ACDRepo commited on
Commit
5e5ae2e
·
verified ·
1 Parent(s): 90c8f50

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +130 -3
README.md CHANGED
@@ -1,3 +1,130 @@
1
- ---
2
- license: cc-by-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-2.0
3
+ pretty_name: Coefficients on Kazhdan–Lusztig polynomials for permutations of size 6
4
+ ---
5
+
6
+ ---
7
+ license: cc-by-2.0
8
+ pretty_name: Coefficients on Kazhdan–Lusztig polynomials for permutations of size 6
9
+ ---
10
+
11
+ # The Coefficients of Kazhdan-Lusztig Polynomials for Permutations of Size 6
12
+
13
+ Kazhdan-Lusztig (KL) polynomials are polynomials in a variable \\(q\\) and
14
+ with integer coefficients that (for our purposes) are indexed by a pair of permutations [1].
15
+ We will write the KL polynomial associated with permutations \\(\sigma\\) and \\(\nu\\) as
16
+ \\(P_{\sigma,\nu}(q)\\). For example, the KL polynomial associated with permutations
17
+ \\(\sigma = 1 \; 4 \; 3 \; 2 \; 7 \; 6 \; 5 \; 10 \; 9 \; 8 \; 11\\) and
18
+ \\(\nu = 4 \; 6 \; 7 \; 8 \; 9 \; 10 \; 1 \; 11 \; 2 \; 3 \; 5\\) is
19
+
20
+
21
+ \\(P_{\sigma,\nu}(q) = 1 + 16q + 103q^2 + 337q^3 + 566q^4 + 529q^5 + 275q^6 + 66q^7 + 3q^8\\)
22
+
23
+ (see [here](https://gswarrin.w3.uvm.edu/research/klc/klc.html) for efficient software to compute
24
+ these polynomials). KL polynomials have deep connections throughout several areas of mathematics. For example,
25
+ KL polynomials are related to the dimensions of intersection homology in Schubert calculus,
26
+ the study of the Hecke algebra, and representation theory of the symmetric group. They
27
+ can be computed via a recursive formula [[1]](https://link.springer.com/article/10.1007/BF01390031),
28
+ nevertheless, in many ways they remain mysterious. For instance, there is no known closed
29
+ formula for the degree of \\(P_{\sigma,\nu}(q)\\).
30
+
31
+ One family of questions revolve around the coefficients of \\(P_{\sigma,\nu}(q)\\).
32
+ For instance, it has been hypothesized that the coefficient on the largest possible monomial term
33
+ \\(q^{\ell(\sigma) - \ell(\nu)-1/2}\\) (where \\(\ell(x)\\) is a statistic of the
34
+ permutation \\(x\\) called the *length* of the permutation), which is known as the
35
+ \\(\mu\\)-coefficient, has a combinatorial interpretation but currently this is not
36
+ known. Better understanding this and other coefficients is of significant
37
+ interest to mathematicians from a range of fields.
38
+
39
+ ## Dataset details
40
+
41
+ Each instance in this dataset consists of a pair of permutations of \\(n,x \in S_n\\)
42
+ along with the coefficients of the polynomial \\(P_{x,w}(q)\\). If \\(x = \;1 \;2 \;3\; 4\; 5\; 6\\),
43
+ \\(w=4 \;5\; 6\; 1 \;2 \;3\\) and \\(P_{v,w}(q) = 1 + 4q + 4q^2 + q^3\\)
44
+ then the coefficients field is written as `1, 4, 4, 1`. Note that coefficients are listed
45
+ by increasing degree of the power of \\(q\\) (e.g., the coefficient on \\(1\\) comes first,
46
+ then the coefficient on \\(q\\), then the coefficient on \\(q^2\\), etc.)
47
+
48
+ We summarize the limited number of values coefficients on \\(P_{x,w}(q)\\) take when \\(x, w \in S_6\\).
49
+
50
+ **Constant Terms:**
51
+
52
+ | | 0 | 1 | Total number of instances |
53
+ |----------|----------|----------|----------|
54
+ | Train | 336,071 | 78,649 | 414,720 |
55
+ | Test | 83,922 | 19,758 | 103,680 |
56
+
57
+ **Coefficients on \\(q\\):**
58
+
59
+ | | 0 | 1 | 2 | 3 | 4 | Total number of instances |
60
+ |----------|----------|----------|----------|----------|----------|----------|
61
+ | Train | 397,386 | 13,253 | 3,483 | 535 | 63 | 414,720 |
62
+ | Test | 99,354 | 3,311 | 883 | 117 | 15 | 103,680 |
63
+
64
+ **Coefficient on \\(q^2\\):**
65
+
66
+ | | 0 | 1 | 2 | 3 | 4 | Total number of instances |
67
+ |----------|----------|----------|----------|----------|----------|----------|
68
+ | Train | 412,707 | 1,705 | 242 | 40 | 26 | 414,720 |
69
+ | Test | 103,177 | 441 | 46 | 8 | 8 | 103,680 |
70
+
71
+ **Coefficient on \\(q^3\\):**
72
+
73
+ | | 0 | 1 | Total number of instances |
74
+ |----------|----------|----------|----------|
75
+ | Train | 414,688 | 32 | 414,720 |
76
+ | Test | 103,670 | 10 | 103,680 |
77
+
78
+
79
+ ## Data Generation
80
+
81
+ Datasets were generated using C code from Greg Warrington's
82
+ [website](https://gswarrin.w3.uvm.edu/research/klc/klc.html). The code we used can be found
83
+ [here](https://github.com/pnnl/ML4AlgComb/tree/master/kl-polynomial_coefficients).
84
+
85
+ ## Task
86
+
87
+ **Math question:** Generate conjectures around the properties of coefficients appearing on KL polynomials.
88
+
89
+ **Narrow ML task:** Predict the coefficients of \\(P_{x,w}(q)\\) given \\(x\\) and \\(w\\).
90
+ We break this up into a separate task for each coefficient though one could
91
+ choose to predict all simultaneously. Since there are generally very few
92
+ different integers that arise as coefficients (at least in these small examples),
93
+ we frame this problem as one of classification.
94
+
95
+ While the classification task as framed does not capture the broader math question exactly,
96
+ illuminating connections between \\(x\\), \\(w\\), and the coefficients of \\(P_{x,w}(q)\\)
97
+ has the potential to yield critical insights.
98
+
99
+ ## Small model performance
100
+
101
+ Since there are many possible tasks here, we did not run exhaustive hyperparameter searches.
102
+ Instead, we ran ReLU MLPs with depth 4, width 256, and learning rate 0.0005.
103
+
104
+ ### Kazhdan-Lusztig Polynomials for Permutations of \\(6\\) elements
105
+
106
+ Accuracy predicting coefficients for permutations of 6 elements:
107
+
108
+ | Coefficient | MLP | Transformer | Guessing largest class |
109
+ |----------|----------|-----------|------------|
110
+ | \\(1\\) | \\(99.9\% \pm 0.0\%\\) | \\(100.0\% \pm 0.0\%\\) | \\(80.9\%\\)|
111
+ | \\(q\\) | \\(99.9\% \pm 0.0\%\\) | \\(99.9\% \pm 0.0\%\\) | \\(95.8\%\\) |
112
+ | \\(q^2\\) | \\(99.9\% \pm 0.0\%\\) | \\(99.9\% \pm 0.0\%\\) | \\(99.5\%\\) |
113
+ | \\(q^3\\) | \\(99.9\% \pm 0.0\%\\) | \\(99.9\% \pm 0.0\%\\) | \\(99.9\%\\) |
114
+
115
+ The associated macro F1-scores are:
116
+
117
+ | Coefficient | MLP | Transformer |
118
+ |----------|----------|-----------|
119
+ | \\(1\\) | \\(99.9\% \pm 0.0\%\\) | \\(100.0\% \pm 0.0\%\\) |
120
+ | \\(q\\) | \\(99.0\% \pm 1.5\%\\) | \\(98.0\% \pm 3.7\%\\) |
121
+ | \\(q^2\\) | \\(97.4\% \pm 5.2\%\\) | \\(98.0\% \pm 3.7\%\\) |
122
+ | \\(q^3\\) | \\(87.9\% \pm 4.5\%\\) | \\(88.3\% \pm 17.1\%\\) |
123
+
124
+ The \\(\pm\\) signs indicate 95% confidence intervals from random weight initialization and training.
125
+
126
+ ## References
127
+
128
+ \[1\] Kazhdan, David, and George Lusztig. "Representations of Coxeter groups and Hecke algebras." Inventiones mathematicae 53.2 (1979): 165-184.
129
+
130
+ \[2\] Warrington, Gregory S. "Equivalence classes for the μ-coefficient of Kazhdan–Lusztig polyno