability across populations
Every object built so far fixes the numbers: a matrix whose entries are settled rate constants, a Gramian assembled from them, a Hankel spectrum read off once those numbers are chosen. This Part lets the numbers vary. We collect the rate constants into a parameter vector and ask what varying it changes. Section V.1 builds a Gramian for parameters exactly as Part I built one for states, and shows the observability Gramian of Section I.4 is the single instance in which the parameter is the state itself. Section V.2 reads that parameter Gramian as the Fisher information of a sampled noisy experiment and imports its Riemannian meaning from this site's probability and statistics article. Section V.3 separates the failure that the structural test of Section I.7 catches without numbers from the failure that only the spectrum sees, and closes the sloppy-model question that Section IV.4 left open. Sections V.4 through V.7 then fit many such systems at once, one per subject, and ask what changes when a population of trajectories replaces a single one.
V.1. Sensitivity equations and the parameter Gramian
Every theorem of Part I fixes the system matrices and varies the input or the initial state. We now fix a trajectory and vary the matrices. The matrix depends on a parameter vector that collects the rate constants a modeller estimates. The dosing matrix and the sensor matrix stay fixed throughout this article's worked examples because the experiment settles the route of administration and the choice of assay. Differentiating the state equation in each parameter produces a linear system with the same dynamics as the original. The machinery of Section I.2 transfers without a new idea.
Let depend smoothly on , with and independent of , and write for the sensitivity of the state to the th parameter. Then solves the sensitivity equation
whose homogeneous part is the matrix of [eq:state-eq-linear] unchanged. Reading [eq:sensitivity-ode-eq] as a linear system driven by the forcing in place of , [variation-of-parameters] gives
Proof
Differentiate [eq:state-eq-linear], , in . The input and the matrix have no , and the mixed partials commute, so
which is [eq:sensitivity-ode-eq] once is written. The initial state is fixed independently of , so . Applying [variation-of-parameters] to this linear system, with the same , zero initial state, and forcing , gives [eq:sensitivity-solution].
Stack the sensitivities into a single matrix , whose th column records how the whole state trajectory moves when the th parameter is nudged. The sensitivity of the output is then , and the weight each parameter direction has at the sensor, integrated over the experiment, is a Gramian built from exactly as Section I.4's was built from .
For a fixed dosing input and initial state, the parameter Gramian of the system [eq:state-eq-linear] is
with the stacked sensitivities of [eq:sensitivity-solution]. Its entry is the overlap, at the output, between the trajectory's response to parameters and . A direction of parameter space in is one along which no infinitesimal parameter change moves the measured output.
The construction is the observability Gramian of Section I.4 with replaced by the sensitivity matrix , and one parametrisation makes the two objects coincide.
Take the parametrisation , the sensitivity of the trajectory to its own initial state, with , and the input all independent of . Then exactly, and
the observability Gramian of [kalman-rank-o] (defined at [eq:gramian-o-finite], solving [eq:lyapunov-o] in the stable case). The parameter Gramian strictly generalises Part I's observability Gramian: is the one instance in which the parameter under study is the state itself.
The proof is a single line. Under the solution [eq:variation-of-parameters-eq] reads , and only the first term has , so column by column , so . Substituting into [eq:parameter-gramian-def] gives [eq:parameter-gramian-is-Wo-eq].
The rate-constant parametrisation underlying the Fisher information of Section V.2 runs the same construction with and the forcing term of [eq:sensitivity-solution] switched back on. It is the general [parameter-gramian] at a different parametrisation of the same object. Everything Part I proved about (its Lyapunov characterisation, its kernel, its place in the Hankel spectrum) is a statement about [eq:parameter-gramian-def] evaluated at .


V.2. Fisher information matrix as a Riemannian metric
Section V.1's parameter Gramian integrates over a continuous horizon. A real experiment samples the output at finitely many times and reads each sample through measurement noise. Restoring both features turns the parameter Gramian into the Fisher information matrix, and the statistical geometry this site's probability and statistics article built for the Fisher information transfers to it.
Add a Gaussian observation model: the output is sampled at times and each sample is read through independent noise,
This article's worked examples all read a single compartment () and is a scalar variance. The vector-output case replaces by the inverse noise covariance at every occurrence and changes nothing else.
For the observation model [eq:obs-model], the Fisher information matrix of the parameter is
the discretely sampled, noise-weighted counterpart of the parameter Gramian [eq:parameter-gramian-def], with the sensitivity matrix of [eq:sensitivity-solution] evaluated at the th sample time.
Two reasons favour the discrete sum over the integral of [eq:parameter-gramian-def]. A continuous-time integral would have an extra factor of that spoils the dimensional statement of the remark below. A real experiment also samples at finitely many times, so [eq:fisher-information-def] is the quantity that a population-pharmacokinetic optimal-design calculation forms.
The matrix [eq:fisher-information-def] is the Fisher-Rao metric, pulled back along the map from the Gaussian observation model [eq:obs-model] onto this article's parameter space. It is therefore a -tensor on that space (the parameter set is the statistical manifold, whose metric is the infinitesimal Kullback-Leibler divergence): writing the old parameters as a smooth function of new ones, with Jacobian , it transforms as
as a tensor transforms.
Proof
For the Gaussian model [eq:obs-model], the log-likelihood of the sampled data is . Its score in the th parameter, using from Section V.1, is
The Fisher information is the expected outer product of the score. Since , the cross terms drop and
which is [eq:fisher-information-def] and matches the general Fisher metric of the statistics article. Under the chain rule gives with , so each and [eq:fisher-information-def] maps to . This is [eq:fim-transform].
The entry has physical units , exactly those that make [eq:fim-transform] dimensionally consistent. When a reparametrisation preserves each parameter's physical type, the Jacobian entry has units , and the two factors of in cancel against 's units to leave on both sides. A units-checked implementation of [eq:fim-transform] is enforcing the transformation law of [fim-is-metric] at the level of arithmetic, the same coordinate-independence discipline [hankel-singular-values] imposed on the state space in Section IV.4, applied now to parameter space.
V.3. Practical vs. structural identifiability
Section V.1's Gramian and Section V.2's metric grade parameters on a continuous scale. Two failures of that grading separate cleanly: the graph test of Section I.7 catches one without numbers and only the spectrum catches the other. The distinction is the parameter-space form of the state-space distinction drawn in Section IV.4.
Let be a stable compartmental matrix whose diagonal entries are all free, with every elimination and exchange rate constant a parameter. If the structured pair fails [structural-observability], then the Fisher matrix [eq:fisher-information-def] of the compartmental rate-constant parametrisation is singular for every parameter value and every sampling design . Structural observability is a necessary condition for a nonsingular Fisher matrix.
Proof
Failure of [structural-observability] means for every numerical realisation (Section I.7), so the unobservable subspace is nonzero, -invariant, and satisfies for all . With every diagonal entry free, the identity is a saturating matching (as in Section IV.2), so the failing condition is output connectivity. Let be the set of compartments with no directed path to the sensor, which contains the internalised sink Section IV.7 found at . Everything downstream of lies in , so is -invariant and lies in . Stability forces some to eliminate, and its elimination rate constant enters only in the entry , so for every state . By [eq:sensitivity-solution], its output sensitivity is
since is -invariant and annihilated by . Then at every sample time, so for every design and every , and is singular.
The condition is necessary only. Its converse fails: a structurally observable system can still have a Fisher matrix that is nonsingular yet arbitrarily close to singular, and the Cramer-Rao bound (the statistics article's Section 2.5) reads the consequence: the covariance of any unbiased estimator of is bounded below by . A parameter combination near has asymptotic variance that grows without bound as its Fisher eigenvalue shrinks, however finely the output is sampled. That is practical unidentifiability, the sloppy-model regime of Gutenkunst et al.[11], and is a statement about the numbers in that no graph test sees. Structural failure is a knife-edge of the wiring. Practical failure is a matter of degree in the spectrum.
By [parameter-gramian-is-Wo], the initial-condition Fisher matrix [eq:fisher-information-def] is times a sampled quadrature of the integrand of . Its null space contains , the unobservable subspace, and equals it once the samples are dense enough. The Hankel spectrum of [hankel-singular-values] is built from and together in Section IV.4 and also weights each direction by its reachability from the dose. A zero Hankel value marks a direction that is unreachable or unobservable, and only the unobservable case forces a zero Fisher eigenvalue for the initial condition. Small Hankel values and small eigenvalues of are related without being equal, since the first are invariant under a change of state coordinates and the second are not. The Hankel cliff of Section IV.4 is therefore the state-space analogue of a sloppy spectrum. The rate-constant Fisher spectra below are the direct parameter-space measurement.
The empirical spectra show the split. We compute the parameter Gramian [eq:parameter-gramian-def] for the rate-constant parametrisations of the networks of [qsp-k3] and [scaling-family], at a single bolus into and a single assay on , and read off the eigenvalues. Up to the factor and the sample spacing, this Gramian is the dense-sampling limit of [eq:fisher-information-def]. Every spectrum falls off a pronounced cliff. One, two and four eigenvalues respectively collapse to the numerical floor of the Lyapunov solve. These are parameter combinations that no single-dose single-sensor experiment can resolve at any measurement precision. The and networks are structurally observable (Section IV.2), so their zeros are failures of parameter identifiability with every state observable. Counting forces at least one in each: the output determines only the transfer function, whose minimal order- form with has free coefficients, against and parameters at and . At the degradation rate of the unobservable sink adds a zero of the kind [structural-forces-singular-fim] predicts. Several further eigenvalues sit orders of magnitude below the leading one. These are the sloppy directions that the Cramer-Rao bound penalises with large but finite variance. The number of floor-level eigenvalues grows with the parameter count from to , and the qualitative shape is the rate-constant counterpart for parameters of the Hankel cliff of [scaling-family].

V.4. Population extension
Everything in Part V so far fits one subject's data to one parameter vector. A population study measures many subjects at once, each metabolising the same drug through the same compartmental wiring but with rate constants of their own, and asks for the distribution of those rate constants across the population. We lift [state-space] to a hierarchy: a population-typical parameter shared by everyone and a subject-specific deviation drawn around it.
For subjects , the population state-space model endows each subject with a parameter vector
where is the entrywise product, acts entrywise, is the population-typical parameter, is subject 's random effect, and is the between-subject covariance. Each subject then runs the state equation of [eq:state-eq-linear] at its own parameter, sampled through the Gaussian observation model of [eq:obs-model],
for sample times and residual variance . The parameters to estimate are the population triple , with the integrated out.
The log-scale placement of the random effect in [eq:population-hierarchy] answers a constraint from Section I.6. Every entry of is a rate constant, and [metzler] together with [perron-frobenius-positivity] needs each of them nonnegative for the flow to stay entrywise nonnegative. No subject's model then predicts a negative concentration. The multiplicative form holds for every real , and the Gaussian random effect on the log scale keeps each rate constant inside the positive cone whatever value it takes. An additive random effect would place positive probability on . That violates the Metzler sign pattern of [eq:metzler-def] and forfeits the positivity guarantee that holds for every compartment count . Positivity alone admits any between-subject law supported on the positive half-line, such as a gamma law or a Box–Cox transform of a Gaussian. The log-normal form is the convention because it keeps Gaussian, the structure that the Laplace and FOCE approximations of Section V.5 expand around.
The likelihood of the population parameters marginalises each subject's random effect against its prior,
with the Gaussian data density read off [eq:population-obs] and the density. This integral has no closed form because depends nonlinearly on through the exponential of [eq:population-hierarchy] feeding , even though [eq:state-eq-linear] is linear in : the intractability lies entirely in how enters the system matrix.

V.5. FOCE and SAEM as approximations to the population likelihood
The marginal likelihood [eq:population-likelihood] has no closed form and every population-pharmacokinetic estimator approximates the same -dimensional integral. FOCE and SAEM, the two in routine use, either linearise the integrand and integrate it exactly or leave it alone and sample from it. The Fisher information of Section V.2 reappears in the first because the curvature that a Laplace approximation needs at each subject's mode is that Fisher information pulled back to the random-effect coordinate.
Write each subject's log-integrand of [eq:population-likelihood] as , and let be subject 's conditional mode, the random effect most consistent with that subject's own data. The Laplace approximation of the per-subject integral is
whose curvature splits into a data term and the random-effect prior term as
with the Fisher information of [fisher-information] evaluated at , and its pullback to the random-effect coordinate along [eq:population-hierarchy] by the transformation law [eq:fim-transform] of [fim-is-metric]. First-order conditional estimation (FOCE) maximises the resulting approximate marginal log-likelihood
over and recomputes each and each as the outer optimisation moves. The estimator is the sensitivity-and-Fisher machinery of Sections V.1 and V.2 read at a mode.
Proof
The mode is a stationary point of , so the first-order term of the Taylor expansion vanishes and
Substituting this quadratic into and evaluating the Gaussian integral over gives the factor of [eq:laplace-integral], the negative Hessian being positive definite at a maximum. The Hessian splits along the two summands of . The prior contributes , whose Hessian adds to . The data term is at the mode; to first order in the residuals, equivalently replacing the observed information by the expected information exactly as the expectation over in the proof of [fim-is-metric] discarded the residual-weighted second-derivative term, this is the Fisher information in the random-effect coordinate. The chain rule through [eq:population-hierarchy] gives , so by the tensorial transformation law [eq:fim-transform] the Fisher information in is . This is [eq:foce-hessian]. Taking logs of [eq:laplace-integral] and summing over the subjects with gives [eq:foce-loglik].[12]
The identity [eq:foce-hessian] is the main result of Part V. The quantity a population fitting routine forms at every iteration, the curvature of each subject's conditional log-likelihood at its mode, is the parameter Gramian of Section V.1 sampled and noise-weighted into the Fisher information of Section V.2, pulled to the log scale by the same transformation law [eq:fim-transform] that made that Fisher information a metric. The random-effect prior adds the single term , the only ingredient the single-subject theory of Section V.2 lacked.
FOCE is exact only to the order at which [eq:laplace-integral] holds, and the approximation degrades when is strongly nonlinear in or when a subject is sampled too sparsely for its conditional mode to be sharp. Stochastic approximation expectation-maximisation (SAEM) is the alternative that never linearises the integrand. It draws each from its conditional posterior by Markov chain Monte Carlo, then updates by a stochastic-approximation EM step whose sufficient statistics are averaged over the sampled random effects with a decreasing gain, and under regularity conditions it converges to a local maximiser of the exact likelihood [eq:population-likelihood] even in the regimes where the Laplace approximation of [foce-laplace] is poor.[13] It is a named alternative to the estimator that this article's machinery produces and is used when the Laplace curvature [eq:foce-hessian] is too crude a summary of a subject's likelihood.
V.6. Software instantiation and typed units
The two directions of Part V each have a standard instantiation in the population-pharmacokinetics software stack, one package per arrow. The hierarchy of Section V.4 describes the forward map from a population triple to simulated data. The algorithms of Section V.5 compute the inverse map from data back to the triple.
mrgsolve runs the forward direction. It compiles a user's ODE model specification to a shared library with an LSODA integrator, with a C++ backend, and simulates [population-state-space] at whatever compartment count the model declares: fix a population triple and a dosing design, draw each subject's random effect by [eq:population-hierarchy], form each , and integrate [eq:population-obs] forward for every subject.[14] It is the single-subject object of [state-space] from Part I, replicated across a population and pushed forward in time, evaluated exactly as [eq:population-obs] is written.
nlmixr2 runs the inverse direction. The original nlmixr package is archived. Its maintained successor nlmixr2 continues the estimation on the compiled-ODE engine rxode2 and adds the nlmixr2est estimation layer.[15] That layer implements the FOCE, FOCEI and SAEM approximations of [foce-laplace] and [saem] and solves the inverse problem of the marginal likelihood [eq:population-likelihood]: recover from the observed subject data . Where mrgsolve maps a triple to data, nlmixr2 maps data to a triple, along the same population hierarchy read in opposite directions.
Both packages are driven from R, a dynamically typed language, and compile the model's ODE right-hand side to C or C++ from plain numeric code. Neither checks the dimensions of that code before the model runs: a rate constant in added to one in goes through without complaint. Correct units in a fitted model rest on the modeller's conventions. A units library such as Python's pint or a statically typed units crate enforces dimensional consistency mechanically and moves that bookkeeping from the modeller to the compiler. It covers the part of the transformation law of [fim-is-metric] that units can see: a change of unit rescales by a constant and by the reciprocal factors ([units-are-tensoriality]). A general reparametrisation, such as the switch to log rate constants, has a Jacobian that depends on , and applying [eq:fim-transform] correctly remains the modeller's job.
The Rust uom crate encodes physical dimensions in the type system by way of typenum, so a dimensionally inconsistent combination is rejected before the program runs. Scaling a concentration by a dimensionless number stays a concentration and compiles. Adding a concentration to a time fails to typecheck because the two inhabit distinct dimensional types.
use uom::si::f64::{Time, MolarConcentration};
use uom::si::time::hour;
use uom::si::molar_concentration::millimole_per_liter;
fn main() {
let t = Time::new::<hour>(2.0);
let conc = MolarConcentration::new::<millimole_per_liter>(0.5);
// Compiles: scaling a concentration by a dimensionless factor
// is still a concentration.
let scaled: MolarConcentration = conc * 0.5;
// Rejected at compile time: concentration and time are distinct
// dimensional types, so `uom` refuses their sum.
// let bad = conc + t;
let _ = (t, scaled);
}The compile error that uom raises for conc + t is the dimensional check of [units-are-tensoriality], which assigns the units , made mechanical in the type system.
V.7. Synthesis
One object has run through the whole article. Part I fixed the state-space object and its controllability, observability and stability once for arbitrary compartment count. Part II read the mass-action sign pattern as a Metzler structure and the reversible/irreversible split as a decomposition. Part III lifted the generator from a matrix to a rate tensor and the split to the GENERIC bracket. Every Part I tool transfers across the lift unchanged. Part IV used that generality: the structural screen, the empirical Gramian, the Hankel spectrum and the low-rank solver are one body of control theory, and running them on a QSP network imports settled results without rebuilding them mechanism by mechanism. Part V then let the numbers vary, and the same object answered a second family of questions with no new construction.
The parameter derivatives of that object are its Fisher information. Differentiating the state trajectory in the rate constants (Section V.1) builds a Gramian for parameters exactly as Part I built one for states, and recovers the observability Gramian of Section I.4 as the single instance in which the parameter under study is the initial state itself. Sampling that Gramian through a noisy assay (Section V.2) turns it into the Fisher information matrix, a -tensor on parameter space whose Riemannian meaning is the Fisher-Rao metric of this site's probability and statistics article.
Structural observability (Section I.7) is necessary but not sufficient for the practical identifiability that machinery grades on a continuous scale. The graph test catches without numbers a coordinate that no assay on can recover. The Cramer-Rao bound, read off , penalises the directions resolvable in principle and sloppy in practice, with variance growing as their Fisher eigenvalue shrinks (Section V.3). The Hankel cliff of Section IV.4 is the state-space analogue of the sloppy spectra of Gutenkunst et al.[11]: by [hankel-graded-identifiability] its unobservable zeros are zeros of the initial-condition Fisher matrix, and the rate-constant Fisher spectra measure sloppiness in parameter space directly.
The population layer is that same object copied across subjects. The hierarchy of Section V.4 replicates [state-space] once per subject, its log-scale random effect forced by the positivity of Section I.6. The estimators that fit it, FOCE and SAEM (Section V.5), are a Laplace approximation and a stochastic expectation-maximisation assembled from the sensitivity-and-Fisher machinery of Sections V.1 and V.2. Their per-subject curvature is that Fisher information pulled to the random-effect coordinate by the transformation law that made it a metric. The software running these two arrows, mrgsolve forward and nlmixr2 back (Section V.6), simulates and estimates the single-subject object of [state-space] pushed across a population and read in either direction.
One discipline runs under all of it. Correctness at every level of this construction, from a compartment's sign pattern (Section I.6) to a population parameter's asymptotic variance (Section V.3), is a statement about which transformations an object is invariant under. The positivity of Section I.6 is invariance of the state cone under the flow; the coordinate-independence of Section IV.4 is invariance of the Hankel spectrum under a change of state basis; the tensoriality of the Fisher information ([fim-is-metric]) is invariance of the metric under reparametrisation. A type system that tracks physical dimension (Section V.6) is a computational instance of that discipline: refusing the sum of a concentration and a time before the program runs enforces, at the level of arithmetic, the one invariance that the article reads off the object at every scale.
References
- R. E. Kalman (1963). Mathematical description of linear dynamical systems. Journal of the Society for Industrial and Applied Mathematics Series A: Control, 1(2), 152-192. DOI
- C.-T. Lin (1974). Structural controllability. IEEE Transactions on Automatic Control, 19(3), 201-208. DOI
- H. C. Öttinger (2005). Beyond Equilibrium Thermodynamics. Wiley-Interscience. DOI
- R. Hermann, A. J. Krener (1977). Nonlinear controllability and observability. IEEE Transactions on Automatic Control, 22(5), 728-740. DOI
- H. J. Sussmann, V. Jurdjevic (1972). Controllability of nonlinear systems. Journal of Differential Equations, 12(1), 95-116. DOI
- M. Feinberg (2019). Foundations of Chemical Reaction Network Theory. Springer. DOI
- D. E. Mager, W. J. Jusko (2001). General pharmacokinetic model for drugs exhibiting target-mediated drug disposition. Journal of Pharmacokinetics and Pharmacodynamics, 28(6), 507-532. DOI
- S. Lall, J. E. Marsden, S. Glavaški (2002). A subspace approach to balanced truncation for model reduction of nonlinear control systems. International Journal of Robust and Nonlinear Control, 12(6), 519-535. DOI
- B. Moore (1981). Principal component analysis in linear systems: controllability, observability, and model reduction. IEEE Transactions on Automatic Control, 26(1), 17-32. DOI
- K. Glover (1984). All optimal Hankel-norm approximations of linear multivariable systems and their L-infinity error bounds. International Journal of Control, 39(6), 1115-1193. DOI
- R. N. Gutenkunst, J. J. Waterfall, F. P. Casey, K. S. Brown, C. R. Myers, J. P. Sethna (2007). Universally sloppy parameter sensitivities in systems biology models. PLoS Computational Biology, 3(10), e189. DOI
- M. J. Lindstrom, D. M. Bates (1990). Nonlinear mixed effects models for repeated measures data. Biometrics, 46(3), 673-687. DOI
- E. Kuhn, M. Lavielle (2005). Maximum likelihood estimation in nonlinear mixed effects models. Computational Statistics & Data Analysis, 49(4), 1020-1038. DOI
- K. T. Baron (2026). mrgsolve: Simulate from ODE-Based Models. CRAN. Link
- M. Fidler, J. J. Wilkins, R. Hooijmaijers, T. M. Post, R. Schoemaker, M. N. Trame, Y. Xiong, W. Wang (2019). Nonlinear mixed-effects model development and simulation using nlmixr and related R open-source packages. CPT: Pharmacometrics & Systems Pharmacology, 8(9), 621-633. DOI