Table of Contents
ChangeType
The function ResponseFunction.ChangeType(G, type, …) can be used to transform the representation of a response function. In Quanty, there are 5 different types for the response function:
-
list of poles (ListOfPoles)
-
tri-diagonal (Tri)
-
Anderson (And)
-
natural impurity orbital (Nat)
-
double tri-diagonal (DoubleTri)
See the first page about response functions for more information on how response functions are internally represented.
These types are related to each other by unitary transformations. ResponseFunction.ChangeType(G) allows one to transform between types.
Note that
-
Internal functions can change the type of a response function.
-
Internal functions accept all different types of a response function.
-
All transformations are unitary, i.e. bijective, however they take time and might lead to significance loss.
-
You can add or subtract response functions of different types.
Input
-
G: the response function to transform
-
type: (string) the representation to transform to, one of “ListOfPoles” (or “LP”), “Tri”, “And”, “NaturalImpurityOrbital” (or “Nat”), “DoubleTri” (or “DTri”), and furthermore “Dense”, “Userdata” and “Table”
(Optional) Third argument, a list of options
-
NTriMax : (integer) the largest chain length the result may have. Accepted by the transformations to Tri, Nat and DoubleTri, which are the representations that have a chain (default: $\infty$)
-
NTriMaxVal : (integer) the largest length of the valence chain. Accepted by the transformations to Nat and DoubleTri, which have a valence and a conduction chain. Overrules NTriMax for that chain (default: the value of NTriMax)
-
NTriMaxCon : (integer) the largest length of the conduction chain, as NTriMaxVal (default: the value of NTriMax)
-
EffectivekBT : (real) the temperature of the Fermi function that decides whether a pole belongs to the valence or to the conduction chain. Accepted by the transformations to Nat and DoubleTri
-
SingularValue : (real) smallest singular value of a block that is considered different from zero. A direction whose singular value is smaller is removed, which reduces the size of the blocks of a matrix valued chain (default: $1000 \, \epsilon_{\mathrm{machine}} \approx 2.2 \cdot 10^{-13}$)
-
MinimalPoleDistance : (real) two poles closer together than this are merged into one
A chain of length $N$ is $a_0 \ldots a_N$ with $b_0 \ldots b_{N-1}$, so NTriMax is the number of bath sites (blocks) behind the observed one, and NTriMax $=0$ leaves $G(\omega) = a_0 + b_0 \frac{1}{\omega - a_1} b_0$ with no bath at all. A chain that comes out shorter than NTriMax on its own is left as it is.
A chain that comes out longer is cut, and it is cut without looking at the values further down: the $a_i$ and $b_i$ that are kept are exactly the ones the transformation without a ceiling would have produced. This is not the same as reducing the number of poles with ResponseFunction.ReducePoles, which merges neighbouring poles and conserves their total weight and their centre of gravity. Cutting a chain instead keeps the first $2N$ moments $\mu_0 \ldots \mu_{2N-1}$ of the spectral function exactly and says nothing about the rest, so the truncated function converges to the full one from the low moments upwards, as the example below shows.
Where the transformation is built on a recursion — the (block) Lanczos that turns a list of poles into a chain — the ceiling stops the recursion rather than letting it run to the end and cutting afterwards. Transforming a list of $10^4$ poles into a chain of $10$ sites then costs ten steps and not $10^4$. The two give the same answer; only the one is much cheaper.
Output
-
ResponseFunction: the same response function in the requested representation
Example
A flat band given as a list of 200 poles, transformed into a tri-diagonal chain of a given length. The last row is the transformation without a ceiling, for comparison.
Input
- ChangeTypeNTriMax.Quanty
-- A hybridisation function given as a list of 200 poles on a flat band Npoles = 200 a = {0} b = {} for i = 1, Npoles do a[i+1] = -3 + 6 * (i - 0.5) / Npoles b[i] = 0.25 * 6 / Npoles end G = ResponseFunction.New({a, b, mu = 0, type = "ListOfPoles", name = "Delta"}) Gfull = ResponseFunction.ChangeType(G, "Tri") print("chain length without a ceiling : " .. #(ResponseFunction.ToTable(Gfull)[2])) print("") print(" NTriMax chain -Im G(0 + 0.3 I) -Im G(2 + 0.3 I)") for _, N in ipairs({4, 10, 20, 40, 80}) do local Gs = ResponseFunction.ChangeType(G, "Tri", {{"NTriMax", N}}) print(string.format(" %7d %5d %16.6f %16.6f", N, #(ResponseFunction.ToTable(Gs)[2]), -Gs(0, 0.3).I, -Gs(2, 0.3).I)) end print(string.format(" %7s %5d %16.6f %16.6f", "none", #(ResponseFunction.ToTable(Gfull)[2]), -Gfull(0, 0.3).I, -Gfull(2, 0.3).I))
Output
chain length without a ceiling : 200
NTriMax chain -Im G(0 + 0.3 I) -Im G(2 + 0.3 I)
4 4 0.149765 0.192410
10 10 0.353597 1.168149
20 20 0.581795 0.727876
40 40 0.734219 0.747012
80 80 0.760022 0.740689
none 200 0.760419 0.740681
Cutting a chain is a real approximation and not a free lunch: a chain of four sites carries four poles and cannot represent a band. It is the right tool when the chain is longer than the problem needs — a bath built from a dense energy grid, or a self-energy carried along at full length through a chain of arithmetic — and the wrong tool when the number of poles is what matters, in which case ResponseFunction.ReducePoles keeps the spectral weight where it is instead of throwing the far end away.