Chapter II: Polyhedral constraints enable holistic analysis of bioregulation
3.5 Log derivative as transform between two coordinate charts
Note that although the theorem states that this log derivative πlogπ₯
π(logπ‘,logπ) can always be calculated at any pointπ₯in this fashion through matrix inversion, we have not proved this, as all previous calculations assumedππ₯πΉ is invertible. To show this, we can prove something stronger, namely the mapπ : logπ₯β¦β (logπ‘,logπ)is a diffeomorphism. This map can be explicitly written as follows:
π(π§) =
β‘
β£
logπΏexpπ§ π π§
β€
β¦, (3.20)
where the exponential map is applied component wise. We can show this map is a diffeomorphism fromRπtoRπ. By Hadamard-Caccioppoli Theorem,πis a diffeomorphism ifπ is proper andππ is bΔ³ective at all points. π is proper because for every sequence ofπ§ escaping to infinity,π(π§)also escapes to infinity, sinceπΏis non-negative and every column is nonzero. So it is left to show thatππ is bΔ³ective at all points, i.e. ππ, the log derivative
π(logπ‘,logπ)
πlogπ₯ is invertible for allπ₯. From (3.19), we see that this is equivalent to the matrix π(π₯;πΏ) :=
β‘
β£
πΏΞπ₯ π
β€
β¦ (3.21)
is invertible for allπ₯. We prove this in the following proposition.
Proposition 3.5.2. π(π₯;πΏ)is invertible for allπ₯βRπ>0.
Proof. Proof by contradiction. If it is not invertible, then there exists a nonzero vectorπ£s.t.
π(π₯;πΏ)π£ = 0. This impliesπ π£ = 0, i.e. π£βrowspanπΏ, so there exists a nonzero vectorπ s.t. π£ =πΏβΊπ. SoπΏΞπ₯π£ =πΏΞπ₯πΏβΊπ= 0. But this is impossible, sinceΞπ₯is positive definite, soπβΊπΏΞπ₯πΏβΊπ>0.
Hence,π : logπ₯ β¦β (logπ‘,logπ)is a diffeomorphism, and the log derivative πlogπ₯
π(logπ‘,logπ) is well defined on all points inβ³. The formula3.19is always applicable.
We can write the relationship of chartlogπ₯and chart(logπ‘,logπ)in the following diagram:
logπ₯ββββββββββββββββββββββ½βββββββββββββββββββββπ(logπ₯)=(logπΏπ₯,πlogπ₯) β
ππβ1=π(logπlogπ‘,logπ₯π)=
β‘
β£
Ξβ1π‘ πΏΞπ₯ π
β€
β¦
β1 (logπ‘,logπ). (3.22)
Now we have established that for a generic equilibrium manifoldβ³of a binding network, in addition to the natural chart logπ₯, we have an alternative chart in terms of totals and equilibrium constants (logπ‘,logπ), which is also a one-chart atlas. Recall that for isomer-atomic binding networks, we have another natural chart using atomic species (logπ₯π,logπ). We study how this chart relates to the total chart(logπ‘,logπ)below.
Log derivatives and chart transform for isomer-atomic binding networks
Recall that from the previous section, we established that if the binding network is isomer- atomic, then we have another chart using the atomic species, namely(logπ₯π,logπ), where π₯= (π₯π,π₯π)is split into the atomic species and the complex species. To relate this atomic chart to the total chart(logπ‘,logπ), since the equilibrium constantsπ)is kept the same, we just need to study how π₯π is mapped to π‘. The chart transform in Eq (3.16) tells us logπ₯π =πΏβΊ2logπ₯π+π2β1logπ, so we have
π‘ =πΏπ₯ =π₯π+πΏ2π₯π=π₯π+πΏ2exp(πΏβΊ2logπ₯π+π2β1logπ).
The inverse map from (logπ₯π,logπ)to(logπ‘,logπ)again requires solving an intractible polynomial system, so we resort to differentials.
Theorem 3.5.3(Isomer-atomic log derivative formula). Givenβ³ β R2π>0, the equilibrium manifold of an isomer-atomic binding network with transpose-reduced stoichiometry matrixπ βRπΓπand conservation law matrixπΏβRπΓπ, with an atom-first ordering so that the firstπspecies are atomic species, i.e. πΏ = [οΈIπ πΏ2
]οΈ
,πΏ2 β RπΓπ, and π = [οΈπ1 π2
]οΈ
, withπΏβΊ2 = βπ2β1π1. Then at any pointπ= (π₯,π‘,π)β β³, we have
πlogπ₯π
π(logπ‘,logπ) =[οΈ(πΏΞπ₯πΏβΊ)β1Ξπ‘ β(πΏΞπ₯πΏβΊ)β1πΏ2Ξπ₯ππ2β1]οΈ,
πlogπ₯π
π(logπ‘,logπ) =πΏβΊ2 πlogπ₯π
π(logπ‘,logπ) +[οΈ0 π2β1]οΈ,
(3.23)
whereIπ is the identity matrix of dimensionπ.
Proof. The second formula is immediately obtained by using chain rule and Eq (3.16):
πlogπ₯π
π(logπ‘,logπ) = ππΏβΊ2logπ₯π+π2β1logπ
π(logπ‘,logπ) =πΏβΊ2 πlogπ₯π
π(logπ‘,logπ) +π2β1[οΈ0 Iπ]οΈ.
The first formula is obtained by block-matrix inversion of Eq (3.19). Block-matrix inversion satisfies
β‘
β£
π΄ π΅ πΆ π·
β€
β¦=
β‘
β£
(π΄βπ΅π·β1πΆ)β1 0
0 (π·βπΆπ΄β1π΅)β1
β€
β¦
β‘
β£
I βπ΅π·β1
βπΆπ΄β1 I
β€
β¦, if blocksπ΄andπ·are both invertible. So
[οΈπ΄ π΅]οΈ= (π΄βπ΅π·β1πΆ)β1[οΈIβπ΅π·β1]οΈ.
Applying this to our case,
πlogπ₯π
π(logπ‘,logπ) =[οΈIπ 0]οΈπ(logπ₯π,logπ₯π)
π(logπ‘,logπ) =[οΈIπ 0]οΈ
β‘
β£
Ξπ₯π πΏ2Ξπ₯π π1 π2
β€
β¦
β1β‘
β£
Ξπ‘ 0 0 Iπ
β€
β¦
=(Ξπ₯π βπΏ2Ξπ₯ππ2β1π1)β1[οΈI βπΏ2Ξπ₯ππ2β1]οΈ
β‘
β£
Ξπ‘ 0 0 Iπ
β€
β¦
=(Ξπ₯π +πΏ2Ξπ₯ππΏβΊ2)β1[οΈΞπ‘ βπΏ2Ξπ₯ππ2β1]οΈ. This is the desired formula by recognizingΞπ₯π+πΏ2Ξπ₯ππΏβΊ2 =πΏΞπ₯πΏβΊ.
We remark that the atomic log derivative formula in Eq (3.23) highlights an internal symmetric structure not obviously seen in the more general log derivative formula in Eq (3.19). In particular, the symmetric positive-semidefinite matrix(πΏΞπ₯πΏβΊ)β1constitute the core structure from which all the log derivatives arise.
In addition to characterizing how the atomic chart(logπ₯π,logπ)is mapped from the total chart(logπ‘,logπ), Eq (3.23) also implies a simpler way to calculate the log derive πlogπ₯
π(logπ‘,logπ)
for isomer-atomic binding entworks. Namely, we can first calculate just the log derivative of π₯π, then use this to obtain the log derivative ofπ₯π. For large scale problems, the bottleneck in log dervative computation would be matrix inversion, which often hasπ(π3)complexity to invert a square matrix with dimensionπ(could be lower by more advanced algorithms, but still larger thanπ(π2.3)). We see that in this formula, we can compute all log derivatives by inverting only one matrix with dimension π and one with dimension π, instead of inverting a matrix with dimensionπin Eq (3.19), so roughly reducing complexity fromπ3 toπ3+π3. If we are only interested in the log derivative ofπ₯with respect toπ‘, we only need to compute πlogπ₯
π
πlogπ‘ = (πΏΞπ₯πΏβΊ)β1Ξπ‘, which only requires to invert aπΓπmatrix, and then compute πlogπ₯
π
πlogπ‘ =πΏβΊ2ππloglogπ₯π‘π =πΏβΊ2(πΏΞπ₯πΏβΊ)β1Ξπ‘.
We summarize the three charts studied so far and the transform between them in Figure 3.4.
Other alternative charts
The two alternative charts we have studied so far have equilibrium constants fixed, and the remainingπdegrees of freedom are represented asπ₯π =π΄ππ₯whereπ΄π :=[οΈIπ 0]οΈin atomic chart, andπ‘=πΏπ₯in total chart. This brings the question about what are the other charts of the form(logπ΄π₯,logπ), for some non-negative matrixπ΄βRπΓπ?
We know(logπ΄π₯,logπ)is a chart if and only if the maplogπ₯β¦β(logπ΄π₯,logπ)is invertible for allπ₯. From log derivative formula (3.19), we know this is equivalent to matrixπ(π₯;π΄)
Figure 3.4The three charts of equilibrium manifoldβ³of a binding network, and the transform between them.
is invertible for allπ₯. Hence we defineπ ={οΈπ΄βRπΓπβ₯0 : detπ(π₯;π΄)ΜΈ= 0,βπ₯}οΈ, the set of π΄such that(logπ΄π₯,logπ)are charts.
Topologically, since determinant is a continuous function, a matrixπ΄β πhas determinant of π(π₯;π΄) either positive for all π₯, or negative for all π₯. This separates π into two connected components,
π+(π) :={οΈπ΄βRπΓπβ₯0 : detπ(π₯;π΄)>0,βπ₯}οΈ, (3.24) and correspondinglyπβ(π).
For the atomic chart,detπ(π₯;π΄π) = detπ2. So
[οΈ
Iπ 0]οΈ β π+(π)ifdetπ2 >0. Similarly, we show below that the total chartβs sign is determined by the same condition for isomer- atomic CRNs.
Proposition 3.5.4. Assume isomer-atomic, thenπΏβ π+(π)if and only ifdetπ2 >0. Proof. Calculate
β‘
β£
πΏ π
β€
β¦
β‘
β£
Iπ 0 0 π2βΊ
β€
β¦=
β‘
β£
Iπ βπ1βΊπ2ββΊ π1 π2
β€
β¦
β‘
β£
Iπ 0 0 π2βΊ
β€
β¦=
β‘
β£
I βπ1βΊ π1 π2π2βΊ
β€
β¦. Take determinant of both sides,
det
β
β
β‘
β£
πΏ π
β€
β¦
β
β det(π2βΊ) = det(π2π2βΊ+π1π1βΊ)>0.
The right hand side is a positive definite matrix, so determinant is positive. Since detπ(1;πΏ)>0if and only ifdetπ(π₯;πΏ)>0for allπ₯βRπ>0from previous Lemma, we have the desired result.
So we know the atomic chart with matrixπ΄=π΄π=[οΈIπ 0]οΈand the total chart withπ΄=πΏ both resides in the same connected component ofπ. Exactly which one depends on the choice of which direction is used for the stoichiometry vectors. Thus, without loss of generality, we assumeπ΄π,πΏβ π+(π).
To find other alternative chart, we note that given aπ΄β π+(π), then for anyπ βGL+π(R) such thatππ΄βRπΓπβ₯0 , thenππ΄β π+(π). HereGL+π(R)is the identity component of the general linear group ofπΓπmatrices, i.e. πΓπmatrices with positive determinant. This states that any invertible matrix with positive determinant can be left-multiplied toπ΄, and if the resulting matrix is non-negative, then it is inπ+(π). There are several elementary matrices of GL+π(R)worth noting. It includes positive scaling ΞπΌ, where πΌ β Rπ>0 is a positive vector inRπ. It also includes permutations with positive sign, i.e. consisting of an even number of transpositions. It also includes row additionsI+ππΈππ, whereπ ΜΈ=π, π, π β {1, . . . , π}, πΈππ has 1at (π, π)entry and zero everywhere else, and π β R is a real number. I+ππΈππ takes theπth row ofπ΄, multiply byπ, and adds to theπth row ofπ΄. We caution that combinations of these elementary operations may go out ofGL+π(R), and the resulting matrix may no longer be non-negative.
As for right multiplication, givenπ΄β π+(π), thenπ΄ΞπΌ β π+(π)for any positive vector πΌ β Rπ>0. This is because this multiplication is just a scaling of variables π₯, without changing its domainRπ>0.
These operations give a basic approach to explore other alternative charts ofβ³. It is an interesting question for further research to characterize the setπ+(π).