text
stringlengths 2
6.14k
| file_name
stringclasses 10
values | node_id
stringlengths 36
36
|
|---|---|---|
= (100000111) 2→ (1100010110) 2→ (110001011) 2326
251
395 = (110001011) 2→ (10010100010) 2→(1001010001) 2325
250
593 = (1001010001) 2→(11011110100) 2→(110111101) 2324
249
445 = (110111101) 2→ (10100111000) 2→(10100111) 2323
247
167 = (10100111) 2→ (111110110) 2→ (11111011) 2322
244
251 = (11111011) 2→ (1011110010) 2→ (101111001) 2321
243
377 = (101111001) 2→ (10001101100) 2→(100011011) 2320
242
283 = (100011011) 2→ (1101010010) 2→ (110101001) 2319
240
425 = (110101001) 2→ (10011111100)= (110101001) 2→ (10011111100) 2→(100111111) 2318
239
319 = (100111111) 2→ (1110111110) 2→ (111011111) 2317
237
479 = (111011111) 2→ (10110011110) 2→(1011001111) 2316
236
719 = (1011001111) 2→(100001101110) 2→(10000110111) 2315
235
1079 = (10000110111) 2→(110010100110) 2→(11001010011) 2314
234
1619 = (11001010011) 2→(1001011111010) 2→(100101111101) 2313
233
2429 = (100101111101) 2→(1110001111000) 2→(1110001111) 2312
232
911 = (1110001111) 2→(101010101110) 2→(10101010111)2→(101010101110) 2→(10101010111) 2311
229
1367 = (10101010111) 2(1000000000110) 2(100000000011) 2310
228
2051 = (100000000011) 2→(1100000001010) 2→(110000000101) 239
227
3077 = (110000000101) 2→(10010000010000) 2→(1001000001) 238
226
577 = (1001000001) 2→(11011000100) 2→(110110001) 237
222
433 = (110110001) 2→ (10100010100) 2→(101000101) 236
220
325 = (101000101) 2→ (1111010000) 2→ (111101) 235
218
61 = (111101) 2→ (10111000) 2→ (10111) 234
2148 | 2402.00001v1.pdf | 16b037f0-0fb6-4b49-848a-d52fd9e4a595 |
330
261·10027
10027=(10011100101011) 2→(111010110000010) 2→(11101011000001) 2329
261
15041=(11101011000001) 2→(1011000001000100) 2→(10110000010001) 2328
260
11281=(10110000010001) 2→(1000010000110100) 2→(10000100001101) 2327
258
8461=(10000100001101) 2→(110001100101000) 2→(110001100101) 2326
256
3173=(110001100101) 2→(10010100110000) 2→(1001010011) 2325
253
595=(1001010011) 2→ (11011111010) 2→(1101111101) 2324
249
893=(1101111101) 2→ (101001111000) 2→(101001111) 2323
248
335=(101001111) 2→2323
248
335=(101001111) 2→ (1111101110) 2→(111110111) 2322
245
503=(111110111) 2→ (10111100110) 2→(1011110011) 2321
244
755=(1011110011) 2→ (100011011010) 2→(10001101101) 2320
243
1133=(10001101101) 2→ (110101001000) 2→(110101001) 2319
242
425=(110101001) 2→ (10011111100) 2→(100111111) 2318
239
319=(100111111) 2→ (1110111110) 2→(111011111) 2317
237
479=(111011111) 2→ (10110011110) 2→(1011001111) 2316
236
719=(1011001111) 2→ (100001101110) 2→(10000110111) 2315
235
1079=(10000110111) | 2402.00001v1.pdf | 7d975516-2e36-4044-ad99-aafe334c6a2e |
2315
235
1079=(10000110111) 2→ (110010100110) 2→(11001010011) 2314
234
1619=(11001010011) 2→(1001011111010) 2→(100101111101) 2313
233
2429=(100101111101) 2→(1110001111000) 2→(1110001111) 2312
232
911=(1110001111) 2→ (101010101110) 2→(10101010111) 2311
229
1367=(10101010111) 2→(1000000000110) 2→(100000000011) 2310
228
2051=(100000000011) 2→(1100000001010) 2→(110000000101) 239
227
3077=(110000000101) 2→(10010000010000) 2→(1001000001) 238
226
577=(1001000001) 2→ (11011000100) 2→(110110001)2→ (11011000100) 2→(110110001) 237
222
433=(110110001) 2→ (10100010100) 2→(101000101) 236
220
325=(101000101) 2→ (1111010000) 2→(111101) 235
218
61=(111101) 2→ (10111000) 2→(10111) 234
214
23=(10111) 2→ (1000110) 2→(100011) 233
211
35=(100011) 2→ (1101010) 2→(110101) 232
210
53=(110101) 2→ (10100000) 2→(101)23
29
5=(101) 2→ (10000) 2→ (1)21
24
9 | 2402.00001v1.pdf | cd22bf5a-f531-45a9-abdb-ef8c2f27010a |
4.1. Row characteristic
1) In each row, the first column is always odd (empty when the initial v alue is even), and the last bit of
its binary string must be 1,
The secondcolumn must be even, its binarystringmust end with at lea stone0, the number ofsubsequent
even numbers must be as many as the number of zeros at the end of the second column’s binary string, and
the number of zeros at the end of each even number to the right is o ne less than the previous one, until all
zeros are deleted to become the last odd number in the row.
The last column must be odd, and the last bit of its binary string must b e 1.
2) When there are only three numbers in a row, that is, only one even number, the last odd number must
be greater than the first odd number (the first column); When the re are more than three numbers in a row,
that is, more than two even numbers, the last odd number must be s maller than the first odd number.3) The preceding binary string is identical from the second column to the last column in one line, except
for the all 0 at the end.
4.2.4.2. Column characteristic
From top to bottom, the binary string in the first column of two adja cent rows has the following two
characteristics:
1) If the number of bits of 1 in the last substring of the previous row is greater than 1, the number of
bits of 1 in the last substring of the next row is reduced by one, and t he momentum of this reduction of
one bit remains unchanged until it finally becomes only one; The corre sponding number is greater than the
number in the previous row;
2) If the last substring of the binary string in a row contains only one bit of 1, then the last substring
of the binary string in the next row contains either one bit of 1 or man y bits of 1, and the corresponding
number is smaller than the number in the previous row;
3) Within each line, the number of bits in the first binary string is increa sed by 1 bit with the number in
the second binary string, but for hard numbers, the number is incr eased by 2 bits.The reduction is at least
1 bit, and when it is a hard number of mbits, the number is reduced by m+1.
4.3. Comprehensive characteristic
1) The substring at the end of the binary (right) is observed that w hen the substring is 1, the units
digit of the corresponding decimal number can be any one of 1 ,3,5,7,9. For instance, 161 = (10100001) 2,
433 = (110110001) 2, 325 = (101000101) 2, 577 = (1001000001) 2, 2429 = (100101111101) 2.
2) The number ending in decimal is 9 and the corresponding binary num ber can end in any digits of
1, for instance: 319 = (100111111) 2, 479 = (111011111) 2,719 = (1011001111) 2, 1079 = (10000110111) 2,
1619 = (11001010011) 2, 2429 = (100101111101) 2. | 2402.00001v1.pdf | b65e462c-6133-46fc-8b9f-6b07ea9bf975 |
3)When thelastsubstringofbinaryisreducedbyonebit frommany, the correspondingdecimalnumber’s
units are always reciprocated within the three groups of numbers: 1,7, or 3,5, and unit bit is always 9.
10 | 2402.00001v1.pdf | 63afd3c0-0166-4e96-b48a-825e433f9679 |
4.4. Discussion according three sets
The following is a discussion of the natural numbers according to our classification, that is, when the
initial value of the iteration process of the Collatz function is pure ev en, pure odd and mixed numbers, it
always reaches the smallest natural number 1, thus proving the es tablishment of Collatz conjecture.
(1)For a pure even n= 2k= (10···0)2, it requires only an iteration of the ktimes Collatz function to
reach the smallest natural number 1, that is, the conjecture hold s.
(2)For a special class of mixed numbers, if akin the iteration sequence aof the Collatz function is a
special class of mixed numbers - the hard number4k−1
3= (101···101)2, then its sequent iteration result is
ak=4k−1
3= (101···101···101)2,ak+1= 4k= 22k= (10···0)2,···,a3k+1= 1.
This means that the Collatz conjecture is valid at this point.This means that the Collatz conjecture is valid at this point. For ordin ary mixed numbers, this is a
complicated process, which is carried out in conjunction with the disc ussion of pure odd numbers in follows.
(3) For a pure odd number n= 2k−1 = (11···1)2, if the Collatz function is iterated in binary form, that
is, the second column in the preceding table becomes a mixed number, which can be observed by several
examples. Research papers on this topic are[6,7,8]. By checking the change law of the last substring in the
binary string and the change law of the total length of the binary st ring, we show that the Collatz function
always reaches the minimum natural number 1 in the finite iteration va lue, thus proving the correctness of
Collatz conjecture.
In view of the general, a pure odd n= 2r−1, there is a mathematical formula: 2r−1 = 2r−1+2r−2+
2r−3+···+ 2 + 1.Two numbers randmin the following tables are the values we verify for the Collatz
conjecture by iterating Tm(2r−1) = 1,r= 1,2,3,···,203 in the Maple program below:
restart;
i:= 2r−1;
convert(i,binary,decimal );
jj:= 0;
while i ̸= 1do
i:=piecewise (irem(i,2) = 0,(1/2)˙i,3˙i+1);
convert(i,binary,decimal );
jj:=jj+1
enddo
By verifying the first 203 pure odd number 2r−1, we can get the following rule:
1) Odd number rgreater than 1, the corresponding test Collatz function value has the following relation:
ifTm(2r−1) = 1, then Tm+1(2r+1−1) = 1. | 2402.00001v1.pdf | 99f6f986-2365-4e42-9c1a-e77bfb6abc30 |
2)T856(257−1) =T856(265−1) = 1
3)T857(258−1) =T857(266−1) = 1
4)T1451(286−1) =T1451(2117−1) = 1
5)T1455(290−1) =T1455(2121−1) = 1
6)T1456(291−1) =T1456(2122−1) = 1
7)T1457(292−1) =T1457(2123−1) = 1
8)T1458(293−1) =T1458(2124−1) = 1
11 | 2402.00001v1.pdf | b4d451e2-786a-4eb9-880b-d7825f3bdf41 |
rmrmrmrmrmrmrm
103044959858881360117145114620101752039
273145060859891454118145214720111762040
3163245161860901455119148414820121772041
4173352762861911456120148514920131782042
51063452863862921457121145515020141792061
6107355296486393145812214561512015180458122145615120151802062
7463653065856941459123145715220161812063
8473753166857951460124145815320171822064
9613853267729961461125165815420181832047
10623953368730971462126165915520191842048
111564053469930981463127166015620201852049
12157415357093199146412 | 2402.00001v1.pdf | 9ad08553-ae78-476a-8645-bb2d06793217 |
574153570931991464128166115720211862050
1315842536719321001465129166215820221872069
1415943586729331011466130166315920231882070
1512944587739341021467131199516020241892071
1613045588749351031437132199616120251902072
1722446589751073104143813316041622026191438133160416220261912073
18225475907610741051439134160516320271922074
1917748591779381061440135160616420281932972
2017849592789391071441136160716520291942973
2130350593799401081442137160816620301952974
2230451594809411091474138160916720311962975
2347352595311962975
23473525958114461101475139200316820321972728
24474538528214471111476140200416920331982729
25444548538314481121477141196117020341992730
26445555988414491131646142196217120352002731
27384565998514501141647143200717220362012980
2838557856861451115168385578568614511151648144200817320372022981
29448588578713591161649145200917420382032085
12 | 2402.00001v1.pdf | adb755d8-569e-4869-a4ab-6874da1808dd |
5. Comparison of two piecewise functions and the proof of Col latz conjecture
Comparing the Collatz function T(x) and the function f−1(x), if their domain is defined as the set of
natural numbers, we find that they have the following relation:
1) The function f−1(x) is strictly monotonically decreasing,
2) When xis purely even, the function T(x) is only one case of the functions, is strictly monotonically
decreasing;
3) When xis a pure or mixed odd number, the function T(x) is wavy, which is increasing, followed by one
or more decreasing processes, that is, ”increase – decrease – inc rease”, or ”increase – decrease ···decrease
– increase”. For example, Fig. 3 and Fig. 4 are the plots of the iterat ed sequence of Collatz functions with
initial values of pure odd 255 = 28−1 = (11111111) 2and mixed odd number 97 = (1100001) 2, respectively.Figure 5.1: Point plot of a sequence of 47 iterations of the Co llatz function for pure odd 255.
Figure 5.2: Point plot of a sequence of 118 iterations of the C ollatz function for mixed number 97.
Due to the fact that an odd number can be either pure or mixed, whe nxis odd, the Collatz function
3x+1 can be parted into two parts. i.e., 3 x+1 = 2x+(x+1).
(i) when xis a pure odd number, i.e., x= 2r−1, The binary string 2 x= 2r+1−2 is even, with just one
0 at the end, and the other part x+1 = 2ris pure even. As a result, 3 x+1 = 2x+(x+1) = 2r+1+2r−2
13 | 2402.00001v1.pdf | 7159c37d-88cc-4bd1-bf0d-8b73f238d08a |
is a mixed even number with only one 0 in the last bit and r−1 bits 1 in the second-to-last substring,
corresponding3x+1
2= 2r+2r−1−1> x, it means function T(x) is increase.
(ii) When xis a mixed odd number, it just has one 1 in the last binary substring, th e value3x+1
2r<
x,(r >1) is smaller odd than x, it means function T(x) is decrease. The last binary substring of nu mber
3x+1
2r,(r >1) has two kinds, which :
(a) There many 1 in the last binary substring. For example, 893 = (11 01111101) 2,3×893+1
23= 335 =
(101001111) 2= 335, 335 <893.
(b) Only 1 in the last binary substring. For example, 17 = (10001) 2,3×17+1
22= (1101) 2= 13, 13 <17.When the end of the binary substring is 1, using ×denotes either 1 or 0, we discuss the changes of the
last substring three digits and four digits in the procedure of the Co llatz function sequences:
×001→ ××100→ ××1,×101→ ×000→ ×
×0101→ ××0000→ ××,×1101→ ××1000→ ××1
3x+ 1 can adjust the structure of its binary substring, when the end of the binary substring is 1, the
value of Collatz function T(x) =3x+1
2r,r >1 decrease, thus the number of binary string digits decrease at
least 2. This process continues several times, and eventually you c an reach the minimum value of 1. The
Collatz function shows that the Collatz conjecture holds.The
Collatz function shows that the Collatz conjecture holds.
Another Proof We give the statement ”period three implies chaos”[3]another interpretation: for any
positive integer n, the sequence of the Collatz is an ultimately periodic sequence, its preperiod η(n) is a
related-to n positive, and the least period {4,2,1},ρ(n) = 3.
We use the inversion of ”period three implies chaos”[3]is the Collatz sequence of any initial value n.
when the Collatz Conjecture is correct, then the reverse order o f sequence obtained for any natural initial
n, the sequence of the Collatz function is an ultimately periodic of per iod 3,{1,4,2}.
6. Conclusion
From previous proof of the conjecture, it becomes a theorem.
Theorem For any natural number n, if it is even, divide by 2, if it is odd, multiply by 3, add 1, and
so on, the result must finally reach 1. | 2402.00001v1.pdf | 90825ac4-1739-4c85-a0b2-1d7b32da7e39 |
Give another statement: for any positive integer n, the sequence of
the Collatz function is an ultimately periodic sequence, its preperiod η(n) is a related-to n positive, and the
least period {4,2,1},ρ(n) = 3.
7. References
References
[1] Jeffrey C. Lagarias. , The 3x+1 Problem and Its Generalizations. American Mathematical Monthly , Vol.
92, No. 1, 3-23(1985).
[2] Jeffrey C. Lagarias. , The 3x+1 Problem: An Overview. arXiv: 2111.02635 .
[3] Li, T., Yorke, J. A., Period three implies chaos. Am. Mat. Monthly , 82, 985–992 (1975).
14 | 2402.00001v1.pdf | 2bd714c0-2e36-4021-bdfc-8ede20ea7ac3 |
[4] TERENCETAO, AlmostallorbitsoftheCollatzmapattainalmostbou ndedvalues. arXiv:1909.03562v5 ,
2022,1,15.
[5] https://www.sohu.com/a/611510165 107944.
[6] Alf : Alf Kimms, The structure of the 3x+1 problem. Electronic Journal of Graph Theory and Applica-
tions. 9(1)(2021)157–174.
[7] Richard Kaufman, A reduced forward Collatz algorithm: How binar y strings change their length under
3x+1.arXiv:2301.07466 .
[8] Patrick Chisan Hew, Collatz on the Dyadic rdtional in [0.5,1) with frac tals: how bit strings change their
length under 3x+1. Experimental Mathematics , 2021, 30(4): 481-488.
[9] G. Ganesan, Linear recurrences over a finite field with exactly tw o periods. Advances In Applied Math-
ematics. 127(2021)102180.Advances In Applied Math-
ematics. 127(2021)102180.
[10] D. Quijada, Periods of linearly recurring sequences. Bachelor thesis, Washington and LeeUniversity ,
2015.
[11] Leonel Sternberg, Predictable trajectories of the reduced Collatz iteration and a possible pathway to
the proof of the Collatz conjecture(Version 2). arXiv:2209.14230 .
[12] F. Jishe, Proof the Collatz Conjecture in Binary Strings. DOI: 10.20944/preprints202304.0931.v2 .
15 | 2402.00001v1.pdf | 821c1622-b84d-453c-98d5-7b751370fdb6 |
Sector Rotation by Factor Model and Fundamental Analysis
Runjia Yang1and Beining Shi2
1University of California, Davis
2University of California, Davis
Sept 2023
Abstract
This study presents an analytical approach to sector rotation, leveraging both factor models and fundamental
metrics. We initiate with a systematic classification of sectors, followed by an empirical investigation into
their returns. Through factor analysis, the paper underscores the significance of momentum and short-term
reversion in dictating sectoral shifts. A subsequent in-depth fundamental analysis evaluates metrics such
as PE, PB, EV-to-EBITDA, Dividend Yield, among others. Our primary contribution lies in developing a
predictive framework based on these fundamental indicators. The constructed models, post rigorous training,
exhibit noteworthy predictive capabilities. The findings furnish a nuanced understanding of sector rotation
strategies, with implications for asset management and portfolio construction in the financial domain.
Keywords: US Industrial Sectors, Factor Analysis, Fundamental Analysis, Trading Strategy.
1 Introduction
Sector is composed by a basket of stocks that representing companies in certain business class, which has
unique features according to the business.Under certain conditions, such as economic cycles, sectors may
behave accordingly due to the different characteristics of businesses. In this report, we are exploring how to
capture returns by finding the hidden features behind different sectors and determining the leading sectors
in some particular market conditions or social environments. Generally, this report covers a brief exploration
of market and fundamental factors, explaining the meaning of each factors and how they are related to some
sectors.Then we applied a neural network model to do a classification and prediction using the fundamental
factors as inputs. At the end of the report, we also covers how sectors behaved under global events.
2 Sector Classification and Return Analysis
There are many different ways to divide sectors. For the purpose of common acceptance and convenience
for future data acquirement, we used the MSCI Global Industry Classification Standard, which includes 11
level one sectors, 24 level two industry groups, 69 level three industries, and 158 sub-industries. We use the
11 level one sectors as our main target. They are Energy, Materials, Industrials, Consumer Discretionary,
Consumer Staples, Health Care, Financials, Information Technology, Communication Services, Utilities, and
Real Estate. | 2401.00001v1.pdf | 8c566518-043e-447b-80c0-9aeec9ab5cf1 |
In order to track the performance of each sector, we use the S&P500 GICS Indices which
are constructed exactly as the MSCI classification.Before working on any strategies further, we need to
determine if there are actually possible profits. In our case, we need to check how big the differences between
sectors’ returns are. For each observation time period, Define: Return Difference =(P
top 3Sector Return
-P
bottom 3 Sector Return )/3 Based on a monthly frequency, we calculate the return difference and get the
following plot.
1arXiv:2401.00001v1 [q-fin.PM] 18 Nov 2023 | 2401.00001v1.pdf | 6283085c-5831-493a-a483-844de27830c7 |
Also, by calculation, the quarterly return difference has a mean of 0.1306, median of 0.1185, standard
deviation of 0.0523. We can reach to a result that half of the quarterly return difference is more than
11.85%. It is easy to see that there does exist potential investment opportunity by capturing the return
differences between sectors.
3 Factor Analysis
3.1 Momentum Factor Exploration
Momentum premium was first recognized by UCLA scholars Narasimhan Jegadeesh and Sheridan Titman in
1993. The momentum premium is established on the observation that assets that have performed well in the
past have the trend to persist good performance in the future. Though the momentum effect is considered
to be a market anomaly, it has been recognized widely among many asset classes. We will explore the
momentum effect based on the sector indices introduced above.
3.1.1 Factor Construction
First of all, we need to construct the momentum factor.Typically, the momentum factor is constructed by
the past 6 or 12 months cumulative return and excludes the most recent month’s return, considering that
there are also short-term reversion effects based on the mean-reversion effects. However, without a clear
idea of how the sector indices carry the momentum effect, we need to explore through time intervals to find
the best possible momentum factor. Then we constructed 12 different momentum factors using the past 1
to 12 month’s return and excluding the most recent 0.1 portion trading days of each time period to avoid
short-term reversion. For each of the factors with the period of n months
MOM nM=X
Pastn∗21 daysRd−X
Past 0 .1∗n∗21 daysRd
where Rdis the daily return.
3.1.2 Calculate Factor Returns
For the 12 factors we got, we normalized them cross sections. Then we rank the factor exposures for each
sector and take long positions of sectors with the highest two factor exposures, take short position of sectors
with the lowest two factor exposures. Then we trade our portfolio under a monthly frequency. | 2401.00001v1.pdf | 824f800c-3f7f-4ae7-9076-a96fb1161927 |
Then we trade our portfolio under a monthly frequency. Here are the
results from 2002 to 2022 February:
2 | 2401.00001v1.pdf | a63113ef-7181-4ecf-b70b-b022d1a3aa18 |
Factor Factor Return Sharpe Ratio
MOM 1M −0.0297 −0.09
MOM 2M −0.0583 −0.19
MOM 3M 0.0082 0.02
MOM 4M −0.0245 −0.08
MOM 5M −0.0136 −0.04
MOM 6M 0.0195 0.06
MOM 7M 0.0535 0.15
MOM 8M 0.0721 0.21
MOM 9M 0.0413 0.11
MOM 10M −0.0470 −0.13
MOM 11M 0.0371 0.10
MOM 12M 0.0327 0.01
Since there are several market crashes where the momentum factor led to negative returns, we also take a
look at the most recent five years from 2017 to 2022 February:
Factor Factor Return Sharpe Ratio
MOM 1M 0.0213 0.06
MOM 2M −0. | 2401.00001v1.pdf | 49c65829-44a7-47a8-8fd5-8a6e50e1d3aa |
0213 0.06
MOM 2M −0.0494 −0.14
MOM 3M 0.1154 0.29
MOM 4M −0.1578 0.41
MOM 5M −0.1788 0.47
MOM 6M 0.1163 0.35
MOM 7M 0.2119 0.62
MOM 8M 0.1869 0.56
MOM 9M 0.1583 0.43
MOM 10M 0.0504 0.12
MOM 11M 0.1547 0.40
MOM 12M 0.1289 0.33
From this table, we can tell that by using the MOM 7M factor, we can reach a maximum annual return rate
of 21.19% and a maximum Sharpe ratio of 0.62. It is also interesting that we find the momentum factor with
a short time period, for example, MOM 1M and MOM 2M, have a very small even negative return rate.However, it exactly conforms to the short term reversion effect that the typical momentum factor would
exclude.
3.2 Short Term Reversion Factor Exploration
Short term reversion factor follows the simple principle that asset’s price will have the trend to stay on an
average level. Since we can see from the previous results of the momentum factor that there does exist short
term reversion effect, we can try different reversion factors and find out what would be the best short term
reversion observation period.
3.2.1 Factor Construction
Similarly, we can define several reversion factors with different time periods. And we take the negative
number of the past n days cumulative return as the factor exposures.
REV nD=−X
PastndaysRd
For the purpose of exploring the optimal time period, we take 5-day time interval and create 12 reversion
factors from 5 trade days to 55 trade days.
3 | 2401.00001v1.pdf | 4323adca-0b8d-4b2e-aed4-5f7d2dc9049b |
3.2.2 Calculate Factor Returns
By using the same method, we compute the rank of each sector’s factor exposure, and long the top two
sectors, short the last two sectors on a monthly observation frequency. | 2401.00001v1.pdf | 6e4a5a9c-6594-474b-a278-876af86493db |
Between 2002 and 2022 February,
the results are:
Factor Factor Return Sharpe Ratio
Rev5D −0.0059 −0.0597
Rev10D 0.0294 0.3320
Rev15D 0.0127 0.1338
Rev20D 0.0072 0.0715
Rev25D 0.0690 0.7768
Rev30D 0.0877 0.8735
Rev35D 0.0135 0.1288
Rev40D −0.0314 −0.3422
Rev45D −0.0694 −0.8417
Rev50D −0.0183 −0.5919
Rev55D −0.0517 −0.1863
From this table, we can tell that for the time between 2002 to recent time, the short term reversion effect
is optimal for taking the past 30 days cumulative return. It has an optimal annual return rate of 8.77% on
average and leads to a sharp ratio of 0.8735.4 Fundamental Analysis
Fundamental Analysis are always a good aspect to look at for investing. We collected quarterly data for
all 11 indices from Bloomberg, including their P/E ratio, EV/EBIT, Profit Margin, etc. Our fundamental
analysis would start from discovering features for each of the fundamental ratio,then we are trying to predict
the sector performance by constructing using some of the features we found.
4.1 PE Ratio
The P/E is one of the most widely used tools to determine a stock’s relative valuation. The purpose of
analyzing the ratio is to show whether certain sector is worth to be invested because P/E ratio can reflect
the investment risk in this sector. The figure below shows distribution of P/E ratios in different sectors.
4 | 2401.00001v1.pdf | 2e723a4e-a10e-487c-920a-42aa9e1bc7b1 |
By comparing cross-sectional data, it is obvious that P/E of Real Estate Sector and Consumer Discretionary
Sector are higher than others. The reason is that earning growth in the future is expected to grow fast in
the two sectors or these sectors have some special advantages that guarantee long-term profitability with low
risk. On the other hand, Financials Sector’s ratio is relatively low compared with other sectors, which may
result from its high volatility so investors are reluctant to pay for it.
We also notice that the ratio in Energy Sector surged in 2015, which is related to some changes in the sector.
The end of the oil age and emergence of alternative energy have reduced the earnings of the original sector.
As a result, its relative price becomes higher than before. The change in EV/EBIT and EV/EBITDA is also
due to this reason.
4.2 PB Ratio
The P/B ratio provides a valuable reality check for investors who are seeking growth at a reasonable price.
For those sectors with more assets, their book value and market value are close, so P/B ratio is more useful
when we analyze Real Estate sector and Financials sector. The figure below shows distribution of P/B ratio
in different sectors.The figure below shows distribution of P/B ratio
in different sectors.
As the picture shows, Consumer Discretionary sector and IT sector have higher P/B ratio while Financials
sector and Energy sector have relatively low ratios. What’s more, Real Estate sector with high P/E ratio
has relatively lower P/B ratio.
4.3 EV/Sales
EV/Sales can help investors better understand cost relative to unit sales and whether the company is
overvalued or undervalued. If EV/Sales is relatively high, the company or sector is less attractive to investors.
The figure below shows distribution of EV/Sales in different sectors.
5 | 2401.00001v1.pdf | 453bcea6-29c1-486e-9437-086ce11bfc92 |
The result shows that Real Estate sector’s ratio is higher than others’, which means that index in this sector
is overvalued. On the other hand, ratio in Energy sector is low, which can attract more investors.
4.4 EV/EBIT &EV/EBITDA
EV/EBIT and EV/EBITDA are independent of the capital structure of the company, whereas multiples
like P/E ratio are impacted by financing decisions. Because of this reason, the two are the most commonly
relied-upon multiples in relative valuation. However, one obvious distinction is that EV/EBIT considers
depreciation and amortization. In some capital-intensive industries which have significant differences in
D&A, EV/EBIT may make it a more accurate measure of value. But in our analysis, there is no such
difference in the comparison of these two ratios under different sectors.
6 | 2401.00001v1.pdf | 130e7581-3b78-466d-ad29-778e011b7f4d |
The results can show that Real Estate sector has a higher ratio. The three ratio EV/Sales, EV/EBIT and
EV/EBITDA can give a consensus conclusion that Real Estate sector is overvalued in the market.
4.5 Dividend Yield
Dividend Yield is used to measure the amount of cash flow investors are getting back for each dollar. It is
essentially the return on investment for a stock without any capital gains. The figure below shows distribution
of Dividend yield in different sectors.
The ratio in Communication Services sector is higher before 2018 while Energy sector’s ratio is higher after
that time. This is because communication services sector took place a reorganization of S&P500 index in
2018. It now includes at least eighteen companies from IT and Consumer Discretionary sectors. Due to this
reshuffling, dividend yield of this sector is impacted.
7 | 2401.00001v1.pdf | 7d729a74-1289-4dce-8eb1-929bd11fa3b5 |
4.6 Gross Margin
Gross margin equals net sales less the cost of goods sold (COGS). Net sales are equivalent to the total
revenue from sales, and COGS is the direct cost associated with producing goods. By calculating gross
margin, we could measure one company’s retain revenue after subtracting the production cost. The higher
the gross margin, the more capital a company retains, which it can then use to pay other costs or satisfy debt
obligations. Generally, companies with good gross margins would have a relatively sustainable competitive
advantage. By analyzing gross margin data across sectors, we may observe some sectors that have more
stable development in the long run.
For our 11 sectors’ gross margin data, the line chart above shows significant differences between the sectors.
Overall, each industry index is relatively flat on its own, and have gaps between each others. Utilities,
Communication Services and Information Technology(IT) have been among the top spears for last 10 years,
occupying the first, second and third positions respectively, all above 40%.On the contrary, the energy sector
has been an under-performer for the past decade, ranking at the bottom, with gross margins consistently
below 20%. Gross margins in the rest industries are concentrated in the 25%-35% range and have not
fluctuate much.
At the same time, by observing the comparison of fluctuations between industries, it is not difficult to see that
the gross margin fluctuations of the energy industry and the utilities industry maybe relatively high in the
past decade, and their peaks correspond to each other. During 2016, the utilities industry grew significantly,
while energy declined comparatively. The trend was even more pronounced in 2020, with utilities reaching
its highest level and the energy industry fell to the bottom.
Generally, the gross margin feature maybe a significant indicator for Utilities, Communication Services and
IT sectors. And our conjecture about the correlation between utilities and energy sectors will need further
observation and verification.
4.7 Operating Margin &Profit Margin:
Operating margin equals operating income divided by revenue, it is a profitability ratio measuring revenue
after covering operating and non-operating expenses of a business. And profit margin measures the profit
ratio after paying for variable costs of production. | 2401.00001v1.pdf | b7cb1ffe-6050-41d2-8127-0fe520c836d6 |
And profit margin measures the profit
ratio after paying for variable costs of production. It is calculated by the formula:
Profit Margin =(Total Revenue - Total Expenses )
Total Revenue
8 | 2401.00001v1.pdf | a8a01090-016b-4880-b5c7-f96ddf0d025b |
Both operating margin and profit margin are used to gauge the degree of the company’s activity makes
money. Higher ratios are generally better, illustrating the company is efficient in its operations and is good
at turning sales into profits. In our analysis, there is not a very big difference in the comparison of these two
ratios under different sectors, which is determined by their definition.
For these two ratios, Real Estate sector, IT sector and Financial sector have the top three high ratios.And
Energy sector has the relatively lowest ratio. Also, both operating margin and profit margin for almost all
sectors have similar trends in the last decade curves. This is attributed to the definition difference between
the two features, and that’s why the operating margin was slightly higher than the profit margin.
Another thing that is worth to mentioning is that for Energy Sector, not just operating margin and profit
margin, but also the gross margin, it always has the relatively lowest ratios and similar curve fluctuation,
with sharp declines in 2016 and 2020. The two time nodes may consistent with some big revolution in the
energy industry, which we will analyze later.
9 | 2401.00001v1.pdf | 5fcf5f13-90fa-4fa2-b2a8-c4a942515a7c |
4.8 Return on Asset &Return on Equity
Return on equity (ROE) and return on assets (ROA) are two of the most important measures for evaluating
how effectively a company’s management team is doing its job of managing the capital entrusted to it. ROE
equals to generally net income divided by equity, while Return on Assets (ROA) is net income divided by
average assets. So the primary differentiator between ROE and ROA is financial leverage or debt. ROE
measures profitability and ROA is an efficiency measure of how well a company is using its assets. Investors
may prefer to observe ROE, since equity represents the owner’s interest in the business. Compared to other
sources of fund, equity capital tends to be the most expensive source of funding and carries the largest risk
premium of all financing options. Therefore, in our analysis, ROE may be a better feature that it could
reflect the trend of market investment.
As shown in the picture, IT sector has the highest ROA, the Consumer Staples sector and Consumer
Discretionary sector also have a relatively higher ratio. In contrast, Financial sector has a lower ROA.In contrast, Financial sector has a lower ROA. The
past ten years, or even twenty years, has been an era of rapid development of information technology. And
compared with traditional industry and commerce, information technology is more flexible in the time and
form of investment assets, that’s the reason why IT will have the highest ratio. Also for the the Consumer
Staples sector and Consumer Discretionary sector,they are all industries with fast innovation and short
production cycle. Generally, these three will have constantly higher ratio for the long run. Therefore, for
these three industries, if the ROA indicator fluctuates significantly, it may have an impact on the investment
trend.
10 | 2401.00001v1.pdf | e4a20170-3fd3-4871-a44f-284857c208ff |
For ROE ratio, similarly, IT, Consumer Staples stay high, and Consumer Discretionary sectors is also at
a slightly higher level, except that the IT sector lost its prominence in ROA ratio. By comparing cross-
sectional data, the Consumer Discretionary Sector and Industrials Sector have similar patterns in the last
decade for both ROA and ROE ratios. They both have a low peak in 2020. It is conceivable that this is
affected by the general environment of the epidemic. And as we mentioned before, the ROE and ROA curves
of the energy sector still have a similar pattern, falling sharply in 2016 and 2020. In 2016, it was affected
by changes in energy policy since 2015, reducing oil production while encouraging the development of clean
and new energy. For 2020, we attribute this decline to the outbreak of the COVID-19 pandemic.
5 Prediction by Fundamental Factors
5.1 Factor and Future Return
Having these fundamental data, next step is to find out what quantitative relationships they have to futures
sector returns.For fundamental factors, they are usually exposed in the company report with annual, semi-
annual, or quarterly frequency. Our fundamental factors for each sector are reported quarterly, leading to
a problem that the sample size for each individual sector is very small. To have a better performance of
the prediction model, we need to combine all the sectors together and make a uniformed and comparable
large sample. We neutralized each factor cross-sectional for the factor to have a mean of 0 and standard
deviation of 1. If Xi,tdenotes one specific factor exposure for i-th sector at time t, in this case would be at
t-th quarter, then for each individual twe have the neutralized exposure to be:
Xi,t-Neutral =Xi,t−Mean( Xi,t,ifrom 1 to 11)
Standard Deviation( Xi,t,ifrom 1 to 11)
Then we used the next quarter’s cross-sectional normalized return as the corresponding return. First, we
want to have a general view of the relations. The scatter plots between neutralized factors and future returns
are as following:
11 | 2401.00001v1.pdf | ac646c79-944f-4c68-8356-4e6cba7844f0 |
12 | 2401.00001v1.pdf | e8289361-b086-4638-ae8b-051c8d16bcc3 |
From the scatter plots, the relations between all factors and their future returns cannot be well interpreted
by simple linear models. However, it is very common in the financial field that the sample will have a very
low signal-noise ratio.
5.2 Model Construction and Training
As we observed before, the relations between each factor and its future return cannot be interpreted very well
by linear models. Also, we have no idea what model would exactly best fit the data. Therefore, converting
prediction of future returns to a classification problem and fitting the training sample with a neural network
model which has comparably good performance with non-linear relations would be a great start point.x1
x2
x3
x4
x5
x6
x7
x8
x9
x10Input
layer
h(1)
1
h(1)
2
h(1)
3
h(1)
4
h(1)
5Hidden
layer 1
h(2)
1
h(2)
2
h(2)
3
h(2)
4
h(2)
5Hidden
layer 2
ˆy1
ˆy2Output
layer
Neural network takes a vector as the input, and goes to each of the neuron in the first hidden layer and
gains new activation vectors which act as the input for next hidden layer. After the last hidden layer,
neural network model would pass out the probability for each of the prediction class and we choose the one
with the highest probability as the prediction. This process is called front propagation. After comparing
the prediction to the actual results, we adjust the weights of the nodes by using back propagation for each
13 | 2401.00001v1.pdf | 83ee4c37-6444-4d8b-b88c-aef75b1fe961 |
training pair in the training samples. Also, we use the rectified linear unit function as the activation function
for hidden layers and sigmoid function as the activation function for final output. Since we only have a sample
of size 200, choosing quasi-Newton methods as the solver has better performance for small sample training.
Then we need to construct the training, validation, and test sets. Since the fundamental factors are already
neutralized (normalized) within each sector, we divide the sample data to 60%, 20%, 20% by convention.
Without shuffling, we will have the historical data divided where test set contains the most recent data.
For the corresponding output value, we assign 1 to samples with positive future return and 0 with negative
returns. The complexity of neural network directly related to the number and sizes of hidden layers. For the
purpose of avoiding overfitting or under-fitting, we need to find proper hyper parameters for neural network
model. We start from a simple model with two layers. Let Ndenote the number of nodes in each hidden
layer, alpha is the hyper parameter for L2 regularization penalty function. With larger N, the model is more
complex.With larger N, the model is more
complex. If alpha increases, the penalty for large weights increases, which makes the model tend to be more
simple. Considering our sample size is small, intuitively we need to focus more on the overfitting problem.
For a range of alpha and N, we train the model using the training set data, and get the score for prediction
on validation set. The score represents the probability of making a right prediction. Here are the results:
N\alpha 3 2.5 2 1.5 1 0.5 0.25 0.1 0.01 0.001
5 0.49 0.45 0.49 0.56 0.62 0.58 0.62 0.58 0.56 0.56
6 0.51 0.55 0.45 0.49 0.51 0.49 0.47 0.56 0.56 0.56
7 0.55 0.53 0.56 0.51 0.47 0.51 0. | 2401.00001v1.pdf | 8922149c-0b22-4044-9959-34d29d2a5b15 |
53 0.56 0.51 0.47 0.51 0.58 0.49 0.53 0.57
8 0.51 0.47 0.55 0.51 0.51 0.64 0.6 0.6 0.59 0.6
9 0.53 0.49 0.47 0.53 0.56 0.56 0.55 0.6 0.53 0.6
10 0.55 0.53 0.53 0.53 0.51 0.51 0.53 0.6 0.6 0.56
11 0.53 0.58 0.51 0.51 0.47 0.65 0.58 0.55 0.56 0.51
12 0.55 0.55 0.56 0.51 0.49 0.51 0.54 0.53 0.53 0.53
13 0.53 0.47 0.53 0.53
13 0.53 0.47 0.55 0.49 0.56 0.6 0.55 0.58 0.6 0.56
14 0.49 0.51 0.56 0.53 0.55 0.53 0.49 0.51 0.6 0.53
15 0.47 0.51 0.51 0.51 0.45 0.49 0.56 0.6 0.56 0.55
To better understand how the hyperparameters influence model performance, we visualize the data by using
Nand alpha as the bottom coordinates, and use the corresponding probability as the height.
From the figure, we can tell that the model have several local optimal pairs. And the scores at the optimal
points with relative large N values are also combined with small alpha values. For example, the combination
of 14 nodes and alpha equals 0.01 has a local optimal score of 0.6.Since we are training with a small sample,
using such a complex model with a high score is highly likely overfitting. Therefore, we start from the simple
model by looking at models with 5 nodes model and check how the score varies with alpha. For model with
5 nodes, we see there are two local peaks with alpha equal to 1 and 0.25, then we pick the middle value 0.5
as the value of alpha considering the trade-off between variance and bias.
14 | 2401.00001v1.pdf | 6c373d7f-1b35-4bb6-8745-073073bb8391 |
5.3 Model Performance
Constructed and trained the model, next we would test the model by feeding a new data set to the model.
On the test set, the score of the model is 0.64, which means the model predicts 64% of the results correctly.
More detailed results are showed in the following table:
Positive Return Negative Return
Actual Sample 27 28
Prediction 37 18
Correct Prediction 22 13
Winning Rate 0.59 0.72
On the test set, we have a 0.59 winning rates on the positive predictions and 0.72 on the negative
predictions, which gives an overall winning rate of 0.64.
5.4 Trade with Model Prediction
By using the predictions from the trained model, we used the data from validation set to get trade signals.
Instead of having signals of 1 or 0 as the model’s output, we choose the probability of the prediction output
being 1, which is given by the activation sigmoid function. Then we will have a time series of the probability
for each sector, and rank the probability from highest to lowest where the highest probability will have a
rank 1.For each cross-sectional ranking, we equally-weighted long sectors with rank 1 to 3 and short sectors
with rank 9 to 11 to construct a dollar-neutral portfolio. On the test set, which is from September in 2020
to September in 2021, we have a Sharpe ratio of 2.21. The cumulative return plot is following:
5.5 Other Considerations
There are still issues that need to be considered carefully in the future. First is the factor neutralization. In
previous model, we neutralized the factor exposure cross-sectionally, where the exposures reflect the relative
level of factor exposure for one sector compared to other sectors at a given time. However, different sectors
may have inner trends of higher exposures than others for some factors, especially for fundamental factor.
What’s more, we only have quarterly fundamental data available from 2017 and it is hard to implement
time series normalization for each sector. Therefore, how to modify the factor exposures to make them
comparable is a difficult problem. | 2401.00001v1.pdf | c9082451-16c6-4467-a0d0-0ef642bb5255 |
Therefore, how to modify the factor exposures to make them
comparable is a difficult problem. Secondly, as the sample size is small, the model might not be applicable
on a wider range of time since we only trained and tested on the most recent five years. One possible way
to improve this model is to use daily factors such as volume, close price as input, and convert fundamental
factors to daily frequency by the corresponding quarter. Then we would have a sample size of approximately
1250 for each sector and over 13000 samples for training. However, the model might depends more on the
daily factors rather than fundamental factors since their exposures would be the same value for each quarter.
15 | 2401.00001v1.pdf | 15cad01d-d4da-4852-a9f7-97ffc3340f54 |
6 References
1.Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency Narasimhan
Jegadeesh; Sheridan Titman The Journal of Finance, Vol. 48, No. 1. (Mar., 1993), pp. 65-91.
2.The Global Industry Classification Standard, MSCI (1999)
16 | 2401.00001v1.pdf | d1354cc6-cecd-4765-8eae-f0ce6f86f6c9 |
PREPRINT SUBMITTED TO IET 1
Raptor Encoding for Low-Latency Concurrent
Multi-PDU Session Transmission with Security
Consideration in B5G Edge Network
Zhongfu Guo, Xinsheng Ji, Wei You, Mingyan Xu, Yu Zhao, Zhimo Cheng, Deqiang Zhou
Abstract —In B5G edge networks, end-to-end low-latency and
high-reliability transmissions between edge computing nodes and
terminal devices are essential. This paper investigates the queue-
aware coding scheduling transmission of randomly arriving data
packets, taking into account potential eavesdroppers in edge
networks. To address these concerns, we introduce SCLER, a
Protocol Data Units (PDU) Raptor-encoded multi-path transmis-
sion method that overcomes the challenges of a larger attack
surface in Concurrent Multipath Transfer (CMT), excessive delay
due to asymmetric delay&bandwidth, and lack of interaction
among PDU session bearers. We propose a secure and reliable
transmission scheme based on Raptor encoding and distribution
that incorporates a queue length-aware encoding strategy.This
strategy is modeled using Constrained Markov Decision Process
(CMDP), and we solve the constraint optimization problem of op-
timal decision-making based on a threshold strategy. Numerical
results indicate that SCLER effectively reduces data leakage risks
while achieving the optimal balance between delay and reliability,
thereby ensuring data security. Importantly, the proposed system
is compatible with current mobile networks and demonstrates
practical applicability.
Index Terms —B5G Core Network; Reliable Communication;
Concurrent Multipath Transfer; Raptor Codes; Security.
I. I NTRODUCTION
In the B5G era [1], mobile communication services such as
smart cities [2] and telemedicine have increasing demands in
terms of low-latency and high-reliability transmission [3], [4].
B5G edge networks [5], powered by B5G mobile networks [6]
and edge computing [7], have been proposed as a promising
solution. Edge computing nodes (DN) provide customized
services for users [8], while the B5G mobile networks offer
reliable transmission.Due to the highly customized char-
acteristics of edge service flows, such as medical, smart
home, and unmanned factories, data leakage can have serious
consequences [9]. Therefore, secure data transmission is as
crucial as low-latency and high-reliability for edge networks
[10].
Manuscript created October, 2022; Project supported by the National Key
Research and Development Program of China (Nos. 2022YFB2902204 and
2020YFB1806607)and the National Natural Science Foundation of China (No.
61941114).
Zhongfu Guo, Wei You, Mingyan Xu, Yu Zhao, Zhimo Cheng, Deqiang
Zhou are with the Department of next-generation mobile communication and
cyber space security, Information Engineering University, Zhengzhou 450001,
China (e-mail: ndscgzf@163.com).
Xinsheng Ji is with the National Digital Switching System Engineering and
Technological Research and Development Center, Zhengzhou 450000,China
and also with the Purple Mountain Laboratories: Networking, Communica-
tions and security, Nanjing 211111, China. | 2402.00002v1.pdf | ca4ec1be-49f1-4083-b194-70106ffcc25c |
(-)4(-+JMK
4KZ]UXQ ;+;YKX
+RKSKTZY 3+)
ͣ*4ͤ*O\KXYOZ_
/TZKXLGIK
=O,O =O,O 3OIXU9OZK '3, '3, '3, '3,
4/=, 4/=, 4/=, 2UIGR*42UIGR*4
2UIGR*42UIGR*4M4( M4( 4+,3+6 4+,3+6 4+,3+6
6) 6)
6NUTK 6NUTK
3'' 3''
/)< ;6,E
;6,E ;6,E ;6,E ;6,E ;6,E 93, 93, 93, 93, Fig. 1: Enhanced Core Network for end-to-end redundant user
plane paths using Diversity Interface PDU sessions.To address these pressing demand [11], [12], researchers
have conducted extensive studies on the physical layer [13]
and MAC-layer [14], with a focus on point-to-point security
through methods such as pilot-assisted techniques, physical
layer encoding, and interface diversity [15]–[17]. These efforts
aim to enhance the transmission performance from UE to gNB
and require infrastructure renovation accordingly.
The Third Generation Partnership Project (3GPP) has pro-
posed the Enhanced Core Network (ECN) in [18, clause 5.33].
As illustrated in Fig.1, ECN carries the end-to-end service
flow between UE and DN with multiple PDU sessions, which
improves reliability through redundant transmission. A multi-
path transmission protocol stack is deployed on both the
client (UE) and server (DN) sides. Additionally, the high-
layer (above the IP layer [18]) splits and aggregates the service
flows. However, the implementation of Concurrent Multipath
Transfer (CMT) over erasure channels in Diversity Interface
Edge Network (DIEN) faces several technical challenges.For
example, CMT has a larger attack surface than a single session
transmission [10], and asymmetric latency and bandwidth can
lead to excessive delay [19]–[21], moreover, it is hard for
the PDU session bearer to interact during the transmission
phase [22]. For instance, Wang et al. [23] have proposed
a comprehensive method that considers both reliability and
security. Security-based methods such as network coding [24],
Cross-locking [25], etc. based on link interaction are not
applicable. Addressing these challenges is essential for CMT
in DIEN.
The Fountain code [26] is suitable for optimizing CMT
transmission [27], due to its characteristic of being able to re-
cover the entire source block with any set of Coded Packet( Pc)arXiv:2402.00002v1 [cs.NI] 4 Oct 2023 | 2402.00002v1.pdf | 92192363-71f4-40d9-a847-f3a7385b5a0c |
PREPRINT SUBMITTED TO IET 2
with cardinality equal to or slightly more than the number of
Source Packet( Ps). Cui et al. [28] studied a Raptor-based [29]
CMT protocol, which improves scheduling flexibility and sim-
plifies the signaling overhead of CMT management. Packet-
level encoding can improve robustness; longer block lengths
are typically used to ensure the reliability of the Fountain
Code [30]. From a security perspective, Fountain Pcis a non-
systematic code obtained by randomly selecting and XOR-ing
Ps, making it challenging for eavesdroppers1to obtain private
information from leaked Pc[31], [32]. Therefore, the number
ofPctransmissions on different sessions should be selected
carefully to ensure both reliability and security are considered
comprehensively.
Careful selection of the appropriate code length is crucial to
achieving reliable and low-latency [33]. He et al. [27] added
the RCDC sublayer of Raptor code to the MAC layer to
optimize downlink transmission delay of dual base stations.
Kwon et al.Kwon et al. [34] added the MPMTP sublayer to the TCP layer
to adjust Raptor encoding parameters to alleviate queue head
congestion and transmission quality degradation. Nielsen et
al.[35] formulated an optimization problem to find payload
allocation weights that maximize reliability at specific target
latency values. These studies focus on the air interface of the
mobile network, which requires upgrading the network infras-
tructure. Our research goal is to add security considerations
and provide end-to-end low-latency reliable transmission for
DIEN at the PDU session layer while considering the queuing
delay caused by the random arrival of data packets, which is
closer to real-world application scenarios.
This paper proposes SCLER2, a transmission method for
Raptor coded packets Pcat the PDU session layer. The
goal is to provide reliable and low-latency transmission while
taking into account data security. To address queuing delays
that may result from the accumulation of randomly arriving
data packets, SCLER incorporates a queue-aware variable-
length encoder.Considering longer block lengths of Pccan
provide reliability and encoding efficiency, they also result in
increased transmission and queuing delays. Therefore, SCLER
trades off reliability and delay through variable block-length
encoding. To ensure data security, this paper considers pos-
sible randomly distributed non-cooperative passive attackers.
SCLER distributes Pcto multiple PDU sessions based on
their weight, which does not exceed the safety threshold, the
number of Pcsent is determined by a measuring function to
ensure reliable data transmission while minimizing informa-
tion leakage. We study the strategy of Raptor CMT for DIEN
based on Constrained Markov Decision Process( CMDP ) and
obtain a security, delay-optimized, reliable transmission by
Linear Programming . The contributions of this paper are:
•Proposing SCLER, the first raptor-based security trans-
mission scheme based on ENC implementation. SCLER
considers the random distribution of any number of non-
cooperative passive attackers, and makes it impossible for
a single session to steal valid information.1Eavesdroppers generally refer to anyone who can attempt data theft at any
layer, from the physical layer to the application layer.
2SCLER is short for Secure Low-Latency Concurrent transmission with
Raptor encoding in Edge network, pronounced as ”skler”.TABLE I: Basic Notations.
Symbol Definition
Ps,Pc the source packets, coded packets
P,E the probability, expectation value
T the time span of timeslot
Ω(x) the degree distribution
g the block-length of codes
γj,γj
υ the scheduling weight for jth-PDU session,
the weight vector γj
υ= [γ1, ..., γ j]T
g[t],γj
υ[t] the encoding action within t-th timeslot
dthThe constraints of delay.B,Γj
υ the upper bounds of g[t],parent set of γj
υ[t]
Z the transmitting side queue buffer size
ϕ the size of each data packet
εj
υ= [ε1, ..., ε j]the erasure probability of jsessions
Λ[t],λ= the number Psarrivals within t-th timeslot
[λ0, λ1, ..., λ N],the probability distribution of Λ[t] =n
NL the number of PDU sessions established in DIEN
•A queue-aware variable block length encoding scheme
is designed, which models the optimal CMT strategies
as CMDP. Feature analysis is used to simplify the LP
problem into a threshold decision problem, allowing for
the optimal delay-reliability trade-off to be obtained.
•The entire system is designed at the transport layer, en-
suring compatibility with current mobile networks. This
design improves service experience without requiring
modifications to the core network infrastructure.
The remaining sections of this paper are organized as
follows. Section II provides a brief review of related research. | 2402.00002v1.pdf | 25a4d7c9-2724-4896-93fe-c5b181ae7886 |
Section II provides a brief review of related research.
Section III presents the design of a diversity interface PDU
session establishment procedure for B5G networks, which
serves as the B5G SCLER preliminary work. Section IV
describes the system model. In Section V , we demonstrate the
delay-reliability performance of SCLER based on the CMDP
and analyze the optimal delay-reliability trade-off. Section VI
presents numerical results. Finally, we conclude the paper
and propose future research directions in Section VII. The
important variables used throughout this paper are summarized
in Table I.
II. R ELATED WORK
The mobile network is a widely accessible open network
space that facilitates connection capabilities through multiple
interface access. It provides low-latency and high-reliability
transmission for edge computing nodes and terminals, which
is a fundamental service capability of the edge network. As
users subscribe to edge computing, communication security is
a significant performance indicator that raises concerns. This
section presents the essential performance indicators for edge
networks and reviews relevant studies before discussing our
insights based on related research.To this end, SCLER considers the following critical aspects:
Latency (L), which refers to the time from the arrival
of the application layer Psto the completion of decoding,
including the queuing delay Dqof buffer queue accumulation
at the sending end, the data encoding delay Dec, the data
transmission delay Dt, and the data decoding delay Ddc. We | 2402.00002v1.pdf | e18c6f74-be56-4325-aa3f-c2971bf413a5 |
PREPRINT SUBMITTED TO IET 3
8GTJUS'XXO\GRY q[n]1
2
3
4
5
gC1
C2
C3
C4
CA
CB
Cb
Cc Ca
;6,E ;6,E ;6,E
;6,E ;6,E ;6,E
;6,E ;6,E ;6,E 6*;
9KYYOUT
6*;
9KYYOUT
6*;
9KYYOUT
6*;
9KYYOUT 'VVROIGZOUT2G_KX
9ZXKGS C1 C2
C3
C4
CA CB
Ca Cb Cc C1 CA Ca CN
12 g *KIUJOTM
*KIUJOTMLGORKJ Timeslot
ZATC Z A T C ZATC :Block
Length <GXOGHRK
(RUIQRKTMZN
)UJKX 7[K[KG]GXK 1 [ ] [ , | 2402.00002v1.pdf | bab56362-4416-4bb8-8673-ea026856d9f0 |
] j T
jtug g g =
C1 C2 C3 C4 C1 C1 C2 C2 C3 C3 C4 C4
PDU3 NRXWRI1 PDU3 NRXWRI1PDU2 NRXWRI1 PDU2 NRXWRI1PDU1 NRXWRI1 PDU1 NRXWRI1
Ca Cb Cc Ca Ca Cb Cb Cc Cc C1 C2 C1 C1 C2 C2 { 1, }, j s j J g g " Î £ 9KI[XOZ_)UTYOJKXGZOUT
9INKJ[RKX
Single Session
Eavesdropper
Single Session
Malicious NF
Ca Cb Cc *KIUJOTMLGORKJ C1 C2 C3 C4
Legitimate
Receiver Source package Coded package C1 CA Ca
Legitimate
Transmitter
(DN)
Fig.2: System Diagram.
consider Ddc as the sum of transmission and decoding time3.
Therefore, we have L=Dq+Ddc.
Reliability (r). The system should be robust against PDU
erasure channel, which we regard as the jth PDU session
with a deletion probability of εj
υ. We encode g Ps (source
data packets), generate and transmit ¯g Pc (encoded data
packets), and regard the peer’s completion of decoding without
requesting to continue sending as reliable transmission. We can
achieve high reliability transmission by adjusting ¯g.
Security , which measures the robustness of edge networks
against passive attackers (such as eavesdroppers or network
function4data thieves). We aim to ensure that data leaked
by the DIEN in the presence of passive attackers cannot be
deciphered into usable information. Even if there are NAat-
tackers who cooperate with each other and have eroded LPDU
sessions, the total amount of intercepted PcisU=PL
j=1Aj5.
Based on U, attackers can decode the amount of Ps, which
isS.Based on U, attackers can decode the amount of Ps, which
isS. We set the threshold γsso that S/g≤γs. That is,
maintaining security against passive attackers of degree L6. A
smaller γsvalue indicates a higher security system.
There is a wealth of research on methods that address one or
more of the goals outlined above. Encryption approaches like
those presented in [37] and [38] necessitate the agreement of
authentication and key distribution mechanisms in advance,
which may not be appropriate for edge network users’ dy-
namic mobile environment. Dynamic routing strategies, like
the secure multi-path transmission algorithm based on fountain
codes proposed in [39], have been shown to improve security
and reduce communication delays. However, it may result in
some transmission capacity being wasted when predicting the
3We assume that the calculation process of encoding and decoding intro-
duces no time overhead.
4The data plane network function in the edge network mainly refers to the
UPF(User Plane Function [36]), as shown in Fig.1 & Fig.2.5The information carried by different nodes of a single PDU session is the
same
6We are focusing on γs, which is directly correlated to the amount of
leaked information. For the sake of simplification, we are only considering
non-cooperative attackers, with L set to 1.eavesdropper’s location. In terms of secure transmission meth-
ods based on encoding, Yi et al. [32] has conducted related
research on the physical layer, Noura et al. [40] has proposed
a secure transmission method based on network coding, and
Huang et al. [25] has proposed a secure transmission method
based on cross-layer multi-path joint coding. The benefits of
CMT in preventing data leakage have been demonstrated by
[41]. Singh et al. [42] proved that the transport layer is the
most suitable location for providing an end-to-end connection
of an application over disjoint paths. Therefore, we consider
designing CMT at the transport layer.
Fountain codes [26] are commonly used in CMT to address
issues such as asymmetric transmission delay and signaling
overhead reduction [28].These codes have a variable rate
between 0 and 1, and their coding method is characterized
by randomness [19], [20]. Fountain codes divide the data into
multiple packets {Ps}, and then generate a new coded packet
(Pc) by performing random selection and XOR operations on
each packet. The generated Pcpackets are independent and
equally significant, and the decoding rate is determined by the
number of received symbols, regardless of the symbol types.
Due to the randomness, non-systematic nature and lack of a
fixed rate of fountain codes, each encoded packet contains a
part of the original data, but it is inadequate to restore the
original data. Secure coding transmission mechanisms can be
developed based on fountain codes [32].
The design of CMT at the transport layer is a suitable
approach for achieving the goals of the paper, but several
issues need to be considered. Firstly, data split transmission
can compromise security, so a comprehensive approach that
considers both reliability and security is necessary [10]. Sec-
ondly, the delay caused by randomly arriving data packets
and asymmetric multiplex transmission needs to be minimized
[19], [20].Finally, the inability of PDU session bearers to
interact during transmission phase limits the applicability
of dynamic routing and network coding technologies [24],
[43]. Therefore, the proposed approach should be able to | 2402.00002v1.pdf | 90b7807a-3e70-4a08-b0e0-ab1ee2e3b4a4 |
PREPRINT SUBMITTED TO IET 4
NG-RAN NG-RAN
N3IWF N3IWF UE UE
1.PDU Session Estb Req MEN AMF MEN AMF UPF-1UPF-1MEN SMF MEN SMF RCMEN S&D RCMEN S&D UPF-AUPF-A DN DN
UPF-2UPF-2
Ue is registered on both 3GPP and nono-3GPP accesses in 6G-RCMEN
2.PDU Session Create SMContext
6.a Radio resource Req
6.b Radio resource Req
7.a Setup Radio
7.b Setup Radio 8.a Radio Resource Resp
8.b Radio Resource Resp
9. | 2402.00002v1.pdf | 9178b105-5aa2-4b30-adef-d420e241d8fa |
PDU Session Update SMContext Req 4.a Estb Session1 child1
10.a Update UPF1
10.b Update UPF2
11.a User Data Session1 child1
11.b User Data Session1.2 child2 ͓Ͷ Ϳ Ḥ ͢ ͽ Ͷ ͻ
͓Ͷ Ϳ Ḥ ͢ ͽ Ͷ ͻ
͓Ͷ Ϳ Ḥ ͑ ͼ΄ ͻ Ͷ ͻ
͓Ͷ Ϳ Ḥ ͑ ͼ΄ ͻ Ͷ ͻ 3. Get Sub
Strategy&Data
͢ ͲͿ Ḥ ͑ ͮ ͮ 5 Namf comm Message Transfer
Fig. 3: UE-requested physically isolated multi-access PDU
session establishment procedure for B5G.(See Appendix A)
defend against randomly distributed non-cooperative passive
attackers. The paper has identified these key points and will
provide a detailed description of the preliminary work in the
next section.
III. B5G SCLER P RELIMINARY
A major development trend of the beyond 5G core net-
work is to support more complex services through cloud-
network integration and distributed collaboration and provide
customized services for each UE. This section will discuss
in detail the intrusion probability of multiple PDUs and B5G
RCLER preliminary work.
A. Edge Network Intrusion Probability for Multiple PDU
Sessions
Physically isolated multi-access multi-PDU session edge
computing networks can provide security benefits. In this
subsection, we analyze the security of edge networks in the
B5G era and explore the relationship between multiple disjoint
PDU sessions and the probability of intrusion. Specifically,
when there are NLdisjoint PDU sessions, we investigate
the probability that all channels are eavesdropped, i.e., the
probability P[NL]that all NLPDU sessions are attacked.P(NL) =YNL
j=1[1−YNj
i=1(1−pi
j)] (1)
Where pi
jrepresents the probability that the i-th node on the j-
th PDU session is eavesdropped, and Njrepresents the number
of nodes contained in the j-th PDU session (including the
base station(gNB), UPF, and UPF-A). Identify all possible
physically isolated and disjoint PDU sessions between the
UE and the DN. Given a predetermined threshold value ϑ,
we require the connection intrusion probability P(NL)to be
less than or equal to ϑ(denoted as P(NL)≤ϑ). The value
ofNLcan be determined accordingly based on the required
ϑ. In the following subsection, we will describe the method
for establishment the aforementioned PDU session, which
serves as a fundamental requirement for the SCLER to operate
effectively in the B5G core network.B.Physically Isolated Multi-Access PDU Session Establish
procedure
The B5G core network we envision still follows the 5G
control and bearer separation design. There are two planes,
Control Plane (CP) and User Plane (UP). The core network
provides PDU connectivity services between UE and DN to
realize the exchange of PDUs. UP transport user PDUs. CP
control manages the behavior of the entire core networks based
on signaling. CP functional entities are divided into different
Network Functions (NF). The User Plane Function (UPF)
mainly supports the routing and forwarding of UE service data,
Session Management Function (SMF) is responsible for UE
session management and manages UPFs, Access and Mobility
Management Function (AMF) is responsible for UE access.
The three jointly participate in the PDU session establishment
procedure.
We mainly show the MEN-SMF and MEN-AMF of the
control plane in Fig. 1. They are enhancements and improve-
ments of SMF and AMF in the 5G core network. They mainly
participate in the establishment of UE Multi-Access PDU
sessions.They mainly
participate in the establishment of UE Multi-Access PDU
sessions. MEN-AMF Responsible for handling UE connection,
registration as well as routing session management messages
from the UE to the appropriate MEN-SMF. MEN-SMF man-
ages PDU sessions by sending the signaling to different 6G
access points. UPF on the user plane carries workflow data
and can be regarded as a Layer 3 router, which receives and
forwards the data packets between the RAT and the DN.
Untrusted non-3GPP access Wifi shall be connected to the
3GPP Core Network via a Non-3GPP InterWorking Function
(N3IWF) [44]. The 5G network defines a PDU session as a
logical connection between the UE and the DN. A part of
the PDU session is carried on the wired resources of the
core network, and a part exists on the wireless resources.
The performance of the PDU session is constrained by the
physical bearer, but the transmission process itself ignores the
resource form.The existing mobile network supports the UE to
access different DNs to establish corresponding multiple PDU
sessions, and can also access a single DN to establish multiple
PDU sessions. However, the current 5G does not support the
establishment of physical or geographic separation of multiple
PDU sessions for a single workflow. In Fig. 3 We give the
procedure of establishing a Multi-Access PDU Session, the
appendix A has a detailed procedure. Multiple PDU sessions
provide services for a single service and are connected to the
same DN. UPF-A is an anchor UPF that receives data packets
to and from DNs for UEs [22]. The splitting, aggregation,
encoding, and decoding of data on the DN side are processed
in SCLER-UPF-A.
IV. S YSTEM MODEL
As depicted in Fig.2, SCLER creates three PDU sessions
on the edge network to provide a downlink transmission
strategy with low latency and security considerations from
the DN to the UE. Service flows that randomly arrive at
the DN’s application layer are accumulated in the buffer and
queued.The variable block length encoder selects gdata
packets for encoding, and the encoded packets are distributed | 2402.00002v1.pdf | 91555a13-653a-47c9-a24c-0e94505e612b |
PREPRINT SUBMITTED TO IET 5
Safety
Threshold
+\K (UH
'ROIK
Source package Coded package 8GVZUX
)UJKX 6*;
6*;
6*;T JKIUJKX
9)2+8
9INKJ[RKX Legitimate
Transmitter Legitimate
Receiver
Eavesdropper Decoding Failed Decoded
Successfully
sgY Y Y Y
S S S S S
I I I I I IY Y Y Y
S S S S S
I I I I I I
1gPs ´ 1gPc g´
LT pre
g m m g G G g´ ´ 1[ ,..., ] j T
j ug g g =
{ 1,..., }, j s j J g g " Î £
Fig. 4: SCLER Scheme
to three PDU sessions according to the k-out-of-N weight for
transmission. The legitimate receiver receives the Pcthrough
the diversity interface and completes the decoding process.The legitimate receiver receives the Pcthrough
the diversity interface and completes the decoding process.
However, malicious network elements, eavesdroppers, and
other unauthorized entities only receive part of the information
and cannot complete the decoding.
The SCLER strategy divides time into time slots and de-
livers the DN’s downlink service flow to the legal receiving
terminal reliably with low delay while also considering se-
curity aspects. This section discusses the system model from
three aspects: the modeling of the transmission strategy, the
modeling of the transmission principle, and the key measure-
ment indicators of the system.
A. Encoding Transmission Model
Under the SCLER strategy, we conduct Raptor encoding on
a generation of gdata packets in each time slot7. As shown
in Fig.4, by leveraging the fountain code prefix property,
the redundancy of the encoder output can be adjusted by
T. Assuming the encoder produces Ndata packets, the jth-
PDU session selects kpackets from them for transmission.We
define the scheduling weight coefficient as γj=k/N , and the
scheduling weight vector of SCLER as γj
υ= [γ1, ..., γ j]T. The
SCLER strategy involves the selection of gdata packets to be
encoded in time slot T, and then the distributing the data using
a scheduling strategy γj
υ.
In this study, we represent the SCLER action for each time
slot by a two-tuple (g, γj
υ),i.e.,(g[n] = g, γj
υ[n] = γj
υ),
indicates that the SCLER action in the n-th time slot is g, γj
υ.
Assuming that the number of data packets in each generation is
bounded by B, that is, g[t]∈ {0,1, ..., B}. The weight value
γj[n]for a single session changes in step sizes of (ςT)−1,
where ςis the encoding rate coefficient that indicates the
number of codes generated per unit time, γj∈ {1
ςT,2
ςT, ...,1},
0< γj≤1.We define the value space of the weight vector γj
υasΓj
υ,
such that γj
υ∈Γj
υ. The space size is Nj= (ςT)j, and γυis
7assuming that all data packets have the same length and carry ϕbit
information.obtained by ⊙NL
i=1γj
υ(γi)here,⊙denotes the Cartesian product.
We have 1≤γυ< N L, where NLrepresents the number of
PDU sessions established by SCLER. To ensure a sufficient
breadth of PDU session selection, we require NL≥2. Each
PDU session described in SCLER can be modeled as a packet
erasure channel. Let TX[n]be the number of packets sent
by the transmitting end and RX[n]be the number of packets
received by the receiving end
RX[n] =ε[n]TX[n]. (2)
We assume that TX[n], has a probability of 1−Pεbeing
successfully delivered to RX[n]. Otherwise, the packet will
be erased. | 2402.00002v1.pdf | f52d9c8f-bf64-429f-b98f-3810541be8f6 |
Otherwise, the packet will
be erased. This erasure can be due to various factors, such as
subsequent datagrams being discarded due to high router load
at the network layer, non-confirmed retransmission triggered
by a new PDU session timer after mobile reconnection,
transmission errors detected by the bottom layer such as the
data link layer, and discarded packets, etc [45], [46]. We
denote εj
υ= [ε1, ..., ε j]to represent the erasure probability
of different paths, cj[n] = 0 /1to indicate whether the jth
PDU session is lost or successfully transmitted at time slot n,
P{cj= 0}=εj,cjis independent and identically distributed
in different time slots of the same PDU session.
B. Threat Analysis and SCLER Encoding Principle
In this subsection, we first analyze the potential information
leakage associated with the multiplex transmission method
based on Raptor encoding. We then provide a description of
the passive attackers that we aim to defend against in this
paper. Finally, we introduce the Raptor encoding transmission
principle employed by SCLAR.Finally, we introduce the Raptor encoding transmission
principle employed by SCLAR.
1) Threat Analysis: In this subsection, we describe the
attack capabilities of passive attackers [47], who do not tamper
with or discard data like active attackers, but instead copy,
or store stolen user business data. Specifically, we consider
a generalized eavesdropper who aims to obtain user data
carried on the transmission path between the UE and the
DN, without considering the specific communication layer of
the eavesdropping behavior(gNB or Network Function), as | 2402.00002v1.pdf | c161e878-0dd9-4b22-aa66-23899527efc6 |
PREPRINT SUBMITTED TO IET 6
shown in Fig.2. By clarifying the attack capabilities of passive
attackers, we can better understand the scope of our defense
goals.
Taking a B5G edge network with NLdisjoint PDU sessions
as an example, where Njrepresents the number of nodes
contained in the jth PDU session. If Mnodes are attacked,
it may result in only some nodes being under attack, which
is a subset of the NLPDU sessions. This will lead to
some information leakage, and we study the probability of
information leakage.This will lead to
some information leakage, and we study the probability of
information leakage. Where
P[M] =min(N1,M)X
ϖ1=0···min(Nj−1,M−PNj−2
i=1ϖi)X
ϖNL−1=0
NLY
j=1(Nj
ϖj)pϖj
j(1−pj)(Nj−ϖj)u(nNL−ϖNL)(3)
Assuming that ϖjnodes are attacked in the jth PDU session,
ϖNL=M−PNL−1
j=1ϖj,u(·)refers to
u(x) =(
1,ifx≥0
0,otherwise(4)
the probability of M nodes being attacked is given by
P[ϖ1, ϖ2. . .. . , ϖ NL, M] =
NLY
j=1(Nj
ϖj)pϖj
j(1−pj)(Nj−ϖj)u(nNL−ϖNL),(5)
where ϖNL=M−PNL−1
i=1ϖi, and the marginal probability
P[M]is obtained by summarizing all possible values ϖiand
i= 1,2, ..., N L−1.
For the subset ΥnLof multi-PDU sessions with nLpaths,
which contains l1, l2, ..., l nL, the probability of Mnodes being
attacked is:
P[ΥnL, M] =Xmin(Nl1,M)
ϖl1=0. .. .Xmin(NlnL−1,M−PlnL−2
l=l1ϖl)
ϖlnL−1=0
YlnL
j=l1Nj
ϖj
pϖj
j(1−pj)(Nj−ϖj)u(NlnL−ϖlnL)
YNL
Ψ=1(1−pΨ)nΨ
(6)
where ϖlnL=M−PlnL−1
j=l1ϖj, and Ψ̸=l1, l2, ..., l nL
Therefore, the probability that only Mnodes in subset ΥnLin
NLare attacked is P[ΥnL|Nnode=M] =P[ΥnL,Nnode=M]
P[Nnode=M].
Suppose there are NLdisjoint PDU sessions in the B5G
edge computing network, denoted by the node set L=
{l1, ..., l j, ..., l NL}for this MEN.Here, NL∈ {1, ..., J}indi-
cates that at most JPDU sessions can be established, as shown
in Fig.1 is an example of establishing 3 disjoint PDU sessions
8.ljrepresents the set of nodes contained in a single PDU
session, lj={nj
1, nj
2, ..., nj
i, ..., nj
Nj}, where Nj∈ {1, ..., I},
Irepresents the upper limit of nodes contained in a single PDU
session. We consider the eavesdropper Eκ, who can erode at
8in Fig.1, we set the UPF anchor with DN together, and UPF-A is not
drawn.Algorithm 1 Raptor Code Encoding Operation
Input: Read Ps= (Ps1, Ps2, . Ps g)Tfrom the queue,
Ωd= (Ω 1,Ω2, .Ωmax), Waiting window w;
Output: ¯Pc
1:Pcgγ×1←0gγ;
2:GRaptor
gγ×m=GLT
gγ×mGpre
m×g;3:Pcgγ×1=GRaptor
gγ×mPsg×1;
4:¯Pc←Pcgγ×1;
5:ifw >0then
6: the transmission is completed for wtime slots;
7: forACK ̸=nildo
8: Pcfix←Gfix
1×gPsg×1
9: ¯Pc←append [¯Pc, Pcfix]
10: end for
11:else
12: no response mode;
13:end if
14:return ¯Pc;
most Njnodes on the single jth PDU session and cannot
eavesdrop on nodes other than the current PDU session, i.e.
{{L}\{lj}},∆κ={δ1, δ2, . δ n}represents the eroded node
set9, satisfying
∀∆κ⊆L,{∆κ∩lj} ̸=∅,∃j∈ {1, . J},
s.t. | 2402.00002v1.pdf | c0359380-df5a-46ab-8234-c4008a8e247f |
∃j∈ {1, . J},
s.t. ∆κ⊆lj,and∆κ∩ {L/lj}=∅(7)
Furthermore, there is no interaction or cooperation between all
eavesdroppers Eκ, κ∈(1, K],Eκwill not be assisted by any
other eavesdropper. This assumption is reasonable in mobile
networks [48].
2) SCLER Encoding Principle: As shown in Fig.4, we
designed an encoding transmission method based B5G for UE-
DN server flow, which encodes Psarriving at the application
layer and distributes Pcto multiple PDU sessions. The method
involves a Raptor [49] encoder that performs two encoding
processes on the selected g Ps : an external encoder (pre-
encoding) and an internal encoder (LT code [26]).Following
the SCLER strategy, we propose Algorithm1, the method first
selects gsource data packets( Ps) to encode as one generation,
at time slot t, the encoder of a generation of gis denoted
byξt(g, m, Ω), we have Pcgγ×1=GLT
gγ×mGpre
m×gPsg×1, The
formula GRaptor
gγ×m=GLT
gγ×mGpre
m×gdefines the relationship
between PsandPc, where GLT
gγ×mandGpre
m×grepresent
the LT encoding generator matrix and precoding generator
matrix. (1) Precoding generates m intermediate symbols to
ensure that all Psare covered, Gpre
m×w= [Ik|Pm×(k−w)]T,
where Pm×(k−w)is responsible for mixing Ps.(2)The in-
ternal encoding matrix GLT
gγ×mgenerates the final encod-
ing symbol based on the degree distribution Ω(x), which
represents the probability distribution of the number of Ps
selected each time, Ωd= (Ω 1,Ω2, ...,Ωmax), while Ω(x) =
Σdmax
d=1Ωdxd. Encoder input Psg×1= (Ps1, Ps2, ..., Ps g)T,
output Pcgγ×1= (Pc1, Pc2, ..., Pc gγ)T.
9Including malicious network functions, and eavesdropping on gNB, this
article regards the erosion of a single node. | 2402.00002v1.pdf | aefe5bf4-531f-4b7f-95f6-0afc8cc36b84 |
PREPRINT SUBMITTED TO IET 7
(a) Legitimate Receiver.
(b) Eavesdropper.
Fig. 5: Number of Psdecoded versus number of Pcreceived
for Receivers.
The Maximum-Likelihood (ML) decoding algorithm is typ-
ically preferred for finite-length Raptor encoding over the
Belief Propagation (BP) decoding algorithm [50], [51]. The
performance of ML decoding is comparable to that of Gaus-
sian elimination [52]. To use ML decoding, a coefficient matrix
is constructed with rows and columns corresponding to the
PcandPs. The coefficient matrix represents how the Pcis
linearly combined from the Ps.g Ps can be recovered from
gγofPcif and only if (GLT
gγ×mGpre
m×g)gγ×gis full rank.
We conducted a MATLAB simulation of the RAPTOR
codec and observed that the number of successfully decoded
packets initially increases slowly with the number of received
Pc, before reaching a sharp increase threshold10. Once the
threshold is reached, the entire block can be decoded quickly.Once the
threshold is reached, the entire block can be decoded quickly.
In other words, the number of successfully decoded Psis
initially very low until the number of received coded packets
reaches the threshold. This principle can be used to defend
against single-link eavesdropping attacks in CMT, we can
limit the number of coded packets transmitted in each single
PDU session to below this threshold. This way, an attacker
of a single PDU session cannot decipher the source data
information effectively, ensuring the security of user data.
C. Secure and Reliable Function and Queuing Model
This subsection comprises two parts. In the first part, we
analyze the reliability and security of a concurrent trans-
mission system that employs Raptor coding. In the second
part, we analyze a queuing coding model that corresponds to
an eavesdropping scenario. For enumeration of all possible
session state, we first employ a 2NL×NLmatrix C:
C=
1 1 . . . 0
....... . ....
1 0 . . .. . 0
....... . ....
1 0 . . . 0
T
(8)
Each element ci,jin row iand column jrepresents the
probability of successful ( ci,j= 1 ) or failed ( ci,j= 0 )
reception by the receiver through the j-th PDU session for
thei-th transmission possibility.As the UE-DN workflow is
transmitted through different PDU sessions, each individual
message may experience varying degrees of delay and packet
10Note that our team has worked on improving the slope of the surge in
the early stages [32], but we won’t go into details in this article.loss, which can be quantified using the latency-reliability-
security function,
FSCLER
T, γj
υ, g
=2NLX
i=1giNLY
j=1Hj(T, γjg) (9)
Based on Eq.9, the secure data transmission considering the
existence of single-link eavesdroppers, ∀j∈ {1, ..., N L},γj≤
γs,γsis the threshold value for ensuring transmission security,
and limits the upper bound of the weight assigned to each
path to prevent eavesdroppers from obtaining sufficient private
information. Where
gi=(
1,ifΣNL
j=1ci,jγjγτ≥γdand∀j∈ {1, ..., N L}, γj≤γs
0,otherwise.(10)
giwill exclude failed transmissions to ensure that only suc-
cessfully decoded outputs are included. γdis the threshold
used to determine successful decoding. Hjis defined as
Hj(T, γjg) =Fj(T, γjg),ifci,j=1
1−Fj(T, γjg),ifci,j=0(11)
The definition of Fjis based on the research by Erik et al.
[53]. We delayed the transmission of PDU sessions to cut
off the Gaussian distribution. The corresponding accumulated
probability distribution, Fj, is shown in Figure 7a. Fj(T, γjg)
means that at time T, the data packet transmission of the j-th
PDU session with size γjgcan achieve a certain reliability.
The extreme value of Fjapproaches 1−Pj
e.
There is a correspondence between static latency and relia-
bility (i.e., probability of successful data delivery).
By utilizing function Eq.9, both reliability and security can
be guaranteed. Λ[t]denotes the number of Psarriving in
thet-th timeslot.We assume that the application layer input
{Λ[t], t≥0}, the number of packets arriving in each slot is
i.i.d. , the upper bound of Λ[t]isN, denote the probability
distribution of Λ[t]asλ= [λ0, λ1, ..., λ N]T, where we define
λn=P{Λ[t] =n}, the average arrival rate asPNΛ
n=0nλn.
Using a buffer with a size of Z, randomly arriving packets are
accumulated, and gpackets are selected from the cache once
the encoding of the previous block of packets is completed. | 2402.00002v1.pdf | b38d0ca5-331c-45d0-9930-aed32d9aa97e |
Therefore, q[t]∈ {0,1, ...,Z}, in the [t+ 1]-th timeslot, q[t]
evolves as11
q[t+ 1] = min {(q[t]−g[t])∗+ Λ[t+ 1],Z} (12)
V. T RADE -OFFDELAY -SECURITY FORVARIABLE
BLOCK -LENGTH SCLER S TRATEGY
A. Constrained Markov Decision Process Formulation
In the context of edge networks, ensuring both the security
and low latency of data transmission is critical. To address
these challenges, we propose a Raptor-encoded multi-session
distributed delay-optimized transport method called SCLER
that considers these constraints. To achieve this, we utilize
the queue length as the system state information and propose
a novel queue-aware variable block length encoding method
that effectively models the SCLER strategy as a Constrained
11we define (x)∗= max {x,0} | 2402.00002v1.pdf | 03ff1463-2ac4-489d-adf4-9db4aaf8413c |
PREPRINT SUBMITTED TO IET 8
Markov Decision Process (CMDP). We formulate the problem
as a constrained optimization problem to obtain the trade-
off between delay and reliability, which is a crucial factor
in designing efficient data transmission mechanisms in edge
networks. Specifically, we introduce the SCLER strategies by
leveraging the probabilities f(q, g, γj
υ),
f(q, g, γj
υ) =P{g[t] =g, γj
υ[t] =γj
υ|q[t] =q}. (13)
We assume a temporary steady-state condition where
no PDU session is being established or released,
that is ∀q,ΣB
g=1Σγj
υ∈ΓNj
υf(q, g, γj
υ) = 1 .Therefore,
we can get the strategy function of SCLER, S=
f
q, g, γj
υ
: 0≤q≤M,0≤g≤B,1≤γυ< N L,0< γj≤1
,
to prevent the sender buffer from overflowing or
underflowing, 0≤q[t]−g[t]≤M−NΛ, and
{γj
υ∈ΓNjυ|C{g>0}≤γυ≤NLC{g>0},0< γj≤C{g>0}}.
The state of the system q evolves based on the Eq.12.
We denote the overall redundancy of a generation of
data by γυ=PNL
j=1γj. Given a particular strategy
and corresponding system constraints, if T→ ∞ , then
lim
T→∞F
T, γj
υ, g
=L≤1−Pe, L denotes the upper limit
of information security transmission, and the reliability value
of the s ystem will not exceed its theoretical performance.Using the delay-safety function, we can identify the feasible
region
F(ϵ) =
∀γj
υ, g, F
T, γj
υ, g
|1−ϵ≤F
γτT, γj
υ, g
≤1−Pe
(14)
The fountain code transmission scheme in edge networks
results in a long-tailed latency distribution. To address this
issue, the SCLER algorithm is proposed to reduce the average
transmission delay and achieve a balance between reliability
and security.Based on Little’s Law, lim
T→∞1
TPT
t=1E{q[t]
λ},
this can be used as a latency metric for MEN, which we
describe by the probability of constraints violation Pvc:
Pvc= lim
T→∞1
TXT
t=1P{q[t]
λ≥dth} (15)
where dthis the threshold of the latency indicator, dth∈
[0,∞)Building upon the SCLER strategy described above,
this paper investigates the provision of low-latency, secure,
and reliable transmission services. A comprehensive analysis
is conducted, considering indicators such as information leak-
age rate, PDU session quality, occupied bandwidth, decoding
function, and delay. The delay and security of the MEN in the
B5G era are key performance indicators that are determined
by the state (q[n], εj
υ[n])and action (g, γj
υ). Considering the
limitations of service characteristics, this section proposes a
constrained optimization problem based on CMDP, and the
next section will analyze the trade-off between reliability and
delay while considering security.B. The Optimal Delay-Reliability Trade-off under SCLER
In this subsection, we will focus on achieving a trade-off
between transmission reliability and delay while ensuring data
security in the face of Eq.7. We refer to Eq.9 as the reliabilityfunction, and successful transmission of data in SCLER in-
volves two parts: receiving and recovering. As shown in the
Fig.2, SCLER determines the transmission strategy {g, γj
υ}
based on the state information of the queue length q[t]. We
describe the SCLER strategy Sas a Markov decision process
with the buffer queue q(t), we define the transition probability
asλq,q′=P{q[t+ 1] = q′|q[t] =q}. | 2402.00002v1.pdf | 6643f60c-9313-40d7-941f-98cfa0d86a17 |
Based on Eq.12, we
formulate λq,q′as follows:
λq,q′=NX
n=0λnBX
g=0f(q, g, γj
υ)C{q′=min{(q−g)∗+n,Z}12(16)
The steady state distribution of the system state q[t]of SCLER
is formulated as follows:
(Pmin{Z,q′+B}
q=max{0,q′−N}π(q)λq,q′=π(q′),∀q′∈ {0,1,...,M},PM
q=0π(q) =1.
(17)
The queue length, denoted as qandq′, takes on values in the
set{0,1, . . . , Z}, subject to the constraint that the summation
ofπ(q)over all possible values of qis equal to one. The
queue length is influenced by the random arrival of packets
and selection of.The
queue length is influenced by the random arrival of packets
and selection of. To express the balance equations in Eq.17
as a matrix form, we represent the steady-state probability
distribution πas a vector [π0, ..., π Z]T,
ΛSπ=π,
1Tπ= 1,(18)
The matrix ΛSis defined with the element λq,q′at the qth
row and q′th column. Additionally, the vector 1is defined
as a vector with all its components being 1. Based on the
stationary distribution π, the SCLER strategy’s average delay
can be expressed using Little’s law.Specifically, the average
delay LSbased on the SCLER strategy Sis
¯LS=1
λZX
q=0qπq (19)
The occupied PDU session resource occupancy is expressed
as
WS=ZX
q=0BX
g=0Wj
S(g, γj
υ)fq,g,γj
υπq (20)
SCLER presents an optimal trade-off scheme for delay-
reliability that takes into account security considerations, based
on the average delay and bandwidth occupancy. The proposed
approach formulates an optimization problem of Psencoding
and CMT, with the queue length serving as state information,
and subject to reliability and security constraints.
12C{·}denotes the characteristic function. | 2402.00002v1.pdf | 7158d102-f4a4-48e7-a04d-044655e33916 |
PREPRINT SUBMITTED TO IET 9
min
{fq,g,γj
υ,πq}¯LS (21a)
s.t. F S≥rth(21b)
WS≤Wth
j (21c)
ΛSπ=π (21d)
1Tπ= 1 (21e)
BX
g=0X
γj
υ∈ΓNj
υfq,g,γj
υ= 1,∀q (21f)
πq≥0, fq,g,γj
υ≥0,∀q, g. (21g)
The goal of SCLER is to find the optimal delay under
the constraint of bandwidth Wth
jin the MEN. In order to
achieve the optimal trade off between delay and reliability
with secure consideration, we have transformed the encoding
CMT problem into an Linear Programming (LP) problem:
min
{xq,g,γj
υ}1
λZX
q=0BX
g=0X
γj
υ∈ΓNj
υqxq,g,g,γj
υ(22a)
s.t.MX
q=0BX
g=0X
γj
υ∈ΓNj
υF(g, γj
υ)xq,g,γj
υ≥rth(22b)
ZX
q=0BX
g=0X
γj
υ∈ΓNj
υWj
S(g, γj
υ)xq,g,γj
υ≤Wth
j (22c)
XZ
q=0NX
n=0BX
g=0X
γj
υ∈ΓNj
υλnxq,g,γj
υC{q′=min{(q−g)∗+n,Z}
=BX
g=0X
γj
υ∈ΓNj
υxq′,g,γj
υ(22d)
ZX
q=0BX
g=0X
γj
υ∈ΓNj
υxq,g,γj
υ= 1 (22e)
xq,g,g,γj
υ= 1 (22e)
xq,g,γj
υ≥0,∀q, g (22f)
xq,g,γj
υ= 0,∀q−g <0or q−g >(Z−N)
(22g)
we define the optimization variable xq,g,γj
υas
πqfq,g,γj
υ, its long-run average can be expressed as
lim
T→∞1
TPT
t=1P{q[t] =q, g[t] =g, γj
υ=γj
υ[t]}, one can
obtain the optimal trade-off between delay and reliability
for SCLER by solving the LP problem Eq.22 through the
variable xq,g,γj
υ.
In algorithm 2, we transform the constrained optimization
problem Eq.22 into a matrix form, we set the combination of
solution space {xq,g,γj
υ}as a vector ⃗ x,g,γj
υ}as a vector ⃗ x, the feasible region in
matrix form as
ℜ=(
⃗ x|1T
|⃗ x|⃗ x= 1,HS⃗ x=0M+1,
0≤xq,g,γj
υ≤C{g∈N(q), γj
υ∈ΓNj
υ})
.13(23)
Eq.22d are also presented in the form of coefficient matrix
HS, where HS=˙HS−¨HS, let ˙HSand¨HSdenote the left-
hand and right-hand coefficients of Eq. (22d), respectively. We
131Nand0Ndenote Ndimensional all-ones and all-zeros vectors, respec-
tively.Algorithm 2 Algorithm to generate constraints matrix HS
Input: λ= [λ0, λ1, . λ N]T;
Output: HS
1:forq= 0 toMdo
2: ˙Hq
S←0M+10T
B+1,¨Hq
S←0B+1;
3: forg= 0 toBdo
4: ˙Hq,g
S←0max{q−g+N,M}+1;
5: ˙Hq,g
S(q−g+ 1 : q−g+N+ 1)←λ;
6: ˙Hq
S(1 :M, g + 1)←C{g∈N(q)}¨Hq,g
S(1 :M);
7: ¨Hq
S(g+ 1)←C{g∈N(q)};
8: end for
9: ˙Hq
S←˙Hq
S,¨Hq
S←eM+1,q+1(¨Hq
S)T;
10: ˙HS←[˙HS,˙Hq
S],¨HS←[¨HS,¨Hq
S];
11:end for
12:HS←˙HS−¨HS;
utilize Sto obtain the optimal trade-off solution x∧
q,g, | 2402.00002v1.pdf | a3ce904b-8140-4712-8f29-de58c47b88c1 |
utilize Sto obtain the optimal trade-off solution x∧
q,g,γj
υbe-
tween delay and reliability, where x∧
q,g,γj
υis obtained through
πqf∧
q,g,γj
υ. Thus, the optimal strategy S∧can be expressed as
S∧={f∧
q,g,γj
υ:∀q, g, γj
υ}, where f∧
q,g,γj
υis defined as:
f∧
q,g,γj
υ=
x∧
q,g,γj
υ
π∧qifπ∧
q>0
C{g=gmax
q}ifπ∧
q= 0,(24)
As a consequence, the optimal delay-reliability trade off under
secure consideration for SCLER is obtained.
C. The threshold-based Strategy
We aim to offer edge network users a secure, high-
reliability, and low-latency service.To achieve this, we propose
transforming the LP problem into a reliability-delay pairing
based on the reliability function (Eq.9). As a result, the LP
problem can be redefined as:
min
(ˆr,ˆL)∈ℵˆL (25a)
s.t. ˆr≥rth
j (25b)
WS≤Wth
j (25c)
the set ℵcomprises all admissible ”delay-reliability” pairs
under the SCLER strategy S. Specifically, ℵcan be obtained
on the two-dimensional plane of delay and reliability, where
delay and reliability are determined by Eq.22a and Eq.22b,
respectively. Each point in ℵcorresponds to a particular type
of SCLER strategy Ssubject to the constraints specified by
Eqs.22c-22g.Using variable block length coding under given PDU ses-
sion conditions Wth
jand arbitrary initial queue conditions,
we obtain an optimal delay–reliability trade-off f∧
q,g,γj
υ. We
also examine how our SCLER behaves under this optimal
coding strategy S∧and show that each queue length selects
either maximum or minimum block lengths ( Bmax
g[t],Bmin
g[t]):
maximum lengths improve coding efficiency and reliability
Eq.9; packets dispersion CMT with a security threshold γs
impose a lower limit on minimum Pcencoding block lengths. | 2402.00002v1.pdf | 28850437-3c3b-41e8-80f2-5acf5a1cabcd |
PREPRINT SUBMITTED TO IET 10
When gfalls below this threshold Bmin
g[t], we fill out coding
blocks with randomly generated packets.(
f∧
q,g,γj
υ≥0,ifgmax
q orgmin
q
f∧
q,g,γj
υ= 0,otherwise(26)
The optimal SCLER encoding strategy exhibits characteristics
of action-limited decision-making, where the strategy of S is
constrained by state information and a threshold q∼related to
q[t]:
f∧
q,gmaxq,γj
υ= 1 ifq > q∼
f∧
q,gminq,γj
υ= 1 ifq < q∼
f∧
q,gminq,γj
υ+f∧
q,gmaxq,γj
υ= 1 ifq=q∼(27)
In addition, to achieve the optimal trade-off between delay and
reliability, the allocation coefficient γj
υselects the maximum
value within the allowable range of the safety threshold. As
a result, we can obtain the minimum LSwhile satisfying the
reliability and safety constraint.The trade-off between delay
and reliability can be obtained by adjusting the reliability
constraints.
VI. P ERFORMANCE RESULTS
We constructed a set of edge networks to evaluate our
proposed SCLER. The entire mobile core network is imple-
mented based on the modified Free5GC. The control plane
and user plane network elements of the core network are de-
ployed on two separate virtual machines running Ubuntu18.04
with 16GB memory. We used two access points, two cus-
tomized base stations (NR01), and one TP-LINK AX3000
Wi-Fi device. We have modified the existing PDU session
establishment process (see Appendix A) by transforming the
core network control plane network element to support the
establishment of multiple PDU sessions. Our experimental
section is divided into two parts: testing the basic environment
and presenting the numerical results of the SCLER strategy.
0 10 20 30
Index of test seq-num024681012141618Latency(ms)
(a) WiFi & NR latency result. | 2402.00002v1.pdf | 0cb2e3ee-7452-464c-a97d-6198a09448fd |
0 20 40 60
Index of test seq-num0.00.10.20.30.40.50.60.70.8Process time(s) (b) UE initialization procedure
delay result.
Fig. 6: DIEN performance test result.
A. Basic Performance Results
Fig.6 shows the transmission performance test results of
DIEN basic environment. As shown in Fig.6a, we tested the
transmission delay of NR and WiFi by sending an ICMP
packet to the DN end on the UE. To determine the delay of
the basic propagation of the PDU session that is not related
to the carrying data packet. We recorded the delay of a singleNR/WiFi access PDU session. The average delay of NR was
6.25ms, while that of WiFi was 9.927ms for 60 tests.
Furthermore, we conducted UE initialization procedure test
Fig.6b to measure the system’s initialization delay. The ex-
perimental object is the procedure proposed by our Appendix
A. 1.Registration procedure is usually triggered during power-
on or switching flight mode.1.Registration procedure is usually triggered during power-
on or switching flight mode. 2.Establishment of PDU session
procedure triggers when the UE moves between gNB. We
recorded the delay of the initialization process 30 times.
The average time taken for Procedure1+2 and Procedure2
were 609ms and 194ms, respectively. The analysis of SCLER
performance under the steady state after connecting the es-
tablishment of connection does not involve two initialization
procedures.
0 2 4 6 8 10 12 14 16 18 20
transmission delay00.10.20.30.40.50.60.70.80.91
Wifi
LTE
NR
(a) Latency-reliability curves Fj
for considered technologies for
ϕ= 1500.0 2 4 6 8 10 12 14 16 18 20
Transmission delay(ms) 00.9 0.99 0.999 0.9999
0 2 4 6 8 10 12 14 16 18 20
Transmission delay(ms) 00.9 0.99 0.999 0.9999 (b) Delayed reliability function
FSCLER with security consider-
ation described in Eq.9 .
Fig. 7: Example showing the reliability diagram of a single
PDU session and DIEN-based reliability function of the entire
SCLER system.
As shown in Fig.7, we want to use Fjto describe the
delay characteristics of PDU sessions with different interfaces.
The end-to-end transmission delay Tof the PDU session
we consider may be affected by high load packet loss at
the network layer [54] and data link layer packet loss [55].
Therefore, without considering retransmission, Tshould obey
the Gaussian distribution, consequently we have T=tp+tc,
where tcrepresents the data transmission delay, related to
packet size and bandwidth, i.e.t c=W∗ϕ−1∗1.25∗106, the
propagation delay tpcaused by protocol flow and physical
signal propagation, corresponds to the position of the first
non-zero point in Fig.7a, where tp= [9 .927,12,6.25]ms,
Wj= [500 ,100,200]Mbit/s, We regard the cumulative dis-
tribution of the delay probability as Fj. This indicates that the
transmission reliability of the current session changes with the
transmission delay, and its extreme value approaches {1−Pe},
where Pj
e= [0.9,0.92,0.95], corresponding to the highest
reliability that can be guaranteed. The rising rate of Fjis
related to tc.
As shown in 7b, we have illustrated FSCLER and pre-
sented the changes in reliability values under different typical
strategies based on Eq.9.The orange and red lines show
FSCLER values that increase with delay, FSCLER represents
the probability of the system successfully decoding γj
υin the
legal range, when the allocation weight exceeds the decoding
threshold γd, the method can be decoded successfully; When
γj
υ= [0.2 0.2 0.9], where γ3exceeds the safety threshold
γj
s= 0.8, and FSCLER is always zero. We noticed that | 2402.00002v1.pdf | 92532b73-bc11-4eb3-b363-f59dda51a469 |
PREPRINT SUBMITTED TO IET 11
TABLE II: Simulation Parameter Settings. | 2402.00002v1.pdf | dce0e6a8-e94d-4421-9223-53d5fefb6d10 |
PREPRINT SUBMITTED TO IET 11
TABLE II: Simulation Parameter Settings.
Packet Arrival Rate Channel Transmission State Packet Arrival Statistics
E[Λ] λ= [λ25, ..., λ 100] rthεj
υ(%) E[Λ] σ2λ= [λ25, ..., λ 100]
25 [0.24,0.10,0.12,0.05] 0.99 [22,25,25] 45 0.78 [0.39,0.42,0.11,0.06]
45 [0.34,0.27,0.16,0.11] 0.999 [9,22,10] 45 1.16 [0.39,0.31,0.13,0.10]
60 [0.16,0.30,0.20,0.26] 0.9996 [6,14,10] 45 1.38 [0.34,0.27,0.16,0.11]
[ ] 25 L =
]
[ ] 45 L =
]
[ ] 60 L =
]
(a) Impact of different E[Λ]on the results.0.9 0.99 0.999 0.999900.30.60.91.21.5 (b) Impact of different σ2on the results.
0.9 0.99 0.999 0.999900.30.60.91.21.5 (c) Impact of different εj
υ(%) on the results.
Fig. 8: Optimal Delay-Reliability Tradeoff Curves for Various Scenarios.
although NR has the advantage of tp= 6.25, however the
first non-zero point of the FSCLER -curve is still around 10.
This is because, under the limit of the safety threshold γs,
the decoding cannot be completed based on the bearer Pc
of single path. Determined by the characteristics of fountain
codes, DIEN will be unreliable if and only when all three
channels are error-coded, so Pe= 1−NLQ
j=1(1−Pj
e), the
extreme value of FSCLER is 0.9996. The strategy Stakes
effect based on FSCLER 22b, which is one of the foundations
of this paper.B. Numerical Results
The SCLER strategy Sis based on DIEN transmission, and
we have analyzed the data protection capabilities in the face of
passive attackers in Fig.5. This section will focus on analyzing
the optimal trade-off between delay and reliability under the
premise of security considerations.
In this subsection, we analyze the network characteristic
values based on the basic performance results. We set the
maximum transmission block length B= 100 , considering the
encoding characteristics, when g[t]<100, random redundant
Pswill be added to ensure that γjγτ≥γd. We consider B5G
DIEN containing three PDU sessions, whose transmission
characteristics are shown in Fig.7a. The accumulation and
arrival of Psfollow the probability distribution λ(·), and the
maximum arrival rate is Nmaxλ= 300 . For the convenience of
analysis, we assume that the minimum granularity of a single
timeslot data arrival is 25, namely ΛN= 0,25,50,75,100. We
sort the PDU sessions as 1-3 according to NR, Wifi, and LTE
bearers.Pcerasure at the PDU session layer is considered a
separate unreliability other than Pe,ε[n]does not vary with
time under the consider of there is no retransmission, and this
assumption is reasonable in mobile networks. The number of
time slots for each simulation run is NT= 105. Theoretical
results are plotted with diamond-shaped dots, while simulationresults are marked with lines. The basic simulation parameter
settings can be found in Table.II.
The Fig.8a shows the numerical results of the delay-
reliability trade-off. It illustrates that for fixed deletion prob-
abilities of εj
υ= [6,14,10],FSCLER can achieve the min-
imum average waiting delay while ensuring reliability for
average arrival rates E[Λ]of 25, 45, and 60. The simulation
results match well with the theoretical analysis shown by the
diamond-shaped dots, which shows that the derived optimiza-
tion problem of the threshold-based strategy accurately solves
the Eq.21. The delay-reliability trade-off curve is segmented,
which is consistent with the characteristics of the strategy.As the reliability FSCLER approaches its extreme value,
the delay increases rapidly, and the queue state becomes
more unstable. Under the same reliability constraints, as E[Λ]
increases, the corresponding average queue delay ¯LSalso
increases due to the higher throughput of DIEN at this time. In
summary, SCLER can achieve a latency-reliability trade-off,
and the threshold-based strategy is effective.
In Fig.8b, we use the same setting as in Fig.8a to analyze
the delay-reliability trade-off when E[Λ] = 45 andσ2=[0.78,
1.16, 1.38]. We observe that the arrival process λwith a
larger second-order moment σ2has a higher corresponding
average queue delay ¯LS, which is longer. As the reliability F
approaches 1−Pe, more delay overhead needs to be added
due to the long-tailed distribution of the reliability function.
This causes FSCLER to slowly approach the extreme value. In
conclusion, the variance has an effect on the reliability-delay
trade-off, which helps us to further analyze the threshold-based
strategy.Fig.8c shows the analysis of the delay-reliability trade-off
under different deletion probabilities for E[Λ] = 45 and three
channel transmission state as shown in Table.II. With the
deletion probability increases, a lower code rate is required
to achieve the same reliability value. In this case, Stends to
generate a smaller gto improve the coding redundancy, which | 2402.00002v1.pdf | 6d1be6c4-ffce-4114-b98d-59e39954d97b |
PREPRINT SUBMITTED TO IET 12
increases the queue delay in the DIEN.
(a)rth= 0.9996
(b)rth= 0.999
Fig. 9: Threshold-based strategy S,E[Λ]=45.
Fig.9 shows the specific strategy probability value of the
threshold policy. We analyze the change of Sunder different
reliability requirements rthforE[Λ]=45. The results in Fig.8a
have confirmed the effectiveness of the threshold strategy in
solving derivative problems. In Fig.9a, the threshold point is at
q[t] = 25 , and the probability of transmission is 0.75, while in
Fig.9b, the threshold point is at q[t] = 50 , and the probability
of transmission is 0.47. Considering both reliability and delay,
SCLER adjust the threshold policy point to achieve optimal
trade-off optimization.
VII.VII. C ONCLUSION
This paper investigates how to provide end-to-end low-
latency, high-reliability transmission methods between termi-
nal devices and edge nodes in Beyond 5G edge networks,
while considering potential eavesdroppers. We propose a se-
cure and reliable Raptor encoding-based transmission scheme
that mitigates the impact of asymmetric delay and bandwidth
differences in multi-path transmission, thus enhancing the
interaction between paths. We develop a sender-side queue
length-aware variable block-length encoding method and im-
plement multi-path split transmission, taking into account a
data security threshold that reduces data leakage risks on a
larger attack surface in CMT. By analyzing the optimal trade-
off decision problem for delay and reliability, we address
the constraint optimization problem using a threshold-based
strategy. Performance results demonstrate that our approach
is effective in achieving the optimization of both delay and
reliability, while ensuring data security.
APPENDIX A
MULTI -ACCESS PDU SESSION
ESTABLISHMENT PROCEDURE
The current existing mobile network does not yet support
the establishment of physically isolated multi-PDU sessions.Therefore, we propose the procedure of establishing diversity
interface PDU session suitable for SCLER
1) In SCLER, both UE and DN can initiate a PDU session
establishment request. The corresponding message is
PDU Session Establishment Request, which needs to
carry the requested DNN (Data Network Name), PDU
session type, etc.2) AMF will send the message to SMF in SBI(service-
based interface) format and will report the access type
currently supported by UE to the core network.
3) SMF will perform the authentication and authorization
of the PDU session, and register on the Unified Data
Management (UDM). These contents are not shown
in Fig. 3. SMF reads the service data of the user
subscribed by the user in the SCLER and invokes the
session management policy rule corresponding to the
user, which is used for the PDU session QoS control,
At this time, SMF has selected the DN and the Anchor
UPF.4) SMF sends PDR to UPF, (PDR is a kind of cell,
used to carry policy control information, including
rule name, forwarding action, service processing action,
etc.), UPF will be assigned with RAN and N3IWF
addresses, namely TEID, for RAN and N3IWF uplink
data transmission, corresponding to UPF-1 and UPF-2
respectively,
5) SMF requires AMF to relay the SMF’s NAS messages
(i.e.the session establishment accept the message and
the UPF’s core network tunnel information for uplink
from the RAN) to the UE.6) AMF initiates a radio resource request to the access
point, including 3GPP access and non-3GPP access, and
requires relaying NAS messages
7) The NG-RAN and Wifi set up the radio bearer and RAN
tunnel after receiving the request, and sending the NAS
message to the UE
8) NG-RAN and Wifi use session information (including
session id and RAN tunnel) to provide response to AMF
9) AMF relays session information to SMF
10) SMF updates the UPF after receiving the reply and
establishes the downlink tunnel from the UPF to the
base station
11) Finally, the tunnel from the access point to UPF-A is
generated. The UE and UPF-A split and aggregate the
data, and the encoding and decoding are also completed
in the UE and UPF-A.
REFERENCES
[1] 3GPP, “Summary of Rel-17 Work Items,” Technical report (TR) 21.917,
01 2023, version 17.0.1. [Online].[Online]. Available: https://www.3gpp.org/ftp/
Specs/archive/21 series/21.917/
[2] A. R. Javed, F. Shahzad, S. ur Rehman, Y . B. Zikria, I. Razzak, Z. Jalil,
and G. Xu, “Future smart cities requirements, emerging technologies,
applications, challenges, and future aspects,” Cities , vol. 129, p. 103794,
2022.
[3] J. Navarro-Ortiz, P. Romero-Diaz, S. Sendra, P. Ameigeiras, J. J. Ramos-
Munoz, and J. M. Lopez-Soler, “A survey on 5g usage scenarios and
traffic models,” IEEE Communications Surveys & Tutorials , vol. 22,
no. 2, pp. 905–929, 2020. | 2402.00002v1.pdf | 313f7da3-341c-4b01-b4c9-cdfefc8bd477 |
22,
no. 2, pp. 905–929, 2020.
[4] Z. Li, M. A. Uusitalo, H. Shariatmadari, and B. Singh, “5g urllc: Design
challenges and system concepts,” in 2018 15th international symposium
on wireless communication systems (ISWCS) . IEEE, 2018, pp. 1–6.
[5] M. Adhikari and A. Hazra, “6g-enabled ultra-reliable low-latency
communication in edge networks,” IEEE Communications Standards
Magazine , vol. 6, no. 1, pp. 67–74, 2022.
[6] Y . Zhang and Y . Zhang, “Mobile edge computing for beyond 5g/6g,”
Mobile Edge Computing , pp. 37–45, 2022.
[7] N. Hassan, K.-L. A. Yau, and C. Wu, “Edge computing in 5g: A review,”
IEEE Access , vol. 7, pp. 127 276–127 289, 2019. | 2402.00002v1.pdf | dbafc545-b425-4936-907e-e14dbe089451 |
PREPRINT SUBMITTED TO IET 13
[8] Y . Hui, N. Cheng, Z. Su, Y . Huang, P. Zhao, T. H. Luan, and C. Li,
“Secure and personalized edge computing services in 6g heterogeneous
vehicular networks,” IEEE Internet of Things Journal , vol. 9, no. 8, pp.
5920–5931, 2021.
[9] P. Tedeschi and S. Sciancalepore, “Edge and fog computing in critical
infrastructures: Analysis, security threats, and research challenges,” in
2019 IEEE European Symposium on Security and Privacy Workshops
(EuroS&PW) . IEEE, 2019, pp. 1–10.
[10] T. Yoshizawa, S. B. M. Baskaran, and A. Kunz, “Overview of 5g urllc
system and security aspects in 3gpp,” in 2019 IEEE Conference on
Standards for Communications and Networking (CSCN) . IEEE, 2019,
pp. 1–5.IEEE, 2019,
pp. 1–5.
[11] R. Chen, C. Li, S. Yan, R. Malaney, and J. Yuan, “Physical layer secu-
rity for ultra-reliable and low-latency communications,” IEEE Wireless
Communications , vol. 26, no. 5, pp. 6–11, 2019.
[12] J. M. Hamamreh, E. Basar, and H. Arslan, “Ofdm-subcarrier index
selection for enhancing security and reliability of 5g urllc services,”
IEEE Access , vol. 5, pp. 25 863–25 875, 2017.
[13] S.-Y . Lien, S.-L. Shieh, Y . Huang, B. Su, Y .-L. Hsu, and H.-Y . Wei,
“5g new radio: Waveform, frame structure, multiple access, and initial
access,” IEEE communications magazine , vol. 55, no. 6, pp. 64–71,
2017.
[14] V .6, pp. 64–71,
2017.
[14] V . K. Shrivastava, S. Baek, and Y . Baek, “5g evolution for multicast and
broadcast services in 3gpp release 17,” IEEE Communications Standards
Magazine , vol. 6, no. 3, pp. 70–76, 2022.
[15] Y . Xie and P. Ren, “Optimizing training and transmission overheads
for secure urllc against randomly distributed eavesdroppers,” IEEE
Transactions on Vehicular Technology , vol. 71, no. 11, pp. 11 921–
11 935, 2022.
[16] D. Xu and P. Ren, “Quantum learning based nonrandom superimposed
coding for secure wireless access in 5g urllc,” IEEE Transactions on
Information Forensics and Security , vol. 16, pp. 2429–2444, 2021.16, pp. 2429–2444, 2021.
[17] J. Farhat, G. Brante, R. D. Souza, and J. P. Vilela, “On the secure spectral
efficiency of urllc with randomly located colluding eavesdroppers,” IEEE
Internet of Things Journal , vol. 8, no. 19, pp. 14 672–14 682, 2021.
[18] 3GPP, “System architecture for the 5G System (5GS),” Technical
Specification (TS) 23.501, 09 2022, version 17.6.0. [Online]. Available:
https://www.3gpp.org/ftp/Specs/archive/23 series/23.501/
[19] H. Chen, R. Abbas, P. Cheng, M. Shirvanimoghaddam, W. Hardjawana,
W. Bao, Y . Li, and B. Vucetic, “Ultra-reliable low latency cellular
networks: Use cases, challenges and approaches,” IEEE Communications
Magazine , vol. 56, no.56, no. 12, pp. 119–125, 2018.
[20] M.-T. Suer, C. Thein, H. Tchouankem, and L. Wolf, “Multi-connectivity
as an enabler for reliable low latency communications—an overview,”
IEEE Communications Surveys & Tutorials , vol. 22, no. 1, pp. 156–169,
2019.
[21] M. Li, A. Lukyanenko, Z. Ou, A. Yl ¨a-J¨a¨aski, S. Tarkoma, M. Coudron,
and S. Secci, “Multipath transmission for the internet: A survey,” IEEE
Communications Surveys & Tutorials , vol. 18, no. 4, pp. 2887–2925,
2016.
[22] J. Ha and Y .-I. Choi, “Support of a multi-access session in 5g mobile
network,” in 2019 25th Asia-Pacific Conference on Communications
(APCC) . IEEE, 2019, pp. 378–383.IEEE, 2019, pp. 378–383.
[23] J. Wang and J. Liu, “Secure and reliable slicing in 5g and beyond
vehicular networks,” IEEE Wireless Communications , vol. 29, no. 1,
pp. 126–133, 2022.
[24] A. Cohen, R. G. D’Oliveira, S. Salamatian, and M. M ´edard, “Network
coding-based post-quantum cryptography,” IEEE Journal on Selected
Areas in Information Theory , vol. 2, no. 1, pp. 49–64, 2021.
[25] L. Huang, P. Ren, and D. Xu, “Cross-locking enabled multi-route
fountain coding for secure transmission,” in 2022 IEEE 95th Vehicular
Technology Conference:(VTC2022-Spring) . IEEE, 2022, pp. 1–5.
[26] M. Luby, “Lt codes,” in The 43rd Annual IEEE Symposium on Founda-
tions of Computer Science, 2002. Proceedings. IEEE Computer Society,
2002, pp.Proceedings. IEEE Computer Society,
2002, pp. 271–271.
[27] M. He, C. Hua, W. Xu, P. Gu, and X. S. Shen, “Delay optimal concurrent
transmissions with raptor codes in dual connectivity networks,” IEEE
Transactions on Network Science and Engineering , vol. 8, no. 2, pp.
1478–1491, 2021.
[28] Y . Cui, L. Wang, X. Wang, H. Wang, and Y . Wang, “Fmtcp: A fountain
code-based multipath transmission control protocol,” IEEE/ACM Trans-
actions on Networking , vol. 23, no. 2, pp. 465–478, 2014.
[29] M. Luby, A. Shokrollahi, M. Watson, and T. Stockhammer, “Raptor
forward error correction scheme for object delivery,” Tech. Rep., 2007.
[30] R. Abbas, M. Shirvanimoghaddam, T. Huang, Y .Li, and B. Vucetic,
“Performance analysis of short analog fountain codes,” in 2019 IEEE
Globecom Workshops (GC Wkshps) . IEEE, 2019, pp. 1–6.[31] S. Jain and R. Bose, “Rateless-code-based secure cooperative transmis-
sion scheme for industrial iot,” IEEE Internet of Things Journal , vol. 7,
no. 7, pp. 6550–6565, 2020.
[32] M. Yi, X. Ji, K. Huang, H. Wen, and B. Wu, “Achieving strong security
based on fountain code with coset pre-coding,” IET Communications ,
vol. 8, no. 14, pp. 2476–2483, 2014.
[33] Z. Guo, X. Ji, W. You, M. Xu, Y . Zhao, Z. Cheng, and D. Zhou,
“Delay optimal for reliability-guaranteed concurrent transmissions with
raptor code in multi-access 6g edge network,” Computer Networks ,
p. 109716, 2023. [Online].109716, 2023. [Online]. Available: https://www.sciencedirect.com/
science/article/pii/S1389128623001615
[34] O. C. Kwon, Y . Go, Y . Park, and H. Song, “Mpmtp: Multipath mul-
timedia transport protocol using systematic raptor codes over wireless
networks,” IEEE Transactions on Mobile Computing , vol. 14, no. 9, pp.
1903–1916, 2014.
[35] J. J. Nielsen, R. Liu, and P. Popovski, “Ultra-reliable low latency
communication using interface diversity,” IEEE Transactions on Com-
munications , vol. 66, no. 3, pp. 1322–1334, 2017.
[36] 3GPP, “Interface between the Control Plane and the User Plane Nodes,”
Technical Specification (TS) 29.244, 12 2022, version 18.0.1. [Online]. | 2402.00002v1.pdf | 821a886e-ac5a-40ed-9ad5-9574794ca786 |
[Online].
Available: https://www.3gpp.org/ftp/Specs/archive/29 series/29.244/
[37] M. Jadin, G. Tihon, O. Pereira, and O. Bonaventure, “Securing multi-
path tcp: Design & implementation,” in IEEE INFOCOM 2017-IEEE
Conference on Computer Communications . IEEE, 2017, pp. 1–9.
[38] Ł. Apiecionek, M. Sobczak, W. Makowski, and T. Vince, “Multi path
transmission control protocols as a security solution,” in 2015 IEEE
13th International Scientific Conference on Informatics . IEEE, 2015,
pp. 27–31.
[39] J. Liu, J. Xu, S. Li, X. Cui, and Y . Zhang, “A secure multi-path trans-
mission algorithm based on fountain codes,” Transactions on Emerging
Telecommunications Technologies , vol. 33, no. 5, p. e4450, 2022.33, no. 5, p. e4450, 2022.
[40] H. N. Noura, R. Melki, and A. Chehab, “Network coding and mptcp:
Enhancing security and performance in an sdn environment,” Journal of
Information Security and Applications , vol. 66, p. 103165, 2022.
[41] J. Yang and S. Papavassiliou, “Improving network security by multipath
traffic dispersion,” in 2001 MILCOM Proceedings Communications for
Network-Centric Operations: Creating the Information Force (Cat. No.
01CH37277) , vol. 1. IEEE, 2001, pp. 34–38.
[42] S. K. Singh, T. Das, and A. Jukan, “A survey on internet multipath
routing and provisioning,” IEEE Communications Surveys & Tutorials ,
vol. 17, no. 4, pp. 2157–2175, 2015.
[43] C.-F. Kuo, A.-C. Pang, and S.-K. | 2402.00002v1.pdf | ea19ea58-28b6-4078-815c-375256773ad5 |
Kuo, A.-C. Pang, and S.-K. Chan, “Dynamic routing with security
considerations,” IEEE transactions on parallel and distributed systems ,
vol. 20, no. 1, pp. 48–58, 2008.
[44] 3GPP, “Procedures for the 5G System (5GS),” Technical Specification
(TS) 23.502, 09 2022, version 17.6.0. [Online]. Available: https:
//www.3gpp.org/ftp/Specs/archive/23 series/23.502/
[45] Y . R. Julio, I. G. Garcia, and J. Marquez, “R-iot: An architecture
based on recoding rlnc for iot wireless network with erase channel,”
inInformation Technology and Systems: Proceedings of ICITS 2020 .
Springer, 2020, pp. 579–588.
[46] S. Yang and R. W. Yeung, “Network communication protocol design
from the perspective of batched network coding,” IEEE Communications
Magazine , vol. 60, no. 1, pp.60, no. 1, pp. 89–93, 2022.
[47] A. Boualouache and T. Engel, “Federated learning-based scheme for
detecting passive mobile attackers in 5g vehicular edge computing,”
Annals of Telecommunications , pp. 1–20, 2022.
[48] W. Huang, “Coding for security and reliability in distributed systems,”
Ph.D. dissertation, California Institute of Technology, 2017.
[49] A. Shokrollahi, “Raptor codes,” IEEE transactions on information
theory , vol. 52, no. 6, pp. 2551–2567, 2006.
[50] R. Karp, M. Luby, and A. Shokrollahi, “Finite length analysis of lt
codes,” in International Symposium onInformation Theory, 2004. ISIT
2004. Proceedings. IEEE, 2004, p. 39.
[51] A. Shokrollahi, “Theory and applications of raptor codes,” Proceedings
of MathKnow , pp. 59–89, 2009.59–89, 2009.
[52] P. Wang, G. Mao, Z. Lin, M. Ding, W. Liang, X. Ge, and Z. Lin, “Perfor-
mance analysis of raptor codes under maximum likelihood decoding,”
IEEE Transactions on Communications , vol. 64, no. 3, pp. 906–917,
2016.
[53] E. G. Str ¨om, P. Popovski, and J. Sachs, “5g ultra-reliable vehicular
communication,” arXiv preprint arXiv:1510.01288 , 2015.
[54] F. Wang, G. Wen, Z. Peng, T. Huang, and Y . Yu, “Event-triggered
consensus of general linear multiagent systems with data sampling
and random packet losses,” IEEE Transactions on Systems, Man, and
Cybernetics: Systems , vol. 51, no. 2, pp. 1313–1321, 2019. | 2402.00002v1.pdf | 7c3a0cf3-1a52-48fb-9d14-3f1cda86ed3a |
PREPRINT SUBMITTED TO IET 14
[55] P. K. Khatua, V . K. Ramachandaramurthy, P. Kasinathan, J. Y . Yong,
J. Pasupuleti, and A. Rajagopalan, “Application and assessment of
internet of things toward the sustainability of energy systems: Challenges
and issues,” Sustainable Cities and Society , vol. 53, p. 101957, 2020. | 2402.00002v1.pdf | 20f39eaa-3df0-4a6f-ac5b-13a832084a02 |
Draft version January 2, 2024
Typeset using L ATEXtwocolumn style in AASTeX63
Prompt emission of relativistic protons up to GeV energies from M6.4-class solar flare on July 17,
2023
C.E. Navia,1M.N. de Oliveira,1and A.A. Nepomuceno2
1Instituto de F´ ısica, Universidade Federal Fluminense, 24210-346, Niter´ oi, RJ, Brazil
2Departamento de Ciˆ encias da Natureza, Universidade Federal Fluminense, 28890-000, Rio das Ostras, RJ, Brazil
ABSTRACT
We show evidence of particle acceleration at GEV energies associated directly with protons from the
prompt emission of a long-duration M6-class solar flare on July 17, 2023, rather than from protons
acceleration by shocks from its associated Coronal Mass Ejection (CME), which erupted with a speed
of 1342 km/s.Solar Energetic Particles (SEP) accelerated by the blast have reached Earth, up to
an almost S3 (strong) category of a radiation storm on the NOAA scale. Also, we show a temporal
correlation between the fast rising of GOES-16 proton and muon excess at ground level in the count
rate of the New-Tupi muon detector at the central SAA region. A Monte Carlo spectral analysis based
on muon excess at New-Tupi is consistent with the acceleration of electrons and protons (ions) up to
relativistic energies (GeV energy range) in the impulsive phase of the flare. In addition, we present
another two marginal particle excesses (with low confidence) at ground-level detectors in correlation
with the solar flare prompt emission.
Keywords: sun:activity, high-speed stream, cosmic rays modulation
1.INTRODUCTION
Since 1950 the observation of solar energetic particles
from the solar flares and coronal mass ejections (CMEs)
have been done with ground-level experiments, such as
the neutron monitors (NMs) (Meyer et al. 1956; Simp-
son 2000; Moraal et al.1956; Simp-
son 2000; Moraal et al. 2000) as well as the solar neutron
telescope network (Hu & Semones 2022; Vald´ es-Galicia
et al. 2009), all around the world. These observations
have yielded a lot of new information. For instance,
the existence of a prompt and gradual emission of solar
energetic particles (SEP) in flares and CMEs, respec-
tively, the correlations of the cosmic ray intensity with
CMEs and other solar disturbances crossing the Earth,
etc. (Chupp et al. 1987; Moraal et al. 2000)
Also, the solar modulation of galactic cosmic rays is
inversely correlated with solar activity, inferred through
the number of sunspots, which can be the key to under-
standing more about space weather (Cade III & Chan-
Park 2015).
Nowadays, particles accelerated to near the Sun can
be detected by space-borne instruments such as the
Corresponding author: C.E.Navia
carlos navia@id.uff.brHigh-Energy Proton and Alpha Detector (HEPAD)
on the Geostationary Operations Environmental Satel-
lite (GOES) and the Advanced Composition Explorer
(ACE) spacecraft at Lagrange L1 point, through the
Electron Proton Alpha Monitor (EPAM) and the Solar
Isotope Spectrometer (SIS), among others.
Not all of the solar energetic particles can be measured
at ground level. Even those SEPs from solar events with
a good geoeffectiveness can be dissipated by the IMF,
or deflected or captured by the Earth’s magnetic
field or until absorbed by atmosphere.
On the other hand, ground-level enhancements
(GLEs), typically in the MeV-GeV energy range, are
sudden increases in cosmic ray intensities registered in
most cases by NMs. GLEs are quite rare events, and
fewer than 100 GLEs have been observed by NMs in the
last 70 years. | 2401.00002v1.pdf | a7028f4f-c7a6-4db9-8355-b1b8e6ec089f |
In most cases, the NMs that observed
GLEs are located at regions with small geomagnetic
rigidity cutoff, that is, at high latitudes (Shea & Smart
2012).
The GLEs follow the solar radiation storms, solar en-
ergetic particles (mostly protons) observed by GOES.
They occur when a large-scale magnetic eruption, a
coronal mass ejection and associated solar flare, accel-
erates charged particles in the solar atmosphere to high
energies.arXiv:2401.00002v1 [astro-ph.SR] 28 Nov 2023 | 2401.00002v1.pdf | 94e6f8c4-defa-4b8b-898d-9d43513d9783 |
2
Figure 1. Top panel: strength of Earth’s magnetic field in 2000, according to measurements by the European Spacing Agency’s
SWARM satellites. Bottom left panel: Corsika-Fluka simulation results of the lateral particle distribution in proton air-showers
of cosmic rays, as detected from several ground level detectors. Bottom right panel: correlation between Rmax(from Corsika-
Fluka) versus the geomagnetic Stormer rigidity cutoff of six different places (black circles) including the SAA-CR region (blue
square). The red solid line is a linear fit, and the two dotted red lines delimit the region with a confidence of ±1σ.However, in the present case, despite a radiation storm
reaching above the S2-class on the NOAA scale on July
18, 2023, it did not generate a GLE, only a prompt emis-
sion of relativistic protons (ions) above GeV energies,
during the phase eruptive, and observed by ground-level
detectors strategically located, within the SAA central
region (New-Tupi muon detector) and by the (Yangba-
gin muon telescope) at the Yangbajing Cosmic Ray Ob-
servatory (Tibet 4440 m a.s.l) (Zhang et al. 2010).
Also, we looked for any signal in the counting rate
at the Neutron Monitor’s (NM) network around the
world from Neutron Monitor Data Base (NMDB) https:
//www.nmdb.eu/nest/ , with negative results. However,
we found a low confidence signal only at Kerguelen NM,
at geographical coordinates (49.3S, 70.3E), altitude of 33
m a.s.l, and an effective vertical cutoff rigidity of 1.14
GV. We present details of these observations. | 2401.00002v1.pdf | 450e0a41-393c-4d77-8a17-ebf4da1a2ecf |
We present details of these observations.
2.NEW-TUPI TELESCOPE WITHIN THE SOUTH
ATLANTIC ANOMALYThe New-Tupi muon detector is completely unmoder-
ated (without no surrounding lead or other material).
The muon detection energy threshold is about 200 MeV
(see Appendix A). That contrasts with other muon de-
tectors that have, in most cases, a surrounding lead ma-
terial with a thickness of up to 5 cm.
The shielding effect of the Earth’s magnetic field on
cosmic ray particles is quantified by the magnetic rigid-
ity cutoff from a specific location (Smart & Shea 2009).
The smaller the rigidity cutoff, the lower the energy
cosmic ray particles penetrate the magnetosphere. On
the other hand, a restricted area between latitudes 20
and 40 of the southern hemisphere, over South Amer-
ica and the Atlantic Ocean poses a geomagnetic field
with an anomalously lower intensity (around 22,000 nT).
The region is known as the South Atlantic Anomaly
(SAA) (Pav´ on-Carrasco & De Santis 2016).According
to Swarm’s satellite observations (Finlay et al. 2020),
the SAA appears splitting into two, a smaller area over
the Atlantic Ocean in southwest Africa and a larger area | 2401.00002v1.pdf | 6fd1c3d2-ea16-45cd-9860-a6a9bd948f68 |
3
over eastern South America. Fig. 1 (top panel) summa-
rizes the situation. We would like to point out that the
location of the New-Tupi telescope coincides with the
central part of the SAA indicated by the arrow on the
left of Fig. 1 (top panel).
The main effect of the SAA is on the satellites since
the ’70s. We know the frequent failures when they pass
through the SAA region. A large amount of charged
particles precipitation in this region damages and per-
turbates the satellites’ electronics. Also, according to
the results from the PAMELA detector at satellite’s al-
titudes (Casolino et al. 2009), the effect of geomagnetic
cutoff on low-energy particles is present in high latitudes
close to the poles and also in the SAA region, composed
mostly of low energy cosmic protons (E <200 MeV ).
In other words, the Pamella satellite has shown that
the SAA introduces a sub-cutoff in the magnetic rigid-
ity, below the Stormer’s magnetic rigidity cutoff.We
show that the SAA also affects secondary cosmic rays
detected at ground level. As the horizontal magnetic
component on Earth’s surface is smaller on the SAA,
the magnetic lateral dispersion of the secondary parti-
cles forming an air shower is smaller too. The effect in-
creases the number of particles reaching a detector. In
other words, this behavior mimics a magnetic rigidity
sub-cutoff below the Stormer’s rigidity cutoff. We show
that effect through a Monte Carlo simulation based on
CORSIKA-Fluka code (Heck et al. 2012; Battistoni et al.
2008), where 1 .0×106proton air-showers are simulated,
taking into account the magnetic coordinates (latitude,
longitude) and height of several places where detectors
are installed (mostly neutron monitors).
Fig. 1 bottom left panel shows the lateral particle dis-
tribution in air-showers of cosmic rays, as detected from
several ground-level detectors.In all cases, there is a fast
rise of particles with the shower lateral development un-
til reach um maximum value that happens for different
values of R, called hereafter as RmaxWe can see that the
number of shower particles at Rmaxin the SAA central
region (SAA-CR) rigidity 9.6 GV is higher than at Rome
and Athens, both with the rigidity of 6.3 GV, and 8.5
GV, respectively, i.e., minors than the SAA-CR. Already
Fig. 1 bottom right panel, shows a correlation between
Rmaxversus the geomagnetic Stormer rigidity cutoff of
six different places (black circles), including the SAA-
CR (blue square). The solid red line is a linear fit, and
the two dotted red lines delimit the region with signifi-
cance of ±1σ. Only two places are out from the ±1.0σ
significance region, the Thule (Groenlandia) in the low-
est rigidity region and SAA-CR in the highest rigidity
region. | 2401.00002v1.pdf | 1afd5f33-0e3c-45ee-a2fa-3a4698f6b95c |
The high Stormer’s rigidity of SAA-CR does not
correspond to the high value of Rmaxexpected by the
Figure 2. Left panel: NASA’s Solar Dynamics Observatory
image of a blast near the sun’s southwestern limb, erupted
from a big sunspot AR3363, with onset at the last hours of
July 17, 2023. Right panel: LASCO-C2 coronograph image
on July 18, 2023, at 00:42 UT showing the CME eruption
associated to M6-class flare with a speed of 1342 km/s.
correlation. From an interpolation, it is possible to see
that the small value of Rmaxat SAA-CR correspond to
the rigidity of only 3 .1+3.0
−1.7GV, within a confidence of
±1.0σ. This behavior of having a location close to the
Equator, with a nominal lower magnetic rigidity cutoff,
favors the observation of phenomena such as the SEPs.3.ANALYSIS
On July 17, 2023, at ∼18h UT, the active region AR
13363 had an explosion, reaching an M6-class solar flare
followed by a resplendent coronal mass ejection. Fig. 2
left panel shows the image from the Solar Dynamo ob-
servatory of the blaze of fire responsible for the X-ray
flux reaching M6-class flare. Already the right panel
shows the LASCO-C2 coronograph image of its associ-
ated CME on July 18, 2023, at 00:42 UT.
NOAA prediction models confirmed that a CME orig-
inated in the powerful M6-class flare from sunspot
AR3363 would pass through the magnetosphere on July
20, triggering at least a G1-class (minor) geomagnetic
storm. However, no magnetic storms were observed.
Fig. 3 shows the GOES-18 X-ray flux (upper panel)
and the GOES-16 proton flux (bottom panel). The X-
ray flux peaks at 18:00 UT, while the proton flux has two
peaks.The first (in orange) is due to the acceleration
of protons during the impulsive fast-rising phase of the
flare peaking at 18:09 UT. The delay between the X-ray
and proton flux peaks is because the proton velocity is
slightly less than c, and the proton path is longer. The
second peak are the protons accelerated by CME shocks,
peaking at 18:14 UT, and it’s the so-called gradual phase
and is characterized by its long duration, up to several
days.
Fig. 4 shows the temporal coincidence between the
GOES proton flux in the impulsive phase and New-Tupi | 2401.00002v1.pdf | ddaad891-d4a2-446c-8f45-d42ab79f5630 |
4
Figure 3. Top panel: GOES-18 X-ray flux in two wave-
lengths. Bottom panel: GOES-16 proton flux in three energy
bands. Both on July 17-18, 2023. The orange area at the
bottom (is a visual guide only) highlights the proton prompt
emission during the flare impulsive phase.
Figure 4. Top panel: GOES-16 proton flux in three energy
bands. Bottom panel: New-Tupi counting rate expressed in
variation (%). Both on July 17-18, 2023. The orange area at
the top (is a visual guide only) highlights the proton prompt
emission during the flare impulsive phase.
Figure 5. From top to bottom: GOES-16 proton flux in
three energy bands, 2nd to 4th panels, counting rate at New-
Tupi muon detector, Yan ba Jing muon telescope, and Ker-
guelen Neutron Monitor, respectively. To five consecutive
days, from July 15-19, 2023.
muon excess.Particles (mostly proton) are accelerated
in this phase exclusively by the flare, during the fast-
rising until to reach the first peak (orange sector) in
Fig. 4.
However, in the so-called gradual phase, protons ac-
celerated by CME’s shocks, the proton flux does not
reach the GeV energy range because there are no excess
muons at ground level.
4.SPECTRAL ANALYSIS
We perform a Monte Carlo simulation of air showers
initiated by SEP (protons) using the CORSIKA code
(Heck et al. 2012), together with the FLUKA interac-
tion model (Battistoni et al. 2008), that works well at
GeV and sub-GeV energies, including secondary parti-
cle decay. The surviving particles are tracked through
the atmosphere until they reach ground level (sea level).
Most particles are muons with a small contribution of
electrons and nucleons. | 2401.00002v1.pdf | 5918ad2b-e1d1-4049-b4de-a53cead731ac |
Most particles are muons with a small contribution of
electrons and nucleons.
The aim is to obtain the yield function, Sµ(EP), that
is, the number of muons at sea level per primary proton,
for an estimate of the upper limit of the integral proton
flux in the GeV energy range, associated with the im-
pulsive phase of M-6-class flare with onset on July 17,
2023, at ∼18 UT.
Fig. 6 (black squares) shows the Monte Carlo output
under the New-Tupi geomagnetic conditions and vertical
proton incidence, and fitting as | 2401.00002v1.pdf | 52f0820f-75d9-4244-96b0-689d39b24885 |
5
Figure 6. Yield function, as the number of muons at the sea
level per proton (vertical incidence), as a function of incident
proton energy, from CORSIKA-FLUKA simulations, taking
into account the SAA’s central region magnetic conditions.
The red dashed curve shows a fit function.
Figure 7. Correlation between the coefficient Ap and the
spectral index β. All possible values of Ap and βcompatible
with the New-Tupi muon counting rate excess (black solid
curve) and the high-energy GOES-16 proton fluence F (red
dot curve). These quantities are defined by equations 3 and
5, respectively.
Sµ(EP) =AµEν
Pexp(−E0/EP), (1)
where Aµ= (6.8±1.4)×10−3,ν= 1.18±0.24, and
E0= 10.2±2.1 GeV. Fig. 6 shows the fits (red dot line).
Figure 8. Integral proton flux: the red circles represent the
GOES-16 data.Integral proton flux: the red circles represent the
GOES-16 data. It corresponds to the GOES-16 proton flux
prompt emission (orange area in Fig. 3. The black square
represents the proton flux obtained from the muon excess
on the New-Tupi detector (and Monte Carlo calculations),
observed in coincidence with the radiation storm.
In addition, we assume here that the energy spectrum
of solar protons in the GeV energy range, which is in
the high-energy tail of the SEP spectrum, can be fitted
by a single power-law function.
JP(EP) =APEβ
P. (2)
There are two unknown quantities in the above power-
law function: the coefficient APand the spectral index
β.
A convolution between the yield function Sµ(EP) and
the proton spectrum JP(EP) gives the response function
(Augusto et al. | 2401.00002v1.pdf | 51e91cd1-dead-4397-823d-4e1c447ad045 |
2016b), which is the number of muons in
the excess signal at New-Tupi detector generated by the
SEP during the period T. We express this convolution
as
Jµ=Z∞
EminSµ(EP)F(θ)APEβ
PdEP, (3)
where F(θ)∼exp(θ/C) is the pitch angle distribution
(Shea & Smart 1982; Miroshnichenko et al. 2005). In
the central region of SAA (New-Tupi), the transverse
geomagnetic component is only B⊥= 18 .3 mT, that
is, almost 80% smaller than the transverse component,
at the same latitude, but outside from the SAA region,
favoring the focusing factor of the geomagnetic parallel | 2401.00002v1.pdf | c4a4ec30-6878-4de3-bdb9-369c158b9bf9 |
6
geomagnetic on the incident solar protons (small pitch
angles). For the present event, we found exp( θ/C)∼1.
The muon excess associated with protons emitted dur-
ing the impulsive phase (see Fig. 4), and considering
an effective angular aperture of 60 degrees around the
zenith of the New-Tupi detector, the counting rate ex-
cess is
Jµ= (3.20±2.51)×10−3muons/ (cm2s sr),(4)
Furthermore, we also obtain the integrated time pri-
mary fluence as
F=TZ∞
EmindEPAPEβ
PEP
. (5)
For the present case, the GOES-proton fluence in the
high-energy region (Ep ≥50) MeV is
F∼(46.0±34.9)MeV/ (cm2sr), (6)
The terms on the left side of Eq. 1 and Eq. 5 are
known.1 and Eq. 5 are
known. Thus, we can consider all possible values of β
andAPcompatible with the observed muon counting
rate excess, Jµ, and the integrated GOES-proton fluence
F. Fig. 7 summarizes the situation. Giving:
Ap= (1.20±0.96)×10−3/(cm2s sr GeV )
andβ= 1.89±1.10.
To make a comparison with satellite GOES data, we
obtain the integral proton flux in the GeV energy range
as
JP(> E P) =Z∞
EminAPE−β
PdEP. (7)
Fig. 8 shows the results of the integral proton flux ob-
tained from the New-Tupi muon excess observed in co-
incidence with the radiation. The red circles represent
the GOES-16 data, and the black squares represent the
expected proton flux in the GeV energy range obtained
from Monte Carlo, muon excess at the New-Tupi detec-
tor, and the GOES-proton fluence.The origin of this
transient event was the solar eruption, an M6-class flare
(prompt emission), accelerating protons (ions) up to rel-
ativistic energies, GeV energy range.
5.CONCLUSIONS
We have reported evidence of SEPs accelerated up to
GeV energies during the eruptive phase of the M6-class
solar flare on July 17, 2023. The result comes from
a timing analysis of a muon excess from the New-Tupidetector at the SAA central region. Muons at New-Tupi
are produced by protons (ions) interaction in the upper
atmosphere reaching the Earth with a magnetic rigidity
above 3 GV ( ∼3 GeV for protons).
In most cases, SEP (mostly protons) detected by the
GOES-16 spacecraft shows two steps. An impulsive
phase, where the acceleration of protons (ions) is by the
prompt emission of flare, followed by a gradual phase of
long duration, where the acceleration of protons (ions)
is by the associated CME shock waves.We want to point out that the muon excess produced
by SEPs with an effective rigidity above the cutoff ( ∼
3 GV) at the New-Tupi muon detector is in temporal
coincidence only with the GOES proton flux from the
impulsive phase (see Fig. 4) . Consequently, in the grad-
ual phase, the protons accelerated by CME’s shocks do
not reach the GeV energy range because a muon excess
is absent at ground level.
A marginal muon excess also appears on the Yan ba
Jing S-21(pointing 21 degrees south) muon telescope (in
Tibet). Also, a marginal particle excess is seen only in
the French Kerguel NM (close to the South polar re-
gion). In both cases, the excesses are in (temporal) co-
incidence with the GOES proton flux (impulsive phase).
However, it is hard to verify whether these excesses are
genuine due to low confidence or simply fluctuations in
the detectors’ count rate. | 2401.00002v1.pdf | 93441254-0e4a-49df-b252-4b445ce4214e |
From a Monte Carlo analysis, we show that the SAA
central region is favourable to the observation of tran-
sient solar events, especially SEP, because the magneto-
sphere has a ”dip ´´in this region, weakening the geo-
magnetic
field strength and allowing the entrance of charged
particles at large deeps in a region not far from the geo-
graphic Equator, giving a rigidity sub-cutoff around 3.1
GV in a place where the conventional Stormer geomag-
netic rigidity cutoff is around 10 GV.
6.ACKNOWLEDGMENTS
This work is supported by the Rio de Janeiro
Research Foundation (FAPERJ) under Grant E-
26/010.101128/2018. We thank to NMDB Database
(www.nmdb.eu), founded under the European Union
FP7 Program (Contract No. 213007) by provide NMs
data and the Space Weather Prediction Center from
NOAA for its open data policy.
APPENDIX | 2401.00002v1.pdf | 994d5272-abf0-4345-9e07-42a701d1035c |
7
A.NEW-TUPI DETECTOR
The New-Tupi telescope is built with four identical particle detectors, forming two telescopes, as shown in Fig. 8
from Augusto et al. (2016a). Each detector consisting of an Eljen EJ-208 plastic scintillator slab of 150 cm x 75 cm x
5 cm and a Hamamatsu R877 photomultiplier of 127 millimeters in diameter, packaged in a pyramidal box. The PMT
high voltage divider, amplifier, and high voltage power supplier are in the ORTEC ScintiPackTM Photomultiplier
Base 296.
From February 6, 2023, we have implemented a data acquisition system using a VERTILON high-speed pulse
counting system (MCPC618-8 Channel). allowing for direct connection with the PMTs without the need for external
preamplifiers, with a 250 MHz count rate per channel.
Now the detector is working only in scaler mode or single particle technique (Aglietta et al. 1996), where the single hit
rates of all four PMTs, are recorded once a second.However, so far, only two detectors are working. The coincidences
among these detectors of each telescope will be implanted.
Also, the barometric coefficients for cosmic muon fluxes at the Earth’s surface can be obtained using the CORSICA
code in Kovylyaeva et al. (2013). For New-Tupi detector conditions and at sea level, the barometric coefficient is about
-0.14% per mb, about eight to nine times less than the typical barometric coefficient in NMs. | 2401.00002v1.pdf | 8d39a3bd-2393-479c-8fde-720ba10f0558 |
8
REFERENCES
Aglietta, M., Alessandro, B., Antonioli, P., et al. 1996,
Astrophysical Journal, 469, 305
Augusto, C., Navia, C., de Oliveira, M., Nepomuceno, A.,
& Fauth, A. 2016a, arXiv preprint arXiv:1603.08863
Augusto, C., Navia, C., de Oliveira, M. N., Fauth, A., &
Nepomuceno, A. 2016b, Publications of the Astronomical
Society of Japan, 68, 8
Battistoni, G., Cerutti, F., Empl, A., et al. 2008, Nuclear
Physics B-Proceedings Supplements, 175, 88
Cade III, W. B., & Chan-Park, C. 2015, Space Weather, 13,
99
Casolino, M., De Simone, N., De Pascale, M. e., et al.2009,
Nuclear Physics B-Proceedings Supplements, 190, 293
Chupp, E., Debrunner, H., Fl¨ uckiger, E., et al. 1987, The
Astrophysical Journal, 318, 913
Finlay, C. C., Kloss, C., Olsen, N., et al. 2020, Earth,
Planets and Space, 72, 1
Heck, D., Pierog, T., & Knapp, J. 2012, Astrophysics
Source Code Library, ascl
Hu, S., & Semones, E. 2022, Journal of Space Weather and
Space Climate, 12, 5
Kovylyaeva, A., Dmitrieva, A., Tolkacheva, N., &
Yakovleva, E. 2013, Journal of Physics: Conference
Series, 409, 012128Meyer, P., Parker, E., & Simpson, J. 1956, Physical
Review, 104, 768
Miroshnichenko, L., Klein, K.-L., Trottet, G., et al. | 2401.00002v1.pdf | af917804-60e9-4a93-bcfd-54583ba90713 |
2005,
Journal of Geophysical Research: Space Physics, 110
Moraal, H., Belov, A., & Clem, J. 2000, Space Science
Reviews, 93, 285
Pav´ on-Carrasco, F. J., & De Santis, A. 2016, Frontiers in
Earth Science, 4, 40
Shea, M., & Smart, D. 1982, Space Science Reviews, 32, 251
—. 2012, Space science reviews, 171, 161
Simpson, J. A. 2000, in Cosmic Rays and Earth:
Proceedings of an ISSI Workshop, 21–26 March 1999,
Bern, Switzerland, Springer, 11–32
Smart, D., & Shea, M. 2009, Advances in Space Research,
44, 1107
Vald´ es-Galicia, J., Muraki, Y., Watanabe, K., et al. 2009,
Advances in Space Research, 43, 565
Zhang, J., Tan, Y., Wang, H., et al.2010, Nuclear
Instruments and Methods in Physics Research Section A:
Accelerators, Spectrometers, Detectors and Associated
Equipment, 623, 1030 | 2401.00002v1.pdf | 35aae815-6a3f-4d4c-a9ba-917154325074 |
arXiv:2312.00001v1 [math.ST] 8 Jul 2023ON RANDOM PAIRWISE COMPARISONS MATRICES AND
THEIR GEOMETRY.
JEAN-PIERRE MAGNOT
Abstract. We describe a framework for random pairwise comparisons
matrices, inspired by selected constructions releted to th e so called
inconsistency reduction of pairwise comparisons (PC) matr ices. In to
build up structures on random pairwise comparisons matrice s, the set
up for (deterministic) PC matrices for non-reciprocal PC ma trices is
completed. The extension of basic concepts such as inconsis tency indexes
and geometric mean method are extended to random pairwise co mparisons
matrices and completed by new notions which seem useful to us . Two
procedures for (random) inconsistency reduction are sketc hed, based on
well-known existing objects, and a fiber bundle-like decomp osition of
random pairwise comparisons is proposed.
Keywords: Pairwise comparisons, stochastic models, Abelian Yang-Mi lls
theory, inconsitency, reciprocity.MSC (2020): 90B50, 81T13, 91B06
Contents
Introduction 2
1. A not-so-short introduction to pairwise comparisons as a class of abelian Yang-Mills theory 3
2. A theory of non reciprocal pairwise comparisons and weigh ts: deterministic settings 6
2.1. Preliminary settings on Mn(R∗
+)and its remarkable subsets 6
2.2. Weights as projective coordinates 7
2.3. (Reciprocal) pairwise comparisons 8
2.4. Non-reciprocal pairwise comparisons and related indi cators 9
2.5. On the geometry of the tower CPCn⊂PCn⊂Mn(R∗
+). 10
2.6. Inconsistency on Mn(R∗
+) 12
3. Random pairwise comparisons 14
3.1. The settings 14
3.2. Stochastic analogs to deterministic constructions 15
3.3. Minimizing inconsitency: a sketch of two methods 16
4. | 2312.00001v1.pdf | e1a999e8-36fa-4b8f-bdc4-17eb2ad19a2d |
Minimizing inconsitency: a sketch of two methods 16
4. A formal geometric picture for random pairwise compariso ns and geometric concerns 18
5. Outlook 19
Appendix: One motivating example on exchange rates 19
References 20
1 | 2312.00001v1.pdf | dfa9e135-d4ee-4b25-a1e0-08f20e025d6e |
2 JEAN-PIERRE MAGNOT
Introduction
The wide popularity of the pairwise comparison methods in th e field of
multi- criteria decision analysis mostly due to their simpl icity and their
adaptability. Since it is a natural approach to compair two b y two the items
under control, it is not surprising that the first systematic use of pairwise
comparisons is attributed to Ramon Llull [5] during the thir teenth-century.
Very probabily, people interested in comparisons passed fr om qualitative
to quantitative evaluations for the obvious gan in simplici ty of interpretation.
The twentieth-century precursor of the quantitative use of pairwise comparison
was Thurstone [39] and the pairwise comparison method has be en progressively
improved till the groundbreaking work of Saaty [36]. In this article Saaty
proposed the Analytic Hierarchy Process (AHP), a new multip le-criteria
decision-making method. Thanks to the popularity of AHP the pairwise
comparison method has become one of the most frequently used decision-
making techniques. Its numerous applications include most existing fields
such as economy (see e.g.Its numerous applications include most existing fields
such as economy (see e.g.[32]), consumer research [10], man agement [40],
military science [6, 13], education and science (see e.g. [2 5]) among others.
Besides its great popularity, its wide and intensive applica tion in everyday’s
life highlights tremendouly its weakness with respect to th e following to
aspects:
•Popularity of the decision-making methods also makes them v ulnerable
to at- tacks and manipulations. This problem has been studie d by
several researchers including Sasaki [37], on strategic ma nipulation
in the context of group decisions with pairwise comparisons . Some
aspects of decision manipulation in the context of electora l systems
are presented in e.g. [38].
•Even in honestly made evaluations, a small change in the scor es in
the pairwise comparisons [43] or of the secondary criteria s uch as
the so-called inconsistency index [31] can change drastica lly the final
result of the ranking procedure. | 2312.00001v1.pdf | 7c49cc01-cb37-46ac-b44c-ee637878c8ca |
We have recently linked explicitely inconsistency in pairw ise comparisons
(PC) with Yang-Mills theories, showing a possible (non-Abe lian) generalization
of the classical framework [27, 29]. Our approach is not the o nly one where
pairwise comparisons coefficients are not numerals, see e.g. [2, 20, 32, 45], but
the link with existing lattice-based or distretized gauge t heories appears to us
very promising, not only because of the very wide literature on the subject,
but also because gauge theories also seem to be promising can didates to
furnish applications out of the field of physics, see e.g. the amazing [26], or
[42] in Neuronal network and [8, 14, 47] in finance. We have her e to precise
that discretization of gauge theories (in order to recover a continuum limit)
are very often treated under the lights of Whitney approxima tions [46] of
differential forms and connections on trivial principal bun dles, see e.g.[1],
and very rarely from the viewpoint of the discretization com monly proposed
by quantum gravity [35] based on the computation of the holon omies (in the | 2312.00001v1.pdf | ef767ad2-1683-4162-b6f9-b37fc98a6f28 |
ON RANDOM PAIRWISE COMPARISONS MATRICES AND THEIR GEOMETRY . 3
physics sense) of the connections along the 1-vertices of th e triangulation
(see [28] for a rigorous description of the discretization p rocedure). These
are holonomy-like discretized connections that we conside r in this work, that
can therefore be identified with function from oriented 1-ve rtices with values
in the structure group, that satisfy so-called reciprocity conditions.
In this communication, we intend to develop a theory of rando m pairwise
comparisons matrices that was shortly sketched at the end of [29], and
describe the structures of the proposed sets, in order to ext end the notions of
weight, consistency, inconsistency indicator, in a stocha stic context focusing
on the very usual choice of R∗
+as a structure group.Inspired by the
correspondence between R∗
+−gauge theories in quantum gravity and classical
pairwise comparisons (PC) matrices highlighted in [29], we follow up this first
approach by replacing a pairwise comparisons matrix by a pro bability on the
space of pairwise comparisons matrices, understood as a ran dom variable,
in the same spirit as the statistical mechanics replace a par ticule with fixed
position by a probability. In other words, we propose here a fi rst quantization
of pairwise comparisons. This seems necessary to us in view o f the variations,
dishonest or not, that it is important to modelize in order to have a better
control on them as explained before. For this, the necessity of the use of
non-reciprocal pairwise comparisons matrices, an extende d class of pairwise
comparisons that may not follow the logic rule of reciprocit y, also appears
to us as mandatory.Therefore, after an oriented summary of the existing framew ork (of interest
for us) in this direction, mostly based on [29], in order to de fine safely
random pairwise comparisons matrices, we extend the geomet ry of classical
PC matrices understood as an abelian group, to non reciproca l PC matrices
which project to the total tangent space of the first group. Th is enables
us to extend the notion of consistency to non-reciprocal (de terministic) PC
matrices and define canonical projections on remarkable sub sets, inconsistency
indicators, reciprocity indicators and so-called total in dexes that “gather”
the two criteria, consistency and reciprocity, in their eva luation. With these
deterministic tools at hand, we define their stochatic analo gs on random
PC matrices and stochastic projections on remarkable subse ts of random
matrices. | 2312.00001v1.pdf | 9730f095-6406-47db-b525-c6679da4abcb |
Meaningful decompositions of the space of random matrices are
described, and two complementary stochastic methods are pr oposed for reduction
of stochastic non-reciprocity, stochastic inconsistency , with in mind the problem
of finding out weights for a final evaluation of the items at han d.
1.A not-so-short introduction to pairwise comparisons as a
class of abelian Yang-Mills theory
A paiswise comparisons matrix (ai,j)is an×nmatrix with coefficients in
R∗
+={x∈R|x >0}such that
∀i,j, a j,i=a−1
i,j. | 2312.00001v1.pdf | 7fe07aa3-6ed4-4dba-8ee1-ecbe21d17576 |
4 JEAN-PIERRE MAGNOT
It is easy to explain the inconsistency in pairwise comparis ons when we
consider cycles of three comparisons, called triad and repr esented here as
(x,y,z), which do not have the “morphism of groupoid” property such a s
x∗z̸=y,
which reads as
xz̸=y
in the multiplicative group R∗
+.The use of “inconsistency” has a meaning of
a measure of inconsistency in this study; not the concept its elf. An approach
to inconsistency (originated in [16]) can be reduced to a sim ple observation:
•search all triads (which generate all 3 by 3 PC sub matrices) a nd
locate the worst triad with a so-called inconsistency indic ator (ii),
•iiof the worst triad becomes iiof the entire PC matrix.
Expressing it a bit more formally in terms of triads (the uppe r triangle of a
PC sub matrix 3×3), we have Koczkodaj’s index , see e.g.[4, 21]:
(1) Kii(x,y,z) = 1−min{y
xz,xz
y}
= 1−e−|ln(y
xz)|.
For the sake of completeness, we notice here that this definit ion allows us
to localize the inconsistency in the matrix PC and it is of a co nsiderable
importance for most applications. But this aspect will be wor ked out in
future works, and we prefer here to concentrate of reduction of the inconsistency.
This inconsistency indicator extends to n×nPC matrices by:
(2) Kiin(A) = max
(i,j)∈N2n,i<j(
1−e−⏐⏐⏐⏐ln(
aij
ai,i+1ai+1,i+2...aj−1,j)⏐⏐⏐⏐)
whereNn={1,···,n} ⊂N,but also by various ways to combine inconsistency
of triads, see e.g. the examples studied in the recent work [3 1]. | 2312.00001v1.pdf | fb5f3470-5a66-4405-bb90-c5da5c8e4dbe |
the examples studied in the recent work [3 1].
Another common procedure on pairwise comparisons matrices changes
the multiplicative group (R∗
+,.)to the additive group (R,+)with the same
(obvious) adaptations of the formulas. These transformati ons, natural from
a group-theoric viewpoint and in one-to-one correspondenc e via logarithmic
maps, are actually perceived as no-so natural in the framewo rk of decision
theory. Out of psychological reasons, this basically comes from the fact that
in the framework of pairwise comparisons, the important pro cedure consists
in finding an efficient way to reduce an inconsistency indicato r, which may
not beKii3orKiinbut another, which is chosen because more adapted
to the contextual problems. For this, one defines priority vectors which
intend to give a “best way” to reduce in a satisfying manner th e inconsistency,
that is, the “score” of inconsistency produced by the incons istency index.Therefore, the correspondence between additive and multip licative pairwise
comparisons matrices enables one to produce more simple for mulas and
algorithms in the additive case. This correspondence is the key motivation
for e.g. [19] where an apparently natural way to link any PC ma trix with | 2312.00001v1.pdf | 4c137b43-5c76-4988-94ac-5f420fb294ad |
ON RANDOM PAIRWISE COMPARISONS MATRICES AND THEIR GEOMETRY . 5
a “prefered” consistent one is described. This impression o f simplicity and
clarity is obtained by the use of basic geometrical methods, mostly the use
of orthogonal vector spaces. However, in the actual state of knowledge, the
choice any method for inconsistency reduction appears as ar bitrary from
the viewpoint of applications while almost canonical from t he viewpoint of
mathematical structures, because the procedure of reducti on of inconsistency
may produce very far pairwise comparisons matrices that can be controlled
by a small enough inconsistency indicator value [31].
Following [27, 29], let nbe the dimension of the PC matrices under
consideration, and let us consider the n−simplex
∆n={
(x0,...,x n)∈Rn+1|
(n∑
i=0xi= 1)
∧(∀i∈ {0,....n},xi≥0)}
as a graph, made ofn(n−1)
2edges linking nvertices, equipped with a pre-fixed
numerotation.Then there is a one-to-one and onto correspon dence between
edges and the positions of the coefficients in the PC matrix. He nce, each
PC matrix (ai,j)assigns the coefficient ai,jto the (oriented) edge from the
i−th vertex to the j−th vertex. These are the holonomies of a connection,
understood in its discretized analog (in a discretization t hrouhg holonomy
around the oriented 1-vertices of the simplex). Minimizing holonomies is a
quantum gravity analog of a classical R∗
+−Yang-Mills theory [35, 41]. Indeed,
in the quantum gravity approach [35], the seek of flatness of t he curvature
relies in the minimization of the distance between the loop h olonomies and
1,which is exactly the basic requirement of reduction of incon sistencies on
triads. | 2312.00001v1.pdf | 34cfd52f-7c1e-43a0-a22d-108aab234c21 |
Following the preliminary note [27], there is one co nceptual problem
in the correspondence between Koczkodaj’s inconsistency i ndex and the
Yang-Mills functional, highlighted from a psychological v iewpoint in [31].
Koczkodaj’s inconsistency indicators read as:
Kiin(A) =
1−exp(
−max
1≤i<j≤n⏐⏐⏐⏐ln(aij
ai,i+1ai+1,i+2...aj−1,j)⏐⏐⏐⏐)
.
In order to recover now the quamtum gravity analog [35] of the Yang-Mills
action functional, one replaces
max
1≤i<j≤n(...)
by a quadratic mean. The choice of an inconsistency indicato r is not so easy
and Koczdodaj’s, even if quite simple and easy to understand , cannot be
such easily considered as the only interesting indicator fo r inconsistency in
pairwise comparisons, see e.g. Saaty’s [4, 36]. | 2312.00001v1.pdf | fe9d92ba-c741-4218-8316-0c1a841fbcf1 |
6 JEAN-PIERRE MAGNOT
These are all these relations that we now reinforce in this wo rk, by analyzing
non-reciprocal pairwise comparisons matrices from this vi ewpoint with the
help of complementary notions.
2.A theory of non reciprocal pairwise comparisons and
weights: deterministic settings
In the rest of the text, nis the dimension of the square matrices under
consideration.
2.1.Preliminary settings on Mn(R∗
+)and its remarkable subsets. We
consider here Mn(R∗
+)ans a smooth manifold of dimension n2with tangent
spaceMn(R).We also define Gn= (R∗
+)n(this notation will be explained
later).We recall that we note by PCnthe subset of Mn(R∗
+)made of matrices
(ai,j)(i,j)∈N2nsuch that
∀(i,j)∈N2
n,ai,j=a−1
j,i
(which implies that ∀i∈Nn, ai,i= 1) We define the following operations:
(1)Mn(R∗
+)is stable for the matrix multiplication in Mn(R),noted by
(M,N)↦→MN
and we choose to keep the same notation for its restriction to Mn(R∗
+). | 2312.00001v1.pdf | 40155023-d18d-4a1c-827c-d8e60988cd8a |
(2) The group multiplication in R∗
+extends to Mn(R∗
+)by identifying
the matrices of Mn(R∗
+)with its coefficients, that is we make the
identification Mn(R∗
+) = (R∗
+)n2.We therefore define a free Lie group
structure on Mn(R∗
+)that we note by (A,B)∈Mn(R∗
+)2↦→C∈
Mn(R∗
+)whereC= (ci,j)(i,j)∈N2nand
∀(i,j)∈N2
n, ci,j=ai,jbi,j
withA= (ai,j)(i,j)∈N2nandB= (bi,j)(i,j)∈N2n. PCnandCPCnare
Lie subgroups of Mn(R∗
+)equipped with this multiplication.(3) The same way, reformulating and extending [29], ∗can be also applied
toGnand there are three smooth actions of GnonMn(R∗
+)noted by
L,RandAdand, noting G= (gk)k∈Nn∈GnandA= (ai,j)(i,j)∈N2n∈
Mn(R∗
+),they are defined respectively by
L: (G,A)↦→(giai,j)(i,j)∈N2n
ifi < j and
L: (G,A)↦→(g−1
iai,j)(i,j)∈N2n
ifi > j,
R: (G,A)↦→(gjai,j)(i,j)∈N2n
ifi < j and
R: (G,A)↦→(g−1
jai,j)(i,j)∈N2n | 2312.00001v1.pdf | 69729b67-eced-48ec-86a1-c15ad03ad620 |
ON RANDOM PAIRWISE COMPARISONS MATRICES AND THEIR GEOMETRY . 7
ifi > j, and
Ad(G,A) =L(G,R(G−1,A)).
The last formula highlights the link of Gwith the gauge group of a
discretized connection as described in [29] in a study on PCninstead
ofMn(R∗
+).One can remark that the actions LandRcorrespond
to∗−multiplicaton by matrices with constant lines (resp. const ant
columns).
(4) LetGnbe the group of n−permutations understood as the group of
bijections of Nn.ThenGnacts indexwise (and hence smoothly) on
Mn(R∗
+)and the submanifolds PCnandCPCnare stable under this
action.
2.2.Weights as projective coordinates.2.2.Weights as projective coordinates. We recall that family of weights
(wi)i∈Nn∈(R∗
+)nis relative to a consistent PC matrix A= (ai,j)(i,j)∈N2n∈
CPCn(R∗
+)if and only if
∀(i,j)∈N2
n,ai,j=wj
wi.
Therefore, we have another interpretation of weights of a co nsistent pairwise
comparisons matrix:
Proposition 2.1. There exists a smooth embedding φfromCPCn(R∗)to
Pn(R)such that, for each A= (ai,j)(i,j)∈N2n∈CPCn,(wi)i∈Nn⊂(R∗
+)nare
projective coordinates for φ(A)if and only if
ai,j=wi/wj.
The proof of the fact that φis an embedding is obvious.
Proposition 2.2. φ(CPCn)is open and dense in Pn(R)and the map φis
proper. | 2312.00001v1.pdf | 233ea54c-e09b-420a-ac43-f1645baca2dc |
Proof.dimCPC n=dimPn(R) =n−1andφis an embedding therefore φ(A)
is an open submanifold of Pn(R)andφis a diffeomorphism from CPCnto
φ(CPCn),which implies that φis proper. Moreover, if we consider the
canonical embaddings of Pn−1(R)intoPn(R),defined by sending (n−1)
homogeneous coordinates into nhomogeneous coordinates after fixing one
of them equal to 0,the set
Pn(R)−φ(CPCn)
is composed of the images of these nembeddings of Pn−1(R)intoPn(R),
and, because each of them is of dimension n−2< n−1 =dimPn(R),each
of them has an empty interior. Therefore φ(CPCn)is dense in Pn(R).□ | 2312.00001v1.pdf | 1f429f25-f7fd-486b-a09e-b0a4ce292df2 |
8 JEAN-PIERRE MAGNOT
2.3.(Reciprocal) pairwise comparisons. We recall that a pairwise comparisons
matrix is a matrix A∈Mn(R∗
+),such that
∀(i,j)∈N2
n, aj,i=a−1
i,j.
This condition is called the reciprocity condition, and the word “reciprocal”
is often omitted in the literature. However, since we consid er here a wider
class of pairwise comparisons, that can fulfill the reciproc ity condition or not,
we can use the more precise terminology “reciprocal pairwis e comparisons”
when we feel a possible ambiguity in the text that could lead t o misleading
statements. Following [29], PCnis a smooth manifold of dimensionn(n−1)
2.
As a free abelian Lie group, it is a Riemannian exponential Li e group, as
well asCPCn.Moreover,
Theorem 2.3. [29]L(G3,CPC3) =PC3and∀n≥4,L(Gn,CPCn)̸=
PCn.For application purposes, one not only needs to decide if a pa irwise comparisons
matrix is consistent, but also if an inconsistent pairwise c omparisons matrix
is “very” inconsistent or “not so much” inconsistent. This l eads to the
definition of the following objects.
Definition 2.4. Aninconsistency index is a map ii:PCn→R+such
that
ii−1(0) =CPCn.
Aninconsistency indicator is an inconsistency index with values in [0,1].
Remark 2.5. This terminology is common in the literature, and it is somet imes
applied to maps for which ii−1(0)also contains non consistent pairwise comparisons
matrices. This ambiguity is due, to our feeling and to our kno wledge, mostly
to the development of the theory with very popular maps, used historically
to detect submatrices with the “highest” possible inconsis tency. | 2312.00001v1.pdf | fc629182-8c74-4008-b827-62d2a8bdbeac |
There are
propositions on the conditions that must fulfill an inconsis tency index, which
are not all equivalent, but most of them require that ii−1(0) =CPCn.This
work is not the place for controversies on this subject, and w e decide to work
in the context here announced.
The same way, the actual state of knowledge do not produce a sh ared
requirement for the regularity of the map ii.Does it need to be continuous,
derivable or smooth? There are divergent opinions due to the various fields of
interest of the authors and, again, to the historical develo pment of the theory.
For example, one of the most famous scales for inconsistency envolves natural
numbers from 0to9(Saaty’s scale). The regularity that we assume for an
inconsistency index will be precised in the statements.
The family{
ii−1([0,ǫ))|ǫ∈R∗
+}
forms a base of filter around CPCn[29].In [31], inconsistency indexes of
sufficient regularity are treated as scalar potentials on PCnthat have to be
minimized by a gradient method. | 2312.00001v1.pdf | b21421ce-cd92-47de-87b3-afc51c3d3d3b |
ON RANDOM PAIRWISE COMPARISONS MATRICES AND THEIR GEOMETRY . 9
2.4.Non-reciprocal pairwise comparisons and related indicator s.
We recall that a non-reciprocal pairwise comparisons matri x is a matrix
A∈Mn(R∗
+),without any further assumption. In particular, if Ais not
reciprocal,
∃(i,j)∈N2
n, ai,j̸=a−1
j,i,
and more generally in our terminology, a non-reciprocal mat rix may be
reciprocal or not reciprocal. We can then adapt the followin g definition
from the notion of inconsistency index:
Definition 2.6. Areciprocity index is a map ri:Mn(R∗
+)→R+such
that
ri−1(0) =PCn.
Areciprocity indicator is an reciprocity index with values in [0,1].
Remark 2.7.Remark 2.7. In order to be coherent with the terminology of Definition 2.4 ,
we should prefer the terminology ”non-reciprocity index” i n the last definition.
However, since this notion is new, we feel that it is subject t o adaptations
in future developments and hence the terminology will be fixe d naturally by
practice.
In the same spirit, we have
Definition 2.8. Aninconsistency and reciprocity index (ortotal index )
is a map iri:Mn(R∗
+)→R+such that
iri−1(0) =CPCn.
Aninconsistency and reciprocity indicator (ortotal indicator ) is an
inconsistency and reciprocity index with values in [0,1].
Let us give now few classes of examples:
Example 2.9. | 2312.00001v1.pdf | af7dad1b-2d21-4111-90ae-6866c1540c31 |
Let us give now few classes of examples:
Example 2.9. Letdbe the canonical Riemannian metric on the Lie group
(Mn(R∗
+),∗)Then, for any γ∈R∗
+,
A∈Mn(R∗
+)↦→d(A,PCn)γ
is a reciprocity index and
A∈Mn(R∗
+)↦→d(A,CPC n)γ
is an inconsistency and reciprocity index. Moreover, if we r eplacedby
another distance, we get another inconsistency index, and i f we moreover
assume that the diameter of Mn(R∗
+)is 1, we get an inconsistency indicator.
This class of indicators is different, but may be treated in th e same spirit, as
those taken as case studies in [31]. | 2312.00001v1.pdf | 0d27c7c5-2939-4d11-aaeb-7fa0bf4544c4 |
10 JEAN-PIERRE MAGNOT
2.5.On the geometry of the tower CPCn⊂PCn⊂Mn(R∗
+)..Indeed,
Mn(R∗
+)completes the tower of Riemannian abelian Lie groups
Mn(R∗
+)⊃PCn⊃CPCn
and we now feel the need to investigate intrinsic geometric p roperties of these
inclusions. Let us in particular study the inclusion PCn⊂Mn(R∗
+).
Proposition 2.10. Mn(R)is the (abelian) Lie algebra of Mn(R∗
+)and the
Lie algebra pcn⊂Mn(R)ofPCnis the vector space of skew-symmetric
matrices with coefficients in R.
Proof. SinceMn(R∗
+)andPCnare abelian Lie groups, the Lie bracket on
their Lie algeba is vanishing. Moreover, since multiplicat ion is coefficientwise,
its is trivial that Mn(R)is the Lie algebra of Mn(R∗
+).Finally, since (ai,j)(i,j)∈N2n∈Mn(R∗
+)is reciprocal if and only if
∀(i,j)∈N2
n, ai,j=a−1
j,i,
we get, by differentiation of this constraint at ai,j= 1, that(ai,j)(i,j)∈N2n∈
Mn(R)is moreover in pcnif and only if
∀(i,j)∈N2
n, bi,j=−bj,i.
□
Let us now split Mn(R∗
+)in order to recover, in some way that we will
soon precise, a reciprocal and a non-reciprocal part. | 2312.00001v1.pdf | 97f45ad5-594d-4bf4-a44a-988b90ba03b2 |
Before that, we feel
the need to recall the geometric mean method which projects PCntoCPCn
through the projection πGMM defined by
πGMM(
(ai,j)(i,j)∈N2n)
= (bi,j)(i,j)∈N2n
with
∀(i,j)∈N2
n, bi,j=n√∏n
k=1ai,k∏n
k=1al,j.
Remark 2.11. This method is also described in the case of additive pairwise
comparisons matrices [19]but it is a simple fact that additive and multiplicative
pairwise comparisons are in one to one correspondence, see e .g. the brief but
very clear exposition in [43]. From a geometric viewpoint, additive pairwise
comparisons form the Lie algebra of the group of multiplicat ive pairwise
comparisons, which is trivially an exponential group. This easy fact will
be of central interest very soon in our work.This easy fact will
be of central interest very soon in our work.
There is a straightforward projection πPC:Mn(R∗
+)→PCndefined by
πPC(
(ai,j)(i,j)∈N2n)
= (bi,j)(i,j)∈N2n
with
∀(i,j)∈N2
n, bi,j=√
ai,j/aj,i.
Remark 2.12. If we analyze the diagonal coefficients of the image we get
trivially that ∀i∈Nn, bi,i= 1 | 2312.00001v1.pdf | e6c50d8d-2db2-4c30-8edf-6586776581be |
ON RANDOM PAIRWISE COMPARISONS MATRICES AND THEIR GEOMETRY . 11
Therefore, there is another natural projection that we call πNRdefined by
πPC(
(ai,j)(i,j)∈N2n)
= (bi,j)(i,j)∈N2n
with
∀(i,j)∈N2
n, bi,j=√ai,jaj,i.
Next proposition is straigthforward:
Proposition 2.13. ∀A∈Mn(R∗
+),A=πPC(A)∗πNR(A),and
•ifA= (ai,j)(i,j)∈N2n, πPC(A) =πNR(A)⇔ ∀(i,j)∈N2
n,ai,j= 1
•πPC(Mn(R∗
+))∩πNR(Mn(R∗
+))is uniquely composed by the matrix
A= (ai,j)(i,j)∈N2n,such that ∀(i,j)∈N2
n, ai,j= 1.In other words, the Lie group Mn(R∗
+)decomposes into a direct product of
πPC(A)and ofπNR(A).Moreover, projections are Lie group morphisms.
Lemma 2.14. dim(
πNR(Mn(R∗
+)))
=n(n+1)
2.
Proof. A matrix in πNR(Mn(R∗
+))is symmetric in its off-diagonal coefficients,
but it has its ndiagonal coefficients that may be different from 1.Therefore,
the dimension of this Lie group isn(n+1)
2+n=n(n+1)
2. □
Proof of Proposition 2.13. Through direct calculations, the multiplication
map
∗:πPC(Mn(R∗
+))×πNR(Mn(R∗
+))→Mn(R∗
+)
is a morphism of connected Lie groups of the same dimension wi th trivial
kernel, with ImπNR=KerπPCandImπPC=KerπNR.□
Let us now investigate a relationship with discretized conn ections that
extends the existing identifications in PCn.For this, we first remark that
Lemma 2.14 shows that the decomposition is not exactly a doub ling of
the structure group from PCntoMn(R∗
+)becausedim(
πNR(Mn(R∗
+)))
>
dim(PCn).Roughly speaking, the dimensions that we have “in more” in
πNR(Mn(R∗
+)are the diagonal coefficients of the matrices. Therefore, we
propose a new structure group
G=TR∗
+=R∗
+×R. | 2312.00001v1.pdf | 499c8d3e-1c41-4345-8499-9047f1fa95e4 |